JP7272575B2 - Data processing device, data processing system and program - Google Patents

Description

The present invention relates to a data processing device, a data processing system and a program.

In recent years, a system using machine learning has been developed as an anomaly detection system that can be applied to various places such as product manufacturing sites.

For example, in Patent Document 1, log data that can be collected at the site is recorded in advance, and while referring to the log data, the presence or absence of an abnormality at each recording time of the log data is used as teacher data for learning processing. An apparatus for determining abnormality is disclosed.

JP 2018-73258 A

However, the actual situation is that, for example, at a product manufacturing site, the data output by a sensor that collects information such as vibration differs from site to site depending on the surrounding noise conditions. Specifically, there is a large difference in output between a site where manufacturing machines that generate different vibrations are operating adjacent to each other and a site where they are not. In addition, since the output is different when the manufacturing machine that produces different vibrations is stopped and when it is operating, the data is the basis for determining abnormalities depending on various factors such as the time of day even at the same site. It is common for there to be substantial differences in

Because of this situation, equipment that collects log data in advance and learns needs to collect log data for each site, which complicates the equipment configuration and requires many preparatory processes before it can be operated. Was. Moreover, even if the learning process is performed in this way, the environment may differ depending on the time zone as described above, so there are cases where appropriate determination cannot always be made.

SUMMARY OF THE INVENTION The present invention has been made in view of the above circumstances, and provides a data processing apparatus, a data processing system, and a program capable of simplifying the preparation process with a relatively simple configuration and making judgments suitable for the environment. to be one of its purposes.

One aspect of the present invention that solves the problems of the above conventional example is a data processing device that performs a predetermined determination process based on data that is repeatedly input, and receives input data and teacher data, and performs a reciprocal operation on the machine. A learnable guessing means, a learning processing means for machine-learning the guessing means using the input data as input data and teacher data, and a comparison between the output of the guessing means and the input data. and output means for performing the predetermined determination process and outputting the result of the determination process.

According to the present invention, the judgment is performed while performing machine learning using input data, and the preparatory process can be simplified with a relatively simple configuration by using an inference means capable of machine learning by reciprocal arithmetic, It is possible to make a judgment suitable for the environment.

1 is a configuration block diagram showing an example of a data processing device according to an embodiment of the present invention; FIG. 1 is a functional block diagram of an example of a data processing device according to an embodiment of the present invention; FIG. FIG. 4 is a configuration block diagram showing an example of an estimator used by the data processing device according to the embodiment of the present invention; 4 is an explanatory diagram showing an example of parameter storage by the data processing device according to the embodiment of the present invention; FIG. It is a functional block diagram concerning another example of the data processor concerning an embodiment of the invention. It is a functional block diagram concerning another example of the data processor concerning an embodiment of the invention.

An embodiment of the present invention will be described with reference to the drawings. An example of a data processing apparatus 1 according to an embodiment of the present invention includes a control section 11, a storage section 12, an input section 13, and an output section 14, as shown in FIG.

Here, the control unit 11 is a program control device such as a CPU, or a logic device such as an FPGA (Field Programmable Gate Array), or an ASIC (Application Specific Integrated Circuit). implement the means. When a program control device such as a CPU is used as the control section 11, the control section 11 executes a program stored in the storage section 12 to realize the operations of the above sections.

When a logic device such as an FPGA is used as the control section 11, it operates according to programmed logic to realize the operation of each section described above.

That is, in the present embodiment, the control unit 11 receives data that is repeatedly input, uses the input data as input data and teacher data, and executes machine learning processing of a machine learning model that can be machine-learned by reciprocal arithmetic. Also, the control unit 11 obtains the output of the estimator when the input data is input to the estimator represented by the machine learning model as input data. Then, the control unit 11 performs predetermined determination processing based on the comparison between the obtained output of the estimator and the input data, and outputs the result of the determination processing. A detailed example of the operation of the control unit 11 will be described later.

The storage unit 12 includes a memory device and a disk device. The storage unit 12 holds programs executed by the control unit 11 when the control unit 11 is a program control device such as a CPU. This program may be provided by being stored in a computer-readable, non-temporary recording medium and stored in the storage unit 12 .

The storage unit 12 also operates as a work memory that holds information necessary for the processing of the control unit 11, such as model parameters of a machine learning model when performing machine learning on the estimator.

The input unit 13 converts data input from an external sensor or the like into digital data, converts it into vector information of a predetermined dimension, and outputs the vector information to the control unit 11 . The output unit 14 outputs information according to instructions input from the control unit 11 . The output unit 14 is, for example, a display, and displays and outputs information input from the control unit 11 .

In the present embodiment, the control unit 11 functionally includes a data receiving unit 21, an estimator 22, a learning processing unit 23, a determination processing unit 24, and an output processing unit 25, as illustrated in FIG. and

The data receiving unit 21 receives input of data from the input unit 13 . In the example of the present embodiment, the data processing device 1 is arranged, for example, at a product manufacturing site, is attached to a device used for manufacturing the product, and detects various information such as vibration and temperature of the device. It is connected to the sensor that outputs. The input unit 13 repeatedly converts the outputs of these sensors into digital values at predetermined timings (for example, at regular timings) and outputs the digital values to the control unit 11 . The input unit 13 also outputs to the control unit 11 vector values obtained by arranging digital values obtained by converting the outputs of a plurality of sensors in a predetermined order as data. The data receiving unit 21 receives data output by the input unit 13 at predetermined timings and outputs the data to the learning processing unit 23 .

The estimator 22 implements estimating means, and functionally, as illustrated in FIG. It is a combinatorial neural network. In this embodiment, the estimator 22 has the number of nodes in the input layer 31 (the dimension of the input data vector) and the number of nodes in the output layer 33 (the dimension of the vector of output data). It is assumed that there is

, 31n of the input layer 31 and the nodes 32a, 32b, . _{is W} (w _{1 , w 2} , _. , 32L _of the intermediate layer 32 and the nodes _33a , _33b , _. ), these connection weight values W and V and the bias b become parameters of the machine learning model of the estimator 22 .

