JP4198225B2 - Sensor signal classification apparatus and method - Google Patents

Description

【００１３】
のように尤度として算出されることが多い。また、(１)式の分類評価値は各特徴の誤差の分散は未知で、かつクラスタ毎に異なると考えているが、これが既知でかつ全クラスタで同一である場合、σ_p＝σ_ip＝・・・＝σ_Kpとする。
【００１４】
そして例えば、信号ｉをあるクラスタｋ₁からクラスタｋ₂に変更することによって、分類評価値Ｊ(Λ)が小さくもしくは大きくなるように変更していき、分類評価値が小さくもしくは大きくならなくなったらこの分類データを出力する。
【００１５】
この場合、信号ｉの所属するクラスタを変更することによって、あるクラスタに含まれる信号の数が０になったり、新たなクラスタを作成することは認めないので、クラスタ数が変化することはない。
【００１６】
最適クラスタ数推定部６では、分類部５から入力される分類データΛ_Kから、クラスタ数をＫであると仮定した場合のクラスタ数評価値を算出する。この評価値は、ＡＩＣ(Akaike Infomation Criterion)、ＭＤＬ(Minimum Description Length)基準、ＭＤＬ原理、ＦＥＰなどによって算出され、各特徴の誤差の分散は未知である場合、例えば、
【００１７】
【数２】

【００１８】
として算出される。
このクラスタ数評価値を含め、それまでに算出されたクラスタ数評価値と比較し、あるクラスタ数Ｋ'でのクラスタ数評価値が最小であると判断されれば、クラスタ数Ｋ'での分類データΛ_K'を出力する。
【００１９】
最小でないと判断されれば、算出したクラスタ数評価値および分類データを記憶し、さらにクラスタ数Ｋをこれまでクラスタ数評価値を算出していないクラスタ数に変更して出力する。通常は、クラスタ数設定部３でＫ＝１として設定し、最小でないと判断された場合に、Ｋの値を１つ増加して出力する。
【００２０】
信号分類部７では、入力された分類データのクラスタ数をＫ'とすると、センサ信号データＹを分類データΛ_K'に基づき、同一種類の信号を集めて、Ｋ'個の同一種類の信号のみを含むセンサ信号データを作成し出力する。
【００２１】
【発明が解決しようとする課題】
以上のように、構成された従来のセンサ信号分類装置では、特にクラスタ数が多い場合に、分類部から出力される分類データは真の最適解ではなく、局所最適解が求まってしまう可能性があり、この結果、最終的な分類結果も誤ってしまうという問題点があった。
【００２２】
また、前記(１)式のような分類評価値の算出において、各特徴の分散は既知である場合は一定値であるとし、未知である場合は同一クラスタ内では分散の期待値は同じ値であるとして、同一クラスタに含まれる信号の特徴のばらつきから推定していた。しかし、誤差の分散はＳ／Ｎ比などによって、信号毎に変化することも考えらる。この場合、分離性能は劣化するという問題点があった。
【００２３】
本発明は以上のような問題点を解消するためになされたもので、局所最適解が求まった場合に、これを検出し、さらに分類を続行することによって局所最適解が求まりにくくしたり、または各信号の各特徴毎に誤差の分散を推定・算出し、分離性能が劣化しにくくできるセンサ信号分類装置およびその方法を提供することを目的とする。
【００２４】
【課題を解決するための手段】
上記の目的に鑑み、本発明は、センサが受信・検出した複数種類の信号を、各種類毎に分類し、同じ種類の信号の集合であるクラスタを求めるセンサ信号分類装置であって、各信号がどのクラスタに含まれるかを示す分類データの任意の初期値と、各信号の特徴を記録したセンサ信号データから、分類に関する評価値である分類評価値を最小または最大にするような分類データを求める分類部が、前記分類データと前記センサ信号データを入力として、所定の分類アルゴリズムを用いて分類評価値を最小または最大にする分類データを出力する分類データ最適化手段と、前記センサ信号データと前記分類データ最適化手段から出力される分類データからこの分類データが局所最適解となっていることを各クラスタに含まれる信号が特徴空間で所定の分布に従っていないことに基づき判定し、局所最適解と判定した場合にはクラスタ数と信号数を出力し、局所最適解でないと判定した場合には前記分類データを出力する局所最適解検出手段と、この局所最適解検出手段からのクラスタ数と信号数を入力として前記分類データを初期化して作成しこれを前記分類データ最適化手段にフィードバックする分類データ初期化手段と、を備えたことを特徴とするセンサ信号分類装置にある。
【００２５】
またこの発明は、センサが受信・検出した複数種類の信号を、各種類毎に分類し、同じ種類の信号の集合であるクラスタを求めるセンサ信号分類装置であって、各信号がどのクラスタに含まれるかを示す分類データの任意の初期値と、各信号の特徴を記録したセンサ信号データから、分類に関する評価値である分類評価値を最小または最大にするような分類データを求める分類部が、前記分類データと前記センサ信号データを入力として、所定の分類アルゴリズムを用いて分類評価値を最小または最大にする分類データを出力する分類データ最適化手段と、前記センサ信号データと前記分類データ最適化手段から出力される分類データから、この分類データにおいて各クラスタが該クラスタに含まれる信号が特徴空間での仮定した分布・条件を満たしているか否かを判定し、満たしていないクラスタが存在する場合には前記分類データと、仮定した分布・条件を満たさないすべてのクラスタの集合であるクラスタ番号集合を出力し、すべてのクラスタが仮定した分布・条件を満たすと判定した場合には前記分類データを出力する誤クラスタ検出手段と、この誤クラスタ検出手段から出力される分類データとクラスタ番号集合を入力とし、この分類データに関してクラスタ番号集合に記録されたクラスタに含まれる信号のみを任意に初期化して前記分類データ最適化手段にフィードバックする分類データ一部初期化手段と、を備えたことを特徴とするセンサ信号分類装置にある。
【００２６】
またこの発明は、センサが受信・検出した複数種類の信号を、各種類毎に分類し、同じ種類の信号の集合であるクラスタを求めるセンサ信号分類装置であって、各信号がどのクラスタに含まれるかを示す分類データの任意の初期値と、各信号の特徴を記録したセンサ信号データから、分類に関する評価値である分類評価値を最小または最大にするような分類データを求める分類部が、前記センサ信号データを入力して各信号毎に信号の特徴の値の分散値を推定し、これを分散推定値データとして出力する分散推定値算出手段と、前記分類データおよびセンサ信号データ並びに前記分散推定値算出手段からの分散推定値データを入力として、各信号の各特徴毎に異なる分散値を取る分類アルゴリズムを用い、得られた分類評価値を最小または最大にする分類データを出力する異分散分類データ最適化手段と、を備えたことを特徴とするセンサ信号分類装置にある。
【００２７】
またこの発明は、センサが受信・検出した複数種類の信号を、各種類毎に分類し、同じ種類の信号の集合であるクラスタを求めるセンサ信号分類装置であって、各信号がどのクラスタに含まれるかを示す分類データの任意の初期値と、各信号の特徴を記録したセンサ信号データから、分類に関する評価値である分類評価値を最小または最大にするような分類データを求める分類部が、前記センサ信号データを入力して各信号毎に信号の特徴の値の分散値を推定し、これを分散推定値データとして出力する分散推定値算出手段と、前記分類データおよびセンサ信号データ並びに前記分散推定値算出手段からの分散推定値データを入力として、各信号の各特徴毎に異なる分散値を取る分類アルゴリズムを用い、得られた分類評価値を最小または最大にする分類データを出力する異分散分類データ最適化手段と、前記センサ信号データと異分散分類データ最適化手段から出力される分類データからこの分類データが局所最適解となっていることを各クラスタに含まれる信号が特徴空間で所定の分布に従っていないことに基づき判定し、局所最適解と判定した場合にはクラスタ数と信号数を出力し、局所最適解でないと判定した場合には前記分類データを出力する局所最適解検出手段と、この局所最適解検出手段からのクラスタ数と信号数を入力として前記分類データを初期化して作成しこれを前記異分散分類データ最適化手段にフィードバックする分類データ初期化手段と、を備えたことを特徴とするセンサ信号分類装置にある。
【００２８】
またこの発明は、センサが受信・検出した複数種類の信号を、各種類毎に分類し、同じ種類の信号の集合であるクラスタを求めるセンサ信号分類装置であって、各信号がどのクラスタに含まれるかを示す分類データの任意の初期値と、各信号の特徴を記録したセンサ信号データから、分類に関する評価値である分類評価値を最小または最大にするような分類データを求める分類部が、前記センサ信号データを入力して各信号毎に信号の特徴の値の分散値を推定し、これを分散推定値データとして出力する分散推定値算出手段と、前記分類データおよびセンサ信号データ並びに前記分散推定値算出手段からの分散推定値データを入力として、各信号の各特徴毎に異なる分散値を取る分類アルゴリズムを用い、得られた分類評価値を最小または最大にする分類データを出力する異分散分類データ最適化手段と、前記センサ信号データと前記異分散分類データ最適化手段から出力される分類データから、この分類データにおいて各クラスタが該クラスタに含まれる信号が特徴空間での仮定した分布・条件を満たしているか否かを判定し、満たしていないクラスタが存在する場合には前記分類データと、仮定した分布・条件を満たさないすべてのクラスタの集合であるクラスタ番号集合を出力し、すべてのクラスタが仮定した分布・条件を満たすと判定した場合には前記分類データを出力する誤クラスタ検出手段と、この誤クラスタ検出手段から出力される分類データとクラスタ番号集合を入力とし、この分類データに関してクラスタ番号集合に記録されたクラスタに含まれる信号のみを任意に初期化して前記異分散分類データ最適化手段にフィードバックする分類データ一部初期化手段と、を備えたことを特徴とするセンサ信号分類装置にある。
【００２９】
またこの発明は、センサが受信・検出した複数種類の信号を、各種類毎に分類し、同じ種類の信号の集合であるクラスタを求めるセンサ信号分類方法であって、各信号がどのクラスタに含まれるかを示す分類データの初期値と、各信号の特徴を記録したセンサ信号データから、分類評価値を最小または最大にするような分類データを求める分類工程が、分類評価値を良くするように大きくまたは小さくするように前記分類データを変更する分類変更ステップと、前記分類評価値が最小または最大になっているか否かを判定し、なっていないと判断した場合には前記分類変更ステップに戻る分類完了判断ステップと、この分類完了判断ステップで最小または最大になっていると判断した場合に、前記分類データが局所最適解となっていることを各クラスタに含まれる信号が特徴空間で所定の分布に従っていないことに基づき判定し、局所最適解でないと判定した場合に前記分類工程を終了する局所最適解検出ステップと、この局所最適解検出ステップで局所最適解と判定した場合に前記分類データを初期化して前記分類変更ステップに戻す分類データ初期化ステップと、を備えたことを特徴とするセンサ信号分類方法にある。
【００３０】
またこの発明は、センサが受信・検出した複数種類の信号を、各種類毎に分類し、同じ種類の信号の集合であるクラスタを求めるセンサ信号分類方法であって、各信号がどのクラスタに含まれるかを示す分類データの初期値と、各信号の特徴を記録したセンサ信号データから、分類評価値を最小または最大にするような分類データを求める分類工程が、分類評価値を良くするように大きくまたは小さくするように前記分類データを変更する分類変更ステップと、前記分類評価値が最小または最大になっているか否かを判定し、なっていないと判断した場合には前記分類変更ステップに戻る分類完了判断ステップと、この分類完了判断ステップで最小または最大になっていると判断した場合に、前記分類データにおいて、各クラスタが該クラスタに含まれる信号が特徴空間での仮定した分布・条件を満たしているかを判定し、すべて仮定した分布・条件を満たしていると判定した場合にこの分類工程を終了する誤クラスタ検出ステップと、この誤クラスタ検出ステップで満たしていないクラスタが存在する場合、仮定した分布・条件を満たさないすべてのクラスタを求め、これをクラスタ番号集合とし、現在の分類データに関してクラスタ番号集合に記録されたクラスタに含まれる信号のみを任意に初期化して前記分類変更ステップに戻す分類データ一部初期化ステップと、を備えたことを特徴とするセンサ信号分類方法にある。
