JP2020009400A - Data processing device, data processing system and program - Google Patents

Abstract

A data processing device, 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 are provided. A data processing device that performs predetermined determination processing based on data that is repeatedly input, receives input data and teacher data, uses an estimator capable of machine learning by reciprocal arithmetic, and processes the input data. is used as input data and teacher data, the estimator is machine-learned, the output of the estimator is compared with the input data, and judgment processing is performed, and the result of the judgment processing is output. be. [Selection drawing] Fig. 1

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 abnormality detection system applicable to various places such as a product manufacturing site.

  For example, in Patent Literature 1, log data that can be collected in the field is recorded in advance, and learning processing is performed with reference to the log data, using the presence or absence of an abnormality at each recording time of the log data as teacher data, An apparatus for performing abnormality determination is disclosed.

JP 2018-73258 A

  However, for example, in a product manufacturing site, the data output from a sensor that collects information such as vibrations differs from site to site depending on the surrounding noise situation. Specifically, the output differs greatly between a site where a manufacturing machine that generates another vibration adjacent thereto is operating and a site where it is not. In addition, since the output is different even when the manufacturing machine that generates another vibration is stopped and while it is operating, even if it is the same site, it is the data that is the basis for making an abnormality judgment due to various factors such as time zone. There is generally a considerable difference between

  Under such circumstances, a device that collects and learns log data in advance needs to collect log data for each site, which complicates the device configuration and requires many preparation steps before operation. Was. Further, even if the learning process is performed in such a manner, the environment may differ depending on the time zone as described above, so that an appropriate determination may not always be performed.

  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 which can simplify a preparation process with a relatively simple configuration, and can perform determination suitable for an environment. That is one of its purposes.

  One aspect of the present invention that solves the above-described problems of the conventional example is a data processing device that performs a predetermined determination process based on repeatedly input data. Learning estimating means, learning processing means for machine-learning the estimating means using the input data as input data and teacher data, and comparing the output of the estimating means with the input data. And an output unit that performs the predetermined determination process and outputs a result of the determination process.

  According to the present invention, determination is performed while performing machine learning using input data, and by using estimating means capable of performing machine learning by reciprocal operation, the preparation process can be simplified with a relatively simple configuration, It is possible to make a determination that is suitable for the environment.

1 is a configuration block diagram illustrating an example of a data processing device according to an embodiment of the present invention. FIG. 2 is a functional block diagram according to an example of the data processing device according to the embodiment of the present invention. FIG. 3 is a configuration block diagram illustrating an example of an estimator used by the data processing device according to the embodiment of the present invention. FIG. 4 is an explanatory diagram illustrating a storage example of parameters by the data processing device according to the embodiment of the present invention. 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 yet 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. As shown in FIG. 1, the example of the data processing device 1 according to the embodiment of the present invention includes a control unit 11, a storage unit 12, an input unit 13, and an output unit 14.

  Here, the control unit 11 is a program control device such as a CPU, 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 unit 11, the control unit 11 executes the programs stored in the storage unit 12 to realize the operations of the above units.

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

  That is, in the present embodiment, the control unit 11 accepts repeatedly input data, and executes machine learning processing of a machine learning model capable of machine learning by reciprocal operation using the input data as input data and teacher data. The control unit 11 obtains an 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 a predetermined determination process based on a comparison between the obtained output of the estimator and the input data, and outputs a result of the determination process. 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. When the control unit 11 is a program control device such as a CPU, the storage unit 12 holds a program executed by the control unit 11. This program may be provided stored in a non-transitory computer readable and non-transitory recording medium, and may be 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 machine learning is performed on the estimator.

  The input unit 13 converts data input from an external sensor or the like into digital data, converts the data 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 an instruction 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. It is comprised including.

  The data receiving unit 21 receives data input from the input unit 13. In the example of the present embodiment, the data processing device 1 is disposed, for example, at a site of product manufacturing, is attached to a device used for product manufacturing, and detects various information such as vibration and temperature of the device. Connected to output sensor. Then, 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 outputs vector values in which digital values obtained by converting outputs of a plurality of sensors are arranged in a predetermined order to the control unit 11 as data. The data receiving unit 21 receives the data output from the input unit 13 at each predetermined timing and outputs the data to the learning processing unit 23.

