JP3550751B2 - Active control device - Google Patents

Description

子アクチュエータ９を制御する信号Ｏ_ｃ２とを出力する構成となっている。
【００５４】
前記同定ニューラルネット１５は、予め同定したアクチュエータの制御による自動車の応答に基づき制御予測値Ｏ_ｆ を出力する。即ち、制御ニューラルネット１３の出力Ｏ_ｃ１，Ｏ_ｃ２及びマイクロフォン３の出力ｍ等を入力し、制御予測値Ｏ_ｆ を出力する構成となっている。
【００５５】
前記評価・調整部１７は、同定ニューラルネット１５が被制御体の動特性を充分に表した後、即ち、車室１の応答を充分に同定した後、同定ニューラルネット１５の出力Ｏ_ｆ と制御目標値Ｘ_Ｔ との差を演算し、この値を用いてバックプロパゲーションにより制御ニューラルネット１３の結合荷重の修正を行なう。
【００５６】
前記遅延器１９，２１及び図示しない記憶部は、ＡＲＭＡモデルに基づいて制御ニューラルネット１３と同定ニューラルネット１５とを構成するものである。
【００５７】
遅延器１９はマイクロフォン３、ピエゾ素子センサ５が騒音状態を検出した時点ｔよりも前、即ち、（ｔ−１），（ｔ−２），…などの制御信号Ｏ_ｃ１，Ｏ_ｃ２を制御ニューラルネット１３に入力する構成となっている。このため、制御ニューラルネット１３は時間遅れ（過去）の信号を入力するニューロン素子を有している。
【００５８】
遅延器２１は、同定ニューラルネット１５の出力信号Ｏ_ｆ の現時点ｔよりも前の（ｔ−１）などのデータを同定ニューラルネット１５及び制御ニューラルネット１３に入力する構成となっている。従って、両ニューラルネット１３，１５は、同定ニューラルネット１５の出力Ｏ_ｆ の過去のデータを入力するニューロン素子をそれぞれ有している。
【００５９】
【外２】

入力により現時点の制御信号Ｏ_ｃ１，Ｏ_ｃ２を出力し、同定ニューラルネット１５は、前記過去のデータとｍ，Ｏ_ｃ１，Ｏ_ｃ２との信号入力により、検出時点よりも後の（ｔ＋Δｔ）の制御予測値Ｏ_ｆ を予測して出力する。このため、自動車のように外部環境が激しく変化するような場合でも、将来を見越した制御ができるため、素早い対応をとることができる。又、その予測を行なう際、常に最前の値を用いて予測を行なっているため適確な予測を行なうことができる。
【００６０】
図３（ａ），（ｂ）は、ニューラルネットの構成の一部を示したもので（ａ）は全体（ｂ）は中間層の１つ（例えば２５ｃ）を示したものである。
【００６１】
このニューラルネットは、入力層と中間層と出力層とからなっている。入力層はニューロン素子２３ａ，２３ｂの２個を示し、中間層はニューロン素子２５ａ，２５ｂ，２５ｃの３個を示し、出力層はニューロン素子２７の１個を示している。
【００６２】
ここで説明するニューラルネットは、入力層のニューロン素子２３ａ，２３ｂでは入力ｘ_１ ，ｘ_２ に対して結合荷重Ｗ_ｉｊを付け、中間層のニューロン素子２５ａ，２５ｂ，２５ｃでは同様に結合荷重Ｈ_ｊｒを付けている。ｉｊはｉ番目の入力層に対して、ｊ番目の中間層が結合していることを示し、ｊｒはｊ番目の中間層に対してｒ番目の出力層が結合していることを示している。
【００６３】
従って、結合荷重Ｗ_ｉｊによる各ニューロンの内部ポテンシャルＯ_１ は、
Ｏ_１ ＝ΣＷ_ｉｊｘ_ｉ ＋Θ_ｉ …（１）
となる。ここに、Θ_ｉ はニューロン素子の閾値を示している。
【００６４】
また、内部ポテンシャルＯ_１ は、ニューロン素子への出力値と決める。出力は、例えば、図４で示すシグモイド関数で計算される。これは、
【数１】

と表される。
【００６５】
ここに、ｕ_０ は図４の傾きを示すもので、この傾きｕ_０ を決めることにより、ランダム信号に対応させることができる。出力は０から１の間の出力であり、例えば、物理量ｖ＝０〜１００ｍ／ｓｅｃ を出力するとしたら、
【数２】

として、０〜１の信号にするのである。
【００６６】
結合荷重Ｗ_ｉｊ等の修正は、同定ニューラルネット１５の出力Ｏ_ｆ と制御目標値Ｘ_Ｔ との誤差を、
Ｅ＝Σ（Ｘ_Ｔ −Ｏ_ｆ ）^２ …（４）
とし、最急降下法により、誤差Ｅを最小値とするＷ_ｉｊを求め、結合荷重とするのである。
【００６７】
ここで、入力ｘのニューロン素子の出力Ｏ_ｆ は、Ｏ_ｆ ＝ｆ（ｙ）で与えられるから、誤差Ｅはｆ（ｙ）の関数となり、ＥをＯ_ｆ で偏微分し、Ｏ_ｆ をＷ_ｉｊで偏微分して、
【数３】

とし、さらにＯ_ｆ をｆ（ｙ）で偏微分し、ｆ（ｙ）をＷ_ｉｊで偏微分して、
【数４】

となる。従って、ｆ（ｙ）とＯ_ｆ との関係でＥを最小とするＷ_ｉｊが求められる。
【００６８】
上記のような関係でＥを最小とするＷ_ｉｊを求めることは、最急降下法を実装したものと等価であり、以下の式が成り立つ。
【００６９】
【数５】

このような関係において、制御パラメータηがある程度大きいと誤差Ｅは収束せず、またηを小さくするとローカルな極小値で収束してしまい、グローバルな最小値を検出することができなくなる。これを図５を用いて説明すると、図５は入力ｘに対する誤差Ｅの変化として誤差関数ｆ（ｙ）を示している。この誤差関数ｆ（ｙ）において、制御パラメータηを大きくしてしまうと例えば誤差関数ｆ（ｙ）上においてＳからＰまで進んでしまい、最小値Ｒに収束することがなかなかできないものとなる。逆にηが小さいと例えばＰ点からＱ点へ進んでローカルな極小値Ｑに収束してしまい、グローバルな最小値Ｒを検出することができなくなる恐れがある。したがって上記のように、単なる最急降下法を実装したものと等価なニューラルネットワークでは学習速度及び制御精度の向上に限界がある。このため、この発明の一実施例では制御パラメータの一つであるηを以下に説明する勾配式ランダム探索法によって決定するのである。すなわち、この実施例では最急降下法とランダム探索法とを結合したアルゴリズムとなっている。
【００７０】
図６はランダム探索法の概略図を示している。ここでは、有る目的関数ｆ（ｘ）を制約条件のもとで最小にする問題を考える。このランダム探索法ではあらたにステップ数、カウント数の二つの定数を定義する。ステップ数とは探索を行なう近傍領域の個数であり、カウント数とは一つの近傍領域を探索する回数の上限値を表わす。制約条件を満足する解の第１次近似解（初期値）ｘ_０ とし、ｘ_０ の周りに近傍領域Ｎ（ｘ_０ ，ｈ_ｉ ）を設定する。ここにＨ＝ｈ_ｉ （ｉ＝１，…，ｒ）はステップである。この探索法の考え方は、夫々の近傍領域の中でｘをランダムに発生させ、これによるｆ（ｘ）がｆ（ｘ_０ ）より小さければ、その点をその領域内の極小値として記憶し、全近傍領域で選択された極小値の中で最も小さなｘを第２次近似解ｘ_１ とし、ｘ_１ の周りに再度近傍領域を設けて探索を繰り返すものである。
【００７１】
以上の手法を一変数関数の最小値の問題に適用すると以下のようになる。
【００７２】
目的関数ｆ（ｘ）が与えられ、変数ｘの値域が（ａ，ｂ）のとき、ステップを順次１／１０づつ小さくし、ステップ数がｒ個で探索が行われた場合、ステップ（Ｈ）はベクトル表示して次のようになる。
【００７３】
【数６】

すなわち図６の上段のようにｈ_１ ＝ｂ−ａを最大の領域として、１／１０のづつ小さくしｈ_ｒ ＝ｈ_ｒ−１ ／１０．０を最小として、ｒ個の領域に分割される。すなわち、ある起点ｘ_０ から夫々のステップに対し近傍領域（Ｎ）が設定される。次いで図６の中段のように近傍領域（Ｎ）でランダムにｘ_０ｉが発生する。この結果、ｆ（ｘ_０ ）＞ｆ（ｘ_０ｉ）ならばｘ_０ｉは記憶され、そうでない場合はカウント数Ｋをプラスし同じ近傍領域で次のランダムなｘ_０ｉが発生する。またカウント数Ｋが与えられた数より大きければ、他の近傍領域（Ｎ）への探索に移る。そして図６の下段のように、全ての近傍領域Ｎでの探索から極小値ｘ_ｓａｖｅ（１），ｘ_ｓａｖｅ（２），…ｘ_ｓａｖｅ（ｒ）を検出し、これら極小値の中から最小値ｘ_ｂｅｓｔを算出し次解の初期値とするのである。この方法は多変数の場合についても、変数をベクトルと考えることにより、その対応は可能である。
【００７４】
次に図７のフローチャートを用いて、上記ランダム探索を説明する。
【００７５】
なお以下の説明において、ステップＳはランダム探索法において、あらたに定義したステップｈとは異なるものである。
【００７６】
まずステップＳ１では、ステップｈ_ｉ の設定が行われる。このステップｈ_ｉ の設定は、上記のように探索を行なう近傍領域の個数の設定である。
【００７７】
ステップＳ２では、領域の設定が行われる。この領域設定によって、ｘの許容範囲を複数の領域に分割する。
【００７８】
次いでステップＳ３，Ｓ４，Ｓ５において初期設定が行われる。すなわち繰り返し数Ｉ＝０、ステップ数ｉ＝１、カウント数Ｋ＝０に設定される。繰り返し数Ｉは一回の領域設定によって得られた最小値を次解の初期値として、再度探索を繰り返す場合の数である。ステップ数ｉは１として設定し、ステップＳ２で設定した領域の最初の近傍領域を選定する。そして、初回はステップＳ６，ステップＳ７と移行してステップＳ８において最大の領域での探索が行われる。すなわち初期値ｘ_０ｉがランダムに発生し、ステップＳ６において目的関数ｆ（ｘ）との比較が行なわれる。ｆ（ｘ_０ ）＞ｆ（ｘ_０ｉ）ならば初期値ｘ_０ｉはステップＳ９において極小値として記憶される。
【００７９】
ステップＳ６においてＮＯと判断されれば、ステップＳ７においてカウント数Ｋが与えられた数より大きいかどうかが判断され、大きくなければステップＳ８において同じ領域であらためてｘが発生し探索が行われる。このような探索において、極小値が見つからず、しかもカウント数が与えられた数より大きくなるとステップＳ１０へ移行し、ステップｈの個数ｉが設定した個数ｒを上回るかどうかの判断が行われ、ｉがｒを下回っていればｉの数をプラスしてステップＳ５へ移行する。
【００８０】
ステップＳ５では再びＫ＝０の設定が行われ、ステップＳ６，ステップＳ７，ステップＳ８へと移行し次の領域での探索が行われる。ステップＳ９において極小値の記憶が行われた場合にも、ステップＳ１０の判断が行われステップの個数ｉのプラスが行われて次の領域での探索が行われるものである。
【００８１】
このようにして、全てのステップでの探索が終了すると、ステップＳ１１において極小値の中の最も小さい値を最小値ｘ＝ｘ_ｂｅｓｔとして算出し記憶する。次にステップＳ１１において記憶した最小値ｘ_ｂｅｓｔを第２次近似解ｘ_１ とし、ｘ_１ の解りに再度近傍領域が設定され、ステップＳ１３においてＩが繰り返し数を上回ってない限りＩがプラスされ、ステップＳ４へ移行する。したがって、あらたな領域設定において再度探索が行われ最小値の算出が行われる。
【００８２】
そしてステップＳ１３において、繰り返し数を満足したと判断されれば探索は終了する。
【００８３】
このような処理によって、上記制御パラメータηの最小値を探索するものである。すなわち図６，図７において、ｘを制御パラメータηとして置き換え処理を行なうのである。制御パラメータηに置き換えて説明すると、制御装置１１は制御パラメータηの初期値η_０ｉをランダムに設定し、初期値η_０ｉでの制御目標値に対する出力値の感度を計算する。次いで制御パラメータηの許容範囲を複数の領域に分割し、前記感度に基づき各領域の中に制御パラメータηのランダムな値を選んで、制御目標値に対する出力値の差の極小値を各領域で検出し、各領域から得られた極小値の中から最小値を算出し、該最小値を次回の初期値としてηの修正を行なうものである。
【００８４】
したがって上記のように、大きさの異なった多領域に分割してηの解を探索するため、盲目的な探索を避けることができ探索に要する回数を減らすことができる。また多領域において夫々極小値を検出し、その中から最小値を算出するためローカルな極小値へ収束することを防ぐことができ、グルーバルな最小値を求めることができる。こうしてニューラルネットの学習の割合であるηの適正な値を求めることができ、精度良く高速に処理することができる。
【００８５】
次に上記ランダム探索法の効果を目的関数ｆ（ｘ）として、一次元関数を定義し、さらに説明する。
【００８６】
この一次元関数は、
【数７】