When input data is input, the estimator 22 inputs the input data to the input layer 31, multiplies each component of the input data input to the input layer 31 by the corresponding connection weight W, and sums them up. By performing a predetermined operation based on each component of the input data input to the input layer 31 and the connection weight W, the value of each node of the intermediate layer 32 is obtained. Also, for values f(h _i ) (i= ₁ , ₂ , . . . _L ) obtained by applying a predetermined nonlinear function f to values h 1 , h 2 . , is multiplied by the corresponding connection weight V to obtain the value of each node in the output layer 33 . Since this calculation is similar to calculation in a general neural network, detailed explanation is omitted here.

As the nonlinear function here, widely known functions such as a sigmoid function and ReLU may be adopted. The estimator 22 outputs a vector whose components are the values of the nodes of the output layer 33 as output data.

In the present embodiment, the estimator 22 can perform machine learning of the parameters of the machine learning model by reciprocal arithmetic, at least depending on the conditions of the learning process. An example of such an estimator 22 will be described in detail later.

The learning processing unit 23 performs machine learning processing of the estimator 22 using the data received by the data receiving unit 21 as input data and teacher data. Specifically, the learning processing unit 23 inputs data for the latest predetermined number of times (batch size) received by the data receiving unit 21 as input data to the estimator 22 . If the batch size is set to 1, the learning processing unit 23 inputs the data received by the data receiving unit 21 to the estimator 22 as it is as input data.

The learning processing unit 23 compares the output data output by the estimator 22 when the input data is input with the data (teacher data) received by the data receiving unit 21, and calculates the difference (absolute error or mean square error etc.) is calculated as a loss. Then, the learning processing unit 23 updates the parameters (connection weight values) of the estimator 22 so that the loss becomes smaller.

That is, in the example of the present embodiment, the learning processing unit 23 performs machine learning using the estimator 22 as an autoencoder.

The determination processing unit 24 executes predetermined determination processing based on the output data output from the estimator 22 when the learning processing unit 23 inputs input data to the estimator 22 .

As a specific example, the determination processing unit 24 refers to the loss calculated by the learning processing unit 23, and if the magnitude of the loss exceeds a predetermined threshold, it indicates that the input data is abnormal (i.e. , that an abnormality has occurred in the manufacturing apparatus, etc.) is output to the output processing unit 25 . When the magnitude of the loss does not exceed the predetermined threshold value, the judgment processing unit 24 outputs a judgment result indicating that the input data is normal (that is, there is no abnormality in the manufacturing equipment, etc.). You may make it output to the process part 25. FIG.

Note that the determination processing unit 24 may be controlled so as not to perform determination processing until the parameters of the estimator 22 are sufficiently learned by the learning processing unit 23 . Specifically, the determination processing unit 24 determines the number of times input data is input to the estimator 22 (when the parameter is updated by the learning processing unit 23) from the state where the estimator 22 is not performing machine learning (reset state). performed) exceeds a predetermined initialization threshold, and if the initialization threshold is exceeded, it may be determined that the parameters of the estimator 22 have been sufficiently learned. . Here, the initialization threshold value is determined in advance as a number equal to or greater than the number L of nodes in the intermediate layer 32, for example.

In this case, the determination processing unit 24 continues until the learning process is executed the number of times determined as the initialization threshold (until the estimator 22 receives the input data the number of times of the initialization threshold and updates the parameters). ) does not process judgment.

The output processing unit 25 outputs the determination result to the output unit 14 according to the instruction input from the determination processing unit 24 .

Here, a specific machine learning model of the estimator 22 according to the example of this embodiment will be described. In this embodiment, the estimator 22 randomly determines the connection weight W between the input layer 31 and the intermediate layer 32 and the bias b by the operation of the learning processing unit 23, and the intermediate layer 32 and the output layer 33 It is assumed that the connection weight V between is a neural network for machine learning. Specifically, an OS-ELM (Online Sequential-Extreme Learning Machine) is realized and used here by the estimator 22 and the learning processing unit 23 . This OS-ELM is based on N.Y. Liang, G.B. Huang, P.Saratchandran, and N.Sundararajan,”A Fast and Accurate Online Sequential Learning Algorithm for Feedforward Networks,” IEEE Transactions on Neural Networks, Vol. 17, No.6, pp. 1411-1423, Nov. 2006, etc., and is widely known, so detailed description thereof will be omitted here.

When this OS-ELM is used, the learning processing unit 23 initially resets the estimator 22, so the connection weight W and the bias b between the input layer 31 and the intermediate layer 32 are randomly determined (at this time , the connection weight V may also be determined randomly). Then, the learning processing unit 23 inputs the data received by the data receiving unit 21 to the estimator 22 as input data. The learning processing unit 23 compares the output data output by the estimator 22 when the input data is input and the data (teacher data) received by the data receiving unit 21, and the difference (for example, the mean square error ) as a loss. Then, the learning processing unit 23 updates the connection weight V between the intermediate layer 32 and the output layer 33 of the estimator 22 so that the loss becomes smaller.

なお、Ｇは、出力層３３の各ノードの出力を表し、ｘ_i（ｊ）は、ｉ番目のバッチに含まれるｊ番目の入力データであることを表す。なお、ｎ_iは、ｉ番目のバッチにおけるバッチサイズである。 In OS-ELM, the learning process is performed as follows. The learning processing unit 23 divides the input data into n _i (i=1, 2, . . . ) batches, and sequentially uses each batch as training data. Here, let the input data contained _in the i- _th batch be x _i (i=1, ₂ , . _{2 , . _ _ _ _} .

Note that G represents the output of each node of the output layer 33, and x _i (j) represents the j-th input data included in the i-th batch. Note that n _i is the batch size in the i-th batch.

となる。なお、右肩のＴは転置を意味する（以下同じ）。 Also, since the teacher data here is the same as the input data, this teacher data T is

becomes. The right shoulder T means transposition (same below).

を最小とする結合重みβを求めて、これを推定器２２の中間層３２と出力層３３との間の結合重みＶとすることで、推定器２２を最適化することとなる。このとき、

を用い、ｉ＝２の場合を考慮すると、（１）式は、

と変形できる。なお、Ａ^-1は、Ａの疑似逆行列を意味する。 When the i-th batch is input, the learning processing unit 23 determines that the magnitude of the loss is

is obtained to minimize , and is used as the connection weight V between the intermediate layer 32 and the output layer 33 of the estimator 22 , thereby optimizing the estimator 22 . At this time,

and considering the case of i=2, the formula (1) is

and can be transformed. Note that A ⁻¹ means a pseudo-inverse matrix of A.