  The estimator 22 implements an estimating means, and functionally, as illustrated in FIG. 3, includes all three layers including an input layer 31, an intermediate layer (hidden layer) 32, and an output layer 33. It is a connected neural network. In the present embodiment, the estimator 22 determines that the number of nodes in the input layer 31 (the dimension of the vector of the input data) matches the number of nodes in the output layer 33 (the dimension of the vector of the data to be output). Shall be

That is, in the estimator 22, the connection weights between the nodes 31a, 31b,..., 31n of the input layer 31 and the nodes 32a, 32b,. the _{W (w 1, w 2,} ..., w L) and (where w _i is the vector dimension equal to the number of nodes n of the input layer 31), the bias _{b (b 1, b 2,} ... b L) and, node 32a of the intermediate layer 32, 32b, ... 32L and node 33a of the output layer 33, 33b, ..., V the coupling weight between the _{33n (v 1, v 2,} ..., v L) ( where v _i is When a vector having a dimension equal to the number n of nodes in the output layer 33), these connection weight values W and V and the bias b are parameters of the machine learning model of the estimator 22.

When the input data is input, the estimator 22 inputs the input data to the input layer 31 and multiplies each component of the input data input to the input layer 31 by a corresponding connection weight W to sum the components. For example, the value of each node of the intermediate layer 32 is obtained 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 h ₁ of each node of the intermediate layer 32, h ₂ ..., a h _L, the value obtained by applying a predetermined non-linear function _{f f (h i) (i} = 1,2, ... L) to , And the corresponding connection weight V to obtain the value of each node of the output layer 33. This operation is the same as the operation in a general neural network, and a detailed description thereof will be omitted.

  As the nonlinear function here, a widely known function such as a sigmoid function or ReLU may be adopted. The estimator 22 outputs, as output data, a vector having a component of each node of the output layer 33 as a component.

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

  The learning processing unit 23 performs a machine learning process 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 the data for the latest predetermined number of times (for the batch size) received by the data receiving unit 21 to the estimator 22 as input data. Here, when 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 compares the difference (absolute error or mean square error). ) Is calculated as a loss. Then, the learning processing unit 23 updates the parameter (the value of the connection weight) of the estimator 22 so that the loss is reduced.

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

  The determination processing unit 24 performs a predetermined determination process based on the output data output from the estimator 22 when the learning processing unit 23 inputs the 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 when the magnitude of the loss exceeds a predetermined threshold value, indicates that the input data is abnormal (that is, , Indicating that an abnormality has occurred in the manufacturing apparatus, etc.). When the magnitude of the loss does not exceed the predetermined threshold, the determination processing unit 24 outputs a determination result indicating that the input data is normal (that is, there is no abnormality in the manufacturing apparatus or the like). You may make it output to the processing part 25.

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

  In this case, the determination processing unit 24 performs the learning process until the estimator 22 receives the input data the number of times of the initialization threshold and receives the parameter update until the learning process is performed the number of times defined as the initialization threshold. A) Does not perform the judgment process.

  The output processing unit 25 outputs the result of the determination 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 the present embodiment will be described. In the present 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 sets the intermediate layer 32 and the output layer 33 Is a neural network for machine learning. Specifically, here, OS-ELM (Online Sequential-Extreme Learning Machine) is realized and used by the estimator 22 and the learning processing unit 23. This OS-ELM is based on NY Liang, GB 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 are widely known, so detailed description is omitted here.

  In the case of the OS-ELM, the learning processing unit 23 randomly determines the connection weight W and the bias b between the input layer 31 and the intermediate layer 32 in order to reset the estimator 22 initially (at this time, , The connection weight V may be determined at random). 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 with the data (teacher data) received by the data receiving unit 21 and determines the difference (for example, a root mean square error). ) Is calculated 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 is reduced.

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 the 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 sets each batch as training data. Here, the input data included in the i-th batch is x _i (i = 1, 2,...), And the connection weight between the input layer 31 and the intermediate layer 32 is W _k (W _k = w ₁ , w _2, ..., w _L), bias _{b k (b k = b 1} , b 2, ... when a b _L), the output of the output layer 33 for the i-th batch is represented by the following matrix .

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. Incidentally, 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）。 Since the teacher data is the same as the input data, the teacher data T

It becomes. Note that T on the right shoulder means transposition (the same applies hereinafter).

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 the magnitude of the loss.