で定義する。この関数の形状は図８のようになっている。また関数の最小値を与えるｘ（＝ｘ_ｂｅｓｔ）は１７．０３９であり、そのときの関数値ｆ（＝ｆ_ｂｅｓｔ） は０．０９４となっている。この問題を最急降下法、ランダム探索法及び二つを組合せた方法を用いて探索した結果が図９である。ここで誤差Ｅは、
Ｅ＝（ｆ−ｆ_ｂｅｓｔ）／ｆ_ｂｅｓｔ
ｆ_ｂｅｓｔ＝最適値
と定義され、ｘの初期値は各々図に示してある通り１３．５，１３．４，５．０としている。この図９より最急降下法は初期値の設定によって解が収束する場合（グローバルな最小値）とそうでない場合（ローカルな極小値）とがあり、初期値に対する依存性が強くあるのに対し、ランダム探索法ではどのような初期値に対しても精度よく解が収束した。
【００８７】
さらに二つの方法を組合せた場合、最急降下法だけのときに見られた初期値に対する依存性を十分に克服していると共に、ランダム探索法だけのときに比べ収束が早くなっている。この結果からランダム探索法は最適化問題を解く上で有効な方法であり、ランダム探索法と他のアルゴリズムを組合せることにより、さらに有効性が増すことが確認できた。
【００８８】
またランダム探索法に用いる新たに定義した定数であるステップ数，カウント数が最適解への収束情況にどのような影響を及ぼすかを考慮してみる。これら二つの定数を変更したときの解の収束性を比較した結果を図１０，図１１に示す。目的関数は上記ｆ（ｘ）を用いている。これら図１０，図１１から明らかなように、ステップ数が増えるとより精度の良いきめこまかな解の探索が可能となる。またカウント数が増えると誤差が最小、すなわちグローバルな最小値にたどりつく確率が増すことになる。但し、この二つの考え方で注意しなければならないのは、この二つの定数をむやみに大きくすると無駄な探索時間や、プログラムの際、無駄なメモリ領域を使ってしまうので、これらの定数の決定には十分な注意を要する。
【００８９】
図１２から図２０はランダム探索法の有効性をさらに説明するためのものである。すなわち一自由度のばね−質量モデル及び多自由度の梁の数値シミュレーション結果を示している。
【００９０】
先ず図１２から図１６は一自由度系の変位励振による応答を示している。図１２に示す一自由度のばね−質量モデルを考えると、この系に正弦波の外力を加え、ある時間経過した後の過渡応答を図１３に示す。固有角振動数は約１００ｒａｄ／ｓｅｃで、時間ステップは１ｍｓｅｃとした。ニューラルネットワークの入力としては、系の応答速度と加速度とを用い、次の時間ステップの加速度の変化分を出力とした。教師信号を図１４に示し、図１５に収束性、図１６に同定結果を示している。図１５，図１６の結果よりランダム探索法の誤差の収束が極めて早いことが確認できた。これは高精度な学習を極めて短時間で行なうことができることを示し、ランダム探索法を用いたニューラルネットワークの写像能力が最急降下法に比べ、極めて優れていることが確認できた。
【００９１】
図１７から図２０は片持梁のインパルス応答をオンライン同定することを試みた。ニューラルネットワークの入出力及びシステムの概略は図１８に示している。この図１８において、ニューラルネットワークの入力としては梁の先端（ｘ６）の加速度と速度とし、加速度の変換を出力する。ここでも片持梁の応答、すなわち教師信号は数値計算により求めている。誤差の収束性を図１９に示し、シミュレーション時間１００秒付近における同定結果を図２０に夫々示している。図１９，図２０より明らかなように、オンライン同定では最急降下法を用いた場合とランダム探索法及び最急降下法を組合せた方法を用いた場合とでは、学習時間及び誤差の収束性にかなりの違いがあった。同定結果も勾配式ランダム探索法では極めて精度良く行われていることが確認できた。
【００９２】
以上説明したランダム探索法の処理は、一つのニューラルネットワークによって順番に行なうこともできるが、図２１のように並列化ニューラルネットによって分散処理することも可能である。すなわちニューラルネットは夫々入力層，中間層，出力層を備えた複数のユニット１〜ｒを備えている。このユニットの個数は分割設定する領域の個数となっている。したがって各領域ごとに各ユニットにより分散処理制御を行なうことができる。すなわち分散処理制御システム６１では全体領域での探索の情報制御及び各ユニットを分散制御する。そして各ユニットごとに各領域での探索が行われ、グローバルな最小値の算出が６３によって行われる。そして、この最小値に基づきバックプロパゲーションによって各ユニットの制御パラメータが修正されるのである。このような並列化処理によって、処理時間を極めて短縮することができ、より高速な処理を行なうことができる。またこのように複数のニューラルネットを並設した場合、一つのユニットの処理が終了し、他のユニットの処理が終了していない場合には終了したユニットが他のユニットの支援を行なうことができ、さらに高速化処理することができる。
【００９３】
なお上記処理において、制御装置１１は各領域でのカウント数の繰り返しによる検出値に応じて、各領域でのカウント数を変更することも可能である。すなわちカウント数が多く残存していても、検出値が大きく収束している場合には、以後の処理においてカウント数を次第に減少し、処理を早めることができるのである。
【００９４】
また制御装置１１は、各領域でカウント数の繰り返しにより極小値が得られないと判断するとき、該当する領域を制約リストに登録し、該領域での極小値の検出を中止することもできる。このように構成することによって、極小値が得られないことが明らかな領域の探索を中止し、より処理を高速化することができる。
【００９５】
さらに制御装置１１は各領域でのカウント数の繰り返しによる検出値に応じて、各領域で選ぶ制御パラメータのランダムな値の初期値に対する方向性に重み付けを行なうことができる。すなわち図７で示すような領域において、初期値ｘ_０ｉを中心にしてランダムな値を発生させるのであるが、カウント数を繰り返しているうちに初期値に対する極小値の方向が領域の中で次第に明らかになってくるため、その方向に多くのランダムな値を発生させることによって、より精度の良い検出を行なうことができるのである。
【００９６】
図２２は、ＡＲＭＡモデルに基づいたニューラルネットを一般的に示したものである。このニューラルネットは、入力層、中間層、出力層を有しており、時刻ｔにおける入力ｘ（ｔ）によって時刻（ｔ＋１）における出力ｙ（ｔ＋１）を予測して出力する。即ち、予測される出力ｙ（ｔ＋１）は、現在の入力ｘ（ｔ）と過去の状態とに関係する。このため、入力層には３つの短期記憶ユニット２９，３０，３１に接続されたニューロン素子を新たに追加している。
【００９７】
記憶ユニット２９は、入力ｘ（ｔ）の過去のデータｘ（ｔ−１）からｘ（ｔ−ｎ）を記憶し、それぞれ対応するニューロン素子に信号を入力するようになっている。記憶ユニット３０は、出力ｙ（ｔ＋１）の過去のデータｙ（ｔ）からｙ（ｔ−ｌ）のデータを記憶し、それぞれ対応するニューロン素子に入力するようになっている。記憶ユニット３１は、出力ｙ（ｔ＋１）の過去のデータｚ（ｔ−ｌ−１）からｚ（ｔ−ｇ）を記憶し、それぞれ対応するニューロン素子に入力するようになっている。
【００９８】
そして、入力ｘ（ｔ）の信号は予測係数α_０ を付けて中間層のニューロン素子に入力され、記憶ユニット２９の各信号は予測係数α_１ からα_ｎ が付けられ、記憶ユニット３０の信号は予測係数β_０ からβ_１ が付けられ、記憶ユニット３１の信号は予測係数γ_ｌ＋１ からγ_ｇ が付けられ、それぞれ中間層のニューロン素子に入力されるようになっている。
【００９９】
このニューラルネットの出力ｙ（ｔ＋１）と検出した状態量ｘ（ｔ＋１）との関係は、予測誤差ｅ（ｔ＋１）とすると、
ｘ（ｔ＋１）＝ｙ（ｔ＋１）＋ｅ（ｔ＋１）
となる。この場合、出力ｙ（ｔ＋１）は、
【数８】

で表される。
【０１００】
従って、予測係数α，β，γは、誤差系列に最小２乗推定法を適用して適正な値へ修正されていくこととなる。
【０１０１】
次に、アクティブコントロール装置の作用を説明する。
【０１０２】
図２３は、この発明の第１実施例に係る作用の概略を示したものである。アクティブコントロール装置は、大概この図２３のように動作する。図１をも参照すると、まず、ピエゾ素子センサ５及びマイクロフォン３によって車室１内の代表点の応答を正確に把握し、制御装置１１によってシステムの動的特性をより正確に同定する。制御装置１１は、システムの同定に基づいて連成解析を行ない、ピエゾ素子アクチュエータ９及びラウドスピーカ７を制御するのである。即ち、図２を参照し、同定ニューラルネット１５が車室１内の騒音に対する動特性を充分に表した後、評価・調整部１７において同定ニューラルネット１５の出力Ｏ_ｆ と制御目標値Ｘ_Ｔ との差を演算し、この値を用いて制御ニューラルネット１３の結合荷重をバックプロパゲーションにより修正するのである。
【０１０３】
図２４は、車室１内の応答に対する制御のタイムチャートを示している。即ち、時刻ｔにおいて、ピエゾ素子センサ５及びマイクロフォン３で車室１内の状態を感知し、制御装置１１が時刻ｔでラウドスピーカ７及びピエゾ素子アクチュエータ９に信号を出力し、時刻ｔ＋Δｔにおいて車室１内の騒音制御が実現されるのである。
【０１０４】
そして、この制御結果を制御装置１１のニューラルネットにフィードバックして結合荷重を修正するため、同定ニューラルネット１５の入力はある時間ステップｔにおける車室１内の騒音状態量とこれに対する制御量とし、出力は次のステップ（ｔ＋Δｔ）における状態量となる。この同定ニューラルネット１５の出力Ｏ_ｆ に基づき、同定ニューラルネット１５の制御評価関数Ｅ_ｆ 、制御ニューラルネット１３の制御評価関数Ｅ_ｃ は、
【数９】

但し、Ｘ：現在の状態
Ｘ_Ｔ ：のぞましい状態
Ｏ_ｆ ：ニューロが予測した状態
となる。
【０１０５】
Ｘは時間ステップｔ＋Δｔにおける検出状態量であり、Ｘ_Ｔ は、制御目標となる状態量である。従って、同定ニューラルネット１５が車室１内の騒音に対する動特性を正確に同定しているならば、評価関数Ｅ_ｆ は所定値以下となり、同定ニューラルネット１５の出力Ｏ_ｆ と目標状態量Ｘ_Ｔ との誤差Ｅ_ｃ は、全て制御ニューラルネット１３によって発生したと考えられる。
【０１０６】
従って、図２５のように誤差情報を逆伝搬し、制御ニューラルネット１３迄戻すことによって、制御装置１１全体についての調整が可能となるのである。
【０１０７】
次に、図２６のフローチャートによりこの発明の第１実施例の制御作用を説明する。
【０１０８】
まず、ステップＳ２１では、制御ニューラルネット１３と同定ニューラルネット１５との初期化を行なう。この初期化は、結合荷重Ｗと閾値Θ及びシグモイド関数の傾き係数ｕ_０ をランダム関数を用いて決定する。
【０１０９】
ステップＳ２２では、同定ニューラルネット１５の初期学習を行なう。この初期学習はオフラインで行なわれ、予め定められたプログラムで行なわれる。このステップＳ２２での初期学習が完了すると、図２のように同定ニューラルネット１５は制御装置１１に組み込まれる。
【０１１０】
同定ニューラルネット１５の初期学習が完了した後、制御ニューラルネット１３が稼動され、その出力Ｏ_ｃ がステップＳ２３で出力される。この出力によってラウドスピーカ７及びピエゾ素子アクチュエータ９が駆動されると共に、出力Ｏ_ｃ は同定ニューラルネット１５に入力される。これにより、ステップＳ２４において同定ニューラルネット１５は、時刻ｔ＋Δｔにおける予測した状態量Ｏ_ｆ を出力する。
【０１１１】
ステップＳ２５では、時刻ｔ＋Δｔにおける状態量Ｘの実測データがサンプリングされる。また、この実施例では、ＡＲＭＡモデルを利用しているため、過去の状態量を記憶装置に書き込む。
【０１１２】
ステップＳ２６では、同定の判断が行なわれる。即ち、このステップＳ２６は、同定ニューラルネット１５が車室１内の騒音に関する動特性を正確に同定しているかどうかを判断するものである。具体的には、実測された状態量Ｘと予測した出力Ｏ_ｆ とを用いて上式のように演算された同定ニューラルネット１５の評価関数Ｅ_ｆ を収束判定基準Ｔ_ｆ より大きいか否かを判断する。Ｅ_ｆ がＴ_ｆ より大きければ、同定ニューラルネット１５がシステムを充分に同定していないため、ステップＳ２７へ移行し、評価・調整部１７によって同定ニューラルネット１５の結合荷重Ｗの修正、学習が行なわれる。
【０１１３】
次いで、ステップＳ２８へ移行し、適応制御の判断が行なわれる。この判断は、車室１内の騒音状態が目標の状態になっているかどうかを判断するものである。具体的には、目標とする状態量Ｘ_Ｔ と同定ニューラルネット１５の出力Ｏ_ｆ とによって上式のように演算される評価関数Ｅ_ｃ を収束判定基準Ｔ_ｃ より大きいか否かを判断する。Ｅ_ｃ がＴ_ｃ よりも大きければ、制御ニューラルネット１３が充分に同定されていないためステップＳ２９に移行し、バックプロパゲーションによって制御ニューラルネット１３の結合荷重Ｗの修正、学習が行なわれる。
【０１１４】
以後、ステップＳ２３からステップＳ２９が繰り返され、車室１内の騒音がフィードフォワードによって制御されることになる。
【０１１５】
即ち、図１、図２のように同定ニューラルネット１５は、予めオフラインで車室１内の騒音に関する動特性を同定されている。そして、制御ニューラルネット
【外３】

制御ニューラルネット１３の出力信号Ｏ_ｃ１，Ｏ_ｃ２の過去の信号及び同定ニューラルネット１５の出力Ｏ_ｆ の過去の信号が入力され、制御信号Ｏ_ｃ１，Ｏ_ｃ２が出力される。ここで、制御ニューラルネット１３は、過去のデータを反映させるために過去のデータを入力するために予め設けられた入力用のニューロン素子全てにデータが入力された後稼動し出力するのである。
【０１１６】
そして、出力Ｏ_ｃ１によってラウドスピーカ７が駆動され、騒音に対する逆相の信号が車室１内に出力される。また、出力Ｏ_ｃ２によって、ピエゾ素子アクチュエータ９が駆動され、車体の振動モードが制御される。これによって、車室１の非線形の連成空間が制御され、騒音低減を図ることができる。
【０１１７】
また、同定ニューラルネット１５は、制御ニューラルネット１３の出力Ｏ_ｃ１，Ｏ_ｃ２を入力すると共に、車室１内の状態量としてマイクロフォン３の出力ｍ及び出力Ｏ_ｆ の過去のデータを入力し、時刻ｔ＋Δｔにおける出力Ｏ_ｆ を予測して出力する。この出力Ｏ_ｆ は、現時点での車室１内の騒音状態及び外力と過去のデータとに基づいており、ランダムな入力による騒音状態であっても正確に予測しているものと言える。
【０１１８】
そして、出力Ｏ_ｆ と現時点ｔにおける車室１内の騒音状態の実測データＸとを比較し、同定ニューラルネット１５の学習が行なわれる。同定ニューラルネット１５がシステムの動特性を正確に同定しているものとした場合に、同定ニューラルネット１５の出力Ｏ_ｆ と目標の状態量Ｘ_Ｔ とにより評価・調整部１７が制御ニューラルネット１３の学習を行なうのである。
【０１１９】
図２７は、この発明の第１実施例に係るニューラルネットワークの具体的な構成を示している。
【０１２０】
上記のように、このニューラルネットワークは、ＡＲＭＡモデルを考慮しており、制御ニューラルネット１３として音場制御ニューラルネットワーク１３ａと構造制御ニューラルネットワーク１３ｂとを備えている。各ニューラルネットワークは、入力層、中間層、出力層とからなっており、入力層は白丸、中間層は斜線の入った丸、出力層は黒丸で示している。
【０１２１】
音場制御ニューラルネットワーク１３ａの入力は、時刻ｔにおけるｎ個のマイクロフォンの音圧感知信号ｍ＝ｍ_１ ^ｓ （ｔ） 〜ｍ_ｎ ^ｓ （ｔ） と、時刻（ｔ−１）〜（ｔ−ｋ）迄の過去のデータｍ_１ ^ａ （ｔ−１） 〜ｍ_１ ^ａ （ｔ−ｋ） ，ｍ_ｎ ^ａ （ｔ−１） 〜ｍ_ｎ ^ａ （ｔ−ｋ） と、ｎ個のラウドスピーカの（ｔ−１）〜（ｔ−ｋ）迄の過去における制御力Ｓ_１ ^ａ （ｔ−１） 〜Ｓ_１ ^ａ （ｔ−ｋ） ，Ｓ_ｎ ^ａ （ｔ−１） 〜Ｓ_ｎ ^ａ （ｔ−ｋ） とが入力層の各ニューロン素子に入力されている。ここに、上添字のｓは検出信号を示し、同ａはニューラルネットによる処理信号を示している。
【０１２２】
音場制御ニューラルネットワーク１３ａの現在の時刻ｔにおける出力Ｏ_ｃ１は、
Ｏ_ｃ１＝Ｓ_１ ^ａ （ｔ） 〜Ｓ_ｎ ^ａ （ｔ）
であり、ｎ個のラウドスピーカに制御力として出力されると共に、同定ニューラルネットワーク１５の入力層の該当するニューロン素子に入力されている。
【０１２３】
この出力Ｏ_ｃ１＝Ｓ_１ ^ａ （ｔ） 〜Ｓ_ｎ ^ａ （ｔ） の過去のデータは記憶され、遅延器１９により音場制御ニューラルネットワーク１３ａの入力層に入力されることは前述の通りである。
【０１２４】
同定ニューラルネットワーク１５の入力層には、現時点ｔにおけるマイクロフォン音圧感知信号ｍ＝ｍ_１ ^ｓ （ｔ） 〜ｍ_ｎ ^ｓ （ｔ） と音圧感知信号の過去のデータｍ_１ ^ａ （ｔ−１） 〜ｍ_１ ^ａ （ｔ−ｋ） ，ｍ_ｎ ^ａ （ｔ−１） 〜ｍ_ｎ ^ａ （ｔ−ｋ） とが入力されている。
【０１２５】
また、構造制御ニューラルネットワーク１３ｂの入力層には、ｎ個のピエゾ素
【外４】

