JP6969993B2 - Information extraction device - Google Patents

Description

The present disclosure relates to a technique for extracting necessary information from input data.

In Patent Document 1 below, a technique is proposed in which reference values of various sensor signals mounted on a vehicle are learned by a neural net (hereinafter referred to as a monitoring model), and the monitoring model is used to detect signals that deviate from the reference values. Has been done.

Specifically, the error between the input data obtained by vectorizing a plurality of sensor signals and the reconstructed data which is the result of reconstructing the input data using the monitoring model is obtained for each element of the vector, and this error is large. It is determined that there is an error in the sensor signal corresponding to the element.

Japanese Unexamined Patent Publication No. 2016-45861

However, in the prior art described in Patent Document 1, although it is possible to detect an abnormality such as data loss, there is no means for recovering the data in which the abnormality is detected, and the data is used as it is for the subsequent processing such as data analysis. There is a problem that the accuracy and reliability of the processing are lowered.

Further, for example, when the input data is a sensor signal, the abnormality of the sensor signal causes a temporary abnormality or loss of data due to noise or the like, in addition to the unrecoverable case caused by a component failure or the like. There is a case. In particular, since the latter case occurs frequently depending on the usage environment, there is also a problem that the post-stage processing cannot be effectively operated if all the data in which the abnormality is detected is excluded.

One aspect of the present disclosure is to provide a technique for efficiently utilizing data including anomalies.

The information extraction device (10) according to one aspect of the present disclosure includes an acquisition unit (12), a restoration unit (13), and an extraction unit (14).
The acquisition unit acquires the input data represented by the multidimensional vector in which each element has a correlation and the status information indicating the abnormal status of each element in the input data. The restoration unit inputs using a transformation model, which is a neural network in which normal correlations between the elements of the input data are learned, with each of one or more elements indicated by the situation information as an abnormality as the target element. Generate restored data by restoring the target element in the data to the normal value. The extraction unit extracts the input data when the status information indicates that there is no abnormality, the restored data as the target data when the status information indicates that there is an abnormality, and one or more latent features related to the target data as output information.

According to such a configuration, even if the input data contains a partial abnormality such as a defect, it is possible to extract output information that can withstand the use in the post-stage processing in which the influence of the abnormality is suppressed. Furthermore, by using such output information, it is possible to realize robust data analysis and the like that are not affected by data including abnormalities.

In addition, the reference numerals in parentheses described in this column and the scope of claims indicate the correspondence with the specific means described in the embodiment described later as one embodiment, and the technical scope of the present disclosure is defined. It is not limited.

It is a block diagram which shows the structure of an information processing system. It is a flowchart of the learning process executed by the learning unit. It is explanatory drawing about the generation method of input data. It is explanatory drawing which shows the outline of a learning process. It is a flowchart of the acquisition process executed by the acquisition unit. It is explanatory drawing which shows the outline of the acquisition process. It is a flowchart of the restoration process executed by the restoration part. It is explanatory drawing which shows the outline of the restoration process. It is a flowchart of the extraction process executed by the extraction unit. It is explanatory drawing which shows the outline of the extraction process.

Hereinafter, embodiments of the present disclosure will be described with reference to the drawings.
[1. composition]
The information processing system 1 shown in FIG. 1 includes a sensor group 3, an information extraction device 10, an information processing device 5, and a model generation device 20. The sensor group 3, the information extraction device 10, and the information processing device 5 are mounted on the vehicle. Further, the model generation device 20 is provided in a base station having a function of performing wireless communication with a vehicle equipped with a sensor group 3 or the like, or a server on a communication network to which the base station is connected. At least one of the information extraction device 10 and the information processing device 5 may be provided outside the vehicle.

[1-1. Sensor group]
The sensor group 3 includes at least one of a sensor that outputs operation data representing a driving operation on the vehicle and a sensor that outputs behavior data representing the behavior of the vehicle as a result of the driving operation.

Specifically, the operation data includes the operation amount of the accelerator, the brake, the steering, and the like. Examples of the behavior data include speed, acceleration, yaw rate and the like. Further, the operation data may include situation data useful for specifying the operation situation. Specific examples of the situation data include video data of an in-vehicle camera, road attribute data related to a traveling road extracted from map data, position data representing the current position, time data representing the current time, and the like.

Hereinafter, the multidimensional vector generated by arranging the outputs from the sensor group 3 in order is referred to as a detection vector. That is, each element belonging to the detection vector corresponds to the output of each sensor constituting the sensor group 3.