とすることができる（逐次更新式）。ただし、

である。 Generalizing this, if learning up to the i-th batch is completed and the connection weight V between the intermediate layer 32 and the output layer 33 of the estimator 22 at that time is V=β _i , the connection weight _βi +1, which is the result of machine learning based on the i+1-th batch,

(sequential update formula). however,

is.

とすると、これらの逐次更新式は、

として（ただしＩは単位行列）、

と表現できる。 here,

Then these iterative update formulas are

as (where I is the identity matrix),

can be expressed as

は、スカラ値となり（（４）式は、Ｎ_i×Ｎ_iの行列であるため）、従ってこの疑似逆行列は、単なる逆数演算により求められることとなる。 Therefore, the learning processing unit 23 obtains the output H _i of the intermediate layer 32 when the i-th batch is input, and the connection weight V=β between the intermediate layer 32 and the output layer 33 optimized at that time is _i and P _i as an intermediate result are obtained. Then, when the i+1-th batch is input, the next intermediate result P _i+1 is obtained by equation (2), and the connection weight V between the intermediate layer 32 and the output layer 33 is calculated by equation (3). Update to _i+1 . In general, _singular value decomposition is used for _the calculation of the inverse matrix. of formula (2), which is the matrix for which

is a scalar value (because equation (4) is a matrix of N _i ×N _i ), so this pseudo-inverse matrix can be obtained by a simple reciprocal operation.

That is, in the present embodiment, the estimator 22 and the learning processing unit 23, which can be machine-learned by inverse calculation, are neural networks that perform machine learning on the connection weights between the intermediate layer and the output layer by pseudo-inverse matrix calculation. and set the machine learning batch size to 1 (by performing machine learning each time data is received). Specifically, such a neural network is an ELM (Extreme Learning Machine) that randomly determines the connection weight between the input layer and the intermediate layer, and a neural network derived from it (FP (Forgetting Parameters)-ELM, OS (On -Line Sequential)-ELM, EOS (Ensemble of OS)-ELM, FOS-ELM (OS-ELM with forgetting mechanism)), etc. However, the neural network that implements the estimator 22 and the learning processing unit 23 as an estimating means capable of machine learning by reciprocal arithmetic is not limited to these examples.

[motion]
The data processing apparatus 1 of this embodiment has the above configuration and operates as follows. In the following example, it is assumed that the controller 11 is implemented using FPGA, and OS-ELM with a batch size of 1 is used as the estimator 22 .

Also, here, the data processing device 1 receives signals from a plurality of vibration sensors arranged in the vicinity of a product manufacturing device. The vibration sensor outputs an analog electrical signal representing the magnitude of vibration of the site to which it is attached.

The data processing device 1 initially randomly determines the connection weight W and the bias b between the input layer 31 and the intermediate layer 32 of the OS-ELM, which is the estimator 22 . The data processing device 1 then converts the electrical signals output from the sensors into digital values and inputs the digital values to the control unit 11 . The control unit 11 functions as a data receiving unit 21 and outputs to the learning processing unit 23 a vector value obtained by arranging digital values obtained by converting the outputs of a plurality of sensors in a predetermined order.

The learning processing unit 23 inputs the input data to the estimator 22 each time it receives data input, that is, each time it receives input data with a batch size of 1. The learning processing unit 23 uses the output data H of the estimator 22, the input data T as teacher data, the connection weight V= _β between the intermediate layer 32 and the output layer 33 of the estimator 22 at that stage, and the previous P _i (the initial values of both β and P are set in advance) is obtained as an intermediate result of the calculation.

Then, the learning processing unit 23 updates the connection weight V between the intermediate layer 32 and the output layer 33 of the estimator 22 according to equations (2) and (3). The learning processing unit 23 also calculates the mean square error between the output of the estimator 22 and the input data (teacher data) as a loss, and outputs it to the determination processing unit 24 .

The determination processing unit 24 determines whether or not the parameters of the estimator 22 have been sufficiently machine-learned. This determination is made based on whether or not the number of times input data is input after the data processor 1 initializes the estimator 22 exceeds a predetermined initialization threshold value. When the determination processing unit 24 determines that the parameters of the estimator 22 are not sufficiently machine-learned, the determination processing unit 24 does not perform determination processing.

On the other hand, when determining that the parameters of the estimator 22 have been sufficiently machine-learned, the determination processing unit 24 determines that the magnitude of the loss value input from the learning processing unit 23 exceeds the predetermined threshold. Examine whether or not the value exceeds the value, and if the magnitude of the loss value exceeds the predetermined threshold value, it is estimated that the input data is abnormal and the manufacturing equipment, etc. is abnormal. It performs processing such as controlling to output the result of determination indicating that to the display, which is the output unit 14 .

As described above, according to the data processing device 1 of the present embodiment, the sequential machine learning of the estimator is performed as an autoencoder after being installed at the site where actual anomaly detection is performed (that is, without separately preparing teacher data ), and anomaly detection is performed based on the result of the machine learning.

Also, by using an estimator capable of performing computations in the process of machine learning by relatively simple reciprocal computations, a relatively simple configuration can be achieved.

Here, an example of detecting an abnormality in a device for manufacturing a product has been described, but the present embodiment is not limited to this example. It is also possible to detect anomalies related to the amount of current using the current on the wiring as input data, detect anomalies in the heat distribution using the heat distribution (thermography) data of a certain product as input data, and detect human behavior and equipment. It can be applied to various examples such as anomaly detection using operation history as input data, anomaly detection related to the operation of an unmanned aerial vehicle (such as a drone).

[Forget]
In addition, in the example of the present embodiment, if the forgetting process is to be included, the forgetting rate is set to α and P _i+1 in equation (3) is multiplied by 1/α ² . This makes it possible to obtain the forgetting effect by a simple method.

[Partial discard of learning results]
Further, in the present embodiment, when learning processing is performed using a group of input data with a relatively small batch size, such as a batch size of 1, adaptation to abnormal data caused by continuous abnormal data is avoided. In order to prevent this, the following process of partially discarding the learning results may be performed.

In one example of the present embodiment, the learning processing unit 23 directly inputs the data received by the data receiving unit 21 to the estimator 22 as input data (the batch size is set to 1 and sequential learning processing is performed). . Specifically, an OS-ELM neural network may be used as the estimator 22 . In this case, the estimator 22 and the learning processing unit 23 realize an OS-ELM that performs sequential learning processing with a batch size of "1".