） が入力されている。また、ｎ個のピエゾ素子アクチュエータの過去の制御信号Ｐ_１ ^ａ （ｔ−１） 〜Ｐ_１ ^ａ （ｔ−ｋ） ，Ｐ_ｎ ^ａ （ｔ−１） 〜Ｐ_ｎ ^ａ （ｔ−ｋ） が入力されている。
【０１２６】
構造ニューラルネットワーク１３ｂの時刻ｔにおける出力Ｏ_ｃ２は、
Ｏ_ｃ２＝Ｐ_１ ^ａ （ｔ） 〜Ｐ_ｎ ^ａ （ｔ）
である。この出力Ｏ_ｃ２＝Ｐ_１ ^ａ （ｔ） 〜Ｐ_ｎ ^ａ （ｔ） は、ｎ個のピエゾ素子アクチュエータに出力されると共に、同定ニューラルネットワーク１５の入力層に入力されている。
【０１２７】
また、出力Ｏ_ｃ２＝Ｐ_１ ^ａ （ｔ） 〜Ｐ_ｎ ^ａ （ｔ） は前記同様逐次記憶され、過去のデータとして構造制御ニューラルネットワーク１３ｂの入力層に入力されるのは前述の通りである。
【０１２８】
同定ニューラルネットワーク１５の入力層には、時刻ｔにおける検出した車体
【外５】

る。
【０１２９】
そして、同定ニューラルネットワーク１５からは時刻ｔ＋Δｔにおけるマイク検出信号の予測値として
Ｏ_ｆ ＝ｍ_１ ^ａ （ｔ＋Δｔ）〜ｍ_ｎ ^ａ （ｔ＋Δｔ）
が出力され、ピエゾ素子センサの検出の予測値として
【数１０】

が出力される。
【０１３０】
なお、同定ニューラルネットワーク１５の出力Ｏ_ｆ ＝ｍ_１ ^ａ （ｔ＋Δｔ）〜ｍ_ｎ ^ａ （ｔ＋Δｔ）は逐次記憶され、過去のデータとして同定ニューラルネットワーク１５と音場制御ニューラルネットワーク１３ａとの各入力層に入力されている。
【０１３１】
【外６】

_ｎ ^ａ （ｔ＋Δｔ）も逐次記憶され、過去のデータとして同定ニューラルネットワーク１５の入力層及び構造制御ニューラルネットワーク１３ｂの入力層にそれぞれ入力されている。
【０１３２】
従って、同定ニューラルネットワーク１５の評価関数Ｅ_ｆ は、
【数１１】

となる。
【０１３３】
また、音場制御ニューラルネットワーク１３ａの制御評価関数Ｅ_ｃ は、
【数１２】

となる。
【０１３４】
更に、構造制御ニューラルネットワーク１３ｂの制御評価関数Ｅ_ｃ は、
【数１３】

となる。
【０１３５】
従って、これらの評価関数Ｅ_ｆ ，Ｅ_ｃ を用いて上記のように同定ニューラルネットワーク１５と制御ニューラルネットワーク１３との結合荷重Ｗの修正、学習を適格に行なうことができる。
【０１３６】
なお、この第１実施例では、同定ニューラルネット、制御ニューラルネットの双方にＡＲＭＡモデルを考慮したニューラルネットを用いているが、少なくとも制御ニューラルネットにつきＡＲＭＡモデルを考慮すればよい。
【０１３７】
図２８は、この発明の第２実施例を示している。この実施例は、アクティブコントロール装置を自動車のサスペンションの振動制御に適用したものである。
【０１３８】
車体３２は、前輪３３と後輪３５とで路面３７に接地している。このようなサスペンションにおいて前輪３３と車体３２との間のバネ定数はＫ_Ｆ２、減衰係数はＣ_Ｆ１であり、前輪３３と路面３７との間のバネ定数はＫ_Ｆ１、減衰係数はＣ_Ｆ１である。また、後輪３５と車体３２との間のバネ定数はＫ_Ｒ２であり、減衰係数はＣ_Ｒ１である。後輪３５と路面３７との間のバネ定数はＫ_Ｒ１、減衰係数はＣ_Ｒ１である。
【外７】

ある。
【０１３９】
そして、前輪３３と車体３２との間には制御力Ｐ_Ｆ を与えて振動低減を図る制振装置３９が設けられ、後輪３５と車体３２との間には制御力Ｐ_Ｒ を与えて振動低減を図る制振装置４１が設けられている。
【０１４０】
制御装置４０には、前輪３３側における路面３７の凹凸Ｙ_０Ｆを検出する前輪側路面検出器４３及び後輪３５側における路面の凹凸を検出する後輪側路面検出器４５の信号が入力されるようになっている。
【０１４１】
また、前輪３３の振動加速度を検出する前輪振動加速度検出器４７と、後輪３５の振動加速度を検出する後輪振動加速度検出器４９との出力信号が入力されるようになっている。
【０１４２】
更に、車体３２の前部振動加速度を検出する前部振動加速度検出器５１と、後輪側の振動加速度を検出する後部振動加速度検出器５３とからの信号が入力されるようになっている。
【０１４３】
これら、検出器４３，４５，４７，４９，５１，５３は、サスペンションの振動状態を検出する振動状態検出手段を構成している。そして、制御装置４０は、第１実施例同様にニューラルネットが用いられており、検出器４３，４５，４７，４９，５１，５３の出力が入力され、前記制振装置３９，４１を制御する信号を出力する。また、当該出力信号に基づく制御予測値を制御目標値と比較することにより、ニューラルネットの結合荷重の修正を行なう構成となっている。
【０１４４】
この制御装置４０のニューラルネットワークの構成は図２９のようになっている。
【０１４５】
このニューラルネットも前記実施例同様、制御ニューラルネット５５と同定ニューラルネット５７とからなっている。制御ニューラルネット５５は、前輪制御ニューラルネット５５ａと後輪制御ニューラルネット５５ｂとからなっている。各制御ニューラルネット５５ａ，５５ｂは白丸の入力層と、斜線を施した丸で示す２段の中間層と黒丸の出力層とからなっており、同定ニューラルネット５７は白丸の入力層、斜線を施した丸の中間層、黒丸の出力層からなっている。
【０１４６】
そして、この実施例では、制御ニューラルネット５５のみをＡＲＭＡモデルを考慮したニューラルネットとしている。即ち、前輪制御ニューラルネット５５ａの入力層には、時刻ｔにおける前輪側路面検出器４３の検出信号Ｙ_０Ｆ ^ｓ （ｔ） 、前
【外８】

制御力Ｐ_Ｆ ^ａ （ｔ） となる。この出力Ｐ_Ｆ ^ａ （ｔ） は、同定ニューラルネット１５に入力されると共に逐次記憶され、過去のデータＰ_Ｆ ^ａ （ｔ−１） として前輪制御ニューラルネット５５ａの入力層に入力されるようになっている。
【０１４７】
また、後輪制御ニューラルネット５５ｂも同様であり、入力は、検出器４５，
【外９】

４１の制御力Ｏ_ｃ ＝Ｐ_Ｒ ^ａ （ｔ） の過去のデータＰ_Ｒ ^ｓ （ｔ−１） となっている。出力Ｏ_ｃ ＝Ｐ_Ｒ ^ａ （ｔ） は制振装置４１に出力されると共に、同定ニューラルネット１５に入力されるようになっている。同定ニューラルネット１５の入力層には、更
【外１０】

更に、同定ニューラルネット１５の入力層には、中間層の出力が入力される構成となっている。これは、一般にレカレント（ｒｅｃｕｒｒｅｎｔ ）と呼ばれ、信号のノイズを除去することができる。レカレントは、例えば、中間層の出力を再び中間層自体に入力することによっても行なうことができる。
【０１４８】
同定ニューラルネット１５の出力Ｏ_ｆ は、次の時刻ｔ＋１における車体３２の振動加速度
【数１４】