Is determined, and this is used as the connection weight V between the intermediate layer 32 and the output layer 33 of the estimator 22, so that the estimator 22 is optimized. At this time,

And considering the case of i = 2, equation (1) becomes

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。 Assuming that this is generalized, the 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 beta _{i + 1} is the result of machine learning based on the (i + 1) th batch,

(Sequential updating type). However,

It 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 successive updating formulas are

(Where I is an 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。 Thus the learning processing unit 23, when the i-th batch is input, with the output H _i of the intermediate layer 32, the connection weights V = beta of the intermediate layer 32 and output layer 33 are optimized at that point _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 the equation (2), and the connection weight V between the intermediate layer 32 and the output layer 33 is calculated by the equation (3). Update to _{i + 1} . In general, singular value decomposition is used for the calculation of the inverse matrix. However, when the batch size N _i in each batch is set to N _i = 1 (the batch size is set to 1), the pseudo inverse matrix is used. Is a matrix for which is to be obtained.

Is a scalar value (because equation (4) is a matrix of N _i × N _i ), and therefore, this pseudo inverse matrix is obtained by simple reciprocal operation.

  That is, in the present embodiment, the estimating means capable of machine learning by reciprocal operation is a neural network that performs machine learning on the connection weight between the intermediate layer and the output layer by pseudo-inverse matrix operation as the estimator 22 and the learning processing unit 23. And by setting the batch size of the machine learning to 1 (by performing machine learning each time data is accepted). Specifically, such a neural network includes an ELM (Extreme Learning Machine) that randomly determines a connection weight between an input layer and a hidden layer, and a neural network (FP (Forgetting Parameters) -ELM, OS (On -Line Sequential) -ELM, EOS (Ensemble of OS) -ELM, FOS-ELM (OS-ELM with forgetting mechanism), etc. are equivalent. However, the neural network which realizes the estimating means capable of machine learning by the reciprocal operation by the estimator 22 and the learning processing unit 23 is not limited to these examples.

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

  Here, it is assumed that the data processing device 1 receives signals from a plurality of vibration sensors arranged in the vicinity of a device for manufacturing a product. The vibration sensor outputs an analog electric signal indicating the magnitude of vibration of the attached part.

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

The learning processing unit 23 inputs the input data to the estimator 22 every time the data input is received, that is, each time the batch size of the input data is received. The learning processing unit 23 outputs the output data H of the estimator 22, the input data T as teacher data, the connection weight V = β _i between the intermediate layer 32 and the output layer 33 of the estimator 22 at that stage, The calculated intermediate result P _i (the initial value of both β and P is set in advance) is obtained.

  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 the equations (2) and (3). Further, the learning processing unit 23 calculates a mean square error between the output of the estimator 22 and the input data (teacher data) as a loss, and outputs the loss 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 the data processing apparatus 1 has input the input data after initializing the estimator 22 exceeds a predetermined initialization threshold. When determining that the parameters of the estimator 22 are not sufficiently machine-learned, the determination processing unit 24 does not perform the determination processing.

  On the other hand, when determining that the parameters of the estimator 22 are sufficiently machine-learned, the determination processing unit 24 determines that the magnitude of the loss value input from the learning processing unit 23 is a predetermined threshold. It is determined whether or not the input data is abnormal, and if the magnitude of the loss value exceeds the predetermined threshold value, the input data is abnormal, and it is estimated that an abnormality has occurred in the manufacturing apparatus or the like. Processing such as controlling to output the result of the determination indicating the fact to the display as the output unit 14 is performed.

  As described above, according to the data processing device 1 of the present embodiment, the sequential machine learning of the estimator is set as an auto encoder after being installed at the site where actual abnormality detection is performed (that is, without preparing teacher data separately). ), And abnormality detection is performed based on the result of the machine learning, so that the preparation process can be simplified, and it is possible to make a determination suitable for the environment.

  In addition, a relatively simple configuration can be achieved by using an estimator capable of performing the operation of the machine learning process by a relatively simple reciprocal operation.

  Here, an example in which an abnormality is detected in an apparatus that manufactures a product has been described. However, the present embodiment is not limited to this example. It is also possible to detect abnormalities in the amount of current using the current on the wiring as input data, to detect abnormalities in the heat distribution using the data of heat distribution (thermography) of a certain product as input data, The present invention can be applied to various examples such as abnormality detection using an operation history as input data, abnormality detection relating to the operation of an unmanned aerial vehicle (drone, etc.).