In this example, the learning processing unit 23 inputs the input data X to the estimator 22, and each time the machine learning process is performed using the output data output by the estimator 22 and the input data X, that is, the estimator 22 Each time the connection weight V between the intermediate layer 32 and the output layer 33 is updated, the input data X input to the estimator 22, the connection weight β obtained for updating, and data required for machine learning processing (for example, The above P) are associated with each other and stored in the storage unit 12 .

Also, at this time, the learning processing unit 23 deletes the information stored in the storage unit 12 that associates the input data X, the connection weight β, and the intermediate result P stored before M times in the past. may

As a result, in the storage unit 12, the connection weight β _j between the intermediate layer 32 and the output layer 33, which is the machine learning result of the estimator 22 for the most recent M times, P _j as the intermediate result, and each connection Information _{R j (} j=1, 2, .

Each time the learning processing unit 23 performs machine learning processing a plurality of times (here, M times) determined by a predetermined method, the intermediate layer 32 and the output layer 33 stored in the storage unit 12 M times before. , intermediate results P, and input data X are read. The learning processing unit 23 inputs the read input data X to the estimator 22 .

The learning processing unit 23 refers to the output of the estimator 22 and determines whether or not the output satisfies a predetermined condition. Here, the learning processing unit 23 calculates a value (for example, the mean square error) based on the difference between the output of the estimator 22 and the input data X (corresponding to teacher data) as a loss, and the magnitude of this loss is Check whether or not a predetermined threshold is exceeded. A condition that the magnitude of the loss calculated here exceeds a predetermined threshold value corresponds to, for example, an example of the above-described predetermined condition.

When the magnitude of the loss exceeds a predetermined threshold value, that is, when the predetermined condition is satisfied, the learning processing unit 23 stores the data in association with the input data input here. The machine learning state of the estimator 22 is corrected using the connection weight β between the intermediate layer 32 and the output layer 33 M times before, which is the result of machine learning. Specifically, the connection weight V between the intermediate layer 32 and the output layer 33 of the estimator 22 is set to the connection weight β read here (this operation partially discards the most recent machine learning results). (equivalent to

Also, at this time, the learning processing unit 23 resets the intermediate result stored at this time to the intermediate result P read here.

Note that the learning processing unit 23 continues the machine learning processing without correcting the machine learning state of the estimator 22 when the magnitude of the loss does not exceed the predetermined threshold value.

According to this example of the present embodiment, each time machine learning is performed the number of times determined by a predetermined method (for example, it may be determined experimentally), the input data at a predetermined point in the past is used as the current estimation. Input to the device 22 and check whether the magnitude of the loss has increased. If the loss is large, the contents of the most recent machine learning are discarded, and the parameters of the estimator 22 are returned to the parameters of the estimator 22 at the predetermined time in the past.

As a result, time-averaged machine learning is performed as in the case where the batch size is a relatively large value.

[Delayed learning]
Further, in one example of the present embodiment, learning may be delayed instead of partially discarding the learning data. In this example, the learning processing unit 23 directly inputs the data received by the data receiving unit 21 to the estimator 22 as input data (performs sequential learning processing with a batch size of 1). This estimator 22 may use an OS-ELM neural network as in the example already described. That is, here too, the estimator 22 and the learning processing unit 23 implement an OS-ELM that performs sequential learning processing with a batch size of "1".

The learning processing unit 23 receives the input data X and inputs it to the estimator 22, obtains the output data output by the estimator 22, and calculates a value based on the difference between the output data and the input data X (for example, the mean square error ) is calculated as a loss, and the input data X and the calculated loss are stored in the storage unit 12 in association with each other. At this stage, the learning processing unit 23 does not perform machine learning processing for the estimator 22 .

The learning processing unit 23 repeats the above process a predetermined number of times M (where M is a natural number of 2 or more), and obtains the input data X for the past M times and the loss value of the estimation result based thereon, When the data is held in the storage unit 12, it is determined whether or not to perform the machine learning process using the input data X M times before.

Specifically, this determination can be made as follows. That is, the learning processing unit 23 refers to the held information and obtains a statistic value (for example, an average Eav here) based on the loss values for the most recent M times. Then, the learning processing unit 23 refers to the value E of the loss M times before (with respect to the input data X M times before for which it is determined whether or not to perform the machine learning process), and this value E Condition E<a·Eav using average Eav, which is a statistical value
is satisfied or not. This a is a predetermined constant, for example, a=3.0. Although the average is used as the statistical value here, an intermediate value may be used instead of the average. Alternatively, a may be determined based on the variance or standard deviation of the loss values for the last M times instead of a constant.

When the above value E satisfies E<a·Eav, the learning processing unit 23 executes machine learning processing using the input data X M times before. That is, using the input data X of the M times before and the corresponding loss E, the connection weight V between the intermediate layer 32 and the output layer 33 of the estimator 22 is updated.

The learning processing unit 23 may then delete the input data X and the loss information stored M times before.

In the above determination, if the value E does not satisfy E<a·Eav, the learning processing unit 23 does not execute the machine learning process using the input data X M times before. Delete the previously stored input data X and loss information.

According to this example of the present embodiment, it is determined whether or not to perform machine learning by delaying the number of times determined by a predetermined method (for example, it may be determined experimentally), and based on input data that greatly deviates Since control is performed so that machine learning is not performed, it is possible to obtain the same effect as the above-described example in which the most recent learning content is discarded according to conditions, and time-averaged machine learning is performed.

[Parallelization]
Furthermore, according to this embodiment, a plurality of estimators 22 may be provided. The control unit 11 of the data processing device 1 according to this example functionally includes a data receiving unit 21, a plurality of estimators 42-1, 42-2 . . . , a determination processing unit 44, and an output processing unit 25. Note that components having the same configuration as those illustrated in FIG. 2 are denoted by the same reference numerals, and repeated descriptions thereof are omitted.

The estimator 42 according to this example of the present embodiment (if there is no need to distinguish between the estimators, the estimators are collectively referred to as the estimator 42. The same applies to the learning processing unit). ) may be the same as the estimator 22 illustrated in FIG. That is, each estimator 42 may be a neural network corresponding to OS-ELM.