となる。従って、同定ニューラルネット１５の評価関数Ｅ_ｆ は、
【数１５】

となり、制御ニューラルネット５５の評価関数Ｅ_ｃ は、
【数１６】

となる。
【０１４９】
従って、この実施例においても評価関数Ｅ_ｆ が所定以下となり、同定ニューラルネット５７が振動入力に対する車体３２の動特性を正確に表しているならば、評価関数Ｅ_ｃ によって制御ニューラルネット５５の結合荷重を修正、学習し、サスペンション制御の適格なフィードフォワード制御を可能とする。
【０１５０】
なお、第２実施例で設けた同定ニューラルネット５７のレカレントは、制御ニューラルネットに用いることもでき、更に、第１実施例のニューラルネットに適用することも可能である。また第２実施例では過去のデータとして（ｔ−１）における１個のデータで説明したが、第１実施例と同様にいくつかの時間遅れのデータも考慮することが可能である。
【０１５１】
【発明の効果】
以上より明らかなように、この発明の構成によれば以下のような効果を奏する。
【０１５２】
請求項１の発明では、ニューラルネットを用いた制御装置によりシステムの動特性を的確に同定することができる。しかも探索領域を大きさの異なった多領域に分割して探索を行なうため盲目的な探索を避けることができ、探索に要する回数を減らすことができる。しかもランダム探索を用いるため、ローカルな極小値へ収束することを防止し、グローバルな最小値を的確に求めることができる。したがって高速で精度のよい処理を行なうことができる。また、領域の中での探索を複数回繰り返すことによって、極小値を確実に検出することができる。したがって精度のよい処理を行なうことが可能である。また、検出値に応じてカウント数を減少することが可能となり、より高速な処理が可能となる。
【０１５３】
請求項２の発明では、ニューラルネットを用いた制御装置によりシステムの動特性を的確に同定することができる。しかも探索領域を大きさの異なった多領域に分割して探索を行なうため盲目的な探索を避けることができ、探索に要する回数を減らすことができる。しかもランダム探索を用いるため、ローカルな極小値へ収束することを防止し、グローバルな最小値を的確に求めることができる。したがって高速で精度のよい処理を行なうことができる。また、領域の中での探索を複数回繰り返すことによって、極小値を確実に検出することができる。したがって精度のよい処理を行なうことが可能である。極小値が得られないことが明らかな場合にその領域の処理を中止し、処理をより高速に行なうことが可能である。
【０１５４】
請求項３の発明では、ニューラルネットを用いた制御装置によりシステムの動特性を的確に同定することができる。しかも探索領域を大きさの異なった多領域に分割して探索を行なうため盲目的な探索を避けることができ、探索に要する回数を減らすことができる。しかもランダム探索を用いるため、ローカルな極小値へ収束することを防止し、グローバルな最小値を的確に求めることができる。したがって高速で精度のよい処理を行なうことができる。また、領域の中での探索を複数回繰り返すことによって、極小値を確実に検出することができる。したがって精度のよい処理を行なうことが可能である。また、初期値に対する方向性に重み付けを行なうことにより、より高速で精度のよい検出を行なうことができる。
【０１５５】
請求項４の発明では、探索領域を大きさの異なった複数の領域に分割して、ランダム探索を行なうため、音又は振動状態の少なくとも一方を高速かつ精度よく制御することができる。また、領域の中での探索を複数回繰り返すことによって、極小値を確実に検出することができる。したがって精度のよい処理を行なうことが可能である。また、検出値に応じてカウント数を減少することが可能となり、より高速な処理が可能となる。
【０１５６】
請求項５の発明では、探索領域を大きさの異なった複数の領域に分割して、ランダム探索を行なうため、音又は振動状態の少なくとも一方を高速かつ精度よく制御することができる。また、領域の中での探索を複数回繰り返すことによって、極小値を確実に検出することができる。したがって精度のよい処理を行なうことが可能である。極小値が得られないことが明らかな場合にその領域の処理を中止し、処理をより高速に行なうことが可能である。
【０１５７】
請求項６の発明では、探索領域を大きさの異なった複数の領域に分割して、ランダム探索を行なうため、音又は振動状態の少なくとも一方を高速かつ精度よく制御することができる。また、領域の中での探索を複数回繰り返すことによって、極小値を確実に検出することができる。したがって精度のよい処理を行なうことが可能である。初期値に対する方向性に重み付けを行なうことにより、より高速で精度のよい検出を行なうことができる。
【０１５８】
請求項７の発明では、請求項４〜請求項６のいずれか１項の発明の効果に加え、車室内騒音をフィードフォワード制御により的確に制御することができる。したがって車室内騒音がエンジン音、エンジンからの入力、サスペンションからの入力、マフラーからの排気音、風切り音などの連成に基づく場合でも連成を的確に解析して、騒音低減を図ることができる。
【０１５９】
請求項８の発明では、請求項４〜請求項６のいずれか１項の発明の効果に加え、サスペンションの信号制御をフィードフォワードにより的確に制御することができる。したがって高速道は勿論、オフロードであってもサスペンションの振動特性を的確に解析し、車体振動の低減を図ることができる。
【０１６０】
請求項９の発明では、請求項４〜請求項８のいずれか１項の発明の効果に加え、制御ニューラルネットの他に同定ニューラルネットを設けることによって、制御予測値を出力し、制御ニューラルネットの制御パラメータの修正を的確に行なうことができる。
【０１６１】
請求項１０の発明では、請求項４〜請求項９のいずれか１項の発明の効果に加え、検出制御の現時点よりも前のデータを入力することによって、現時点よりも後の制御予測値を予測して出力するため、ランダムな入力であっても音振状態を的確に予測することができ、ランダムな入力に対して的確な制御が可能となる。
【０１６２】
請求項１１の発明では、請求項１〜請求項１０のいずれか１項の発明の効果に加え、ニューラルネットの結合荷重を学習するときの学習する割合を制御パラメータとして修正するため、学習の割合を適切な値にすることを高速かつ精度良く行なうことができ、最小値の検出をより高速かつ高精度に行なうことができる。
【０１６３】
請求項１２の発明では、請求項１〜請求項１１のいずれか１項の発明の効果に加え、ニューラルネットを領域の個数に応じて複数並列に備えたため、処理を高速に行なうことができる。
【図面の簡単な説明】
【図１】この発明の一実施例に係る概略ブロック図である。
【図２】制御装置のブロック図である。
【図３】ニューラルネットの一部を示す説明図である。
【図４】シグモイド関数を示す説明図である。
【図５】最急降下法による収束状況を示す説明図である。
【図６】ランダム探索法の概略図である。
【図７】ランダム探索法のフローチャートである。
【図８】一次元関数の形状を示す図である。
【図９】各初期値における収束状況を説明するグラフである。
【図１０】ステップ数を変更した場合の説明図である。
【図１１】カウント数を変更した場合の説明図である。
【図１２】一自由度系ばね−質量モデルの説明図である。
【図１３】一自由度系ばね−質量モデルの過渡応答の説明図である。
【図１４】教師信号のグラフである。
【図１５】線形モデルの誤差の時刻歴のグラフである。
【図１６】同定結果のグラフである。
【図１７】片持梁にインパルス入力を行なう状況の説明図である。
【図１８】片持梁の同定ニューラルシステムの説明図である。
【図１９】片持梁のリアルタイム同定過程での誤差の履歴を示すグラフである。
【図２０】ニューラルネットによる片持梁の同定結果のグラフである。
【図２１】並列化ニューラルネットによる該略図である。
【図２２】ＡＲＭＡモデルに基づいたニューラルネットの説明図である。
【図２３】作用の概略説明図である。
【図２４】システムの応答図である。
【図２５】バックプロパゲーションの説明図である。
【図２６】フローチャートである。
【図２７】一実施例に係るニューラルネットの構成図である。
【図２８】他の実施例に係る概略ブロック図である。
【図２９】他の実施例に係るニューラルネットワークの構成図である。
【図３０】従来の能動型騒音制御装置の概略ブロック図である。
【図３１】信号の干渉状態を示す説明図である。
【符号の説明】
１ 車室
３ マイクロフォン（騒音状態検出手段、音振状態検出手段）
７ ラウドスピーカ（制御音源、アクチュエータ）
１１ 制御装置
３９ 制振装置（アクチュエータ）
４０ 制御装置
４１ 制振装置（アクチュエータ）
４３ 前輪側路面検出器（振動状態検出手段、音振状態検出手段）
４５ 後輪側路面検出器（振動状態検出手段、音振状態検出手段）
４７ 前輪振動加速度検出器（振動状態検出手段、音振状態検出手段）
４９ 後輪振動加速度検出器（振動状態検出手段、音振状態検出手段）
５１ 前部振動加速度検出器（振動状態検出手段、音振状態検出手段）
５３ 後部振動加速度検出器（振動状態検出手段、音振状態検出手段）[0001]
[Industrial applications]
The present invention relates to an active control device that actively reduces vehicle interior noise and vehicle body vibration.
[0002]
[Prior art]
As a conventional active control device of this type, there is, for example, an active noise control device shown in FIG. 30 of British Patent Publication No. 2149614.
[0003]
This conventional device is applied to a cabin of an aircraft or a similar closed space, and includes loudspeakers 103a, 103b, 103c and microphones 105a, 105b, 105c, 105d in a closed space 101. Control sounds 103b and 103c are used to generate a control sound that interferes with noise, and the microphones 105s, 105b, 105c and 105d measure residual signals (residual noise). The loudspeakers 103a, 103b, 103c and the microphones 105a, 105b, 105c, 105d are connected to a signal processor 107. The signal processor 107 is connected to the fundamental frequency of the noise source measured by the fundamental frequency measuring means and the microphones 105a, 105b. , 105c, and 105d, and outputs a control sound signal to the loudspeakers 103a, 103b, and 103c so as to minimize the sound pressure level in the closed space 101.
[0004]
Here, in the closed space 101, three loudspeakers 103a, 103b, 103c and four microphones 105a, 105b, 105c, 105d are provided. It is assumed that they are provided one by one.
[0005]
Assume that a transfer function from the noise source to the microphone 105a is H, a transfer function from the loudspeaker 103a to the microphone 105a is C, and a sound source information signal generated by the noise source is X._pThen, the residual signal E observed by the microphone 105a is
E = X_p・ H + X_p・ GC
It becomes. Here, G is a transfer function required for silencing. When the noise is completely canceled at the noise reduction target point, E = 0. At this time, G is
G = -H / C
It becomes. Then, a G at which the microphone detection signal E is minimized is obtained, and the filter coefficient in the signal processor 107 is adaptively updated based on the G. As means for obtaining a filter coefficient so as to minimize the microphone detection signal E, there is an LMS algorithm (Least Mean Square) which is a kind of the steepest descent method.
[0006]
When a plurality of microphones are installed as shown in FIG. 30, control is performed so that the sum of signals detected by the microphones 105a, 105b, 105c, and 105d is minimized.
[0007]
By the control as described above, for example, a canceling sound of the opposite phase as shown in FIG. 31B is output with respect to the cabin noise as shown in FIG. Counteract with the cancellation sound.
[0008]
[Problems to be solved by the invention]
By the way, the conventional device is effective in reducing a single noise source consisting of a fixed signal such as a sine component, for example, noise generated from an engine.
[0009]
However, there are various types of vehicle interior noise such as noise associated with engine rotation, road noise transmitted from a suspension, wind noise, exhaust noise, and the like, and many have random characteristics. . Therefore, if the noise control is performed using only the sine component of the noise accompanying the engine rotation as the reference signal, noise components other than the control target may be added each time calculation is performed, causing divergence. May worsen.
[0010]
In addition to the active noise control device shown in FIG. 30, the active control device can be applied to suspension vibration control and the like. However, the vibration input of the suspension has a random property, It was difficult to control.
[0011]
For this reason, the applicant of the present application employs a control device using a neural network to achieve accurate noise control even when noise from a plurality of noise sources is coupled and has random characteristics. An active control device capable of accurately controlling even a random vibration has already been filed (see Japanese Patent Application No. 5-5785).
[0012]
By the way, even if such an active control device using a neural network is effective in a low frequency region, there is a limit to the improvement of the identification learning speed and the control accuracy for a system having strong nonlinearity in a high frequency region. I was invited.
[0013]
More specifically, in a control device using a neural network, generally, an algorithm such as a steepest descent method LMS (Least Mean Square) is used to find a minimum value of an error evaluation function that can be calculated from an output value of the neural network. The control parameters of the network are adjusted.
[0014]
However, the error surface indicated by the error evaluation function of the neural network has a number of local minima in addition to the global minimum, and random extraction is performed. It will be done. For this reason, it is not possible to determine whether the detected value is a local minimum value or a global minimum value, which limits the adjustment of control parameters.
[0015]
On the other hand, recently, a neural network algorithm that obtains a global minimum by adding noise to the original signal has been announced (see US Pat. No. 2,284,152). This is to obtain a global minimum by adding noise to the original signal at a certain level of frequency and gradually reducing the degree of this noise.
[0016]
However, it is not possible to unambiguously determine how much noise should be introduced or how much noise should be reduced, and in some cases the minimum value may not be obtained.
[0017]
Accordingly, an object of the present invention is to provide an active control device that can achieve an accurate learning speed and control accuracy while being an active control device using a neural network.
[0018]
[Means for Solving the Problems]
In order to solve the above-described problem, the invention according to claim 1 includes a control device using a neural network that corrects a control parameter by comparing an output value and a control target value, and the control device uses an initial value of the control parameter. The value is set at random, the sensitivity of the output value to the control target value at the initial value is calculated, the allowable range of the control parameter is divided into a plurality of areas, and the control is performed in each area based on the sensitivity. The local minimum value of the difference between the output value and the control target value is detected in each region by selecting a random value of the parameter, and the minimum value is calculated from the local minimum values obtained from each region. Make the above corrections as initial values.In detecting the minimum value of the difference between the output value and the control target value in each area, a plurality of random values of the control parameter are selected in each area, and the selected number is used as a count number. The detection in the area is repeated a number of times less than or equal to the count number, and the count number in each area is changed according to the detection value obtained by repeating the detection in each area.It is characterized by the following.
[0019]
The invention of claim 2 isOutput value andA control device using a neural network for correcting a control parameter by comparing with a control target value, the control device randomly setting an initial value of the control parameter, and an output value for the control target value at the initial value Is calculated, the allowable range of the control parameter is divided into a plurality of regions, and a random value of the control parameter is selected in each of the regions based on the sensitivity, and a difference between an output value and a control target value is minimized. The valueDetected in each area,A minimum value is calculated from the minimum values obtained from the respective areas, and the correction is performed using the minimum value as a next initial value.I
In detecting the minimum value of the difference between the output value and the control target value in each area, a plurality of random values of the control parameter are selected in each area, and the selected number is used as the count number, and in each area, Is detected a number of times or less and is repeated a plurality of times. When it is determined that a minimum value cannot be obtained during the repetition of the detection in each of the regions, the corresponding region is registered in the restriction list, and the minimum value in the region is determined. Stop detecting valuesIt is characterized by the following.
[0020]
The invention of claim 3 isA control device using a neural network that corrects a control parameter by comparing an output value and a control target value, the control device randomly sets an initial value of the control parameter, and sets a control target value at the initial value. Calculate the sensitivity of the output value with respect to, divide the allowable range of the control parameter into a plurality of regions, select a random value of the control parameter in each of the regions based on the sensitivity, the output value for the control target value The minimum value of the difference is detected in each area, the minimum value is calculated from the minimum values obtained from each area, the correction is performed using the minimum value as the next initial value, and the output value with respect to the control target value is calculated. In detecting the minimum value of the difference in each area, a plurality of random values of the control parameters are selected in each area, and the selected number is used as a count number, and the detection in each area is counted. Repeated as and more than once in a few less, in response to said detected values obtained in the repetition of detection in each area, performs the weighting in a direction against the initial value of the random value of the control parameter to choose in each areaIt is characterized by the following.
[0021]
The invention of claim 4 isAn actuator that controls at least one of a sound or vibration state, a sound and vibration state detection unit that detects at least one of a sound and a vibration state, and an output of the sound and vibration state detection unit that is input and outputs a signal that controls the actuator And a control device using a neural network that corrects a control parameter by comparing a control predicted value based on the output signal with a control target value, and the control device randomly sets an initial value of the control parameter. Set, calculate the sensitivity of the control predicted value to the control target value at the initial value, divide the allowable range of the control parameter into a plurality of regions, and randomize the control parameter in each region based on the sensitivity. The minimum value of the difference between the control predicted value and the control target value is detected in each region in parallel, and the minimum value obtained from each region is determined. The minimum value is calculated from the above, the correction is performed with the minimum value as the next initial value, and the minimum value of the difference between the output value and the control target value is detected in each region. A plurality of random values are each selected, and the selected number is used as a count number, and the detection in each area is repeated a number of times less than the count number and a plurality of times, and is obtained by repeating the detection in each area.It is characterized in that the number of counts in each of the regions is changed according to the detected value.
[0022]
The invention of claim 5 isAn actuator that controls at least one of a sound or vibration state, a sound and vibration state detection unit that detects at least one of a sound and a vibration state, and an output of the sound and vibration state detection unit that is input and outputs a signal that controls the actuator And a control device using a neural network that corrects a control parameter by comparing a control predicted value based on the output signal with a control target value, and the control device randomly sets an initial value of the control parameter. Set, calculate the sensitivity of the control predicted value to the control target value at the initial value, divide the allowable range of the control parameter into a plurality of regions, and randomize the control parameter in each region based on the sensitivity. The minimum value of the difference between the control predicted value and the control target value is detected in each region in parallel, and the minimum value obtained from each region is determined. The minimum value is calculated from the above, the correction is performed with the minimum value as the next initial value, and the minimum value of the difference between the output value and the control target value is detected in each region. A plurality of random values are each selected, and the selected number is set as a count number, and the detection in each area is repeated a number of times less than or equal to the count number, and the detection in each area is performed.When it is determined that the minimum value cannot be obtained during the repetition of the above, the corresponding area is registered in the restriction list, and the detection of the minimum value in the area is stopped.
[0023]
The invention of claim 6 isAn actuator that controls at least one of a sound or vibration state, a sound and vibration state detection unit that detects at least one of a sound and a vibration state, and an output of the sound and vibration state detection unit that is input and outputs a signal that controls the actuator And a control device using a neural network that corrects a control parameter by comparing a control predicted value based on the output signal with a control target value, and the control device randomly sets an initial value of the control parameter. Set, calculate the sensitivity of the control predicted value to the control target value at the initial value, divide the allowable range of the control parameter into a plurality of regions, and randomize the control parameter in each region based on the sensitivity. The minimum value of the difference between the control predicted value and the control target value is detected in each region in parallel, and the minimum value obtained from each region is determined. The minimum value is calculated from the above, the correction is performed with the minimum value as the next initial value, and the minimum value of the difference between the output value and the control target value is detected in each region. A plurality of random values are each selected, and the selected number is set as a count number, and the detection in each area is repeated a number of times less than or equal to the count number, and the detection in each area is performed.The weighting is performed on the directionality of the control parameter selected in each of the regions with respect to the initial value of the random value in accordance with the detection value obtained in the repetition of the above.
[0024]
The invention of claim 7 is the claim4To any one of claims 6 to1 itemThe active control device according to the above,The actuator is a control sound source for reducing a noise by generating a control sound that interferes with a vehicle interior noise, and the sound vibration state detection means is a noise state detection means for detecting a noise state in the vehicle interior.It is characterized by being.
[0025]
The invention of claim 8 is the claim4~ Claim6Any of1 itemThe active control device according to the above,The actuator is a vibration damping device that reduces vibration by applying a control force to a vehicle suspension, and the sound vibration state detection unit is a vibration state detection unit that detects a vibration state of the vehicle suspension.It is characterized by the following.
[0026]
The invention of claim 9 is the claim4To any one of claims 8 to1 itemThe active control device according to the above,The neural network of the control device is composed of a control neural network and an identification neural network.The control neural network receives the output of the sound vibration state detecting means and outputs a signal for controlling the actuator, and outputs the identification neural network. The net outputs the control predicted value based on the response of the vehicle under the control of the actuator identified in advance.It is characterized by the following.
[0027]
The invention of claim 10 is the claim4~ Claim9Any of1 itemThe active control device according to the above,The control device includes a neuron element for inputting at least a control signal of the actuator as data before the current time detected by the sound vibration state detecting means, and outputs a control signal of the current actuator, and outputs the control signal of the current time. Predict and output the control predicted value afterIt is characterized by that.
[0028]
The invention of claim 11 is the claim1To any one of claims 10 to1 itemThe active control device according to claim 1, wherein the controlThe parameter is a learning ratio when learning the connection weight of the neural network.It is characterized by the following.
[0029]
The invention of claim 12 is the claim1~ Claim11Any of the descriptions1 itemAn active control device,A plurality of the neural nets are provided in parallel according to the number of steps which is the number of the regions.It is characterized by the following.
[0030]
[Action]
According to the first aspect of the present invention, a control device using a neural network corrects a control parameter by comparing an output value with a control target value. In this case, the control device randomly sets the initial value of the control parameter, calculates the sensitivity of the output value to the control target value at the initial value, divides the allowable range of the control parameter into a plurality of regions, and Select a random value of the control parameter in each area based on the minimum value of the difference between the output value and the control target value in each area, and calculate the minimum value from the minimum values obtained from each area. The correction can be performed using the minimum value as the next initial value.
[0031]
In other words, the allowable range of the control parameter is divided into a plurality of regions, and the minimum value is detected from each of the regions, and the smallest value is calculated as the minimum value. The control parameter can be corrected by obtaining the value.Further, upon detection of the local minimum value in each area, a plurality of random values of the control parameters can be selected in each area, and the selected number can be repeated a plurality of times as the count number. Therefore, the probability of finding the global minimum increases. Further, the count number in each area can be changed according to the detection value obtained by repeating the detection in each area. For this reason, for example, even if a large number of counts remain, if the detected values are largely converged, the counts can be reduced in the subsequent processing to speed up the processing.
[0032]
According to the invention of claim 2, the control device using the neural network corrects the control parameter by comparing the output value with the control target value. In this case, the control device randomly sets the initial value of the control parameter, calculates the sensitivity of the output value to the control target value at the initial value, divides the allowable range of the control parameter into a plurality of regions, and Select a random value of the control parameter in each area based on the minimum value of the difference between the output value and the control target value in each area, and calculate the minimum value from the minimum values obtained from each area. The correction can be performed using the minimum value as the next initial value. In other words, the allowable range of the control parameter is divided into a plurality of regions, and the minimum value is detected from each of the regions, and the smallest value is calculated as the minimum value. The control parameter can be corrected by obtaining the value.Further, upon detection of the local minimum value in each area, a plurality of random values of the control parameters can be selected in each area, and the selected number can be repeated a plurality of times as the count number. Therefore, the probability of finding the global minimum increases. Further, when the detection value does not converge due to the repetition of the count number in each area, the corresponding area can be registered in the constraint list and the detection of the minimum value in the area can be stopped. Therefore, it is possible to continue the detection by focusing on an area where a local minimum value can be obtained.
[0033]
According to the third aspect of the present invention, the control device using the neural network corrects the control parameters by comparing the output value with the control target value. In this case, the control device randomly sets the initial value of the control parameter, calculates the sensitivity of the output value to the control target value at the initial value, divides the allowable range of the control parameter into a plurality of regions, and Select a random value of the control parameter in each area based on the minimum value of the difference between the output value and the control target value in each area, and calculate the minimum value from the minimum values obtained from each area. The correction can be performed using the minimum value as the next initial value. In other words, the allowable range of the control parameter is divided into a plurality of regions, and the minimum value is detected from each of the regions, and the smallest value is calculated as the minimum value. The control parameter can be corrected by obtaining the value.Further, upon detection of the local minimum value in each area, a plurality of random values of the control parameters can be selected in each area, and the selected number can be repeated a plurality of times as the count number. Therefore, the probability of finding the global minimum increases. In addition, it is possible to weight the directionality with respect to the initial value of the random value of the control parameter selected in each area according to the detection value obtained by repeating the count number in each area. Therefore, in each area, a random value of the control parameter can be set in a direction in which the detected value converges more.
[0034]
In the invention of claim 4,The sound vibration state detecting means detects the sound vibration state of the object, and when the output is input to the control device, a signal generated by the neural network is output. With this signal output, the actuator controls the sound vibration state of the object. Further, the control prediction value based on the output signal is output by the neural network, and the control prediction value is compared with the control target value. As a result, the control parameters of the neural network are corrected, and the learning of the neural network is adaptively performed. Done. At this time, the control device randomly sets the initial value of the control parameter, calculates the count of the control predicted value for the control target value at the initial value, divides the allowable range of the control parameter into a plurality of regions, Select the random value of the control parameter in each area based on the sensitivity, detect the minimum value of the difference between the control target value and the control predicted value in each area, and determine the minimum value from the minimum values obtained from each area. Then, the control parameter can be corrected using the minimum value as the next initial value. That is, when correcting the control parameters using the neural network, the minimum value is detected from the minimum value detected for each region, and the minimum value is calculated. The minimum value can be determined accurately. Then, the control parameter can be accurately corrected by the global minimum value. Further, upon detection of the local minimum value in each area, a plurality of random values of the control parameters can be selected in each area, and the selected number can be repeated a plurality of times as the count number. Therefore, the probability of finding the global minimum increases. Further, the count number in each area can be changed according to the detection value obtained by repeating the detection in each area. For this reason, for example, even if a large number of counts remain, if the detected values are largely converged, the counts can be reduced in the subsequent processing to speed up the processing.
[0035]
In the invention of claim 5,The sound vibration state detecting means detects the sound vibration state of the object, and when the output is input to the control device, a signal generated by the neural network is output. With this signal output, the actuator controls the sound vibration state of the object. Further, the control prediction value based on the output signal is output by the neural network, and the control prediction value is compared with the control target value. As a result, the control parameters of the neural network are corrected, and the learning of the neural network is adaptively performed. Done. At this time, the control device randomly sets the initial value of the control parameter, calculates the count of the control predicted value for the control target value at the initial value, divides the allowable range of the control parameter into a plurality of regions, Select the random value of the control parameter in each area based on the sensitivity, detect the minimum value of the difference between the control target value and the control predicted value in each area, and determine the minimum value from the minimum values obtained from each area. Then, the control parameter can be corrected using the minimum value as the next initial value. That is, when correcting the control parameters using the neural network, the minimum value is detected from the minimum value detected for each region, and the minimum value is calculated. The minimum value can be determined accurately. Then, the control parameter can be accurately corrected by the global minimum value. Further, upon detection of the local minimum value in each area, a plurality of random values of the control parameters can be selected in each area, and the selected number can be repeated a plurality of times as the count number. Therefore, the probability of finding the global minimum increases. Further, when the detection value does not converge due to the repetition of the count number in each area, the corresponding area can be registered in the constraint list and the detection of the minimum value in the area can be stopped. Therefore, it is possible to continue the detection by focusing on an area where a local minimum value can be obtained.
[0036]
In the invention of claim 6,The sound vibration state detecting means detects the sound vibration state of the object, and when the output is input to the control device, a signal generated by the neural network is output. With this signal output, the actuator controls the sound vibration state of the object. Further, the control prediction value based on the output signal is output by the neural network, and the control prediction value is compared with the control target value. As a result, the control parameters of the neural network are corrected, and the learning of the neural network is adaptively performed. Done. At this time, the control device randomly sets the initial value of the control parameter, calculates the count of the control predicted value for the control target value at the initial value, divides the allowable range of the control parameter into a plurality of regions, Select the random value of the control parameter in each area based on the sensitivity, detect the minimum value of the difference between the control target value and the control predicted value in each area, and determine the minimum value from the minimum values obtained from each area. Then, the control parameter can be corrected using the minimum value as the next initial value. That is, when correcting the control parameters using the neural network, the minimum value is detected from the minimum value detected for each region, and the minimum value is calculated. The minimum value can be determined accurately. Then, the control parameter can be accurately corrected by the global minimum value. Further, upon detection of the local minimum value in each area, a plurality of random values of the control parameters can be selected in each area, and the selected number can be repeated a plurality of times as the count number. Therefore, the probability of finding the global minimum increases. In addition, it is possible to weight the directionality with respect to the initial value of the random value of the control parameter selected in each area according to the detection value obtained by repeating the count number in each area. Therefore, in each area, a random value of the control parameter can be set in a direction in which the detected value converges more.
[0037]
In the invention of claim 7, the claim4To any one of claims 6 toItem 1In addition to the action of the invention ofWhen the noise state detecting means detects the noise state in the vehicle interior, the output is input to the control device, and the signal generated by the neural network is output. With this output signal, the control sound source is controlled to generate a control sound that interferes with the vehicle interior noise, thereby reducing the noise. In addition, the control prediction value based on the output signal is output by the neural network, and the control prediction value is compared with the control target value. As a result, the control parameters of the neural network are corrected, and learning of the neural network is performed. it can. That is, even noise having random characteristics can be accurately controlled using a neural network.
[0038]
In the invention of claim 8, the claim4~ Claim6Any ofItem 1In addition to the action ofWhen the vibration state detecting means detects the vibration state of the suspension, its output is input to the control device, and a signal generated by the neural network is output. With this output signal, the vibration damping device is controlled and a control force is applied to the suspension of the vehicle, so that the vibration can be reduced. Further, the control prediction value based on the output signal is output by the neural network, and the control parameters of the neural network are corrected by comparing with the control target value, so that the neural network can be learned. That is, suspension vibration having a random property can be controlled using a neural network.
[0039]
In the invention of claim 9, the claim4To any one of claims 8 toItem 1In addition to the action ofThe output of the sound vibration state detecting means is input to the control neural network, and a signal for controlling the actuator is output. Further, the identification neural network can output a control prediction value based on a response of the vehicle identified in advance. Therefore, by comparing the control predicted value output from the identified neural network with the control target value, the control parameters of the control neural network can be corrected and the neural network can be learned. That is, by employing the identification neural network, it is possible to accurately output a predicted control value for correcting a control parameter of the control neural network.
[0040]
According to the tenth aspect of the present invention,4~ Claim9Any ofItem 1In addition to the action ofIn the control device, at least a control signal of the actuator is input as data before the present time detected by the sound vibration state detecting means, and a control signal of the present actuator is output. Further, it is possible to predict and output a control prediction value after the detection time. That is, since the control signal of the actuator is output based on the past data and the predicted control value is output, the control signal can be adapted to noise and vibration having random characteristics.
[0041]
In the invention according to claim 11, the claim1To any one of claims 10 toItem 1In addition to the action ofThe learning parameter when learning the connection weight of the neural network can be used as the control parameter. Therefore, when detecting the minimum value, the learning rate can be appropriately selected.
[0042]
In the invention of claim 12,1To the effect of any one of claims 11 to 11,A plurality of neural networks can be provided in parallel according to the number of steps, which is the number of regions. Therefore, the search can be performed by each neural net for each area.
[0043]
【Example】
Hereinafter, embodiments of the present invention will be described.
[0044]
FIG. 1 is a schematic diagram showing an entire active noise control device to which an active control device according to a first embodiment of the present invention is applied.
[0045]
In the cabin 1 shown in FIG. 1, noise from the engine, road noise from the suspension, exhaust noise from the muffler, wind noise, vibration of the vehicle body panel, and the like are combined to form a nonlinear coupled space. I have. A microphone 3 and a piezo element sensor 5 as sound vibration state detecting means (noise state detecting means) are provided to detect a noise state as a sound vibration state in the vehicle interior 1 as such a coupled space. I have.
[0046]
As an actuator for controlling the sound vibration state of the vehicle, a loudspeaker 7 which is a control sound source for generating a control sound that interferes with the noise in the vehicle compartment 1 of the vehicle to reduce the noise is provided.
[0047]
Further, in this embodiment, a piezo element actuator 9 is provided as an actuator for controlling the sound vibration state of the vehicle, which controls the vibration mode of the vehicle body by applying an external force to the vehicle body panel.
[0048]
The outputs of the microphone 3 and the piezo element sensor 5 are input to a control device 11 using a neural network, and the control device 11 outputs a signal for controlling the loudspeaker 7 and the piezo element actuator 9. .
[0049]
At the same time, by comparing a control predicted value based on the output signal with a control target value, a parameter of the neural network, for example, a connection weight is corrected.
[0050]
In this embodiment, the piezo element sensor 5 and the piezo element actuator 9 are provided in addition to the microphone 3 and the loudspeaker 7 in order to control the noise in the passenger compartment 1 as described above. This is because the noise in the passenger compartment 1 is a coupling between the vibration response of the vehicle body and the sound field in the passenger compartment 1.
[0051]
Also, for simplicity of description, it is assumed that only one microphone 3, piezo element sensor 5, loudspeaker 7, and piezo element actuator 9 are provided. , The loudspeakers 7 are arranged at the left and right doors respectively corresponding to the front and rear seats of the vehicle, and the piezo element sensors 5 and the piezo element actuators 9 are distributed on the vehicle body roof, the left and right doors and the like. It can also be provided.
[0052]
FIG. 2 is a schematic block diagram showing the control device 11. That is, the control device 11 includes a control neural network 13, an identification neural network 15, an evaluation / adjustment unit 17, a delay unit 19, and a storage unit (not shown).
[0053]
The control neural network 13 controls the output m of the microphone 3 and the piezo element cell.
[Outside 1]