【００３１】
またこの発明は、センサが受信・検出した複数種類の信号を、各種類毎に分類し、同じ種類の信号の集合であるクラスタを求めるセンサ信号分類方法であって、各信号がどのクラスタに含まれるかを示す分類データの初期値と、各信号の特徴を記録したセンサ信号データから、分類評価値を最小または最大にするような分類データを求める分類工程が、各信号毎に、信号の特徴の値の分散値を推定し、これを分散推定値データとする分散推定値算出ステップと、各信号の各特徴毎に異なる分散値を取る分類アルゴリズムを用い、得られた分類評価値を最小または最大にする異分散分類評価値による分類変更を行う分類変更ステップと、前記分類評価値が最小または最大になっているか否かを判定し、なっていないと判断した場合には前記分類変更ステップに戻り、なっていればこの分類工程を終了する分類完了判断ステップと、を備えたことを特徴とするセンサ信号分類方法にある。
【００３２】
またこの発明は、センサが受信・検出した複数種類の信号を、各種類毎に分類し、同じ種類の信号の集合であるクラスタを求めるセンサ信号分類方法であって、各信号がどのクラスタに含まれるかを示す分類データの初期値と、各信号の特徴を記録したセンサ信号データから、分類評価値を最小または最大にするような分類データを求める分類工程が、各信号毎に、信号の特徴の値の分散値を推定し、これを分散推定値データとする分散推定値算出ステップと、各信号の各特徴毎に異なる分散値を取る分類アルゴリズムを用い、得られた分類評価値を最小または最大にする異分散分類評価値による分類変更を行う分類変更ステップと、前記分類評価値が最小または最大になっているか否かを判定し、なっていないと判断した場合には前記分類変更ステップに戻る分類完了判断ステップと、この分類完了判断ステップで最小または最大になっていると判断した場合に、前記分類データが局所最適解となっていることを各クラスタに含まれる信号が特徴空間で所定の分布に従っていないことに基づき判定し、局所最適解でないと判定した場合に前記分類工程を終了する局所最適解検出ステップと、この局所最適解検出ステップで局所最適解と判定した場合に前記分類データを初期化して前記分類変更ステップに戻す分類データ初期化ステップと、を備えたことを特徴とするセンサ信号分類方法にある。
【００３３】
またこの発明は、センサが受信・検出した複数種類の信号を、各種類毎に分類し、同じ種類の信号の集合であるクラスタを求めるセンサ信号分類方法であって、各信号がどのクラスタに含まれるかを示す分類データの初期値と、各信号の特徴を記録したセンサ信号データから、分類評価値を最小または最大にするような分類データを求める分類工程が、各信号毎に、信号の特徴の値の分散値を推定し、これを分散推定値データとする分散推定値算出ステップと、各信号の各特徴毎に異なる分散値を取る分類アルゴリズムを用い、得られた分類評価値を最小または最大にする異分散分類評価値による分類変更を行う分類変更ステップと、前記分類評価値が最小または最大になっているか否かを判定し、なっていないと判断した場合には前記分類変更ステップに戻る分類完了判断ステップと、この分類完了判断ステップで最小または最大になっていると判断した場合に、前記分類データにおいて、各クラスタが該クラスタに含まれる信号が特徴空間での仮定した分布・条件を満たしているかを判定し、すべて仮定した分布・条件を満たしていると判定した場合にこの分類工程を終了する誤クラスタ検出ステップと、この誤クラスタ検出ステップで満たしていないクラスタが存在する場合、仮定した分布・条件を満たさないすべてのクラスタを求め、これをクラスタ番号集合とし、現在の分類データに関してクラスタ番号集合に記録されたクラスタに含まれる信号のみを任意に初期化して前記分類変更ステップに戻す分類データ一部初期化ステップと、を備えたことを特徴とするセンサ信号分類方法にある。
【００３４】
【発明の実施の形態】
以下、本発明の各実施の形態について、図面を参照して説明する。なお、本発明によるセンサ信号分類装置の全体の構成は図１４に示す従来のものと基本的に同一であり、分類部の部分の構成が異なる。また各図において、共通する要素には同一符合を付し、重複する説明を省略する。
【００３５】
実施の形態１．
図１はこの発明の一実施の形態によるセンサ信号分類装置の分類部(分類工程)の構成を示すブロック図を示す。分類部以外の動作は、従来例と同様である。同図において、１１ａは分類データ最適化手段１０の出力する分類データΛ_Kとセンサ信号データＹを入力として、分類結果が局所最適解となっているか否かを判断し、局所最適解でなければ入力された分類データΛ_Kを出力し、局所最適解であれば後述する分類データ初期化手段１２ａにクラスタ数Ｋと信号数Ｎを出力する局所最適解検出手段、１２ａは入力されたクラスタ数Ｋと信号数Ｎに基づいて、任意の方法で分類データを作成し、これを出力する分類データ初期化手段である。
【００３６】
以下、図２に示すフローチャートに沿って図１に示すセンサ信号分類装置の動作を説明する。分類データ最適化手段１０は従来例と同様に、(１)式などの分類評価値を良く(小さくもしくは大きく)するように、分類データを変更し(ステップＳＴ１２)、これ以上、分類評価値が良くならないと判断したら、その分類データを出力する(ステップＳＴ１３)。
【００３７】
この際、分類評価値が良くなるように、分類データを変更していく手法には、Ｋ-means法、ファジィ-c-means法などの最適化手法を用いる。これらの手法は、真に評価値を最良とする分類データ(真の最適解)でなく、局所的に考えた場合に評価値を最良とする分類データ(局所最適解)を求めてしまうことがある。このように、局所最適解を求めてしまうようなアルゴリズムを用いる理由は、必ず真の最適解を求めることができる総当たり法では、演算時間がかかりずぎ、現実的な演算時間で解を求めることができないためである。
【００３８】
局所最適解と真の最適解の例を図３、４に示す。これらの図は、分類データ最適化手段１０で得られた分類データに基づき、クラスタ毎に記号を変えて特徴空間に各信号をプロットした例である。この特徴数は２である。図３の局所最適解ではクラスタ２とクラスタ４の信号が特徴２によって分割された状態であるのに対し、図４の真の最適解では特徴１によって分割されており、見ためにも図４の分割の方が正しく思われる。
【００３９】
実際、(１)式の分類評価値も図４の方が小さくなる。このように、図３のような局所最適解は、例えば複数の信号の所属するクラスタを同時に変更しなければ評価値が小さくならないのに、k-means法では複数の信号の所属するクラスタを同時に変更することができないために求められる。
【００４０】
局所最適解検出手段１１ａでは、分類データ最適化手段１０の出力する分類データΛ_Kとセンサ信号データＹを入力とし、この分類データが局所最適解であるか否かの判定を行い、局所最適解であると判定された場合には分類データ初期化手段１２ａにクラスタ数Ｋと信号数Ｎとを出力し、局所最適解でなかった場合には分類データΛ_Kを出力する(ステップＳＴ１４)。
【００４１】
ただし、仮定しているクラスタが種類数と等しくないと、局所最適解であると判定され続ける可能性があるので、Ｍ(任意の整数)回以上、局所最適解検出(ステップＳＴ１４)処理を行なっても局所最適解であると判定される場合、もしくは、前に局所最適化検出手段１１ａにおいてステップＳＴ１４の処理を行なった際の分類データと分類評価値を記憶しておき、この分類評価値が小さくならなかった場合、局所最適解でないと判断して、分類データΛ_Kを出力する。
【００４２】
局所最適解であるか否かの判定方法について述べる。ここで、分かりやすくするため、各クラスタは特徴空間で十分距離が離れていて、信号から抽出される特徴に含まれる誤差は、正規分布に従う場合について考える。
【００４３】
この場合、仮定したクラスタ数が種類数と等しく、かつ真の最適解を求めることができれば、各クラスタに含まれる信号は、それぞれ特徴空間内で多変量正規分布に従って分布する。よって、各クラスタに含まれる信号が特徴空間で多変量正規分布に従っているか否かを判定し、従っていないクラスタがあれば、局所最適解がもとめられているか、もしくは仮定しているクラスタ数が誤っていることがわかる。判定方法として、仮説通りの分布に従っているか否かを判定することができる適合度検定、Kolmogorov-Smirnov検定などを用いる。
【００４４】
分かりやすくするため特徴の分布を正規分布としたが、これ以外の分布形でも良い。
【００４５】
また、信号から抽出される特徴がそれぞれ独立に正規分布に従い、かつそれらの特徴の中で分散が未知の特徴と既知の特徴がある場合、分散が未知の特徴のみから構成する特徴空間において上記の検定を行ない、分散が既知の特徴のみから構成する特徴空間において各クラスタの分散値を推定し、この値と既知の分散の値が等しいという仮定が成り立つか否かを検定するχ二乗検定を行なうこともできる。
【００４６】
その他にも、特徴空間内で各クラスタ間の距離が十分離れていると仮定したが、逆にこの距離が近く、特徴数が１である場合の特徴空間内での信号の頻度分布の例を図５に示す。この場合、クラスタ１に含まれるはずの信号でも特徴空間内でクラスタ２に近ければ、クラスタ２に含まれるとされる。この結果、この２つのクラスタにおいて特徴の誤差の分散として推定される値は、実際の分散よりも小さい値となる。反面、分散が大きくなることはない。
【００４７】
このため、例えば特徴の誤差の分散が既知の場合に、各クラスタから推定される分散値と既知の分散値が等しいという仮説でなく、推定される分散が既知の分散値より小さいという仮説を立て、検定を行なうこともできる。他にも、適合度検定、Kolmogorov-Smirnov検定を行なう場合に、分布全体を利用するのではなく、図５のようなクラスタの重なりがないと予想される領域のみを用いて検定を行なうこともできる。
【００４８】
分類データ初期化手段１２ａでは、局所最適解検出手段１１ａの出力するクラスタ数Ｋと信号数Ｎを入力として、クラスタ数Ｋ、信号数Ｎである分類データを任意に作成し、これを出力する(ステップＳＴ１５)。この分類データの初期化法は、分類部５に入力される分類データを作成する分類データ初期化部４(図１４参照)の方法と異なる方法とすることにより、再び同じ局所最適解が求められることを防ぐこともできる。
【００４９】
実施の形態２．
図６はこの発明の別の実施の形態によるセンサ信号分類装置の分類部の構成を示すブロック図を示す。この実施の形態にかかる分類部５は、分類データ最適化手段１０と、誤クラスタ検出手段１１ｂと、分類データ一部初期化手段１２ｂを備える。
【００５０】
以下、図７に示すフローチャートに沿って図６に示すセンサ信号分類装置の動作を説明する。
【００５１】
誤クラスタ検出手段１１ｂは、局所最適解検出手段１１ａと同様に考え、仮定した分布・条件に従わないクラスタを検出する。ただしこの際、実施の形態１で述べたχ二乗検定などの方法により仮定した分布・条件と異なると判定されたクラスタの番号を記憶する。この記憶されたクラスタ番号の集合を、クラスタ番号集合と呼ぶことにする。そして、仮定した分布・条件に従わないクラスタがあると判断された場合には、このクラスタ番号集合{k}と分類データΛ_Kを分類データ一部初期化手段１２ｂに出力し、なかったと判断された場合、分類データΛ_Kを出力する(ステップＳＴ２１)。
【００５２】
またこの際、実施の形態１と同様に、Ｍ回以上，誤クラスタ検出処理を行なった場合などでは、誤クラスタはないと考える。
【００５３】
分類データ一部初期化手段１２ｂは、誤クラスタ検出手段１１ｂから出力されたクラスタ番号集合と分類データを入力とし、クラスタ番号集合に記録されたクラスタに含まれる信号のみを任意に初期化し、初期化した分類データを出力する(ステップＳＴ２２)。
【００５４】
実施の形態３．
図８はこの発明の別の実施の形態によるセンサ信号分類装置の分類部の構成を示すブロック図を示す。この実施の形態にかかる分類部５は、分散推定値算出手段２０と異分散分類データ最適化手段１０ａを備える。
【００５５】
以下、図９に示すフローチャートに沿って図８に示すセンサ信号分類装置の動作を説明する。
【００５６】
分散推定値算出手段２０では、センサ信号データＹを入力として、各信号毎に、各特徴の推定誤差の分散値を推定する(ステップＳＴ３１)。
【００５７】