Also, the learning processing unit 43 sequentially updates the parameters of the corresponding estimator 42 by machine learning processing in the same manner as the learning processing unit 23 already described. The learning processing unit 43 also calculates and outputs a value (root mean square error or the like) related to the difference between the output data of the estimator 42 and the input data (corresponding to teacher data) as a loss. Furthermore, in this example of the present embodiment, the learning processing unit 43 outputs the value of the loss, and learns the number of parameter updates (the number of input data inputs) after resetting the corresponding estimator 42. Output as status information. Also in this example, each learning processing unit 43 may execute processing for partially discarding the learning result.

In this example of the present embodiment, the learning processing unit 43-i (i=1, 2, . is reset at each predetermined timing T _i defined in . That is, the learning processing unit 43-i resets the estimator 42-i last time (randomly determines the connection weight W and the bias b between the input layer 31 and the intermediate layer 32), and the time T _i has passed. Each time, the estimator 42-i is reset again.

Here, the timing T _i may be determined by the number of times input data is input (for example, the timing T _i may be set every time input data is input q _i times), or a clock unit (not shown) (which measures or acquires the current time). It may be determined based on the actual passage of time, which is determined based on the time information acquired from the circuit unit that performs the time.

Also, this timing T _i may be a timing at which all the estimators 42 are not reset all at once (at least during the time required for abnormality detection processing). For example, when the timing T _i is determined by the number of inputs of input data, the number of inputs for each timing T _i is assumed to be a prime number. As an example, when two estimators 42 are used, if the respective reset timings are set to T ₁ =27437 and T ₂ =27449 (both are prime numbers), resetting can be performed at the same timing up to 750 million times. disappears.

When the timing T _i is determined by the actual time, T ₁ is set at 00:00:00 every day, T ₂ is set at 1:00:00 every Monday, and so on. will no longer be reset.

As described above, in one example of the present embodiment, the learning processing unit 43 resets each estimator 42 at different timings, for example, so that all the estimators 42 (at least the time required for abnormality detection processing ), so that they are not reset all at once.

In addition, the determination processing unit 44 in this example receives the calculation result of the loss related to the output data output by the corresponding estimator 42 from the plurality of learning processing units 43, and the learning status information (here, the corresponding estimator 42 number of parameter updates since reset).

Then, the determination processing unit 44 refers to the learning status information, refers to the loss output by the learning processing unit 43 corresponding to the estimator 42 whose parameters have been sufficiently learned, and determines that the magnitude of the loss is predetermined. Check if the threshold is exceeded. Here, whether or not the parameters have been sufficiently learned may be determined by, for example, whether or not the number of times the parameters represented by the learning status information have been updated exceeds a predetermined initialization threshold value. In other words, the determination processing unit 44 determines that the parameter has been sufficiently learned if the update count of the parameter represented by the learning status information exceeds a predetermined initialization threshold value.

Of the learning processing units 43 corresponding to the estimators 42 determined to have sufficiently learned parameters, the determination processing unit 44 determines the number Q of estimators 42 determined to have sufficiently learned parameters. , the number q of output losses exceeding a predetermined threshold value is divided, and it is checked whether this value q/Q exceeds a predetermined value, for example, 1/2. When this value q/Q exceeds, for example, 1/2 (when this predetermined value is 1/2, it is determined that a majority of the estimators 42 have detected an abnormality), the input data is abnormal. The output processing unit 25 is caused to output the determination result indicating that there is an abnormality (that is, that there is an abnormality in the manufacturing apparatus or the like).

[Clustering]
Further, when a plurality of estimators 22 are provided and parallelized in this manner, machine learning processing may be performed as follows. Functionally, the control unit 11 of the data processing device 1 according to this example of the present embodiment includes a data receiving unit 21 and a plurality of , a learning processing unit 43', a determination processing unit 44', an output processing unit 25, and a second learning processing unit 45. In addition, the same reference numerals are assigned to the components having the same configuration as those illustrated in FIG. 5, and the repeated description thereof will be omitted.

Each of the estimators 42 according to this example of the present embodiment (also collectively referred to as the estimators 42 when there is no need to distinguish between the estimators) is illustrated in FIG. The same one as the estimator 22 may be used. That is, each estimator 42 may be a neural network corresponding to OS-ELM.

The learning processing unit 43 ′ inputs the input data X received by the data receiving unit 21 to each estimator 42 . Then, the learning processing unit 43' calculates the mean square error between the output data of each estimator 42 and the input data X as a loss, and outputs it to the determination processing unit 44'. The learning processing unit 43 ′ accumulates and holds the input data X in the storage unit 12 .

Further, the learning processing unit 43' specifies the estimator 42 whose judgment result based on the loss related to the output data indicates that the input data X is "normal" among the estimators 42. Information is received from the determination processing unit 44', and parameters of the estimator 42 specified by the information are updated by machine learning processing in the same manner as the learning processing unit 23 already described.

The determination processing unit 44' receives information related to the loss of each estimator 42 from the learning processing unit 43', and determines the normal/abnormality determination result of the input data X by each estimator 42 based on the loss. Output. Specifically, the determination processing unit 44' determines that the input data X is abnormal for the estimator 42 whose corresponding loss magnitude exceeds a predetermined threshold value. Also, it is assumed that the estimator 42 for which the magnitude of the corresponding loss does not exceed the predetermined threshold determines that the input data X is normal. If all the estimators 42 determine that the input data X is abnormal, the determination processing unit 44′ determines that the input data X is abnormal, and instructs the output processing unit 25 to output that effect. .

Further, when at least one estimator 42 determines that the input data is normal, the determination processing unit 44' determines that the input data X is normal and instructs the output processing unit 25 to output that fact.

Further, the determination processing unit 44' outputs information specifying the estimator 42 that has determined that the input data X is normal to the learning processing unit 43'.

The second learning processing unit 45 is activated at a predetermined timing and executes machine learning processing of the estimator 42 . Here, the predetermined timing is the time when the data processing device 1 is started, or once at a predetermined time, every time the input data X is input a predetermined number of times, the input data determined to be abnormal is output a predetermined number of times. It may be determined as the time point when the number of clusters m is exceeded, the time point when the value of the cluster number m determined based on the tendency of the input data is judged to be inappropriate, or the time point when the user instructs. Determining whether or not the value of the number of clusters is appropriate is performed by performing trial clustering with different numbers of clusters, referring to the results, and determining whether or not to change the number of clusters. This is done by repeatedly executing each judgment timing. Here, the specific determination of whether to change the number of clusters is based on the input data belonging to the same cluster in the result of clustering with the number of clusters at the time of the determination process and the result of the above trial clustering. This can be done based on the distance between the input data (cohesion) and the distance between the input data belonging to different clusters (inter-cluster discreteness).