[Forgetting]
Further, in the example of the present embodiment, when the forgetting process is to be included, the forgetting rate may be set to α, and P _{i + 1} in Expression (3) may be multiplied by 1 / α ² . This makes it possible to obtain a forgetting effect by a simple method.

[Partial discard of learning result]
Further, in the present embodiment, when learning processing is performed using a relatively small number of batch size input data groups, such as a batch size of 1, the adaptation to abnormal data caused by continuous abnormal data is considered. To prevent this, the following process of partially discarding the learning result may be performed.

  In one example of the present embodiment, the learning processing unit 23 inputs the data received by the data receiving unit 21 as it is to the estimator 22 as input data (performs a sequential learning process with a batch size of 1). . 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 the batch size set to “1”.

  In this example, the learning processing unit 23 inputs the input data X to the estimator 22 and performs the machine learning process using the output data output from the estimator 22 and the input data X. 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 the update, and data required for the machine learning process (for example, The above P) are stored in the storage unit 12 in association with each other.

  Further, at this time, the learning processing unit 23 deletes the information stored in the storage unit 12 before the previous M times and in which the input data X, the connection weight β, and the intermediate result P are stored in the storage unit 12. You may.

Thus, the storage unit 12 stores 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, and the connection result P _j as the intermediate result. Information R _j (j = 1, 2,...) In which the input data X _j when the weight is obtained is associated with each other is stored (FIG. 4).

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

  The learning processing unit 23 refers to the output of the estimator 22 and determines whether the output satisfies a predetermined condition. Here, the learning processing unit 23 calculates a value (for example, a root-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 It is checked whether or not a predetermined threshold is exceeded. The condition that the magnitude of the loss calculated here exceeds a predetermined threshold value corresponds to, for example, an example of the above-mentioned 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 loss value 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 the 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 out here (this operation partially discards the latest machine learning result). That is).

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

  If the magnitude of the loss does not exceed the predetermined threshold, the learning processing unit 23 continues the machine learning process without correcting the machine learning state of the estimator 22.

  According to this example of the present embodiment, each time machine learning is performed for the number of times determined by a predetermined method (for example, may be determined experimentally), input data at a predetermined time in the past is estimated by the current estimation. To determine whether the magnitude of the loss is large. If the loss is large, the latest machine learning contents are discarded, and the parameters of the estimator 22 are returned to the parameters of the estimator 22 at a predetermined time in the past.

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

[Delayed learning]
Further, in an example of the present embodiment, learning may be performed with delay, instead of partially discarding the learning data. In this example, the learning processing unit 23 inputs the data received by the data receiving unit 21 as it is to the estimator 22 as input data (performs a sequential learning process with a batch size of 1). This estimator 22 may use an OS-ELM neural network as in the example described above. That is, also here, it is assumed that the estimator 22 and the learning processing unit 23 realize an OS-ELM that performs a sequential learning process with the batch size set to “1”.

  The learning processing unit 23 receives the input data X and inputs the data to the estimator 22, obtains output data output from the estimator 22, and obtains a value based on the difference between the output data and the input data X (for example, a root 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 the machine learning process of the estimator 22.

  The learning processing unit 23 repeats the above process a predetermined number of times M (M is a natural number of 2 or more), and calculates the input data X for the past M times and the value of the loss of the estimation result based on the data. When the state is held in the storage unit 12, it is determined whether to perform the machine learning process using the input data X M times earlier.

Specifically, this determination can be made as follows. That is, the learning processing unit 23 refers to the stored information and obtains a statistical value (for example, an average Eav in this case) based on the values of the losses 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 the machine learning process is performed or not), and this value E is Condition E <a · Eav using average Eav which is a statistical value
Is determined. This “a” is a predetermined constant, for example, a = 3.0. Although the average is used as the statistical value here, the average value may be used instead of the average. Also, a may be determined based on the variance or the standard deviation of the loss values of the latest M times, instead of the constant.

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

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

  If the value E does not satisfy E <a · Eav in the above determination, the learning processing unit 23 performs the M learning without executing the machine learning process using the input data X M times before. The previously stored input data X and loss information are deleted.