特徴推定誤差の分散は、センサ固有の値として一定値で決定されていることが多い。しかし、例えば信号がレーダパルスであり、その特徴がパルス幅であった場合、そのレーダパルスが受信された際のＳ／Ｎ比によって測定精度は異なり、Ｓ／Ｎ比が低ければ測定精度は悪化し、パルス幅推定誤差の分散は大きくなる。また、各特徴が各Ｓ／Ｎ比によってどのような推定誤差分散を持つかを事前に調べておくことができる。
【００５８】
このデータを分散推定値算出手段２０に記憶し、信号ｉが検出された際のＳ／Ｎ比から、信号ｉの特徴ｐの値ｙ_ipが持つ誤差の分散の推定値σ_ip ²を算出することができる。すべてのｉ(１〜Ｎ)、ｐ(１〜Ｐ)におけるこの分散の推定値σ_ip ²を記録したデータを分散推定値データσと呼ぶことにする。分散推定値算出手段２０では、この分散推定値データσを出力する。
【００５９】
異分散分類データ最適化手段１０ａでは、分散推定値算出手段２０の出力する分散推定値データσと、分類データΛ_K ⁽⁰⁾、センサ信号データＹを入力として、分類データ最適化手段１０と同様にして、分類データを変更する(ステップＳＴ３２)。ただし、分類評価値は(１)式と異なり、
【００６０】
【数３】

【００６１】
のように、各信号の各特徴毎に異なる分散値を持つようにして算出する。ここで、σ_kpは特徴ｐの誤差の分散がクラスタ毎に異なる場合に正の値を取る。例えば信号がレーダパルスで特徴がパルス幅である場合、レーダとセンサの位置関係によって、パルス幅推定誤差の分散が異なることが考えられるので、この場合は各クラスタにおける信号のばらつきからσ_kpを推定して、分類評価値が算出される。
【００６２】
実施の形態４．
図１０はこの発明の別の実施の形態によるセンサ信号分類装置の分類部の構成を示すブロック図を示す。この実施の形態にかかる分類部５は、図１の分類部５において、分類データ最適化手段１０の代わりに図８の分散推定値算出手段２０と異分散分類データ最適化手段１０ａを備える。
【００６３】
図１１に示すフローチャートは、図１０に示すセンサ信号分類装置の動作を示すものであり、図１０、１１における各手段、ステップの動作は、実施の形態１、３において同じ符合をつけた手段、ステップの動作と同様であり、両実施の形態の合わせた効果が得られる。
【００６４】
実施の形態５．
図１２はこの発明の別の実施の形態によるセンサ信号分類装置の分類部の構成を示すブロック図を示す。この実施の形態にかかる分類部５は、図６の分類部５において分類データ最適化手段１０の代わりに、図８の分散推定値算出手段２０と異分散分類データ最適化手段１０ａを備える。
【００６５】
図１３に示すフローチャートは、図１２に示すセンサ信号分類装置の動作を示すものであり、図１２、１３における各手段、ステップの動作は、実施の形態２、３において同じ符合をつけた手段、ステップの動作と同様であり、両実施の形態の合わせた効果が得られる。
【００６６】
【発明の効果】
以上のようにこの発明によれば、局所最適解であることを検出するので、より正しいセンサ信号の分類結果を出力するセンサ信号分類装置および方法を得ることができる。
【００６７】
また、仮定した分布・条件と異なるクラスタのみを初期化し、再度分類データを最適化するので、演算時間を短縮できるセンサ信号分類装置および方法を得ることができる。
【００６８】
さらに、信号毎に誤差の分散値を設定できるので、より正しいセンサ信号の分類結果を出力するセンサ信号分類装置および方法を得ることができる。
【００６９】
さらにまた、信号毎に誤差の分散値を設定でき、かつ、局所最適解であることを検出するので、より一層、正しいセンサ信号の分類結果を出力するセンサ信号分類装置および方法を得ることができる。
【００７０】
さらにまた、信号毎に誤差の分散値を設定でき、かつ仮定した分布・条件と異なるクラスタのみを初期化し、再度分類データを最適化するので、正しいセンサ信号の分類結果を出力し、かつ演算時間を短縮できるセンサ信号分類装置および方法を得ることができる。
【図面の簡単な説明】
【図１】 この発明の実施の形態１によるセンサ信号分類装置の分類部の構成を示すブロック図である。
【図２】 この発明の実施の形態１における動作を示すフローチャート図である。
【図３】 局所最適解である分類結果を特徴空間にプロットした例を示す図である。
【図４】 真の最適解である分類結果を特徴空間にプロットした例を示す図である。
【図５】 特徴空間における信号の分布の例を示す図である。
【図６】 この発明の実施の形態２によるセンサ信号分類装置の分類部の構成を示すブロック図である。
【図７】 この発明の実施の形態２における動作を示すフローチャート図である。
【図８】 この発明の実施の形態３によるセンサ信号分類装置の分類部の構成を示すブロック図である。
【図９】 この発明の実施の形態３における動作を示すフローチャート図である。
【図１０】 この発明の実施の形態４によるセンサ信号分類装置の分類部の構成を示すブロック図である。
【図１１】 この発明の実施の形態４における動作を示すフローチャート図である。
【図１２】 この発明の実施の形態５によるセンサ信号分類装置の分類部の構成を示すブロック図である。
【図１３】 この発明の実施の形態５における動作を示すフローチャート図である。
【図１４】 従来のこの種のセンサ信号分類装置および本発明によるセンサ信号分類装置の全体的構成を示す図である。
【符号の説明】
１ アンテナ、２ センサ信号データ作成部、４ 分類データ初期化部、５ 分類部、６ 最適クラスタ数推定部、７ 信号分類部、１０ 分類データ最適化手段、１０ａ 異分散分類データ初期化手段、１１ａ 局所最適解検出手段、１１ｂ 誤クラスタ検出手段、１２ａ 分類データ初期化手段、１２ｂ 分類データ一部初期化手段、２０ 分散推定値算出手段。[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a sensor signal classification apparatus and method for classifying a plurality of types of signals received and detected by a sensor for each type.
[0002]
[Prior art]
In order to classify a plurality of types of signals received / detected by a sensor for each type, there has conventionally been a sensor signal classification apparatus and method for extracting features from each signal and classifying them based on the similarity of the features.
Here, the signal received / detected by the sensor is, for example, each radar pulse in the case of a sensor that receives a radar signal, and the number of types of signals depends on the number of radars or the number of observation targets. In addition, if the sensor is an image radar such as a synthetic aperture radar, each pixel constituting the image is a signal, and the number of types of signals depends on the number of vegetation, targets, and observations in the area observed by the image radar. Depends on whether it can be classified as a target.
[0003]
Here, a set of one or two or more signals is called a cluster, a record indicating which cluster each signal belongs to is called classification data, and a record of features extracted from each signal is called sensor signal data. The number of clusters is equal to the number of signal types.
[0004]
FIG. 14 shows the configuration of a sensor signal classification device obtained by analogy with the “pulse classification device” disclosed in Japanese Patent Application Laid-Open No. 8-43520 as a conventional sensor signal classification device having such a function.
[0005]
In FIG. 14, 1 is a receiving antenna, 2 is a sensor signal data creating unit that detects and creates signals to be classified from received signals, extracts the characteristics of these signals, and creates sensor signal data. 4 is a cluster number setting unit for setting the number of types of signals, that is, a cluster number, and 4 is a classification data initializing unit for initializing classification data from the number of input sensor signal data and the input number of clusters. Is a classification data optimizing means for calculating a classification evaluation value, which is an evaluation value related to classification, from input sensor signal data and classification data, and obtaining classification data that minimizes or maximizes the evaluation value.
[0006]
Also, 6 is an optimal value for calculating an evaluation value for the number of clusters of this classification data, that is, an evaluation value for the number of clusters, from the input classification data and sensor signal data, and storing this to estimate the number of clusters considered to be most appropriate. A cluster number estimation unit 7 classifies signals from the input classification data and sensor signal data, and there are signal classification units 8a and 8b that generate several sensor signal data composed of only one type of signal. The sensor signal data is composed of only one type of signal, and 5 is a classification unit composed of only the classification data optimizing means 10.