Signal O for controlling the child actuator 9_c2Is output.
[0054]
The identified neural network 15 calculates a predicted control value O based on the response of the vehicle under the control of the actuator identified in advance._fIs output. That is, the output O of the control neural network 13_c1, O_c2And the output m of the microphone 3 and the like, and the control prediction value O_fIs output.
[0055]
The evaluation / adjustment unit 17 determines the output O of the identified neural network 15 after the identified neural network 15 has sufficiently represented the dynamic characteristics of the controlled object, that is, has sufficiently identified the response of the cabin 1._fAnd control target value X_TThen, the connection weight of the control neural network 13 is corrected by back propagation using this value.
[0056]
The delay units 19 and 21 and a storage unit (not shown) constitute a control neural network 13 and an identification neural network 15 based on an ARMA model.
[0057]
The delay unit 19 controls the control signal O such as (t-1), (t-2),... Before time t when the microphone 3 and the piezo element sensor 5 detect the noise state._c1, O_c2Is input to the control neural network 13. For this reason, the control neural net 13 has a neuron element for inputting a time-delayed (past) signal.
[0058]
The delay unit 21 outputs the output signal O of the identification neural network 15._fIs input to the identification neural network 15 and the control neural network 13 such as (t-1) before the current time t. Therefore, the two neural networks 13 and 15 output the output O of the identification neural network 15._f, Each having a neuron element for inputting past data.
[0059]
[Outside 2]