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

[Parallelization]
Further, according to the present embodiment, there may be a plurality of estimators 22. Functionally, the control unit 11 of the data processing apparatus 1 according to this example includes a data receiving unit 21, a plurality of estimators 42-1 and 42-2,. , A plurality of learning processing units 43-1, 43-2,... Provided in correspondence with each other, a determination processing unit 44, and an output processing unit 25. Components having the same configuration as that illustrated in FIG. 2 are denoted by the same reference numerals, and repeated description is omitted.

  Estimator 42 according to this example of the present embodiment (here, when it is not necessary to distinguish each estimator, each estimator is collectively referred to as 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.

  Further, the learning processing unit 43 sequentially updates the corresponding parameters of the estimator 42 by machine learning processing in the same manner as the learning processing unit 23 described above. The learning processing unit 43 also calculates and outputs a value (root mean square error) relating to the difference between the output data of the estimator 42 and the input data (corresponding to teacher data) as a loss. Further, in this example of the present embodiment, the learning processing unit 43 outputs the value of the loss, and learns the number of times the parameter has been updated (the number of times input data has been input) since the corresponding estimator 42 was reset. Output as status information. Note that, also in this example, each learning processing unit 43 may execute a process of partially discarding the learning result.

In this example of the present embodiment, the learning processing unit 43-i (i = 1, 2,...) Sets a corresponding estimator 42-i (i = 1, 2,...) For each learning processing unit 43-i. Is reset at every predetermined timing T _i defined in the above. That learning processing unit 43-i is (connection weights W and bias b of the input layer 31 and the intermediate layer 32 randomly determined) preceding estimator 42-i reset after the elapsed much time timing T _i Each time, the estimator 42-i is reset again.

Here, the timing T _i may be determined by the number of times of input of the input data (for example, the timing T _i may be each time the input data is input q _i times), or a clock unit (not shown) Information may be determined from the actual elapsed time, which is determined based on the time information obtained from the circuit unit performing the time.

Further, the timing _Ti may be a timing at which all the estimators 42 are not reset at the same time (at least during a time required for the abnormality detection processing). For example, when the timing T _i is determined by the number of times of input data, the number of times of input at each timing T _i is 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 is performed at the same timing up to 7.5 billion times. Disappears.

The cases specified by the reality of the time the timing T _i is, the T ₁ to the daily morning 0 hours, 0 minutes, 0 seconds, be determined so that the T ₂ every Monday morning to 1 hours, 0 minutes, 0 seconds of ..., same timing Will not be reset.

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

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

  Then, the determination processing unit 44 refers to the learning state 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 the magnitude of the loss in advance. Check if the threshold is exceeded. Here, whether or not the parameter has been sufficiently learned may be determined based on, for example, whether or not the number of updates of the parameter indicated by the learning status information exceeds a predetermined initialization threshold value. That is, the determination processing unit 44 determines that the parameter has been sufficiently learned when the number of updates of the parameter represented by the learning status information exceeds a predetermined initialization threshold value.

  Based on the number Q of the estimators 42 that have been determined that the parameters have been sufficiently learned, the determination processing unit 44 includes, among the learning processing units 43 corresponding to the estimators 42 that have been determined that the parameters have been sufficiently learned, Then, it is checked whether or not the value q / Q exceeds a predetermined value, for example, し て by dividing the number q of the output loss exceeding a predetermined threshold value. When the value q / Q exceeds, for example, １ ／ (when the predetermined value is １ ／, when it is determined that the majority of the estimators 42 have detected an abnormality), the input data is abnormal. The output processing unit 25 is caused to output a result of the determination indicating that there is an abnormality (that is, an abnormality has occurred in the manufacturing apparatus or the like).

[Clustering]
When a plurality of estimators 22 are provided and parallelized, machine learning processing may be performed as follows. As illustrated in FIG. 6, the control unit 11 of the data processing device 1 according to the example of the present embodiment includes a data reception unit 21 and a plurality of data reception units, similarly to the example illustrated in FIG. , A learning processing unit 43 ', a determination processing unit 44', an output processing unit 25, and a second learning processing unit 45. Note that components having the same configuration as that illustrated in FIG. 5 are denoted by the same reference numerals, and repeated description will be omitted.

  Each of the estimators 42 according to this example of the present embodiment (here again, when it is not necessary to distinguish each estimator, the respective estimators are collectively described as an estimator 42) are illustrated in FIG. The same 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 the result to the determination processing unit 44'. The learning processing unit 43 ′ accumulates and holds the input data X in the storage unit 12.