[1-2. Model generator]
The model generation device 20 includes an information storage unit 21, a learning unit 22, a model storage unit 23, and a distribution unit 24.

The information storage unit 21 acquires the time series of the detection vectors output from the sensor group 3 mounted on the plurality of vehicles by some method and stores them in a large amount. However, the information storage unit 21 stores a time series of detection vectors (that is, non-missing data) that do not include abnormalities such as data loss.

The learning unit 22 executes a learning process for generating a monitoring model by learning a neural network using the data stored in the information storage unit 21. The details of the learning process will be described later. The monitoring model uses a neural network called a self-encoder that converts the input data into one or more latent features and then reconstructs the original data. In particular, here, a multilayer self-encoder having a plurality of intermediate layers is used.

The model storage unit 23 stores the monitoring model generated by the learning unit 22.
The distribution unit 24 distributes the monitoring model stored in the model storage unit 23 to the information extraction device 10 mounted on each of the plurality of vehicles.

Here, the learning process executed by the learning unit 22 will be described with reference to the flowchart of FIG. The learning process is executed, for example, every time a predetermined amount of information is accumulated in the information storage unit 21, or every time a certain period of time elapses. Further, the learning process may be executed when a command from the outside is received.

When the learning process is started, in S110, the learning unit 22 acquires the time series of the detection vector X stored in the information storage unit 21. The detection vector X is expressed by the equation (1), where D is the number of dimensions of the detection vector X, that is, the number of sensors constituting the sensor group 3.

Ｓ１２０では、学習部２２は、検出ベクトルＸの時系列を、時間窓を用いて順次切り取り、切り取ったデータを多次元ベクトルで表現した多数の学習データを生成する。時間窓は、図３に示すように、Ｔを時間窓の幅とし、Ｓを時間窓のスライド間隔として、Ｗ＞Ｓとなるように設定された、いわゆる滑走時間窓が用いられる。つまり、学習部２２では、一部の時間領域が重複するように設定される時間窓を用いて、多数の学習データが生成される。なお、学習データの系列は、学習データの次元数をＰ（但し、Ｐ＝Ｄ×Ｔ）として、（２）式で表現される。（３）及び（４）式は、個々の各学習データの詳細を表す。但し、Ｔ，Ｓは、検出ベクトル生成間隔（即ち、センサ信号のサンプリング間隔）の整数倍によって表現される。以後、（２）式で表現される時刻ｔでの次元数ＰのベクトルをＹ_ｔとも表現することにする。

In S120, the learning unit 22 sequentially cuts the time series of the detection vector X using a time window, and generates a large number of learning data in which the cut data is represented by a multidimensional vector. As shown in FIG. 3, as the time window, a so-called gliding time window set so that W> S is used, where T is the width of the time window and S is the slide interval of the time window. That is, in the learning unit 22, a large number of learning data are generated by using the time window set so that some time domains overlap. The series of training data is expressed by the equation (2), where P (however, P = D × T) is the number of dimensions of the training data. Equations (3) and (4) represent the details of each individual learning data. However, T and S are expressed by an integral multiple of the detection vector generation interval (that is, the sampling interval of the sensor signal). Hereinafter, the vector of the number of dimensions P at the time t expressed by the equation (2) will also be expressed _{as Y t.}

Ｓ１３０では、学習部２２は、先のＳ１２０で生成された個々の学習データを正規化する。ここでの正規化は、学習データに属する各要素の値が０〜１を値域とする値となるように変換することを意味する。

In S130, the learning unit 22 normalizes the individual learning data generated in S120. The normalization here means that the value of each element belonging to the training data is converted so that the value has a range of 0 to 1.

In S140, the learning unit 22 executes the learning of the neural network using the normalized learning data (hereinafter, simply learning data) generated in the previous S130, and uses the monitoring model generated by the learning as the model storage unit. It is stored in 23, and this process is terminated.

In this learning, as shown in FIG. 4, the training data is used as the input of the neural network, and the learning data used for the input is used as the target value of the output of the network (hereinafter referred to as reconstruction data). Then, the parameters of the monitoring model (that is, the neural network) are adjusted so that the reconstruction error, which is the difference between the input training data and the reconstruction data, is minimized. For such learning, a known learning method in a neural network such as backpropagation can be used. The monitoring model used (ie, neural network) can be a self-encoder such as Stacked Denoising Auto-encoder or Stacked Sparse Auto-encoder, or a variant thereof.