[0007]
The operation of the sensor signal classification device shown in FIG. 14 will be described below. The antenna 1 receives, for example, radar signals from a plurality of radars and outputs the received signals to the sensor signal data creation unit 2. The sensor signal data creation unit 2 detects and creates signals from the received signals input from the antenna 1, extracts features from those signals, and outputs them as sensor signal data Y.
[0008]
If the minimum number of signal types or the approximate number of signal types is known in advance, the cluster number setting unit 3 sets the value as the number of types, that is, the number of clusters. If it is unknown, an appropriate value, generally 1 is set. Then, the set cluster number K is output.
[0009]
In the classification data initialization unit 4, from the number of clusters K input from the cluster number setting unit 3 or the optimum cluster number estimation unit 6 described later and the signal number of the sensor signal data Y input from the sensor signal data creation unit 2, Initial value Λ of classification data indicating which signal is included in which cluster_K ⁽⁰⁾And output this. The initial value may be set in any order, or may be determined using a random number.
[0010]
The classification unit 5 calculates a classification evaluation value from the sensor signal data input from the sensor signal data creation unit 2 and the classification data input from the classification data initialization unit 4. Here, the number of clusters is K, the number of signals is N, the number of features extracted from the signal is P, and the value of the feature p of the signal i is y_ip, The standard deviation of the error of the feature p of the cluster k is σ_kp ²And
[0011]
Classification data Λ_KΛ_ikIs a matrix of N rows and K columns, and the signal i is included in the cluster k, λ_ik= 1, if not included λ_ik= 0. In this case, the classification evaluation value is, for example,
[0012]
[Expression 1]

[0013]
It is often calculated as a likelihood. In addition, the classification evaluation value of equation (1) is considered that the variance of error of each feature is unknown and is different for each cluster. If this is known and is the same for all clusters, σ_p= Σ_ip= ... = σ_KpAnd
[0014]
And, for example, let signal i be a certain cluster k₁To cluster k₂By changing to, the classification evaluation value J (Λ) is changed so as to be small or large. When the classification evaluation value does not become small or large, this classification data is output.
[0015]
In this case, changing the cluster to which the signal i belongs does not allow the number of signals included in a certain cluster to become zero or create a new cluster, so the number of clusters does not change.
[0016]
In the optimum cluster number estimation unit 6, the classification data Λ input from the classification unit 5_KFrom this, the cluster number evaluation value when the number of clusters is assumed to be K is calculated. This evaluation value is calculated by AIC (Akaike Infomation Criterion), MDL (Minimum Description Length) standard, MDL principle, FEP, etc. When the variance of error of each feature is unknown, for example,
[0017]
[Expression 2]

[0018]
Is calculated as
If this cluster number evaluation value is compared with the cluster number evaluation value calculated so far, and it is determined that the cluster number evaluation value at a certain cluster number K ′ is the smallest, the classification at the cluster number K ′ is performed. Data Λ_K'Is output.
[0019]
If it is determined that the number is not the minimum, the calculated cluster number evaluation value and the classification data are stored, and the cluster number K is changed to a cluster number for which no cluster number evaluation value has been calculated so far and output. Normally, the cluster number setting unit 3 sets K = 1, and when it is determined that the number is not minimum, the value of K is increased by one and output.
[0020]
In the signal classification unit 7, if the number of clusters of the input classification data is K ′, the sensor signal data Y is classified into classification data Λ._KBased on ', the same type of signals are collected, and sensor signal data including only K' of the same type of signals is generated and output.
[0021]
[Problems to be solved by the invention]
As described above, in the conventional sensor signal classification device configured as described above, particularly when the number of clusters is large, there is a possibility that the classification data output from the classification unit is not a true optimum solution but a local optimum solution is obtained. As a result, there was a problem that the final classification result was wrong.
[0022]
In the calculation of the classification evaluation value as in the above equation (1), the variance of each feature is assumed to be a constant value if known, and if the variance is unknown, the expected value of variance is the same value in the same cluster. In some cases, the estimation is based on variations in the characteristics of signals included in the same cluster. However, it is conceivable that the variance of the error changes for each signal depending on the S / N ratio. In this case, there is a problem that the separation performance deteriorates.