Machine learning processing by the second learning processing unit 45 is performed as follows. The second learning processing unit 45 clusters the input data accumulated and recorded in the storage unit 12 by the learning processing unit 43'. The clustering method here may be any method as long as it is unsupervised clustering, and any one of various processes such as DBScan, SUBCLU, and k-means can be employed.

The second learning processing unit 45 classifies into the same or smaller number of clusters than the number of estimators 42 . In other words, if the number of estimators 42 is n, the number of clusters is m where m≤n. Alternatively, more estimators 42 than the number of clusters obtained here may be prepared in advance (in that case, the data to be input data are accumulated in advance, clustering is performed, and the number of clusters is determined). ).

The second learning processing unit 45 assigns each estimator 42 to one of the clusters as an estimator that performs machine learning on the input data related to one of the clusters. In addition, when the allocation has been performed before, the second learning processing unit 45 uses the result of the previous allocation as it is without changing the allocation again.

Then, the second learning processing unit 45 sequentially retrieves the input data accumulated in the storage unit 12, and refers to the clustering result of the retrieved input data (information representing the cluster to which it belongs). The second learning processing unit 45 inputs the input data to the estimator 42 assigned to the cluster represented by the information, obtains the output data, and uses the same method as described above. , the estimator 42 is machine-learned. At this time, machine learning processing is not performed for the estimators 42 that are not assigned to the cluster represented by the referenced information.

After completing the above processing for all the input data stored in the storage unit 12, the second learning processing unit 45 deletes the input data stored in the storage unit 12 (does not use it for the next learning). control).

After the machine learning processing by the second learning processing unit 45, the operations of the learning processing unit 43', the determination processing unit 44', and the output processing unit 25 are continued.

That is, in this example, input data is classified into clusters, an estimator 42 corresponding to each classified cluster is prepared, and machine learning is performed on the estimator 42 using input data belonging to the corresponding cluster.

According to this example of the present embodiment, each estimator 42 performs machine learning specialized for the type of input data by the operation of the second learning processing section 45 . As a result, for example, in an example of determining an abnormality in a manufacturing apparatus, a plurality of different types of vibration, such as vibration when the manufacturing apparatus is temporarily stopped, vibration during operation, and so on, can be machine-learned. Become.

In this example of the present embodiment, the user may arbitrarily label each cluster. In this example, the control unit 11 stores the label information.

Then, when at least one estimator 42 determines that the input data is normal, the determination processing unit 44 ′ outputs label information related to the cluster to which the estimator 42 is assigned to the output processing unit 25 . , to output the information of the label. According to this, in addition to judging whether the input data is abnormal or normal, the user can determine what kind of state the input data corresponds to (for example, in the above example, such as "device is stopped"). state) can be identified.

In addition, the judgment processing unit 44' comprehensively judges the judgment results of all the estimators 42 to determine the label corresponding to the input data (what state the input data corresponds to). may output the confidence of For example, when one estimator 42 is determined to be normal and all the other estimators 42 are determined to be abnormal, the certainty (the Confidence in cluster classification) is higher. On the other hand, when a plurality of estimators 42 simultaneously determine that one input data is normal, the reliability of the label information related to the cluster to which each of the plurality of estimators 42 is assigned becomes low.

Therefore, the determination processing unit 44' independently determined that each estimator 42 was normal for each of the plurality of input data (when the other estimators 42 were determined to be abnormal, only the relevant estimator 42 determined to be normal), that is, the number of times it was determined to be normal by itself divided by the number of times it was determined to be normal (including the number of times when other estimators 42 were also determined to be normal). It may be output as the confidence of the label associated with the assigned cluster.

[Detection of behavioral anomalies]
Moreover, as already described, the data processing apparatus 1 of the present embodiment can also be used to determine whether the behavior of the subject is normal or abnormal.

When detecting this behavioral abnormality, the behavior of the subject (for example, if the subject is driving a vehicle, it is the behavior of turning the steering wheel to the left, turning to the right, stepping on the accelerator, etc.). In the case of , a code (for example, one letter of the alphabet) representing the type of input command, etc.) is defined in advance, and a series of actions of the subject is expressed as a code string (a string such as A, C, E, B, etc.) .

The data receiving unit 21 of the data processing device 1 receives, from the input unit 13, multiple times input of code strings representing a series of behaviors of the subject at regular intervals. The data receiving unit 21 then generates a state transition table for each period based on the code string corresponding to each period. Specifically, the data receiving unit 21 obtains the i-th (i is an integer of 0<i<N) code C _i and the i+1-th code C _i+1 from a code string consisting of N codes. Then, a permutation (C _i , C _i+1 ) is created, and the appearance probability for each permutation is calculated.

The data receiving unit 21 generates vector information associated with the appearance probabilities for all possible permutations (assuming that the appearance probabilities for permutations not extracted from the code string are 0) to form a state transition table. The data receiving unit 21 outputs the vector information of the state transition table as input data to the learning processing unit 23 (or the learning processing unit 43 or learning processing unit 43') to obtain an estimator that determines normality/abnormality.

However, it is assumed that the state transition table generated in this way is sparse vector information (most elements are "0").

Therefore, in this example of the present embodiment, the data receiving unit 21 compresses the state transition table obtained for each of the plurality of periods using well-known compression (combining a plurality of elements to reduce the number of elements). ) After the compression processing, the vector information after the compression processing may be used as the input data. A widely known method such as a method based on Candes-Tao's theory can be used for this process, so a detailed description thereof will be omitted here.

Alternatively, when the data receiving unit 21 obtains the state transition table V _j corresponding to the j-th period, if there is a state transition table V _j-1 corresponding to the j-1-th period, the j-th period Using the state transition table V' _j obtained by the above method corresponding to the period, the state transition table V _j to be obtained is:
V _j =V′ _j +α·V _j−1
can be obtained as Here, α is an arbitrary constant, such as α=0.8.

The data processing device 1 of this example may have the configuration illustrated in FIG. In the case of having the configuration of FIG. 6, each estimator 42 detects whether the behavior of the subject driving the vehicle is abnormal or normal. It is expected that it will be differentiated and machine-learned. In this case, the data processing device 1 outputs information indicating that the behavioral abnormality is detected when any of the estimators 42 determines that the behavior is abnormal.