Here, the reconstruction error is represented by a multidimensional vector having the same number of dimensions as the training data (hereinafter, reconstruction error vector), and the equation (5) is an element specified by p in the reconstruction error vector (hereinafter, “reconstruction error vector”). Individual error). In the equation (5), the data represented by the code in which the hat symbol is added to the code representing the learning data is the reconstruction data. In the training, as shown in the equation (6), the result of adding the individual errors of all the elements belonging to the reconstruction error vector is used as the error evaluation value, and the sum of the error evaluation values with respect to the entire training data is used as the objective function. Then, the weight of each node constituting the neural network is adjusted so that this objective function is minimized.

［１−３．情報抽出装置］
図１に戻り、情報抽出装置１０は、センサ群３からの出力である検出データの時系列から検出データの特徴を表す情報を抽出し、出力情報として情報処理装置５に出力する。情報抽出装置１０は、取得部１２と、復元部１３と、抽出部１４とを備える。情報抽出装置１０は、モデル記憶部１１と、受信部１５を備えてもよい。

[1-3. Information extraction device]
Returning to FIG. 1, the information extraction device 10 extracts information representing the characteristics of the detection data from the time series of the detection data output from the sensor group 3, and outputs the information to the information processing device 5 as output information. The information extraction device 10 includes an acquisition unit 12, a restoration unit 13, and an extraction unit 14. The information extraction device 10 may include a model storage unit 11 and a reception unit 15.

The receiving unit 15 receives the monitoring model delivered from the model generation device 20 and stores it in the model storage unit 11. The receiving unit 15 may be omitted. When the receiving unit 15 is omitted, the monitoring model generated by the model generation device 20 may be manually stored in the model storage unit 11.

The acquisition unit 12 acquires the time series of the detection vector from the sensor group 3 to generate input data, and also uses the monitoring model stored in the model storage unit 11 to cause an abnormality in the input data and thus belongs to the sensor group 3. Detects the presence or absence of abnormalities such as defects in the output of each sensor.

When the acquisition unit 12 detects an abnormality in the input data, the restoration unit 13 uses the monitoring model stored in the model storage unit 11 to restore the restored data in which the abnormality in the input data is detected. Generate.

The extraction unit 14 is the restoration data when an abnormality in the input data is detected and the restoration data is generated by the restoration unit 13, and the input data generated by the acquisition unit 12 when an abnormality in the input data is not detected. Is used as the target data, and the target data and the latent features of the target data obtained by inputting the target data into the monitoring model are extracted as output information and output to the information processing apparatus 5. As the latent feature of the target data, for example, the output of the intermediate layer of the neural network representing the monitoring model, particularly the output of the encoder unit of the self-encoder may be used.

[1-3-1. Acquisition process]
Here, the acquisition process executed by the acquisition unit 12 will be described with reference to the flowchart of FIG. The acquisition process is repeatedly executed.

When the acquisition process is started, in S210, the acquisition unit 12 sequentially acquires and stores the detection data from the sensor group 3. Here, the time series of the detected data is stored by the time width T of the time window described above with reference to FIG. It should be noted that the portion exceeding the time width T may be overwritten in order from the oldest stored content.

In S220, the acquisition unit 12 determines whether or not it is the processing timing. If it is the processing timing, the process proceeds to S230, and if it is not the processing timing, the process returns to S210. The processing timing is determined based on whether or not a time corresponding to the shift amount S has elapsed from the previous processing timing.

In S230, the acquisition unit 12 generates input data by multidimensional vectorizing the time series of the detection data stored in S210, and further normalizes the generated input data. The specific processing of normalization is the same as that described in S130 above.

In S240, the acquisition unit 12 generates reconstruction data by applying the normalized input data (hereinafter, simply input data) generated in S230 to the monitoring model stored in the model storage unit 11. ..

In S250, the acquisition unit 12 calculates a reconstruction error vector, which is an error of the reconstruction data with respect to the input data, using the above equation (5).
In S260, the acquisition unit 12 determines whether or not each of the individual errors, which are each element of the reconstruction error vector calculated in S250, is larger than the preset error threshold value, and determines whether or not the individual error is larger than the error threshold value. The element to have (that is, any dimension of the input data) is extracted as the target element.

In S270, the acquisition unit 12 determines whether or not the target element has been extracted in S260. If no target element is extracted, the process proceeds to S280, and if any target element is extracted, the process proceeds to S290.