  In addition, the learning processing unit 43 ′ identifies the estimator 42 of the estimators 42 whose determination result based on the loss related to the output data indicates that the input data X is “normal”. The information is received from the determination processing unit 44 ', and the 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 described above.

  The determination processing unit 44 ′ receives the information on the loss of each estimator 42 from the learning processing unit 43 ′, and based on the loss, determines the result of the normal / abnormal determination of the input data X by each estimator 42. Output. More specifically, it is assumed that the estimator 42 determines that the input data X is abnormal for the estimator 42 whose loss magnitude exceeds a predetermined threshold value. For the estimator 42 whose corresponding loss magnitude does not exceed the predetermined threshold, it is assumed that the estimator 42 has determined that the input data X is normal. When 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 fact. .

  When at least one of the estimators 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 the fact.

  Further, the determination processing unit 44 'outputs information for 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 starts at a predetermined timing and executes the machine learning process of the estimator 42. Here, the predetermined timing is determined at the time of starting the data processing device 1 or once every predetermined time, every time the input data X is input a predetermined number of times, the input data determined to be abnormal is determined by the predetermined number of times. It may be determined as a point in time when the value is exceeded, a point in time when it is determined that the value of the number m of clusters determined based on the tendency of the input data is not appropriate, or a point in time according to a user's instruction. The determination as to whether the value of the number of clusters is appropriate is made by performing trial clustering by changing the number of clusters, and by referring to the result, determining in advance whether to change the number of clusters. It is performed by repeatedly executing each judgment timing. Here, the specific determination as to whether or not to change the number of clusters is based on the input data belonging to the same cluster in the result of the clustering with the number of clusters at the time of the above-described determination processing and the result of the above-described experimental clustering. This can be performed based on the distance between them (cohesion) and the distance between input data belonging to different clusters (inter-cluster discreteness).

  The 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. For example, any of various processes such as DBScan, SUBCLU, and k-means can be adopted.

  The second learning processing unit 45 classifies the cluster into the same or a smaller number of clusters as the number of the estimators 42. That is, if the number of estimators 42 is n, the number of clusters is m, where m ≦ n. Alternatively, a larger number of estimators 42 than the number of clusters obtained here may be prepared in advance (in that case, data to be input data is stored in advance, and a clustering process is performed to determine the number of clusters. ).

  The second learning processing unit 45 assigns each of the estimators 42 to any of the clusters, assuming that each of the estimators 42 is an estimator that machine-learns input data related to any of the clusters. If the assignment has been made before, the second learning processing unit 45 uses the result of the previous assignment without changing the assignment again.

  Then, the second learning processing unit 45 sequentially extracts the input data stored in the storage unit 12, and refers to a result of clustering of the extracted input data (information indicating a cluster to which the data 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 obtains the output data in the same manner as described above. Then, the estimator 42 performs a machine learning process. At this time, the machine learning process is not performed on the estimator 42 not assigned to the cluster represented by the referred information.

  When the above processing is completed for all of 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 (so that the input data is not used 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, the input data is classified into clusters, a corresponding estimator 42 is prepared for each of the classified clusters, and the estimator 42 is machine-learned with the 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 unit 45. Thereby, for example, in an example of performing abnormality determination of a manufacturing apparatus, a plurality of different types of vibrations are machine-learned, such as vibration when the manufacturing apparatus is temporarily stopped, vibration during operation, and so on. Become.

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

  Then, when at least one of the estimators 42 determines that the input data is normal, the determination processing unit 44 ′ outputs label information relating to the cluster to which the estimator 42 is assigned to the output processing unit 25. Then, the information of the label is output. According to this, in addition to determining whether the input data is abnormal or normal, the user can determine which state the input data corresponds to (for example, in the above example, “device is stopped” or the like). State) can be identified.

  Further, the determination processing unit 44 'comprehensively determines the determination results of all the estimators 42, and thereby determines a label corresponding to the input data (what state the input data corresponds to). May be output. For example, when one estimator 42 is determined to be normal and all other estimators 42 are determined to be abnormal, the certainty factor for the label information relating to the cluster to which the estimator 42 determined to be normal is assigned (the The reliability of the cluster classification) increases. On the other hand, when the plurality of estimators 42 simultaneously determine that one input data is normal, the confidence in the label information relating to the cluster to which each of the plurality of estimators 42 is assigned becomes low.