[0023]
The present invention has been made to solve the above problems, and when a local optimal solution is obtained, it is detected and further classification is continued to make it difficult to obtain the local optimal solution, or It is an object of the present invention to provide a sensor signal classification apparatus and method for estimating and calculating the variance of error for each feature of each signal and making it difficult for the separation performance to deteriorate.
[0024]
[Means for Solving the Problems]
  In view of the above-described object, the present invention is a sensor signal classification device that classifies a plurality of types of signals received and detected by a sensor for each type and obtains a cluster that is a set of signals of the same type. Classification data that minimizes or maximizes the classification evaluation value, which is an evaluation value related to classification, from any initial value of classification data that indicates which cluster is included and sensor signal data that records the characteristics of each signal The classification unit to be obtained has the classification data and the sensor signal data as inputs, classification data optimization means for outputting classification data that minimizes or maximizes the classification evaluation value using a predetermined classification algorithm, and the sensor signal data This classification data becomes a local optimum solution from the classification data output from the classification data optimization means.Based on the fact that the signals contained in each cluster do not follow a predetermined distribution in the feature spaceA local optimum solution detection unit that outputs the number of clusters and the number of signals when it is determined to be a local optimal solution, and outputs the classification data when it is determined not to be a local optimal solution, and the local optimal solution detection unit A sensor signal classification device comprising: classification data initialization means for initializing and creating the classification data by receiving the number of clusters and the number of signals from the input and feeding back the classification data to the classification data optimization means is there.
[0025]
  The present invention is also a sensor signal classification device that classifies a plurality of types of signals received and detected by a sensor for each type and obtains a cluster that is a set of signals of the same type, in which cluster each signal is included. A classification unit that obtains classification data that minimizes or maximizes a classification evaluation value that is an evaluation value related to classification, from an arbitrary initial value of classification data that indicates whether or not and sensor signal data that records the characteristics of each signal, Classification data optimization means for inputting classification data and sensor signal data and outputting classification data that minimizes or maximizes a classification evaluation value using a predetermined classification algorithm; and the sensor signal data and classification data optimization From the classification data output from the means,In this classification data,The signal included in the cluster is the feature space.Judgment whether or not the assumed distribution / condition is satisfied. If there is a cluster that does not satisfy the condition, output the classification data and a cluster number set that is a set of all clusters that do not satisfy the assumed distribution / condition. If it is determined that all clusters satisfy the assumed distribution / condition, the erroneous cluster detection means for outputting the classification data, and the classification data and cluster number set output from the erroneous cluster detection means are input, A classification data partial initialization means for arbitrarily initializing only the signals included in the clusters recorded in the cluster number set with respect to the classification data and feeding back to the classification data optimization means. In the signal classification device.
[0026]
  The present invention is also a sensor signal classification device that classifies a plurality of types of signals received and detected by a sensor for each type and obtains a cluster that is a set of signals of the same type, in which cluster each signal is included. A classification unit that obtains classification data that minimizes or maximizes a classification evaluation value that is an evaluation value related to classification, from an arbitrary initial value of classification data that indicates whether or not and sensor signal data that records the characteristics of each signal,Sensor signal dataFor each signalThe value of the signal featureA variance estimated value calculating means for estimating a variance value and outputting it as variance estimated value data; and the variance estimated value data from the classification data and sensor signal data and the variance estimated value calculating means as inputs,Obtained using a classification algorithm that takes different variance values for each feature of each signalA sensor signal classification device comprising: heterovariance classification data optimization means for outputting classification data that minimizes or maximizes a classification evaluation value.
[0027]
  The present invention is also a sensor signal classification device that classifies a plurality of types of signals received and detected by a sensor for each type and obtains a cluster that is a set of signals of the same type, in which cluster each signal is included. A classification unit that obtains classification data that minimizes or maximizes a classification evaluation value that is an evaluation value related to classification, from an arbitrary initial value of classification data that indicates whether or not and sensor signal data that records the characteristics of each signal,Sensor signal dataFor each signalThe value of the signal featureA variance estimated value calculating means for estimating a variance value and outputting it as variance estimated value data; and the variance estimated value data from the classification data and sensor signal data and the variance estimated value calculating means as inputs,Obtained using a classification algorithm that takes different variance values for each feature of each signalThe heterogeneous classification data optimization means for outputting classification data that minimizes or maximizes the classification evaluation value, and the classification data is a local optimal solution from the sensor signal data and the classification data output from the heterogeneous classification data optimization means. BecomeBased on the fact that the signals contained in each cluster do not follow a predetermined distribution in the feature spaceA local optimum solution detecting unit that outputs the number of clusters and the number of signals when it is determined to be a local optimal solution, and outputs the classification data when it is determined not to be a local optimal solution, and the local optimal solution detecting unit Classification data initialization means for initializing and creating the classification data with the number of clusters and the number of signals from the input as input, and feeding back the classification data to the heterogeneous classification data optimization means, sensor signal classification In the device.
[0028]
  The present invention is also a sensor signal classification device that classifies a plurality of types of signals received and detected by a sensor for each type and obtains a cluster that is a set of signals of the same type, in which cluster each signal is included. A classification unit that obtains classification data that minimizes or maximizes a classification evaluation value that is an evaluation value related to classification, from an arbitrary initial value of classification data that indicates whether or not and sensor signal data that records the characteristics of each signal,Sensor signal dataFor each signalThe value of the signal featureA variance estimated value calculating means for estimating a variance value and outputting it as variance estimated value data; and the variance estimated value data from the classification data and sensor signal data and the variance estimated value calculating means as inputs,Obtained using a classification algorithm that takes different variance values for each feature of each signalFrom the classification data output from the heterogeneous classification data optimization means for outputting classification data that minimizes or maximizes the classification evaluation value, and from the sensor signal data and the classification data output from the heterogeneous classification data optimization means,The signal included in the cluster is the feature space.Judgment whether or not the assumed distribution / condition is satisfied. If there is a cluster that does not satisfy the condition, output the classification data and a cluster number set that is a set of all clusters that do not satisfy the assumed distribution / condition. If it is determined that all clusters satisfy the assumed distribution / condition, the erroneous cluster detection means for outputting the classification data, and the classification data and cluster number set output from the erroneous cluster detection means are input, Classification data partial initialization means for arbitrarily initializing only the signals included in the clusters recorded in the cluster number set for the classification data and feeding back to the heterogeneous classification data optimization means, In the sensor signal classification device.
[0029]
  The present invention is also a sensor signal classification method for classifying a plurality of types of signals received / detected by a sensor for each type and obtaining a cluster which is a set of signals of the same type, in which cluster each signal is included. The classification process for obtaining classification data that minimizes or maximizes the classification evaluation value from the initial value of the classification data indicating whether or not the signal is recorded and the sensor signal data recording the characteristics of each signal so that the classification evaluation value is improved. A classification changing step for changing the classification data so as to increase or decrease, and whether or not the classification evaluation value is minimum or maximum is determined. If it is determined that the classification evaluation value has not been reached, the process returns to the classification changing step. When it is determined that the classification completion determination step and the classification completion determination step are minimum or maximum, the classification data becomes a local optimal solution.Based on the fact that the signals contained in each cluster do not follow a predetermined distribution in the feature spaceA local optimum solution detection step for ending the classification process when it is determined that the solution is not a local optimum solution, and the classification data is initialized by initializing the classification data when the local optimum solution is determined in the local optimum solution detection step. And a classification data initializing step for returning to the step.
[0030]
  The present invention is also a sensor signal classification method for classifying a plurality of types of signals received / detected by a sensor for each type and obtaining a cluster which is a set of signals of the same type, in which cluster each signal is included. The classification process for obtaining classification data that minimizes or maximizes the classification evaluation value from the initial value of the classification data indicating whether or not the signal is recorded and the sensor signal data recording the characteristics of each signal so that the classification evaluation value is improved. A classification changing step for changing the classification data so as to increase or decrease, and whether or not the classification evaluation value is minimum or maximum is determined. If it is determined that the classification evaluation value has not been reached, the process returns to the classification changing step. When it is determined that the classification completion determination step and the classification completion determination step are minimum or maximum, in the classification data,The signal included in the cluster is the feature space.Judgment is made whether the assumed distribution / condition is satisfied, and all the assumed distribution / condition are satisfiedWhenIf there is an erroneous cluster detection step that ends this classification process when it is determined, and a cluster that does not satisfy the erroneous cluster detection step, all clusters that do not satisfy the assumed distribution / condition are obtained, and the cluster number set is obtained. And a classification data partial initialization step for arbitrarily initializing only the signals included in the clusters recorded in the cluster number set with respect to the current classification data and returning to the classification changing step. In the signal classification method.