In S280, the acquisition unit 12 outputs the input data generated in S230 as the target data to be processed by the extraction unit 14, activates the extraction unit 14, and ends this processing.
In S290, the acquisition unit 12 outputs the reconstruction data generated in S240 to each of the one or a plurality of target elements extracted in S260 as the restoration target in the restoration unit 13, and activates the restoration unit 13. This process ends.

That is, as shown in FIG. 6, the acquisition unit 12 generates the reconstruction data generated by applying the input data to the monitoring model, and obtains the reconstruction error vector which is the difference between the input data and the reconstruction data. By evaluating each element, it is estimated that the sensor has an abnormality such as data loss.

[1-3-2. Restoration process]
Next, the restoration process executed by the restoration unit 13 will be described with reference to the flowchart of FIG. 7. The restoration process is started by executing S290 described above.

When the restoration process is started, in S310, the restoration unit 13 acquires the input data generated in S230.
In S320, the restoration unit 13 applies the input data to the monitoring model to generate reconstruction data.

In S330, the restoration unit 13 calculates a reconstruction error vector which is an error of the reconstruction data with respect to the input data, and calculates the error evaluation value shown in the equation (6) based on the reconstruction error vector.
In S340, the restoration unit 13 calculates a gradient for reducing the error evaluation value calculated in S330 for the target element of the input data by using the error back propagation method. Specifically, as shown in Eq. (7), the partial derivative of the objective function with respect to the target element is calculated as a gradient. However, p represents the dimension of the input data that specifies the target element. If there are multiple target elements, the same process is executed for each.

なお、誤差逆伝播法は、例えば、Rumelhart,David E.,Hinton,Geoffrey E.,Williams,Ronald J.(8 October 1986).”Learning representations by back-propagating errors”.等に詳述された公知技術であるため、ここでの説明は省略する。

The back-propagation method is known as described in detail in, for example, Rumelhart, David E., Hinton, Geoffrey E., Williams, Ronald J. (8 October 1986). "Learning representations by back-propagating errors". Since it is a technology, the description here is omitted.

In S350, the restoration unit 13 updates the target element of the input data according to the equation (8) using the gradient calculated in S340. However, η is the learning rate of parameter update. The learning rate η may be multiplied by a constant every fixed number of updates, for example, changed from 0.1 to 0.01 to 0.001 every 10 updates, or is known as AdaGrad, RMSProp, Adam, etc. It may be set by using the setting method of. In the following, the input data whose target element has been updated is referred to as restored data.

つまり、復元処理では、学習処理と同様の手法が用いられるが、学習処理では、学習データ全体の誤差評価値の和である目的関数の値を減少させるようにニューラルネットワークの各ノードの重みを更新するのに対して、復元処理では、入力データに対する誤差評価値を減少させるように入力データの対象要素を更新する点で異なる。

In other words, the restoration process uses the same method as the training process, but in the training process, the weights of each node of the neural network are updated so as to reduce the value of the objective function, which is the sum of the error evaluation values of the entire training data. On the other hand, the restoration process differs in that the target element of the input data is updated so as to reduce the error evaluation value for the input data.

In S360, the restoration unit 13 determines whether or not the preset end condition is satisfied. If the end condition is not satisfied, the restored data generated in S350 is used in place of the input data acquired in S310, and the processes of S320 to S350 are repeated. If the end condition is satisfied, the process proceeds to S370. As the end condition, for example, at least one of the fact that the value of the error evaluation value is equal to or less than the preset end threshold value and the number of repetitions of the processes of S320 to S350 reaches a predetermined number of times can be used.

In S370, the restoration unit 13 outputs the restoration data calculated at the time when the end condition is satisfied as the target data to be processed by the extraction unit 14, activates the extraction unit 14, and ends this processing. do.

That is, in the restoration process, as shown in FIG. 8, the value of the target element is restored by repeatedly updating the value of the target element of the input data in the direction of back-propagating the reconstruction error and reducing the reconstruction error. Generate the restored data. By this restoration process, the sensor output (target element) with an abnormality such as a defect realizes the correlation of each element in the normal state held in the monitoring model, that is, its own sensor output and other sensor outputs. It is restored to be consistent with.

[1-3-3. Extraction process]
Next, the extraction process executed by the extraction unit 14 will be described with reference to the flowchart of FIG. The extraction process is started by executing S270 or S370 described above.

When the extraction process is started, in S410, the extraction unit 14 acquires the target data. That is, if it is determined that there is an abnormality in the input data and the restoration data is generated by the restoration unit 13, the restoration data is used as the target data, and if it is determined that there is no abnormality in the input data, the input data is used. It is used as it is as target data.