  Accordingly, the determination processing unit 44 ′ determines that each of the plurality of input data is normal for each of the estimators 42 alone (when the other estimators 42 determine that they are abnormal, only Is determined to be normal), that is, a value obtained by dividing the number of times that it is determined that it is normal alone by the number of times that it is determined to be normal (the number of times that the other estimators 42 also determine that it is normal), It may be output as the certainty factor of the label related to the assigned cluster.

[Detection of abnormal behavior]
As described above, the data processing device 1 of the present embodiment can also be used to determine whether the subject's behavior is normal or abnormal.

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

The data receiving unit 21 of the data processing device 1 receives, from the input unit 13, a plurality of input of a code string representing a series of actions of the subject for each predetermined period. Then, the data receiving unit 21 generates a state transition table for each period based on the code string corresponding to each period. Specifically the data receiving unit 21, i-th from the bit stream consisting of N code (i is, 0 <i <each integer N) taken and the code C _i of the i + 1 th code C _{i + 1} Then, a permutation (C _i , C _{i + 1} ) is created, and the appearance probability of each permutation is calculated.

  For all possible permutations, the data receiving unit 21 generates vector information in which the appearance probabilities are associated (assuming that the appearance probabilities for permutations not extracted from the code sequence are 0), and sets the state transition table. Then, 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 the learning processing unit 43 '), and obtains an estimator that determines normal / abnormal.

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

  Therefore, in this example of the present embodiment, the data receiving unit 21 performs a widely known compression (integrating a plurality of elements to reduce the number of elements) for the state transition table obtained for each of the plurality of periods. After the processing, the vector information after the compression processing may be used as input data. As this processing, a widely known method such as a method based on the Candes-Tao theory can be adopted, and a detailed description thereof will be omitted.

Alternatively, when obtaining the state transition table V _j corresponding to the j-th period, the data receiving unit 21 determines the j-th state transition table if there is a state transition table V _{j- 1} corresponding to the j−1 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
May be obtained as Here, α is an arbitrary constant, for example, 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 shown in FIG. 6, when detecting abnormal or normal behavior of the subject who drives the vehicle, each estimator 42 corresponds to the one during running, the one corresponding to the stopped state, and so on. It is expected that machine learning will be differentiated in this way. Then, in this case, when it is determined that any of the estimators 42 is abnormal, the data processing device 1 outputs information indicating that the behavior abnormality is detected.

[Preprocessing]
Further, data receiving unit 21 of data processing device 1 of the present embodiment may perform preprocessing on data output from input unit 13. This pretreatment, vector data x to be processed (elements (x _0, x _1, x ₂ ..., and x _n)) and performs a predetermined conversion with respect to, the converted vector y ( _{Let the} elements be (y ₀ , y ₁ , y ₂ …, yn))
y _i = Σα _j · x _j
(However, Σ means obtaining the sum of j). Here, α is a 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)
It may be.

Also,
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)
It may be.

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

  In this way, for example, when vector data in which values are arranged in a time series is used as input data, it is possible to make a robust determination against a change with time.

  Further, as described above, since there are a plurality of ways of determining the filter function, the configuration illustrated in FIG. 5 may be used. In this case, an estimator 42 corresponding to each filter function is determined. Then, in this case, the data receiving unit 21 obtains a plurality of data converted by applying each of the plurality of filter functions, and generates an estimator 42 corresponding to each filter function by using the data converted by the corresponding filter function. The corresponding converted data may be output to each learning processing unit 43 so as to cause the learning processing unit 43 to output the converted data.

  Further, as a method of the conversion processing, a method using a 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) is set, the gradient direction and the gradient strength of the region (local data) are calculated, and the histogram is used as a converted vector y, and the converted data 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 process is effective when the vector data x is originally the image data of the predetermined size (w × h). In this case, each component of the vector data x is a luminance value of each pixel of the image data. Also, since a method of expressing the HOG feature amount as a vector y is widely known, a detailed description thereof will be omitted.