[0031]
  The present invention is also a sensor signal classification method for classifying a plurality of types of signals received / detected by a sensor for each type and obtaining a cluster which is a set of signals of the same type, in which cluster each signal is included. The classification process for obtaining classification data that minimizes or maximizes the classification evaluation value from the initial value of the classification data indicating whether or not the sensor signal data recorded the characteristics of each signal, for each signal,The value of the signal featureA variance estimate calculation step that estimates a variance value and uses this as variance estimate value data, and takes a different variance value for each feature of each signalObtained using a classification algorithmA classification change step for performing classification change based on the heterogeneous classification evaluation value that minimizes or maximizes the classification evaluation value, and whether or not the classification evaluation value is minimum or maximum. Is a sensor signal classification method comprising: a classification completion judging step for returning to the classification changing step and ending the classification step if the classification is changed.
[0032]
  The present invention is also a sensor signal classification method for classifying a plurality of types of signals received / detected by a sensor for each type and obtaining a cluster which is a set of signals of the same type, in which cluster each signal is included. The classification process for obtaining classification data that minimizes or maximizes the classification evaluation value from the initial value of the classification data indicating whether or not the sensor signal data recorded the characteristics of each signal, for each signal,The value of the signal featureA variance estimate calculation step that estimates a variance value and uses this as variance estimate value data, and takes a different variance value for each feature of each signalObtained using a classification algorithmA classification change step for performing classification change based on the heterogeneous classification evaluation value that minimizes or maximizes the classification evaluation value, and whether or not the classification evaluation value is minimum or maximum. Returns to the classification change step, and when it is determined that the classification completion determination step is minimum or maximum in the classification completion determination step, the classification data becomes a local optimal solution.Based on the fact that the signals contained in each cluster do not follow a predetermined distribution in the feature spaceA local optimum solution detection step for ending the classification process when it is determined that the solution is not a local optimum solution, and the classification data is initialized by initializing the classification data when the local optimum solution is determined in the local optimum solution detection step. And a classification data initializing step for returning to the step.
[0033]
  The present invention is also a sensor signal classification method for classifying a plurality of types of signals received / detected by a sensor for each type and obtaining a cluster which is a set of signals of the same type, in which cluster each signal is included. The classification process for obtaining classification data that minimizes or maximizes the classification evaluation value from the initial value of the classification data indicating whether or not the sensor signal data recorded the characteristics of each signal, for each signal,The value of the signal featureA variance estimate calculation step that estimates a variance value and uses this as variance estimate value data, and takes a different variance value for each feature of each signalObtained using a classification algorithmA classification change step for performing classification change based on the heterogeneous classification evaluation value that minimizes or maximizes the classification evaluation value, and whether or not the classification evaluation value is minimum or maximum. In the classification data, when it is determined that the classification completion determination step returns to the classification change step and the classification completion determination step is minimum or maximum.The signal included in the cluster is the feature space.Judgment is made whether the assumed distribution / condition is satisfied, and all the assumed distribution / condition are satisfiedWhenIf there is an erroneous cluster detection step that ends this classification process when it is determined, and a cluster that does not satisfy the erroneous cluster detection step, all clusters that do not satisfy the assumed distribution / condition are obtained, and the cluster number set is obtained. And a classification data partial initialization step for arbitrarily initializing only the signals included in the clusters recorded in the cluster number set with respect to the current classification data and returning to the classification changing step. In the signal classification method.
[0034]
DETAILED DESCRIPTION OF THE INVENTION
Embodiments of the present invention will be described below with reference to the drawings. The overall configuration of the sensor signal classification apparatus according to the present invention is basically the same as the conventional one shown in FIG. 14, and the configuration of the classification unit is different. Moreover, in each figure, the same code | symbol is attached | subjected to a common element and the overlapping description is abbreviate | omitted.
[0035]
Embodiment 1 FIG.
FIG. 1 is a block diagram showing a configuration of a classification unit (classification step) of a sensor signal classification device according to an embodiment of the present invention. The operations other than the classification unit are the same as in the conventional example. In the figure, reference numeral 11a denotes classification data Λ output by the classification data optimizing means 10._KAnd sensor signal data Y as input, it is determined whether or not the classification result is a local optimal solution. If it is not a local optimal solution, the input classification data Λ_KIf it is a local optimum solution, the local optimum solution detection means for outputting the cluster number K and the signal number N to the classification data initialization means 12a described later, 12a is based on the input cluster number K and the signal number N. This is classification data initialization means for creating classification data by an arbitrary method and outputting it.
[0036]
The operation of the sensor signal classification device shown in FIG. 1 will be described below along the flowchart shown in FIG. Similar to the conventional example, the classification data optimizing means 10 changes the classification data so as to improve (smaller or larger) the classification evaluation value such as the expression (1) (step ST12). If it is determined that the condition does not improve, the classification data is output (step ST13).
[0037]
At this time, an optimization method such as a K-means method or a fuzzy-c-means method is used as a method of changing the classification data so that the classification evaluation value is improved. These methods may find classification data (local optimal solution) that has the best evaluation value when considered locally, rather than classification data that truly has the best evaluation value (true optimal solution). is there. In this way, the reason for using an algorithm that finds a local optimum solution is that the brute force method, which can always find the true optimum solution, takes less computation time and finds the solution in a realistic computation time. It is because it cannot be done.
[0038]
Examples of the local optimum solution and the true optimum solution are shown in FIGS. These figures are examples in which each signal is plotted in the feature space by changing the symbol for each cluster based on the classification data obtained by the classification data optimizing means 10. This feature number is two. In the local optimum solution of FIG. 3, the signals of the cluster 2 and the cluster 4 are divided by the feature 2, whereas in the true optimum solution of FIG. Seems to be correct.
[0039]
In fact, the classification evaluation value of equation (1) is also smaller in FIG. As described above, the local optimum solution as shown in FIG. 3 does not decrease the evaluation value unless, for example, the clusters to which a plurality of signals belong are changed simultaneously, but the k-means method simultaneously uses the clusters to which a plurality of signals belong. Required because it cannot be changed.
[0040]
In the local optimum solution detection means 11a, the classification data Λ output from the classification data optimization means 10_KAnd the sensor signal data Y are input, and it is determined whether or not the classification data is a local optimal solution. If it is determined that the classification data is a local optimal solution, the classification data initialization means 12a receives the number of clusters K and the signal. When the number N is output and the local optimal solution is not obtained, the classification data Λ_KIs output (step ST14).
[0041]
However, if the assumed cluster is not equal to the number of types, there is a possibility that it will continue to be determined as a local optimal solution, so the local optimal solution detection (step ST14) process is performed M (arbitrary integer) times or more. However, if it is determined that the solution is a local optimal solution, or the classification data and the classification evaluation value obtained when the processing of step ST14 was previously performed in the local optimization detection means 11a are stored, this classification evaluation value is stored. If it is not reduced, it is determined that the solution is not a local optimal solution, and the classification data Λ_KIs output.
[0042]
A method for determining whether or not it is a local optimum solution will be described. Here, for the sake of easy understanding, a case is considered in which each cluster is sufficiently separated in the feature space, and the error included in the feature extracted from the signal follows a normal distribution.
[0043]
In this case, if the assumed number of clusters is equal to the number of types and a true optimum solution can be obtained, the signals included in each cluster are each distributed according to the multivariate normal distribution in the feature space. Therefore, it is determined whether or not the signal included in each cluster follows the multivariate normal distribution in the feature space. If there is a cluster that does not follow, the local optimal solution has been obtained or the assumed number of clusters is incorrect. I understand that. As a determination method, a goodness-of-fit test, a Kolmogorov-Smirnov test, or the like that can determine whether or not the distribution follows a hypothesis is used.
[0044]
For the sake of simplicity, the distribution of features is a normal distribution, but other distribution forms may be used.
[0045]
In addition, if the features extracted from the signal independently follow a normal distribution, and there are features with unknown variance and known features among those features, the above-described feature space in the feature space composed only of features with unknown variance Perform a test to estimate the variance value of each cluster in a feature space consisting only of features whose variance is known, and perform a chi-square test to test whether this value and the known variance value are equal You can also.
[0046]
In addition, although it is assumed that the distance between the clusters is sufficiently large in the feature space, on the contrary, an example of the frequency distribution of signals in the feature space when this distance is close and the number of features is 1. As shown in FIG. In this case, a signal that should be included in cluster 1 is included in cluster 2 if it is close to cluster 2 in the feature space. As a result, the value estimated as the variance of the feature error in these two clusters is smaller than the actual variance. On the other hand, dispersion does not increase.
[0047]
For this reason, for example, when the variance of feature errors is known, the hypothesis that the estimated variance is smaller than the known variance is not the hypothesis that the variance estimated from each cluster is equal to the known variance. You can also perform a test. In addition, when performing the goodness-of-fit test and the Kolmogorov-Smirnov test, the test may be performed using only the region where no cluster overlap is expected as shown in FIG. 5 instead of using the entire distribution. it can.
[0048]
The classification data initialization means 12a receives the number of clusters K and the number of signals N output from the local optimum solution detection means 11a as input, arbitrarily creates classification data having the number of clusters K and the number of signals N, and outputs this ( Step ST15). This classification data initialization method is different from the classification data initialization section 4 (see FIG. 14) for creating classification data input to the classification section 5, whereby the same local optimum solution is obtained again. Can also be prevented.
[0049]
Embodiment 2. FIG.
FIG. 6 is a block diagram showing a configuration of a classification unit of a sensor signal classification device according to another embodiment of the present invention. The classification unit 5 according to this embodiment includes a classification data optimization unit 10, an erroneous cluster detection unit 11b, and a partial classification data initialization unit 12b.