Control signal O_c1, O_c2Is output, and the identification neural network 15 outputs m, O_c1, O_c2(T + Δt) control prediction value O after the detection time_fAnd output. For this reason, even in a case where the external environment changes drastically, such as in a car, control can be performed in anticipation of the future, and a quick response can be taken. In addition, when the prediction is performed, since the prediction is always performed using the foremost value, an accurate prediction can be performed.
[0060]
3 (a) and 3 (b) show a part of the configuration of the neural network. FIG. 3 (a) shows the whole (b) showing one of the intermediate layers (for example, 25c).
[0061]
This neural network is composed of an input layer, an intermediate layer, and an output layer. The input layer shows two neuron elements 23a and 23b, the middle layer shows three neuron elements 25a, 25b and 25c, and the output layer shows one neuron element 27.
[0062]
The neural network described here has an input x in the neuron elements 23a and 23b of the input layer.₁, X₂To the coupling load W_ijThe neuron elements 25a, 25b, and 25c in the intermediate layer similarly have a connection weight H_jrIs attached. ij indicates that the j-th hidden layer is connected to the i-th input layer, and jr indicates that the r-th output layer is connected to the j-th hidden layer. .
[0063]
Therefore, the coupling load W_ijInternal potential O of each neuron₁Is
O₁= ΣW_ijx_i+ Θ_i  … (1)
It becomes. Here, Θ_iIndicates the threshold value of the neuron element.
[0064]
Also, the internal potential O₁Is determined as the output value to the neuron element. The output is calculated by, for example, a sigmoid function shown in FIG. this is,
(Equation 1)

It is expressed as
[0065]
Where u₀Shows the inclination of FIG. 4, and this inclination u₀Can be made to correspond to a random signal. The output is an output between 0 and 1. For example, if a physical quantity v = 0 to 100 m / sec is output,
(Equation 2)

To make the signal of 0 to 1.
[0066]
Coupling load W_ijAre corrected by the output O of the identification neural network 15._fAnd control target value X_TError with
E = Σ (X_T-O_f)²  … (4)
And W that minimizes the error E by the steepest descent method_ijIs determined and used as the connection load.
[0067]
Here, the output O of the neuron element of the input x is_fIs O_f= F (y), the error E is a function of f (y),_fPartial differential with O_fTo W_ijPartially differentiated by
(Equation 3)

And then O_fIs partially differentiated by f (y), and f (y) is_ijPartially differentiated by
(Equation 4)

It becomes. Therefore, f (y) and O_fW to minimize E in relation to_ijIs required.
[0068]
W that minimizes E in the above relationship_ijIs equivalent to implementing the steepest descent method, and the following equation holds.
[0069]
(Equation 5)

In such a relationship, when the control parameter η is large to some extent, the error E does not converge, and when the control parameter η is reduced, the error E converges at a local minimum value, so that a global minimum value cannot be detected. This will be described with reference to FIG. 5. FIG. 5 shows the error function f (y) as a change in the error E with respect to the input x. If the control parameter η is increased in the error function f (y), for example, the error function f (y) progresses from S to P, and it is difficult to converge to the minimum value R. Conversely, if η is small, for example, the point advances from the point P to the point Q and converges on the local minimum value Q, so that the global minimum value R may not be detected. Therefore, as described above, in a neural network equivalent to a simple steepest descent method, there is a limit in improving the learning speed and the control accuracy. Therefore, in one embodiment of the present invention, η, which is one of the control parameters, is determined by the gradient random search method described below. That is, in this embodiment, the algorithm is a combination of the steepest descent method and the random search method.
[0070]
FIG. 6 shows a schematic diagram of the random search method. Here, a problem of minimizing a certain objective function f (x) under a constraint condition is considered. In this random search method, two constants of the number of steps and the number of counts are newly defined. The number of steps is the number of neighboring areas to be searched, and the count number represents the upper limit of the number of times to search one neighboring area. First-order approximate solution (initial value) x of solution satisfying constraints₀And x₀Around the neighborhood N (x₀, H_i) Is set. Where H = h_i(I = 1,..., R) are steps. The idea of this search method is to generate x at random in each of the neighboring areas, and thereby f (x) becomes f (x₀), The point is stored as the local minimum value in the area, and the smallest x among the local minimum values selected in all the neighboring areas is the second approximate solution x.₁And x₁The search is repeated by providing a neighboring area again around.
[0071]
When the above method is applied to the problem of the minimum value of a one-variable function, the following is obtained.
[0072]
When the objective function f (x) is given and the range of the variable x is (a, b), the steps are sequentially reduced by 1/10, and if the search is performed with r steps, the step (H) Is expressed as a vector as follows.
[0073]
(Equation 6)

That is, as shown in the upper part of FIG.₁= Ba is the maximum area, and is reduced by 1/10, h_r= H_r-1It is divided into r regions, with /10.0 being the minimum. That is, a certain starting point x₀, A neighborhood area (N) is set for each step. Next, as shown in the middle part of FIG._0iOccurs. As a result, f (x₀)> F (x_0i) Then x_0iIs stored, otherwise the count K is added and the next random x in the same neighborhood_0iOccurs. If the count number K is larger than the given number, the process shifts to searching for another neighboring area (N). Then, as shown in the lower part of FIG._save(1), x_save(2), ... x_save(R) is detected, and the minimum value x is obtained from these minimum values._bestIs calculated and used as the initial value of the next solution. This method can cope with the case of multiple variables by considering the variables as vectors.
[0074]
Next, the random search will be described with reference to the flowchart of FIG.
[0075]
In the following description, step S is different from step h newly defined in the random search method.
[0076]
First, in step S1, step h_iIs set. This step h_iIs a setting of the number of neighboring areas to be searched as described above.
[0077]
In step S2, an area is set. With this area setting, the allowable range of x is divided into a plurality of areas.
[0078]
Next, initialization is performed in steps S3, S4, and S5. That is, the repetition number I = 0, the step number i = 1, and the count number K = 0 are set. The repetition number I is a number when the search is repeated again with the minimum value obtained by one region setting as an initial value of the next solution. The number i of steps is set as 1, and the first neighboring area of the area set in step S2 is selected. Then, in the first time, the process proceeds to steps S6 and S7, and in step S8, the search in the largest area is performed. That is, the initial value x_0iRandomly occur, and a comparison with the objective function f (x) is performed in step S6. f (x₀)> F (x_0i) Is the initial value x_0iIs stored as the minimum value in step S9.
[0079]
If NO is determined in the step S6, it is determined whether or not the count number K is larger than a given number in a step S7, and if not, x is generated again in the same area in a step S8 and a search is performed. In such a search, when the minimum value is not found and the count number becomes larger than the given number, the process proceeds to step S10, and it is determined whether or not the number i of the steps h exceeds the set number r. Is smaller than r, the number of i is added, and the routine goes to Step S5.
[0080]
In step S5, K = 0 is set again, and the process proceeds to steps S6, S7, and S8, and a search in the next area is performed. Even when the minimum value is stored in step S9, the determination in step S10 is performed, the number i of steps is added, and the search in the next area is performed.
[0081]
In this manner, when the search in all the steps is completed, in step S11 the smallest value among the minimum values is set to the minimum value x = x_bestIs calculated and stored. Next, the minimum value x stored in step S11_bestTo the second approximate solution x₁And x₁Then, the neighborhood area is set again, and I is added unless I exceeds the number of repetitions in step S13, and the process proceeds to step S4. Therefore, the search is performed again in the new area setting, and the minimum value is calculated.
[0082]
If it is determined in step S13 that the number of repetitions is satisfied, the search ends.
[0083]
By such processing, the minimum value of the control parameter η is searched. That is, in FIGS. 6 and 7, the replacement process is performed with x as the control parameter η. Explaining the control parameter η, the control device 11 sets the initial value η of the control parameter η._0iIs set at random and the initial value η_0iCalculate the sensitivity of the output value to the control target value at. Next, the allowable range of the control parameter η is divided into a plurality of regions, a random value of the control parameter η is selected in each region based on the sensitivity, and the minimum value of the difference between the output value and the control target value is determined in each region. The detection is performed, the minimum value is calculated from the minimum values obtained from the respective regions, and η is corrected using the minimum value as the next initial value.
[0084]
Therefore, as described above, since the solution of η is searched by dividing into multiple regions having different sizes, blind search can be avoided and the number of times required for the search can be reduced. In addition, since the local minimum value is detected in each of the multiple regions and the minimum value is calculated from the local minimum value, it is possible to prevent convergence to a local local minimum value, and it is possible to obtain a global minimum value. In this manner, an appropriate value of η, which is a learning rate of the neural network, can be obtained, and processing can be performed accurately and at high speed.
[0085]
Next, a one-dimensional function is defined with the effect of the random search method as an objective function f (x), and further description will be given.
[0086]
This one-dimensional function is
(Equation 7)

Defined by The shape of this function is as shown in FIG. X (= x which gives the minimum value of the function_best) Is 17.039, and the function value f (= f_best) Is 0.094. FIG. 9 shows the result of searching for this problem using the steepest descent method, the random search method, and a method combining the two. Where the error E is
E = (ff_best) / F_best
f_best= Optimal value
And the initial value of x is 13.5, 13.4, 5.0 as shown in the figure. According to FIG. 9, the steepest descent method has a case in which the solution converges by setting the initial value (global minimum value) and a case in which the solution does not converge (local local minimum value). In the random search method, the solution converged with high accuracy for any initial value.
[0087]
Furthermore, when the two methods are combined, the dependence on the initial value, which was observed only in the steepest descent method, is sufficiently overcome, and convergence is faster than in the case of the random search method alone. From these results, it was confirmed that the random search method is an effective method for solving the optimization problem, and that the effectiveness is further increased by combining the random search method with another algorithm.
[0088]
Also consider how the newly defined constants, step and count, used in the random search method will affect the convergence to the optimal solution. FIGS. 10 and 11 show the results of comparing the convergence of the solutions when these two constants are changed. The objective function uses the above f (x). As is clear from FIGS. 10 and 11, as the number of steps increases, a more precise and precise solution search becomes possible. As the count increases, the error is minimized, that is, the probability of reaching the global minimum increases. However, care must be taken with these two ideas because if these two constants are excessively increased, useless search time or useless memory area will be used during programming. Needs great care.
[0089]
FIGS. 12 to 20 are for further explaining the effectiveness of the random search method. That is, numerical simulation results of a spring-mass model having one degree of freedom and a beam having multiple degrees of freedom are shown.
[0090]
First, FIGS. 12 to 16 show responses by displacement excitation in a one-degree-of-freedom system. Considering the one-degree-of-freedom spring-mass model shown in FIG. 12, a transient response after a certain time has passed after applying a sinusoidal external force to this system is shown in FIG. The natural angular frequency was about 100 rad / sec, and the time step was 1 msec. As the input to the neural network, the response speed and acceleration of the system were used, and the change in acceleration at the next time step was output. The teacher signal is shown in FIG. 14, FIG. 15 shows the convergence, and FIG. 16 shows the identification result. From the results of FIGS. 15 and 16, it was confirmed that the error convergence of the random search method was extremely fast. This indicates that highly accurate learning can be performed in a very short time, and it has been confirmed that the mapping ability of the neural network using the random search method is extremely superior to the steepest descent method.
[0091]
17 to 20 attempted to identify the impulse response of the cantilever online. The input / output of the neural network and the outline of the system are shown in FIG. In FIG. 18, the input of the neural network is the acceleration and velocity of the tip (x6) of the beam, and the converted acceleration is output. Here, the response of the cantilever, that is, the teacher signal is obtained by numerical calculation. FIG. 19 shows the convergence of the error, and FIG. 20 shows the identification results in the vicinity of the simulation time of 100 seconds. As is clear from FIGS. 19 and 20, in the case of using the steepest descent method and the case of using the method combining the random search method and the steepest descent method in online identification, the convergence of the learning time and the error is considerably large. There was a difference. It was also confirmed that the identification result was extremely accurate in the gradient random search method.
[0092]
The processing of the random search method described above can be performed in order by one neural network, but can also be distributed by a parallelized neural network as shown in FIG. That is, the neural network includes a plurality of units 1 to r each having an input layer, an intermediate layer, and an output layer. The number of units is the number of areas to be divided and set. Therefore, distributed processing control can be performed by each unit for each area. That is, the distributed processing control system 61 performs information control of search in the entire area and distributed control of each unit. Then, a search in each area is performed for each unit, and a global minimum value is calculated by 63. Then, the control parameters of each unit are corrected by back propagation based on the minimum value. By such parallel processing, the processing time can be extremely reduced, and higher-speed processing can be performed. Also, when a plurality of neural nets are arranged side by side in this way, the processing of one unit is completed, and if the processing of another unit is not completed, the completed unit can support another unit. , Can be further speeded up.
[0093]
In the above processing, the control device 11 can also change the count number in each area according to the detection value obtained by repeating the count number in each area. In other words, even if a large number of counts remain, if the detected values are largely converged, the counts can be gradually reduced in the subsequent processing to speed up the processing.
[0094]
If the control device 11 determines that a minimum value cannot be obtained by repeating the count number in each area, the control apparatus 11 may register the corresponding area in a restriction list and stop detection of the minimum value in the area. With this configuration, it is possible to stop searching for an area in which it is apparent that a minimum value cannot be obtained, and to further speed up the processing.
[0095]
Further, the control device 11 can weight the directionality of the control parameter selected in each region with respect to the initial value of the random value, in accordance with the detection value obtained by repeating the count number in each region. That is, in a region as shown in FIG._0iGenerates a random value around the center, but as the count number is repeated, the direction of the local minimum with respect to the initial value gradually becomes clear in the area, so many random values in that direction , It is possible to perform more accurate detection.
[0096]
FIG. 22 generally shows a neural network based on the ARMA model. This neural network has an input layer, an intermediate layer, and an output layer, and predicts and outputs an output y (t + 1) at time (t + 1) based on an input x (t) at time t. That is, the predicted output y (t + 1) is related to the current input x (t) and past states. For this reason, a neuron element connected to the three short-term storage units 29, 30, 31 is newly added to the input layer.
[0097]
The storage unit 29 stores the past data x (t−1) to x (t−n) of the input x (t) and inputs signals to the corresponding neuron elements. The storage unit 30 stores the past data y (t) to y (t−1) of the output y (t + 1) and inputs the data to the corresponding neuron elements. The storage unit 31 stores the past data z (t−1) to z (t−g) of the output y (t + 1) and inputs them to the corresponding neuron elements.
[0098]
The signal of the input x (t) is a prediction coefficient α₀, And input to the neuron element of the intermediate layer, and each signal of the storage unit 29 has a prediction coefficient α₁From α_nAnd the signal of the storage unit 30 has a prediction coefficient β₀From β₁And the signal of the storage unit 31 has a prediction coefficient γ_{l + 1}To γ_g, Which are input to the neuron elements of the intermediate layer.
[0099]
The relationship between the output y (t + 1) of the neural network and the detected state quantity x (t + 1) is given by a prediction error e (t + 1).
x (t + 1) = y (t + 1) + e (t + 1)
It becomes. In this case, the output y (t + 1) is
(Equation 8)

