WO2023120438A1

WO2023120438A1 - Information processing system, information processing method, and program

Info

Publication number: WO2023120438A1
Application number: PCT/JP2022/046506
Authority: WO
Inventors: 明廣瀬; 准也加藤; 剛平田中; 了昌中根
Original assignee: 国立大学法人東京大学
Priority date: 2021-12-20
Filing date: 2022-12-16
Publication date: 2023-06-29
Also published as: JP2023091563A

Abstract

[Problem] To provide an information processing system or the like which can detect an abnormality of an operation on the basis of a more generic phenomenon. [Solution] According to an aspect of the present invention, an information processing system is provided. This information processing system comprises a processor. The processor is configured to execute a program that executes each of the following steps. In an input processing step, a first signal, which is a time series signal, is input to a neural network. In a reconstruction step, a second signal is generated so that the first signal is reconstructed from a plurality of output signals output from the neural network. In a detection step, an abnormality pertaining to the first signal is detected on the basis of an error between the first signal and the second signal, and a predetermined reference criterion.

Description

Information processing system, information processing method and program

The present invention relates to an information processing system, an information processing method, and a program.

There are technologies that monitor various phenomena such as physical phenomena that can occur along with predetermined operations in factories, etc., and detect operational abnormalities based on those phenomena. For example, Patent Document 1 discloses a prior art that uses sound as a physical phenomenon.

Japanese Patent Application Laid-Open No. 2021-163088

In the technique disclosed in Patent Literature 1, elements other than normal sounds that are normally assumed during operation are identified as abnormal sounds, but it is also possible that normal sounds and abnormal sounds are not linearly combined. . That is, there is a demand for anomaly detection technology based on more general phenomena.

In view of the above circumstances, the present invention provides an information processing system that can detect abnormalities in operation based on more general phenomena.

According to one aspect of the present invention, an information processing system is provided. This information processing system includes a processor. The processor is configured to run a program that causes the following steps to be performed. In the input processing step, the first signal, which is a time-series signal, is input to the neural network. The reconstruction step generates a second signal so as to reconstruct the first signal from the plurality of output signals output from the neural network. The detecting step detects an abnormality in the first signal based on the error between the first signal and the second signal and a predetermined reference standard.

According to such an aspect, an abnormality in operation can be detected based on a more general phenomenon.

1 is a block diagram showing a hardware configuration of an information processing device 1 (information processing system) according to a first embodiment; FIG. 3 is a block diagram showing functions realized by a control unit 13 and the like in the information processing apparatus 1; FIG. 2 is a schematic diagram showing the configuration of a neural network 2; FIG. 3 is an activity diagram showing the flow of information processing executed by the information processing apparatus 1; FIG. FIG. 4 is a schematic diagram representing the time shift between the first signal SG1 and the second signal SG2; FIG. 6A shows the case of inputting normal data, and FIG. 6B shows the case of inputting abnormal data. FIG. 7A shows the error E(n) when evaluating the normal first signal SG1 employing non-dominant combinations with weights w_outjj'<0.05, and FIG. 7B shows the error E(n) with weights w_outjj'≧0. 05 leading combination is used to evaluate the normal first signal SG1 error E(n). FIG. 8A shows the error E(n) when evaluating the abnormal first signal SG1 employing a non-dominant combination with weights w_outjj′<0.05 and FIG. 8B shows the error E(n) when weights w_outjj′≧0.05. 05 dominant combination is used to evaluate the abnormal first signal SG1, the error E(n). It is a block diagram which shows the hardware constitutions of the information processing system which concerns on 2nd Embodiment.

Embodiments of the present invention will be described below with reference to the drawings. Various features shown in the embodiments shown below can be combined with each other.

By the way, the program for realizing the software appearing in this embodiment may be provided as a non-transitory computer-readable medium (Non-Transitory Computer-Readable Medium), or may be downloaded from an external server. It may be provided as possible, or may be provided so that the program is activated on an external computer and the function is realized on the client terminal (so-called cloud computing).

Further, in the present embodiment, the term “unit” may include, for example, a combination of hardware resources implemented by circuits in a broad sense and software information processing that can be specifically realized by these hardware resources. . In addition, various information is handled in the present embodiment, and these information are, for example, physical values of signal values representing voltage and current, and signal values as binary bit aggregates composed of 0 or 1. It is represented by high and low, or quantum superposition (so-called quantum bit), and communication and operation can be performed on a circuit in a broad sense.

A circuit in a broad sense is a circuit realized by at least appropriately combining circuits, circuits, processors, memories, and the like. That is, Application Specific Integrated Circuit (ASIC), programmable logic device (for example, Simple Programmable Logic Device (SPLD), Complex Programmable Logic Device (CPLD), and field It includes a programmable gate array (Field Programmable Gate Array: FPGA)).