[0050]
The operation of the sensor signal classification device shown in FIG. 6 will be described below along the flowchart shown in FIG.
[0051]
The erroneous cluster detection means 11b considers the same as the local optimum solution detection means 11a and detects clusters that do not follow the assumed distribution / condition. However, at this time, the cluster number determined to be different from the distribution / condition assumed by the method such as the chi-square test described in the first embodiment is stored. This set of stored cluster numbers is referred to as a cluster number set. If it is determined that there is a cluster that does not follow the assumed distribution / condition, this cluster number set {k} and the classification data Λ_KIs output to the classification data partial initialization means 12b._KIs output (step ST21).
[0052]
At this time, as in the case of the first embodiment, there is no erroneous cluster when the erroneous cluster detection process is performed M times or more.
[0053]
The classification data partial initialization means 12b receives the cluster number set output from the erroneous cluster detection means 11b and the classification data as input, arbitrarily initializes only the signals included in the clusters recorded in the cluster number set, and initializes them. The classified data is output (step ST22).
[0054]
Embodiment 3 FIG.
FIG. 8 is a block diagram showing a configuration of a classification unit of a sensor signal classification device according to another embodiment of the present invention. The classification unit 5 according to this embodiment includes a variance estimated value calculation unit 20 and a different variance classification data optimization unit 10a.
[0055]
The operation of the sensor signal classification device shown in FIG. 8 will be described below along the flowchart shown in FIG.
[0056]
The variance estimated value calculation means 20 receives the sensor signal data Y as input, and estimates the variance value of the estimation error of each feature for each signal (step ST31).
[0057]
The variance of the feature estimation error is often determined as a constant value as a value unique to the sensor. However, for example, when the signal is a radar pulse and the characteristic is the pulse width, the measurement accuracy differs depending on the S / N ratio when the radar pulse is received, and the measurement accuracy deteriorates if the S / N ratio is low. However, the variance of the pulse width estimation error becomes large. In addition, it is possible to examine in advance what kind of estimated error variance each characteristic has depending on each S / N ratio.
[0058]
This data is stored in the variance estimated value calculating means 20, and from the S / N ratio when the signal i is detected, the value y of the feature p of the signal i_ipEstimated value of error variance σ_ip ²Can be calculated. An estimate σ of this variance at all i (1-N), p (1-P)_ip ²Will be referred to as estimated variance data σ. The variance estimated value calculation means 20 outputs this variance estimated value data σ.
[0059]
In the different variance classification data optimizing means 10a, the variance estimated value data σ output from the variance estimated value calculating means 20 and the classification data Λ_K ⁽⁰⁾Then, using the sensor signal data Y as input, the classification data is changed in the same manner as the classification data optimizing means 10 (step ST32). However, the classification evaluation value is different from equation (1).
[0060]
[Equation 3]

[0061]
As described above, the calculation is performed so as to have a different variance value for each feature of each signal. Where σ_kpTakes a positive value when the error variance of the feature p is different for each cluster. For example, if the signal is a radar pulse and the feature is the pulse width, the variance of the pulse width estimation error may vary depending on the positional relationship between the radar and the sensor. In this case, σ_kpIs estimated, and a classification evaluation value is calculated.
[0062]
Embodiment 4 FIG.
FIG. 10 is a block diagram showing a configuration of a classification unit of a sensor signal classification device according to another embodiment of the present invention. The classification unit 5 according to this embodiment includes the estimated variance calculation unit 20 and the different variance classification data optimization unit 10a in FIG. 8 instead of the classification data optimization unit 10 in the classification unit 5 in FIG.
[0063]
The flowchart shown in FIG. 11 shows the operation of the sensor signal classification device shown in FIG. 10, and the means and steps in FIGS. 10 and 11 are the same as those in the first and third embodiments. The operation is the same as that of the step, and the combined effect of both embodiments can be obtained.
[0064]
Embodiment 5 FIG.
FIG. 12 is a block diagram showing a configuration of a classification unit of a sensor signal classification device according to another embodiment of the present invention. The classification unit 5 according to this embodiment includes the variance estimated value calculation unit 20 and the heterovariance classification data optimization unit 10a of FIG. 8 instead of the classification data optimization unit 10 of the classification unit 5 of FIG.
[0065]
The flowchart shown in FIG. 13 shows the operation of the sensor signal classification device shown in FIG. 12, and the means and steps in FIGS. 12 and 13 are the same as those in the second and third embodiments. The operation is the same as that of the step, and the combined effect of both embodiments can be obtained.
[0066]
【The invention's effect】
As described above, according to the present invention, since it is detected that the solution is a local optimal solution, it is possible to obtain a sensor signal classification device and method for outputting a more correct sensor signal classification result.
[0067]
In addition, since only the clusters different from the assumed distribution / condition are initialized and the classification data is optimized again, it is possible to obtain a sensor signal classification apparatus and method that can reduce the computation time.
[0068]
Furthermore, since a variance value of an error can be set for each signal, a sensor signal classification apparatus and method for outputting a more correct sensor signal classification result can be obtained.
[0069]
Furthermore, since it is possible to set an error variance value for each signal and to detect that the solution is a local optimal solution, it is possible to obtain a sensor signal classification device and method for outputting a correct sensor signal classification result. .
[0070]
Furthermore, error variance values can be set for each signal, and only the clusters that differ from the assumed distribution and conditions are initialized, and the classification data is optimized again. Sensor signal classification apparatus and method can be obtained.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a configuration of a classification unit of a sensor signal classification device according to Embodiment 1 of the present invention.
FIG. 2 is a flowchart showing an operation in Embodiment 1 of the present invention.
FIG. 3 is a diagram illustrating an example in which a classification result which is a local optimum solution is plotted in a feature space.
FIG. 4 is a diagram illustrating an example in which a classification result which is a true optimum solution is plotted in a feature space.
FIG. 5 is a diagram illustrating an example of signal distribution in a feature space.
FIG. 6 is a block diagram showing a configuration of a classification unit of a sensor signal classification device according to Embodiment 2 of the present invention.
FIG. 7 is a flowchart showing an operation in the second embodiment of the present invention.
FIG. 8 is a block diagram showing a configuration of a classification unit of a sensor signal classification device according to Embodiment 3 of the present invention.
FIG. 9 is a flowchart showing an operation in the third embodiment of the present invention.
FIG. 10 is a block diagram showing a configuration of a classification unit of a sensor signal classification device according to Embodiment 4 of the present invention.
FIG. 11 is a flowchart showing an operation in the fourth embodiment of the present invention.
FIG. 12 is a block diagram showing a configuration of a classification unit of a sensor signal classification device according to Embodiment 5 of the present invention.
FIG. 13 is a flowchart showing an operation in the fifth embodiment of the present invention.
FIG. 14 is a diagram showing the overall configuration of a conventional sensor signal classification device of this type and a sensor signal classification device according to the present invention.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 1 Antenna, 2 Sensor signal data preparation part, 4 Classification data initialization part, 5 Classification part, 6 Optimal cluster number estimation part, 7 Signal classification part, 10 Classification data optimization means, 10a Heterogeneous dispersion classification data initialization means, 11a Local optimal solution detection means, 11b erroneous cluster detection means, 12a classification data initialization means, 12b partial classification data initialization means, 20 variance estimated value calculation means.

Claims (10)

A sensor signal classification device that classifies a plurality of types of signals received and detected by a sensor for each type, and obtains a cluster that is a set of signals of the same type,
Classification that minimizes or maximizes the classification evaluation value, which is an evaluation value related to classification, from an arbitrary initial value of classification data that indicates which cluster each signal is included in and sensor signal data that records the characteristics of each signal The classification part that asks for data
Classification data optimization means for outputting classification data that minimizes or maximizes a classification evaluation value using a predetermined classification algorithm, with the classification data and the sensor signal data as inputs.
It is determined from the sensor signal data and the classification data output from the classification data optimization means that the classification data is a local optimal solution based on the fact that the signals included in each cluster do not follow a predetermined distribution in the feature space If it is determined to be a local optimal solution, the number of clusters and the number of signals are output, and if it is determined not to be a local optimal solution, a local optimal solution detection unit that outputs the classification data;
Classification data initialization means for initializing and creating the classification data by inputting the number of clusters and the number of signals from the local optimum solution detection means and feeding back the classification data to the classification data optimization means,
A sensor signal classification device comprising: A sensor signal classification device that classifies a plurality of types of signals received and detected by a sensor for each type, and obtains a cluster that is a set of signals of the same type,
Classification that minimizes or maximizes the classification evaluation value, which is an evaluation value related to classification, from an arbitrary initial value of classification data that indicates which cluster each signal is included in and sensor signal data that records the characteristics of each signal The classification part that asks for data
Classification data optimization means for outputting classification data that minimizes or maximizes a classification evaluation value using a predetermined classification algorithm, with the classification data and the sensor signal data as inputs.