In S420, as shown in FIG. 10, the extraction unit 14 applies the target data to the encoder unit of the monitoring model, and the output of the intermediate layer corresponding to the output of the encoder unit is used as information indicating the latent feature of the target data. Extract. The information to be extracted is not limited to the output of the intermediate layer corresponding to the output of the encoder unit, and the output of another intermediate layer may be used. Further, the information to be extracted may be switched according to the application executed by the subsequent device (for example, the information processing device 5). For example, if the device in the subsequent stage uses the detection data of the sensor group 3 as input, the extraction unit 14 uses partial data of the target data (that is, input data if there is no abnormality, restoration data if there is an abnormality). You may output it as it is.

In S430, the extraction unit 14 outputs the target data and the output of the intermediate layer extracted in S420 to the information processing apparatus 5 as output information, and ends this process.
[1-4. Information Processing Department]
The information processing apparatus 5 performs post-stage processing such as data analysis based on the output information from the information extraction apparatus 10. Specifically, for example, the statistic of the target data may be obtained, or the target data or the statistic may be used to analyze the driving scene, analyze the driver characteristics, and the like. The driver characteristics to be analyzed include, for example, the average accelerator opening degree, steering shake, and the like.

Further, the information processing apparatus 5 may extract information for a specific application based on the output information. Specifically, for example, the target data may be visualized as it is, or information indicating signs such as dangerous driving or involuntary driving may be extracted and used for a process of calling attention to the driver.

Here, the information extraction device 10, the model generation device 20, and the information processing device 5 may all include a microcomputer having a CPU and, for example, a semiconductor memory (hereinafter, memory) such as RAM or ROM. .. In this case, each function of the information extraction device 10, the model generation device 20, and the information processing device 5 is realized by the CPU executing a program stored in a non-transitional substantive recording medium. In this example, the memory corresponds to a non-transitional substantive recording medium in which a program is stored. In addition, when this program is executed, the method corresponding to the program is executed. The information extraction device 10, the model generation device 20, and the information processing device 5 may all include one microcomputer or a plurality of microcomputers.

The method for realizing the functions of each part included in the information extraction device 10, the model generation device 20, and the information processing device 5 is not limited to software, and some or all of the functions use one or more hardware. May be realized. For example, when the above function is realized by an electronic circuit which is hardware, the electronic circuit may be realized by a digital circuit, an analog circuit, or a combination thereof.

[2. effect]
According to the embodiment described in detail above, the following effects are obtained.
(1a) The information processing system 1 uses a monitoring model that learns the normal correlation between each element of the input data, and if there is an abnormality in the input data, the output information is output from the restored data obtained by restoring the abnormal element. Is being extracted. Therefore, even if there is a possibility that an abnormality such as a temporary loss may occur in the input data, it is possible to perform robust information extraction in which the influence of the abnormality is suppressed. As a result, the accuracy and reliability of the post-stage processing using the output information can be improved.

(1b) Since the information processing system 1 uses a multi-layer self-encoder as a monitoring model, latent features in which redundant information of input data is excluded are extracted as output information from the intermediate layer of the monitoring model. can do.

(1c) Since the information processing system 1 uses a multi-layer self-encoder capable of learning non-linear features as a monitoring model, it is possible to accurately reproduce complex correlations between elements.

(1d) Since the information processing system 1 uses a monitoring model for restoring the target element (that is, missing data) of the input data, the input data is input according to the normal correlation between each element of the input data held by the monitoring model. The target element can be restored so that the entire data is consistent. That is, when the value of a specific sensor in the detection data includes an abnormality such as a defect, the target element can be restored so that the entire detection data including the value of another sensor can be matched. Furthermore, since the input data includes a time series of detection data, if the output of the sensor that intermittently causes an abnormality such as a defect in a width narrower than the time window becomes the target element, it will move back and forth on the time axis. The target element can be restored so that it matches the data of the same located sensor.

(1e) In the information processing system 1, it is realized by detecting an element having an abnormal value due to data loss or the like (that is, a target element), restoring the target element, extracting a latent feature, and using a single monitoring model. Therefore, the system configuration can be simplified.

(1g) In the information processing system 1, the reconstruction error is obtained by using the monitoring model, and the element whose individual error is larger than the error threshold value is detected as the target element that needs to be restored. Therefore, it is not necessary to prepare a standard other than the error threshold value in advance for the detection of the target element, and the design effort can be reduced.

[3. Other embodiments]
Although the embodiments of the present disclosure have been described above, the present disclosure is not limited to the above-described embodiments, and can be variously modified and implemented.