[Multilayer]
Further, a plurality of data processing apparatuses 1 of the present embodiment may be used, and each data processing apparatus 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 having no parent node (root node) is set as the highest level. The data processing device 1p corresponding to the node having a child determines the result of the determination output by the data processing device 1f corresponding to the child node (whether the input data input to the data processing device 1f is normal or not). And information about the result of this determination is received from the input unit 13, and as an object of machine learning (input data and teacher data), a learning process is performed using an estimator composed of OS-ELM or the like as an auto encoder. Then, it is determined 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 vibrations of a product manufacturing machine), an abnormality is detected in the details of the system targeted for abnormality detection, and for example, one operation is performed. In a room, the data processing device 1 (corresponding to a parent node) that receives input from a plurality of data processing devices 1 that performs abnormality detection on vibrations of individual manufacturing machines is provided in the entire work room (in the system). Integral abnormality detection in a wider area) is performed.

  As described above, by connecting the data processing apparatuses 1 of the present embodiment in multiple layers, it is possible to perform abnormality detection at various scales of the system.

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

  In this case, the data processing device 1 accumulates and holds at least N recently input data. Then, when the input data determined to be abnormal is input, the control waits until m (m <N, m = N−n) input data is input. After input, when m pieces of input data are input, the held N pieces of input data (n-1 pieces before the occurrence of an error, one piece of input data determined to be abnormal, and (N in total of the m pieces of input data after the above) are sent to the factor estimating apparatus.

  The configuration of the factor estimating apparatus as described above can employ widely known ones, and a detailed description thereof will be omitted.

DESCRIPTION OF SYMBOLS 1 data processing apparatus, 11 control part, 12 storage part, 13 input part, 14 output part, 21 data receiving part, 22 estimator, 23 learning processing part, 24 judgment processing part, 25 output processing part, 31 input layer, 32 Hidden 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 repeatedly input data,
Estimating means that can accept input data and teacher data and perform machine learning by reciprocal operation;
Learning processing means for machine learning the inference means, using the input data as input data and teacher data,
Output means for performing the predetermined determination processing based on a comparison between the output of the estimation means and the input data, and outputting a result of the determination processing;
A data processing device including: The data processing device according to claim 1, wherein
A data processing apparatus comprising a neural network that randomly determines a connection weight between an input layer and a hidden layer and machine-learns a connection weight between the hidden layer and an output layer. The data processing device according to claim 1 or 2,
The data processing device, wherein the learning processing means performs a machine learning process of the estimating means each time the data input is received. The data processing device according to claim 3, wherein
Each time the learning processing unit performs the machine learning process, records the machine learning result of the estimating unit and the input data in association with each other,
Each time a plurality of machine learning processes determined by a predetermined method are performed, reference is made to the output of the estimating means when the recorded input data is input to the estimating means at that time, and the output is determined in advance. A data processing device for correcting a machine learning state of the estimating means based on a machine learning result recorded in association with the input data when a predetermined condition is satisfied. The data processing device according to claim 1, wherein:
Comprising a plurality of said estimating means,
The learning processing means resets the plurality of estimating means at a timing at which all the estimating means are not reset at the same time,
A data processing device that performs the predetermined determination process based on a comparison between each output of the plurality of estimation units and the input data, and outputs a result of the determination process. 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 repeatedly input data,
Estimating means capable of receiving input data and teacher data and performing machine learning by reciprocal operation;
Learning processing means for machine learning the inference means, using the input data as input data and teacher data,
Output means for performing the predetermined determination processing based on a comparison between the output of the estimation means and the input data, and outputting a result of the determination processing;
A data processing system comprising: Computer
A program for functioning as a data processing device that performs a predetermined determination process based on repeatedly input data,
Estimation means that accepts input data and teacher data and can perform machine learning by reciprocal operation,
Learning processing means for machine-learning the estimating means using the input data as input data and teacher data, and performing the predetermined determination processing based on a comparison between the output of the estimating means and the input data. A program that causes a computer to function as an output unit that performs the determination process and outputs the result of the determination process. The data processing device according to claim 1 or 2,
The estimating means is configured to include a neural network that performs machine learning using an output when input data is input and loss information based on a difference between teacher data,
The learning processing means records the loss with respect to the most recent input data for a predetermined number of M times (M is a natural number), and calculates a statistical value calculated based on the recorded loss and the input data M times before. Based on a comparison with the loss at the time of input to the estimating means, it is determined whether or not to perform a machine learning process based on the input data M times before. A data processing device that machine-learns the estimating means by using input data from a previous session as input data and teacher data. The data processing device according to claim 1 or 2,
A plurality of the inference means, 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 input data, and outputs the estimating means corresponding to the determined cluster to the input data. A data processing device that performs machine learning using.

Images