From the classification data output from said sensor signal data and the classification data optimization means, determining whether the signals each cluster in the classification data included in the cluster meets the assumed distribution and conditions in the feature space If there is an unsatisfied cluster, the classification data and a cluster number set that is a set of all clusters not satisfying the assumed distribution / condition are output, and all clusters are assumed to satisfy the assumed distribution / condition. If it is determined that, erroneous cluster detection means for outputting the classification data,
The classification data output from the erroneous cluster detection means and the cluster number set are input, and only the signals included in the clusters recorded in the cluster number set for the classification data are arbitrarily initialized and fed back to the classification data optimization means Classification data partial initialization means to perform,
A sensor signal classification device comprising: A sensor signal classification device that classifies a plurality of types of signals received and detected by a sensor for each type, and obtains a cluster that is a set of signals of the same type,
Classification that minimizes or maximizes the classification evaluation value, which is an evaluation value related to classification, from an arbitrary initial value of classification data that indicates which cluster each signal is included in and sensor signal data that records the characteristics of each signal The classification part that asks for data
A variance estimated value calculating means for inputting the sensor signal data , estimating a variance value of a signal feature value for each signal, and outputting the estimated value as variance estimated value data;
Using the classification data, sensor signal data, and variance estimation value data from the variance estimation value calculation means as inputs, and using a classification algorithm that takes different variance values for each feature of each signal, the obtained classification evaluation value is minimized or Heterogeneous classification data optimization means for outputting classification data to be maximized,
A sensor signal classification device comprising: A sensor signal classification device that classifies a plurality of types of signals received and detected by a sensor for each type, and obtains a cluster that is a set of signals of the same type,
Classification that minimizes or maximizes the classification evaluation value, which is an evaluation value related to classification, from an arbitrary initial value of classification data that indicates which cluster each signal is included in and sensor signal data that records the characteristics of each signal The classification part that asks for data
A variance estimated value calculating means for inputting the sensor signal data , estimating a variance value of a signal feature value for each signal, and outputting the estimated value as variance estimated value data;
Using the classification data, sensor signal data, and variance estimation value data from the variance estimation value calculation means as inputs, and using a classification algorithm that takes different variance values for each feature of each signal, the obtained classification evaluation value is minimized or Heterogeneous classification data optimization means for outputting classification data to be maximized,
Based on the fact that the signals included in each cluster do not follow a predetermined distribution in the feature space, from the sensor signal data and the classification data output from the heterogeneous classification data optimization means, that the classification data is a local optimal solution. A local optimal solution detecting means for outputting the number of clusters and the number of signals when determined as a local optimal solution, and outputting the classification data when determined as not a local optimal solution;
Classification data initialization means for initializing and creating the classification data by inputting the number of clusters and the number of signals from the local optimum solution detection means and feeding back the classification data to the heterogeneous classification data optimization means,
A sensor signal classification device comprising: A sensor signal classification device that classifies a plurality of types of signals received and detected by a sensor for each type, and obtains a cluster that is a set of signals of the same type,
Classification that minimizes or maximizes the classification evaluation value, which is an evaluation value related to classification, from an arbitrary initial value of classification data that indicates which cluster each signal is included in and sensor signal data that records the characteristics of each signal The classification part that asks for data
A variance estimated value calculating means for inputting the sensor signal data , estimating a variance value of a signal feature value for each signal, and outputting the estimated value as variance estimated value data;
Using the classification data, sensor signal data, and variance estimation value data from the variance estimation value calculation means as inputs, and using a classification algorithm that takes different variance values for each feature of each signal, the obtained classification evaluation value is minimized or Heterogeneous classification data optimization means for outputting classification data to be maximized,
Based on the sensor signal data and the classification data output from the heterogeneous classification data optimization means, whether or not each cluster in the classification data satisfies the assumed distribution / condition in the feature space . If there are unsatisfied clusters, the classification data and a cluster number set that is a set of all clusters that do not satisfy the assumed distribution / condition are output. An erroneous cluster detection means for outputting the classification data when determined to satisfy
The heterogeneous classification data optimizing means which receives as input the classification data and cluster number set output from the erroneous cluster detection means and arbitrarily initializes only the signals included in the clusters recorded in the cluster number set with respect to the classification data Classification data partial initialization means to be fed back to
A sensor signal classification device comprising: A sensor signal classification method for classifying a plurality of types of signals received and detected by a sensor for each type and obtaining a cluster that is a set of signals of the same type,
A classification step for obtaining classification data that minimizes or maximizes the classification evaluation value from the initial value of the classification data indicating which cluster each signal is included in and sensor signal data that records the characteristics of each signal.
A classification changing step for changing the classification data so as to increase or decrease the classification evaluation value,
It is determined whether or not the classification evaluation value is minimum or maximum, and when it is determined that it is not, classification completion determination step to return to the classification change step,
If it is determined that at the minimum or maximum in this classification completion determination step, that the classification data does not conform to the distribution signal is given in the feature space included in each cluster that has a local optimal solution Determining based on the local optimal solution detection step of ending the classification step when it is determined that the local optimal solution is not,
A classification data initialization step that initializes the classification data and returns to the classification change step when it is determined as a local optimum solution in this local optimum solution detection step;
A sensor signal classification method comprising: A sensor signal classification method for classifying a plurality of types of signals received and detected by a sensor for each type and obtaining a cluster that is a set of signals of the same type,
A classification step for obtaining classification data that minimizes or maximizes the classification evaluation value from the initial value of the classification data indicating which cluster each signal is included in and sensor signal data that records the characteristics of each signal.
A classification changing step for changing the classification data so as to increase or decrease the classification evaluation value,
It is determined whether or not the classification evaluation value is minimum or maximum, and when it is determined that it is not, classification completion determination step to return to the classification change step,
If it is determined in the classification completion determination step that the minimum or maximum value is obtained, each cluster determines whether the signal included in the cluster satisfies the assumed distribution / condition in the feature space. , a false cluster detection step to terminate the classification process when it is determined that satisfies all assumed distribution and conditions,
If there is a cluster that does not satisfy the erroneous cluster detection step, all clusters that do not satisfy the assumed distribution / condition are obtained, and this is used as the cluster number set. A classification data partial initialization step for arbitrarily initializing only included signals and returning to the classification change step;
A sensor signal classification method comprising: A sensor signal classification method for classifying a plurality of types of signals received and detected by a sensor for each type and obtaining a cluster that is a set of signals of the same type,
A classification step for obtaining classification data that minimizes or maximizes the classification evaluation value from the initial value of the classification data indicating which cluster each signal is included in and sensor signal data that records the characteristics of each signal.
For each signal, a variance estimation value calculating step for estimating a variance value of the signal feature value and using this as a variance estimation value data;
A classification change step for performing classification change by different variance classification evaluation value that minimizes or maximizes the obtained classification evaluation value using a classification algorithm that takes different variance values for each feature of each signal,
It is determined whether or not the classification evaluation value is minimum or maximum, if it is determined that it is not, return to the classification change step, if it is, the classification completion determination step to end this classification step,
A sensor signal classification method comprising: A sensor signal classification method for classifying a plurality of types of signals received and detected by a sensor for each type and obtaining a cluster that is a set of signals of the same type,
A classification step for obtaining classification data that minimizes or maximizes the classification evaluation value from the initial value of the classification data indicating which cluster each signal is included in and sensor signal data that records the characteristics of each signal.
For each signal, a variance estimation value calculating step for estimating a variance value of the signal feature value and using this as a variance estimation value data;
A classification change step for performing classification change by different variance classification evaluation value that minimizes or maximizes the obtained classification evaluation value using a classification algorithm that takes different variance values for each feature of each signal,
It is determined whether or not the classification evaluation value is minimum or maximum, and when it is determined that it is not, classification completion determination step to return to the classification change step,
Based on the fact that the signals included in each cluster do not follow a predetermined distribution in the feature space when the classification completion determination step determines that the classification data is minimum or maximum, that the classification data is a local optimal solution. Determining a local optimal solution detecting step for ending the classification step when it is determined that the solution is not a local optimal solution;
A classification data initialization step that initializes the classification data and returns to the classification change step when it is determined as a local optimum solution in this local optimum solution detection step;
A sensor signal classification method comprising: A sensor signal classification method for classifying a plurality of types of signals received and detected by a sensor for each type and obtaining a cluster that is a set of signals of the same type,
A classification step for obtaining classification data that minimizes or maximizes the classification evaluation value from the initial value of the classification data indicating which cluster each signal is included in and sensor signal data that records the characteristics of each signal.
For each signal, a variance estimation value calculating step for estimating a variance value of the signal feature value and using this as a variance estimation value data;
A classification change step for performing classification change by different variance classification evaluation value that minimizes or maximizes the obtained classification evaluation value using a classification algorithm that takes different variance values for each feature of each signal,
It is determined whether or not the classification evaluation value is minimum or maximum, and when it is determined that it is not, classification completion determination step to return to the classification change step,
If it is determined in the classification completion determination step that the minimum or maximum value is obtained, each cluster determines whether the signal included in the cluster satisfies the assumed distribution / condition in the feature space. , a false cluster detection step to terminate the classification process when it is determined that satisfies all assumed distribution and conditions,
If there is a cluster that does not satisfy the erroneous cluster detection step, all clusters that do not satisfy the assumed distribution / condition are obtained, and this is used as the cluster number set. A classification data partial initialization step for arbitrarily initializing only included signals and returning to the classification change step;
A sensor signal classification method comprising:

Images