[Preprocessing]
Furthermore, the data receiving unit 21 of the data processing device 1 of the present embodiment may preprocess the data output by the input unit 13 . In this preprocessing, vector data x (elements are (x ₀ , x ₁ , x ₂ . . . , x _n )) to be processed is subjected to a predetermined transformation, and vector y ( Let the elements be (y ₀ , y ₁ , y _{2 .} . . , y _n ),
y _i =Σα _j x x _j
(However, Σ means to find the sum of j). where α is the filter function (kernel), for example
α _j =0 (when j<i−1 or j>i+1)
α _j = 1/3 (when i-1≤j≤i+1)
may be

again,
α _j =0 (when j<i−1 or j>i+1)
α _j =1/4 (when j=i−1 or j=i+1)
α _j =1/2 (when j=i)
may be

The data receiving unit 21 receives the data output by the input unit 13, performs conversion processing on the data using the filter function described above, and uses the data after the conversion processing as input data for the learning processing unit 23 ( Alternatively, it may be output to the learning processing unit 43 or the learning processing unit 43').

By doing so, for example, when vector data in which values are arranged in time series is used as input data, it is possible to perform robust determination against fluctuations over time.

Further, as described above, since there are multiple ways of determining the filter function, the configuration illustrated in FIG. 5 may be used. In this case, the estimator 42 corresponding to each filter function is defined. The data receiving unit 21 in this case obtains a plurality of data converted by applying a plurality of filter functions, and the estimator 42 corresponding to each filter function performs machine learning using the data converted by the corresponding filter function. Transformed data corresponding to each learning processing unit 43 may be output to each learning processing unit 43 so as to make it possible.

Further, as a conversion processing method, a method using the HOG feature amount may be adopted. Specifically, after arranging each element of the vector data x in a matrix of a predetermined size (for example, w×h), a predetermined window size Ww×Hw (Ww<w, Hw <h) area is set, the gradient direction and gradient strength of the area (which becomes local data) are calculated, the histogram is used as the vector y after conversion, and the data after the conversion process is used as input data It may be output to the learning processing unit 23 (or the learning processing unit 43 or the learning processing unit 43').

This processing is effective when the vector data x is originally image data of the predetermined size (w×h). In this case, each component of the vector data x is the brightness value of each pixel of the image data. Further, since the method of expressing the HOG feature quantity as a vector y is widely known, detailed description thereof will be omitted here.

[Multilayer]
Further, a plurality of data processing apparatuses 1 of this embodiment may be used, and the data processing apparatuses 1 may be connected to each other in a tree structure network such as a FAT tree.

In this case, the data processing device 1 is arranged at each node of the tree structure network. Then, the data processing device 1 corresponding to a node (root node) having no parent node is set to the highest level. The data processing device 1p corresponding to the child node receives the determination result output by the data processing device 1f corresponding to the child node (whether or not the input data input to the data processing device 1f is normal). information representing whether or not) is received from the input unit 13, the information of the result of this determination is received from the input unit 13, and learning processing is performed using an estimator configured by OS-ELM etc. as an autoencoder as a target of machine learning (input data and teacher data) and determines whether the input data is normal or abnormal.

According to this example, in the data processing device 1 on the lower side (inputting information such as vibration of a product manufacturing machine), an abnormality in details of the system targeted for abnormality detection is detected, and, for example, one task is detected. In the room, a data processing device 1 (corresponding to a parent node) that receives inputs from a plurality of data processing devices 1 that detect anomalies in the vibration of individual manufacturing machines controls the entire work room (system In other words, intensive anomaly detection in a wider area) will be performed.

By connecting the data processing apparatuses 1 of the present embodiment in multiple layers in this way, it becomes possible to perform abnormality detection at various scales of the system.

[Factor estimation]
Further, the data processing apparatus 1 of the present embodiment receives input data determined to be abnormal and a plurality of pieces of input data that have been input to the data processing apparatus 1 before and after that input data, and obtains information representing the cause of the abnormality. It may be connected to a factor estimating device using a deep learning neural network that performs machine learning processing as a correct answer.

In this case, the data processing apparatus 1 accumulates and holds at least N pieces of recently input data. When the input data determined to be abnormal is input, the system waits until m (m<N, m=N−n) input data are input, and the input data determined to be abnormal is input. After inputting, when m pieces of input data are input, N pieces of input data held (n-1 pieces before abnormality occurred, 1 piece of input data judged to be abnormal, A total of N pieces of m pieces of input data after being processed are sent to the factor estimating device.

A well-known configuration can be adopted for the configuration of the factor estimating device as described above, so a detailed description thereof will be omitted here.

1 data processing device 11 control unit 12 storage unit 13 input unit 14 output unit 21 data reception unit 22 estimator 23 learning processing unit 24 determination processing unit 25 output processing unit 31 input layer 32 Intermediate layer, 33 output layer, 42 estimator, 43, 43' learning processing unit, 44, 44' determination processing unit, 45 second learning processing unit.

Claims (9)