Is represented by
[0100]
Therefore, the prediction coefficients α, β, and γ are corrected to appropriate values by applying the least squares estimation method to the error sequence.
[0101]
Next, the operation of the active control device will be described.
[0102]
FIG. 23 schematically shows the operation according to the first embodiment of the present invention. The active control device generally operates as shown in FIG. Referring also to FIG. 1, first, the response of a representative point in the vehicle compartment 1 is accurately grasped by the piezo element sensor 5 and the microphone 3, and the dynamic characteristic of the system is more accurately identified by the control device 11. The control device 11 performs a coupled analysis based on the identification of the system, and controls the piezo element actuator 9 and the loudspeaker 7. That is, referring to FIG. 2, after the identification neural network 15 sufficiently expresses the dynamic characteristic with respect to the noise in the vehicle cabin 1, the evaluation / adjustment unit 17 outputs the output O of the identification neural network 15._fAnd control target value X_TIs calculated, and the connection weight of the control neural network 13 is corrected by back propagation using this value.
[0103]
FIG. 24 shows a time chart of the control with respect to the response in the passenger compartment 1. That is, at time t, the state in the passenger compartment 1 is sensed by the piezo element sensor 5 and the microphone 3, the control device 11 outputs signals to the loudspeaker 7 and the piezo element actuator 9 at time t, and at time t + Δt The noise control within 1 is realized.
[0104]
Then, in order to correct the coupling load by feeding back the control result to the neural network of the control device 11, the input of the identification neural network 15 is a noise state amount in the vehicle compartment 1 at a certain time step t and a control amount corresponding thereto, The output is the state quantity in the next step (t + Δt). The output O of this identification neural network 15_f, The control evaluation function E of the identification neural network 15_f, Control evaluation function E of control neural network 13_cIs
(Equation 9)

Where X: current state
X_T: Nostalgic condition
O_f: State predicted by neuro
It becomes.
[0105]
X is a detection state quantity at time step t + Δt, and X_TIs a state quantity to be a control target. Therefore, if the identification neural network 15 accurately identifies the dynamic characteristics with respect to the noise in the passenger compartment 1, the evaluation function E_fIs equal to or less than a predetermined value, and the output O of the identification neural network 15 is_fAnd target state quantity X_TAnd the error E_cAre all generated by the control neural network 13.
[0106]
Therefore, as shown in FIG. 25, the error information is back-propagated and returned to the control neural network 13, thereby making it possible to adjust the entire control device 11.
[0107]
Next, the control operation of the first embodiment of the present invention will be described with reference to the flowchart of FIG.
[0108]
First, in step S21, the control neural network 13 and the identification neural network 15 are initialized. This initialization is based on the connection weight W, the threshold Θ, and the slope coefficient u of the sigmoid function.₀Is determined using a random function.
[0109]
In step S22, initial learning of the identification neural network 15 is performed. This initial learning is performed off-line and is performed by a predetermined program. When the initial learning in step S22 is completed, the identification neural network 15 is incorporated in the control device 11 as shown in FIG.
[0110]
After the initial learning of the identification neural network 15 is completed, the control neural network 13 is operated and its output O_cIs output in step S23. The loudspeaker 7 and the piezo element actuator 9 are driven by this output, and the output O_cIs input to the identification neural network 15. As a result, in step S24, the identification neural network 15 outputs the predicted state amount O at time t + Δt._fIs output.
[0111]
In step S25, actual measurement data of the state quantity X at time t + Δt is sampled. Further, in this embodiment, since the ARMA model is used, past state quantities are written in the storage device.
[0112]
In step S26, identification is determined. That is, in step S26, it is determined whether or not the identification neural network 15 has accurately identified the dynamic characteristic relating to the noise in the passenger compartment 1. Specifically, the actually measured state quantity X and the predicted output O_fAnd the evaluation function E of the identification neural network 15 calculated as_fIs the convergence criterion T_fDetermine if it is greater than. E_fIs T_fIf it is larger, the identification neural network 15 has not sufficiently identified the system, so the process proceeds to step S27, and the evaluation / adjustment unit 17 corrects and learns the connection weight W of the identification neural network 15.
[0113]
Next, the process proceeds to step S28, and the adaptive control is determined. This determination is for determining whether the noise state in the passenger compartment 1 is at the target state. Specifically, the target state quantity X_TAnd the output O of the identification neural network 15_fAnd the evaluation function E calculated as_cIs the convergence criterion T_cDetermine if it is greater than. E_cIs T_cIf it is larger, the control neural network 13 has not been sufficiently identified, and the flow shifts to step S29 to correct and learn the connection weight W of the control neural net 13 by back propagation.
[0114]
Thereafter, steps S23 to S29 are repeated, and the noise in the passenger compartment 1 is controlled by feedforward.
[0115]
That is, as shown in FIGS. 1 and 2, the identification neural network 15 has previously identified the dynamic characteristics of the noise in the passenger compartment 1 offline. And the control neural net
[Outside 3]

Output signal O of the control neural network 13_c1, O_c2Of the past and the output O of the identification neural network 15_fIs input and the control signal O_c1, O_c2Is output. Here, the control neural network 13 operates and outputs after data is input to all input neuron elements provided in advance to input past data in order to reflect past data.
[0116]
And output O_c1As a result, the loudspeaker 7 is driven, and a signal having a phase opposite to that of the noise is output into the passenger compartment 1. Also, output O_c2Accordingly, the piezo element actuator 9 is driven, and the vibration mode of the vehicle body is controlled. As a result, the non-linear coupling space of the passenger compartment 1 is controlled, and noise can be reduced.
[0117]
Further, the identification neural network 15 outputs the output O of the control neural network 13._c1, O_c2And output m and output O of the microphone 3 as the state quantity in the passenger compartment 1._fIs input and the output O at time t + Δt is input._fAnd output. This output O_fIs based on the current noise state and external force in the cabin 1 and past data, and it can be said that the noise state due to random input is accurately predicted.
[0118]
And output O_fIs compared with the actual measurement data X of the noise state in the passenger compartment 1 at the present time t, and learning of the identification neural network 15 is performed. Assuming that the identification neural network 15 accurately identifies the dynamic characteristics of the system, the output O of the identification neural network 15_fAnd target state quantity X_TThus, the evaluation / adjustment unit 17 learns the control neural network 13.
[0119]
FIG. 27 shows a specific configuration of the neural network according to the first embodiment of the present invention.
[0120]
As described above, this neural network considers the ARMA model, and includes the sound field control neural network 13a and the structure control neural network 13b as the control neural network 13. Each neural network is composed of an input layer, an intermediate layer, and an output layer. The input layer is indicated by a white circle, the intermediate layer is indicated by a hatched circle, and the output layer is indicated by a black circle.
[0121]
The input of the sound field control neural network 13a is a sound pressure sensing signal m = m of n microphones at time t.₁ ^s(T) to m_n ^s(T) and past data m from time (t-1) to (tk)₁ ^a(T-1)-m₁ ^a(Tk), m_n ^a(T-1)-m_n ^a(Tk) and the control force S in the past from (t-1) to (tk) of the n loudspeakers.₁ ^a(T-1)-S₁ ^a(Tk), S_n ^a(T-1)-S_n ^a(Tk) are input to each neuron element in the input layer. Here, the upper suffix “s” indicates a detection signal, and “a” indicates a processed signal by a neural network.
[0122]
Output O of sound field control neural network 13a at current time t_c1Is
O_c1= S₁ ^a(T)-S_n ^a(T)
Are output as control force to n loudspeakers and are also input to corresponding neuron elements in the input layer of the identification neural network 15.
[0123]
This output O_c1= S₁ ^a(T)-S_n ^aAs described above, the past data of (t) is stored and input to the input layer of the sound field control neural network 13a by the delay unit 19.
[0124]
The input layer of the identification neural network 15 has a microphone sound pressure sensing signal m = m at the current time t.₁ ^s(T) to m_n ^s(T) and past data m of the sound pressure sensing signal₁ ^a(T-1)-m₁ ^a(Tk), m_n ^a(T-1)-m_n ^a(Tk) is input.
[0125]
The input layer of the structure control neural network 13b has n piezo elements.
[Outside 4]

) Is entered. Also, the past control signals P of the n piezo element actuators₁ ^a(T-1)-P₁ ^a(Tk), P_n ^a(T-1)-P_n ^a(Tk) has been input.
[0126]
Output O of the structured neural network 13b at time t_c2Is
O_c2= P₁ ^a(T) -P_n ^a(T)
It is. This output O_c2= P₁ ^a(T) -P_n ^a(T) is output to the n piezo element actuators and is input to the input layer of the identification neural network 15.
[0127]
Also, output O_c2= P₁ ^a(T) -P_n ^a(T) is sequentially stored in the same manner as described above, and is input to the input layer of the structure control neural network 13b as past data as described above.
[0128]
The input layer of the identification neural network 15 includes the detected vehicle body at time t.
[Outside 5]

You.
[0129]
Then, from the identification neural network 15, the predicted value of the microphone detection signal at the time t + Δt is obtained.
O_f= M₁ ^a(T + Δt) to m_n ^a(T + Δt)
Is output as the predicted value of the detection of the piezo element sensor.
(Equation 10)

Is output.
[0130]
The output O of the identification neural network 15_f= M₁ ^a(T + Δt) to m_n ^a(T + Δt) is sequentially stored and input to the input layers of the identification neural network 15 and the sound field control neural network 13a as past data.
[0131]
[Outside 6]

_n ^a(T + Δt) is also sequentially stored and input as past data to the input layer of the identification neural network 15 and the input layer of the structure control neural network 13b, respectively.
[0132]
Therefore, the evaluation function E of the identification neural network 15_fIs
(Equation 11)

It becomes.
[0133]
The control evaluation function E of the sound field control neural network 13a_cIs
(Equation 12)

It becomes.
[0134]
Further, the control evaluation function E of the structure control neural network 13b_cIs
(Equation 13)

It becomes.
[0135]
Therefore, these evaluation functions E_f, E_cAs described above, the correction and learning of the connection weight W between the identification neural network 15 and the control neural network 13 can be performed appropriately.
[0136]
In the first embodiment, a neural network considering the ARMA model is used for both the identification neural network and the control neural network.
[0137]
FIG. 28 shows a second embodiment of the present invention. In this embodiment, an active control device is applied to vibration control of an automobile suspension.
[0138]
The vehicle body 32 is grounded on a road surface 37 by a front wheel 33 and a rear wheel 35. In such a suspension, the spring constant between the front wheel 33 and the vehicle body 32 is K_F2, The damping coefficient is C_F1And the spring constant between the front wheel 33 and the road surface 37 is K_F1, The damping coefficient is C_F1It is. The spring constant between the rear wheel 35 and the vehicle body 32 is K_R2And the damping coefficient is C_R1It is. The spring constant between the rear wheel 35 and the road surface 37 is K_R1, The damping coefficient is C_R1It is.
[Outside 7]

is there.
[0139]
The control force P is applied between the front wheel 33 and the vehicle body 32._FA vibration damping device 39 is provided to reduce vibration by giving a control force P between the rear wheel 35 and the vehicle body 32._RAnd a vibration damping device 41 for reducing vibration.
[0140]
The control device 40 includes the unevenness Y of the road surface 37 on the front wheel 33 side._0FOf the front wheel side road surface detector 43 and the rear wheel side road surface detector 45 detecting the unevenness of the road surface on the rear wheel 35 side.
[0141]
Output signals from a front wheel vibration acceleration detector 47 for detecting the vibration acceleration of the front wheel 33 and a rear wheel vibration acceleration detector 49 for detecting the vibration acceleration of the rear wheel 35 are input.
[0142]
Further, signals are input from a front vibration acceleration detector 51 for detecting a front vibration acceleration of the vehicle body 32 and a rear vibration acceleration detector 53 for detecting a rear wheel vibration acceleration.
[0143]
These detectors 43, 45, 47, 49, 51, 53 constitute a vibration state detecting means for detecting the vibration state of the suspension. As in the first embodiment, a neural network is used as the control device 40, and the outputs of the detectors 43, 45, 47, 49, 51, and 53 are input to control the vibration control devices 39 and 41. Output a signal. Further, by comparing a control predicted value based on the output signal with a control target value, the connection weight of the neural network is corrected.
[0144]
The configuration of the neural network of the control device 40 is as shown in FIG.
[0145]
This neural network also comprises a control neural network 55 and an identification neural network 57, as in the above embodiment. The control neural network 55 includes a front wheel control neural network 55a and a rear wheel control neural network 55b. Each of the control neural nets 55a and 55b is composed of a white circle input layer, a two-stage intermediate layer indicated by hatched circles, and a black circle output layer. It consists of an intermediate layer with a circle and an output layer with a black circle.
[0146]
In this embodiment, only the control neural network 55 is a neural network considering the ARMA model. That is, the detection signal Y of the front wheel side road surface detector 43 at the time t is input to the input layer of the front wheel control neural network 55a._0F ^s(T), before
[Outside 8]

Control force P_F ^a(T). This output P_F ^a(T) is input to the identification neural network 15 and sequentially stored, and the past data P_F ^a(T-1) is input to the input layer of the front wheel control neural network 55a.
[0147]
The same applies to the rear wheel control neural network 55b, and the input is the detector 45,
[Outside 9]

41 control force O_c= P_R ^a(T) Past data P_R ^s(T-1). Output O_c= P_R ^a(T) is output to the vibration damping device 41 and is also input to the identification neural network 15. The input layer of the identification neural network 15 has
[Outside 10]

Further, the output of the hidden layer is input to the input layer of the identification neural network 15. This is generally called recurrent, and can remove signal noise. Recurrent can also be performed, for example, by inputting the output of the intermediate layer to the intermediate layer itself again.
[0148]
Output O of identification neural network 15_fIs the vibration acceleration of the vehicle body 32 at the next time t + 1
[Equation 14]

It becomes. Therefore, the evaluation function E of the identification neural network 15_fIs
(Equation 15)

And the evaluation function E of the control neural network 55_cIs
(Equation 16)