(3a) In the above embodiment, the target element is detected by using the monitoring model, but the present invention is not limited to this. For example, a sensor with an abnormality detection function may be used as a sensor constituting the sensor group 3, and the sensor self-confidence may output the presence or absence of an abnormality. Further, the abnormality of the sensor may be detected by an abnormality detector provided separately from the sensor. The configuration for detecting an abnormality in the sensor without using these monitoring models corresponds to the abnormality detection unit.

(3b) In the above embodiment, the case where the sensor group 3 is a plurality of sensors mounted on the vehicle has been described, but the present invention is not limited to this, and the outputs of the plurality of sensors have a correlation with each other. Just do it.

(3c) A plurality of functions possessed by one component in the above embodiment may be realized by a plurality of components, or one function possessed by one component may be realized by a plurality of components. .. Further, a plurality of functions possessed by the plurality of components may be realized by one component, or one function realized by the plurality of components may be realized by one component. Further, a part of the configuration of the above embodiment may be omitted. Further, at least a part of the configuration of the above embodiment may be added or replaced with the configuration of the other above embodiment. It should be noted that all aspects included in the technical idea specified from the wording described in the claims are embodiments of the present disclosure.

(3d) In addition to the above-mentioned information extraction device 10, a system having the information extraction device 10 as a component, a program for operating a computer as the information extraction device 10, a semiconductor memory in which this program is recorded, and the like are non-transitional. The present disclosure can also be realized in various forms such as an actual recording medium and an information extraction method.

1 ... Information processing system, 3 ... Sensor group, 5 ... Information processing device, 10 ... Information extraction device, 11,23 ... Model storage unit, 12 ... Acquisition unit, 13 ... Restoration unit, 14 ... Extraction unit, 15 ... Reception unit , 20 ... model generator, 21 ... information storage unit, 22 ... learning unit, 24 ... distribution unit.

Claims (8)

An acquisition unit (12) configured to acquire input data represented by a multidimensional vector in which each element has a correlation and status information indicating an abnormal status of each element in the input data, and an acquisition unit (12).
Using a monitoring model, which is a neural network in which normal correlations between the elements of the input data are learned, using each of the one or a plurality of the elements indicated as abnormal by the situation information as the target element, the input is performed. A restoration unit (13) configured to generate restoration data obtained by restoring the target element in the data to a normal value, and
When the status information indicates that there is no abnormality, the input data is extracted, when the status information indicates that there is an abnormality, the restored data is used as the target data, and one or more latent features related to the target data are extracted as output information. The configured extraction unit (14) and
An information extraction device equipped with. The information extraction device according to claim 1.
An information extraction device configured to use a self-encoder having a plurality of intermediate layers as the monitoring model. The information extraction device according to claim 2.
The extraction unit is an information extraction device configured to use the output from the intermediate layer of the monitoring model obtained by inputting the target data into the monitoring model as the latent feature. The information extraction device according to any one of claims 1 to 3.
The restoration unit uses the monitoring model to reconstruct the input data input to the monitoring model as reconstruction data, and evaluates an error indicating the magnitude of the error of the reconstruction data with respect to the input data. The gradient of the value and the error evaluation value with respect to the target element is calculated, the target element is updated in the direction in which the error evaluation value becomes smaller according to the gradient, and the same is performed using the input data in which the target element is updated. An information extraction device configured to generate the restored data by repeating the process of. The information extraction device according to any one of claims 1 to 4.
The acquisition unit is configured to continuously generate the input data by dividing the time-series data output from each of the plurality of sensors mounted on the vehicle at regular time intervals and collectively vectorizing the time-series data. Was done,
Information extraction device. The information extraction device according to claim 5.
The acquisition unit is an information extraction device configured to divide the time-series data by using a gliding time window. The information extraction device according to any one of claims 1 to 6.
The acquisition unit calculates a reconstruction error, which is the difference between the input data and the reconstruction data, using the result of reconstructing the input data input to the monitoring model as the reconstruction data using the monitoring model. Then, among the individual errors that are the values of each element of the reconstruction error, the element having the individual error larger than the preset error threshold is set as the target element, and the target element is present. Configured to generate the status information indicating that the element is abnormal,
Information extraction device. The information extraction device according to claim 5 or 6.
The acquisition unit is an information extraction device configured to acquire the detection result of the abnormality detection unit that detects the presence or absence of an abnormality in the output of the sensor provided in each of the plurality of sensors as the status information.

Images