A data processing device that performs a predetermined determination process based on data that is repeatedly input,
a guessing means that accepts input data and teacher data and can be machine-learned by reciprocal arithmetic;
learning processing means for performing machine learning on the inferring means using the input data as input data and teacher data;
output means for performing the predetermined determination process based on the comparison between the output of the estimation means and the input data, and outputting the result of the determination process;
including
A plurality of the estimating means are provided, and each of the estimating means includes a neural network that randomly determines the connection weight between the input layer and the intermediate layer and machine-learns the connection weight between the intermediate layer and the output layer. ,
The learning processing means resets the connection weights and biases between the input layer and the intermediate layer of the plurality of estimating means at a timing that does not reset all the estimating means at the same time,
The data processing device, wherein the output means performs the predetermined determination process based on a comparison between the output of each of the plurality of estimation means and the input data, and outputs a result of the determination process. A data processing device that performs a predetermined determination process based on data that is repeatedly input,
a guessing means that accepts input data and teacher data and can be machine-learned by reciprocal arithmetic;
learning processing means for performing machine learning on the inferring means using the input data as input data and teacher data;
output means for performing the predetermined determination process based on the comparison between the output of the estimation means and the input data, and outputting the result of the determination process;
including
The learning processing means performs machine learning processing of the inference means each time the input of the data is accepted,
Each time a machine learning process is performed, the machine learning result of the inference means and the input data are recorded in association with each other,
Each time the machine learning process is performed a plurality of times determined by a predetermined method, the output of the guessing means when the recorded input data is input to the guessing means at that time is referred to, and based on the output A data processing device for setting and correcting the machine learning state of the estimating means to the machine learning result recorded in association with the input data when the condition that the loss exceeds a predetermined threshold value is satisfied. The data processing device according to claim 2,
The estimation means includes a neural network that randomly determines connection weights between the input layer and the intermediate layer and machine-learns connection weights between the intermediate layer and the output layer,
The learning processing means records the machine learning result including the connection weight between the intermediate layer and the output layer of the inference means and the input data in association each time the machine learning process is performed,
Each time the machine learning process is performed a plurality of times determined by a predetermined method, the output of the guessing means when the recorded input data is input to the guessing means at that time is referred to, and based on the output When the condition that the loss exceeds a predetermined threshold is satisfied, the machine learning state including the connection weight between the intermediate layer and the output layer of the estimation means is recorded in association with the input data. A data processing device for setting and correcting a machine learning result including a connection weight between an intermediate layer and an output layer of the estimation means . A data processing device that performs a predetermined determination process based on data that is repeatedly input,
a guessing means that accepts input data and teacher data and can be machine-learned by reciprocal arithmetic;
learning processing means for performing machine learning on the inferring means using the input data as input data and teacher data;
output means for performing the predetermined determination process based on the comparison between the output of the estimation means and the input data, and outputting the result of the determination process;
including
The learning processing means performs machine learning processing of the inference means each time the input of the data is accepted,
Each time a machine learning process is performed, the machine learning result of the inference means and the input data are recorded in association with each other,
The estimation means includes a neural network that performs machine learning using loss information based on the difference between the output when the input data is input and the teacher data,
The learning processing means records the loss for the most recent input data for a predetermined M times (M is a natural number), and compares the statistical value calculated based on the recorded loss and the input data M times before. Based on the comparison with the loss when inputting to the inference means, it is determined whether or not to perform machine learning processing based on the input data M times before, and when it is determined to perform machine learning processing, the M A data processing device for machine-learning the estimating means using previous input data as both input data and teacher data. The data processing device according to any one of claims 1 to 4,
A plurality of the estimating means are provided, and further comprising means for classifying the input data into clusters,
The learning processing means associates each of the plurality of estimating means with each of the classified clusters, determines a cluster belonging to each piece of input data, and assigns an estimating means corresponding to the determined cluster to the input data. A data processing device that performs machine learning using including a plurality of data processing devices connected to each other in a tree structure network,
each of the data processing devices
A data processing device that performs a predetermined determination process based on data that is repeatedly input,
a guessing means that accepts input data and teacher data and can be machine-learned by reciprocal arithmetic;
learning processing means for performing machine learning on the inferring means using the input data as input data and teacher data;
output means for performing the predetermined determination process based on the comparison between the output of the estimation means and the input data, and outputting the result of the determination process;
including
A plurality of the estimating means are provided, and each of the estimating means includes a neural network that randomly determines the connection weight between the input layer and the intermediate layer and machine-learns the connection weight between the intermediate layer and the output layer. ,
The learning processing means resets the connection weights and biases between the input layer and the intermediate layer of the plurality of estimating means at a timing that does not reset all the estimating means at the same time,
The output means is a data processing device that performs the predetermined determination process based on a comparison between the output of each of the plurality of estimation means and the input data, and outputs the result of the determination process. data processing system. including a plurality of data processing devices connected to each other in a tree structure network,
each of the data processing devices
A data processing device that performs a predetermined determination process based on data that is repeatedly input,
a guessing means that accepts input data and teacher data and can be machine-learned by reciprocal arithmetic;
learning processing means for performing machine learning on the inferring means using the input data as input data and teacher data;
output means for performing the predetermined determination process based on the comparison between the output of the estimation means and the input data, and outputting the result of the determination process;
including
The learning processing means performs machine learning processing of the inference means each time the input of the data is accepted,
Each time a machine learning process is performed, the machine learning result of the inference means and the input data are recorded in association with each other,
Each time the machine learning process is performed a plurality of times determined by a predetermined method, the output of the guessing means when the recorded input data is input to the guessing means at that time is referred to, and based on the output a data processing device that sets and corrects the machine learning state of the inferring means to the machine learning result recorded in association with the input data when the condition that the loss exceeds a predetermined threshold is satisfied; A data processing system that is A program for causing a computer to function as a data processing device that performs predetermined determination processing based on repeatedly input data,
each of which accepts input data and teacher data, is a guessing means capable of machine learning by reciprocal arithmetic, randomly determines connection weights between the input layer and the intermediate layer, and connects between the intermediate layer and the output layer a plurality of guessing means including a neural network that machine-learns the weights;
learning processing means for performing machine learning on the inferring means using the input data as input data and teacher data;
output means for performing the predetermined determination process based on the comparison between the output of the estimation means and the input data, and outputting the result of the determination process;
make the computer function as
When functioning as the learning processing means, each of the connection weights and biases between the input layer and the intermediate layer of the plurality of estimation means is reset at a timing that does not reset all the estimation means at once,
A program that, when functioning as the output means, performs the predetermined determination process based on a comparison between the output of each of the plurality of estimation means and the input data, and outputs the result of the determination process. . A program for causing a computer to function as a data processing device that performs predetermined determination processing based on repeatedly input data,
a guessing means that accepts input data and teacher data and can be machine-learned by reciprocal arithmetic;
learning processing means for performing machine learning on the inferring means using the input data as input data and teacher data;
output means for performing the predetermined determination process based on the comparison between the output of the estimation means and the input data, and outputting the result of the determination process;
function as
When functioning as the learning processing means, each time the input of the data is accepted, the machine learning process of the inference means is performed, and each time the machine learning process is performed, the machine learning result of the inference means and the input data are recorded in association with each other, and each time the machine learning process is performed a plurality of times determined by a predetermined method, the output of the inferring means when the recorded input data is input to the inferring means at that time is recorded. When the condition that the loss based on the output exceeds a predetermined threshold is satisfied, the machine learning state of the inference means is set to the machine learning result recorded in association with the input data. A program that corrects

Images