It becomes.
[0149]
Therefore, also in this embodiment, the evaluation function E_fIs less than or equal to a predetermined value, and if the identification neural network 57 accurately represents the dynamic characteristics of the vehicle body 32 with respect to the vibration input, the evaluation function E_cThus, the connection weight of the control neural network 55 is corrected and learned, thereby enabling appropriate feed-forward control of suspension control.
[0150]
Note that the recurrent of the identification neural network 57 provided in the second embodiment can be used for the control neural network, and can also be applied to the neural network of the first embodiment. In the second embodiment, one piece of data at (t-1) has been described as past data. However, as in the first embodiment, some time-delayed data can be considered.
[0151]
【The invention's effect】
As is clear from the above, according to the configuration of the present invention, the following effects can be obtained.
[0152]
According to the first aspect of the present invention, the dynamic characteristics of the system can be accurately identified by the control device using the neural network. Moreover, since the search is performed by dividing the search area into multiple areas having different sizes, blind search can be avoided, and the number of times required for the search can be reduced. Moreover, since a random search is used, it is possible to prevent convergence to a local minimum value and to accurately obtain a global minimum value. Therefore, high-speed and accurate processing can be performed.Further, by repeating the search in the region a plurality of times, the minimum value can be reliably detected. Therefore, accurate processing can be performed. In addition, the count number can be reduced according to the detected value, and higher-speed processing can be performed.
[0153]
According to the second aspect of the present invention, the dynamic characteristics of the system can be accurately identified by the control device using the neural network. Moreover, since the search is performed by dividing the search area into multiple areas having different sizes, blind search can be avoided, and the number of times required for the search can be reduced. Moreover, since a random search is used, it is possible to prevent convergence to a local minimum value and to accurately obtain a global minimum value. Therefore, high-speed and accurate processing can be performed.Further, by repeating the search in the region a plurality of times, the minimum value can be reliably detected. Therefore, accurate processing can be performed. When it is clear that the minimum value cannot be obtained, the processing of the area can be stopped and the processing can be performed at higher speed.
[0154]
According to the third aspect of the present invention, the dynamic characteristics of the system can be accurately identified by the control device using the neural network. Moreover, since the search is performed by dividing the search area into multiple areas having different sizes, blind search can be avoided, and the number of times required for the search can be reduced. Moreover, since a random search is used, it is possible to prevent convergence to a local minimum value and to accurately obtain a global minimum value. Therefore, high-speed and accurate processing can be performed.Further, by repeating the search in the region a plurality of times, the minimum value can be reliably detected. Therefore, accurate processing can be performed. Further, by weighting the directionality with respect to the initial value, it is possible to perform faster and more accurate detection.
[0155]
In the invention of claim 4,Since the search area is divided into a plurality of areas having different sizes and a random search is performed, at least one of the sound and the vibration state can be controlled at high speed and with high accuracy. Further, by repeating the search in the region a plurality of times, the minimum value can be reliably detected. Therefore, accurate processing can be performed. In addition, the count number can be reduced according to the detected value, and higher-speed processing can be performed.
[0156]
In the invention of claim 5,Since the search area is divided into a plurality of areas having different sizes and a random search is performed, at least one of the sound and the vibration state can be controlled at high speed and with high accuracy. Further, by repeating the search in the region a plurality of times, the minimum value can be reliably detected. Therefore, accurate processing can be performed. When it is clear that the minimum value cannot be obtained, the processing of the area can be stopped and the processing can be performed at higher speed.
[0157]
In the invention of claim 6,Since the search area is divided into a plurality of areas having different sizes and a random search is performed, at least one of the sound and the vibration state can be controlled at high speed and with high accuracy. Further, by repeating the search in the region a plurality of times, the minimum value can be reliably detected. Therefore, accurate processing can be performed. By weighting the directionality with respect to the initial value, faster and more accurate detection can be performed.
[0158]
In the invention of claim 7, the claim4~ Claim6 of any one ofIn addition to the effects of the invention,Vehicle interior noise can be accurately controlled by feedforward control. Therefore, even when the vehicle interior noise is based on a combination of an engine sound, an input from an engine, an input from a suspension, an exhaust sound from a muffler, a wind noise, and the like, the coupling can be accurately analyzed to reduce noise. .
[0159]
In the invention of claim 8, the claim4~ Claim6 of any one ofIn addition to the effects of the invention,Suspension signal control can be accurately controlled by feedforward. Therefore, the vibration characteristics of the suspension can be accurately analyzed not only on the highway but also on the off-road, and the vibration of the vehicle body can be reduced.
[0160]
In the invention of claim 9, the claim4To claim 8Any one ofIn addition to the effects of the invention,By providing an identification neural network in addition to the control neural network, it is possible to output a control predicted value and accurately correct control parameters of the control neural network.
[0161]
According to the tenth aspect of the present invention,4~ Claim9 of any one ofIn addition to the effects of the invention,By inputting data before the present time of the detection control, the control prediction value after the present time is predicted and output, so that even if the input is random, the sound vibration state can be accurately predicted, Precise control is possible for random inputs.
[0162]
In the invention according to claim 11, the claim1To any one of claims 10 to1 itemIn addition to the effects of the invention,Since the learning ratio when learning the connection weight of the neural network is corrected as a control parameter, it is possible to quickly and accurately set the learning ratio to an appropriate value, and to detect the minimum value more quickly and with higher accuracy. Can be performed with precision.
[0163]
In the invention of claim 12,1To any one of claims 11 to1 itemIn addition to the effects of the invention,Since a plurality of neural nets are provided in parallel according to the number of regions, processing can be performed at high speed.
[Brief description of the drawings]
FIG. 1 is a schematic block diagram according to an embodiment of the present invention.
FIG. 2 is a block diagram of a control device.
FIG. 3 is an explanatory diagram showing a part of a neural network.
FIG. 4 is an explanatory diagram showing a sigmoid function.
FIG. 5 is an explanatory diagram showing a convergence state by the steepest descent method.
FIG. 6 is a schematic diagram of a random search method.
FIG. 7 is a flowchart of a random search method.
FIG. 8 is a diagram showing the shape of a one-dimensional function.
FIG. 9 is a graph illustrating a convergence state at each initial value.
FIG. 10 is an explanatory diagram when the number of steps is changed.
FIG. 11 is an explanatory diagram when the count number is changed.
FIG. 12 is an explanatory diagram of a single-degree-of-freedom system spring-mass model.
FIG. 13 is an explanatory diagram of a transient response of a single-degree-of-freedom system spring-mass model.
FIG. 14 is a graph of a teacher signal.
FIG. 15 is a graph of a time history of errors of the linear model.
FIG. 16 is a graph of an identification result.
FIG. 17 is an explanatory diagram of a situation in which an impulse input is performed to a cantilever.
FIG. 18 is an explanatory diagram of a cantilever identification neural system.
FIG. 19 is a graph showing a history of errors in a real-time cantilever identification process.
FIG. 20 is a graph showing a result of identification of a cantilever using a neural network.
FIG. 21 is a schematic diagram of a parallelized neural network.
FIG. 22 is an explanatory diagram of a neural network based on an ARMA model.
FIG. 23 is a schematic explanatory view of the operation.
FIG. 24 is a response diagram of the system.
FIG. 25 is an explanatory diagram of back propagation.
FIG. 26 is a flowchart.
FIG. 27 is a configuration diagram of a neural network according to one embodiment.
FIG. 28 is a schematic block diagram according to another embodiment.
FIG. 29 is a configuration diagram of a neural network according to another embodiment.
FIG. 30 is a schematic block diagram of a conventional active noise control device.
FIG. 31 is an explanatory diagram showing a signal interference state.
[Explanation of symbols]
1 cabin
3 microphones (noise state detection means, sound vibration state detection means)
7 loudspeakers (control sound source, actuator)
11 Control device
39 Damping device (actuator)
40 control device
41 Damping device (actuator)
43 Front wheel side road surface detector (vibration state detection means, sound vibration state detection means)
45 Rear wheel side road surface detector (vibration state detection means, sound vibration state detection means)
47 Front wheel vibration acceleration detector (vibration state detection means, sound vibration state detection means)
49 Rear wheel vibration acceleration detector (vibration state detection means, sound vibration state detection means)
51 Front vibration acceleration detector (vibration state detection means, sound vibration state detection means)
53 Rear vibration acceleration detector (vibration state detection means, sound vibration state detection means)

Claims (12)

Equipped with a control device using a neural network that corrects control parameters by comparing the output value and the control target value,
The control device randomly sets an initial value of the control parameter, calculates a sensitivity of an output value to a control target value at the initial value, divides an allowable range of the control parameter into a plurality of regions, Based on the above, a random value of the control parameter is selected in each of the regions, and a minimum value of a difference between an output value and a control target value is detected in each region, and a minimum value is selected from the minimum values obtained from the respective regions. calculates, the modified lines stomach the outermost small value as next initial values,
In detecting the minimum value of the difference between the output value and the control target value in each area, a plurality of random values of the control parameter are selected in each area, and the selected number is used as the count number, and in each area, Is repeated a number of times less than the count number,
An active control device , wherein the count number in each of the regions is changed according to a detection value obtained by repeating the detection in each of the regions . Equipped with a control device using a neural network that corrects control parameters by comparing the output value and the control target value,
The control device randomly sets an initial value of the control parameter, calculates a sensitivity of an output value to a control target value at the initial value, divides an allowable range of the control parameter into a plurality of regions, The local value of the difference between the control value and the control target value is detected in each area by selecting a random value of the control parameter in each area, and the minimum value is selected from the local minimum values obtained from each area. calculates, the modified lines stomach the outermost small value as next initial values,
In detecting the minimum value of the difference between the output value and the control target value in each area, a plurality of random values of the control parameter are selected in each area, and the selected number is used as the count number, and in each area, Is repeated a number of times less than the count number,
When it is determined that the minimum value cannot be obtained during the repetition of the detection in each of the regions, the corresponding region is registered in the restriction list, and the detection of the minimum value in the region is stopped, and the active control device is characterized in that . Equipped with a control device using a neural network that corrects control parameters by comparing the output value and the control target value,
The control device randomly sets an initial value of the control parameter, calculates a sensitivity of an output value to a control target value at the initial value, divides an allowable range of the control parameter into a plurality of regions, The local value of the difference between the control value and the control target value is detected in each area by selecting a random value of the control parameter in each area, and the minimum value is selected from the local minimum values obtained from each area. Is calculated, and the correction is performed using the minimum value as a next initial value,
In detecting the minimum value of the difference between the output value and the control target value in each area, a plurality of random values of the control parameter are selected in each area, and the selected number is used as the count number, and in each area, Is repeated a number of times less than the count number,
An active control device , wherein the directionality of a random value of a control parameter selected in each region is weighted in accordance with a detection value obtained by repeating detection in each region . An actuator for controlling at least one of a sound or vibration state,
Sound and vibration state detection means for detecting at least one of a sound and a vibration state,
An output of the sound vibration state detecting means is input, and a signal for controlling the actuator is output, and a neural network for correcting a control parameter by comparing a control predicted value based on the output signal with a control target value is used. With a control device,
The control device randomly sets an initial value of the control parameter, calculates the sensitivity of a control predicted value to a control target value at the initial value, divides an allowable range of the control parameter into a plurality of regions, A random value of the control parameter is selected in each of the regions based on the sensitivity, a local minimum value of a difference between a control target value and a control predicted value is detected in each region in parallel, and a local minimum value obtained from each region is detected. Calculate the minimum value from among the above, and perform the correction using the minimum value as the next initial value,
In detecting the minimum value of the difference between the output value and the control target value in each area, a plurality of random values of the control parameter are selected in each area, and the selected number is used as the count number, and in each area, Is repeated a number of times less than the count number,
An active control device, wherein the count number in each of the regions is changed according to a detection value obtained by repeating the detection in each of the regions. An actuator for controlling at least one of a sound or vibration state,
Sound and vibration state detection means for detecting at least one of a sound and a vibration state,
An output of the sound vibration state detecting means is input, and a signal for controlling the actuator is output, and a neural network for correcting a control parameter by comparing a control predicted value based on the output signal with a control target value is used. With a control device,
The control device randomly sets an initial value of the control parameter, calculates the sensitivity of a control predicted value to a control target value at the initial value, divides an allowable range of the control parameter into a plurality of regions, A random value of the control parameter is selected in each of the regions based on the sensitivity, a local minimum value of a difference between a control target value and a control predicted value is detected in each region in parallel, and a local minimum value obtained from each region is detected. Calculate the minimum value from among the above, and perform the correction using the minimum value as the next initial value,
In detecting the minimum value of the difference between the output value and the control target value in each area, a plurality of random values of the control parameter are selected in each area, and the selected number is used as the count number, and in each area, Is repeated a number of times less than the count number,
When it is determined that the minimum value cannot be obtained during the repetition of the detection in each of the regions, the corresponding region is registered in the restriction list, and the detection of the minimum value in the region is stopped. . An actuator for controlling at least one of a sound or vibration state,
Sound and vibration state detection means for detecting at least one of a sound and a vibration state,
An output of the sound vibration state detecting means is input, and a signal for controlling the actuator is output, and a neural network for correcting a control parameter by comparing a control predicted value based on the output signal with a control target value is used. With a control device,
The control device randomly sets an initial value of the control parameter, calculates the sensitivity of a control predicted value to a control target value at the initial value, divides an allowable range of the control parameter into a plurality of regions, A random value of the control parameter is selected in each of the regions based on the sensitivity, a local minimum value of a difference between a control target value and a control predicted value is detected in each region in parallel, and a local minimum value obtained from each region is detected. Calculate the minimum value from among the above, and perform the correction using the minimum value as the next initial value,
In detecting the minimum value of the difference between the output value and the control target value in each area, a plurality of random values of the control parameter are selected in each area, and the selected number is used as the count number, and in each area, Is repeated a number of times less than the count number,
An active control device, wherein the direction of a random value of a control parameter selected in each of the regions is weighted in accordance with a detection value obtained by repeating the detection in each of the regions. An active control apparatus according to any one of Claims 4 to 6,
The actuator is a control sound source that generates a control sound that interferes with vehicle interior noise and reduces noise.
An active control device according to claim 1, wherein said sound vibration state detecting means is a noise state detecting means for detecting a noise state in a vehicle cabin . An active control apparatus according to any one of Claims 4 to 6,
The actuator is a vibration damping device that reduces vibration by applying a control force to a vehicle suspension,
The active control device according to claim 1, wherein said sound vibration state detecting means is a vibration state detecting means for detecting a vibration state of a vehicle suspension . An active control apparatus according to any one of Claims 4 to 8,
The neural network of the control device includes a control neural network and an identification neural network,
The control neural network receives an output of the sound vibration state detection unit and outputs a signal for controlling the actuator.
The active control device , wherein the identification neural network outputs the control prediction value based on a response of the vehicle under control of the actuator identified in advance . An active control apparatus according to any one of Claims 4 to 9,
The control device includes a neuron element for inputting at least a control signal of the actuator as data before the current time detected by the sound vibration state detecting means, and outputs a control signal of the current actuator, and outputs the control signal of the current time. active control device according to claim also be output from predicting a control predicted value later than. An active control apparatus according to any one of claims 1 to 10,
The active control device , wherein the control parameter is a learning rate when learning a connection weight of the neural network . To any one of claims 1 to 11, wherein an active control device,
An active control device , wherein a plurality of the neural nets are provided in parallel according to the number of steps which is the number of the regions .

Images