[First Embodiment]
1. Hardware Configuration This section describes the hardware configuration of the first embodiment. In this embodiment, an information processing system consists of one or more devices or components. Therefore, for example, even the single information processing apparatus 1 is an example of an information processing system. The hardware configuration of the information processing apparatus 1, which is an example of an information processing system, will be described below.

FIG. 1 is a block diagram showing the hardware configuration of an information processing device 1 (information processing system) according to the first embodiment. The information processing apparatus 1 has a communication section 11, a storage section 12, a control section 13, a display section 14, and an input section 15, and these constituent elements communicate with the communication bus 10 inside the information processing apparatus 1. are electrically connected via Each component will be further described.

The communication unit 11 is preferably a wired communication means such as USB, IEEE1394, Thunderbolt (registered trademark), wired LAN network communication, etc., but wireless LAN network communication, mobile communication such as LTE/5G, Bluetooth (registered trademark) communication, etc. may be included as required. That is, it is more preferable to implement as a set of these communication means. That is, the information processing device 1 may communicate various information from the outside via the communication unit 11 and the network. Specifically, the information processing device 1 is connected to the sensor 51 via the communication unit 11 . The type of the sensor 51 is not particularly limited, and microphones (broadly defined voice sensors), cameras (broadly defined vision sensors), temperature sensors, infrared sensors, pressure sensors, photoelectric sensors, color sensors, ultrasonic sensors, displacement sensors, etc. are suitable. can be used.

The storage unit 12 stores various information defined by the above description. For example, it can be used as a storage device such as a solid state drive (SSD) for storing various programs related to the information processing apparatus 1 executed by the control unit 13, or as a temporary storage device related to program calculation. It can be implemented as a memory such as a random access memory (RAM) that stores information (arguments, arrays, etc.) required for . The storage unit 12 stores various programs, variables, etc. related to the information processing apparatus 1 executed by the control unit 13 .

The control unit 13 (processor) processes and controls overall operations related to the information processing device 1 . The control unit 13 is, for example, a central processing unit (CPU) (not shown). The control unit 13 implements various functions related to the information processing device 1 by reading out a predetermined program stored in the storage unit 12 . That is, information processing by software stored in the storage unit 12 can be specifically realized by the control unit 13 which is an example of hardware, and can be executed as each functional unit included in the control unit 13 . These are further detailed in the next section. Note that the control unit 13 is not limited to a single unit, and a plurality of control units 13 may be provided for each function. A combination thereof may also be used.

For example, the display unit 14 may be included in the housing of the information processing device 1 or may be externally attached. The display unit 14 displays a screen of a graphical user interface (GUI) that can be operated by the user. It is preferable to use display devices such as a CRT display, a liquid crystal display, an organic EL display, and a plasma display, for example.

The input unit 15 may be included in the housing of the information processing device 1 or may be externally attached. For example, the input unit 15 may be integrated with the display unit 14 and implemented as a touch panel. With a touch panel, the user can input a tap operation, a swipe operation, or the like. Of course, a switch button, a mouse, a QWERTY keyboard, or the like may be employed instead of the touch panel. That is, the input unit 15 receives an operation input made by the user. The input is transferred as a command signal to the control unit 13 via the communication bus 10, and the control unit 13 can execute predetermined control and calculation as necessary.

2. Functional Configuration This section describes the functional configuration of this embodiment. As described above, information processing by software stored in the storage unit 12 is specifically realized by the control unit 13, which is an example of hardware, so that each functional unit included in the control unit 13 can be executed. In other words, the information processing device 1, which is an example of an information processing system, includes a control unit 13 (processor). The control unit 13 is configured to execute a program that causes the steps of each functional unit to be executed.

FIG. 2 is a block diagram showing functions realized by the control unit 13 and the like in the information processing device 1. As shown in FIG. Specifically, the information processing apparatus 1 includes a reception unit 131, a conversion unit 132, an input processing unit 133, an output processing unit 134, a reconstruction unit 135, a determination unit 136, a detection unit 137, a transmission and a portion 138 .

The reception unit 131 is configured to receive various types of information received from the outside via the communication unit 11 . For example, the receiving unit 131 may receive the original signal SG0, which is a time-series signal, as an example of the receiving step.

The conversion unit 132 is configured to convert various information handled by the information processing device 1 . For example, as an example of the conversion step, the conversion unit 132 may convert the received original signal SG0 into a spectrogram decomposed for each frequency component.

The input processing unit 133 is configured to input necessary information as input to the neural network 2 (see FIG. 3) pre-stored in the storage unit 12 . For example, the input processing unit 133 may input the first signal SG1, which is a time-series signal, to the neural network 2 as an example of the input processing step.

The output processing unit 134 is configured to output various information from the neural network 2 (see FIG. 3) pre-stored in the storage unit 12 .

The reconstruction unit 135 is configured to reconstruct desired information from the information output from the output processing unit 134 . For example, as an example of the reconstruction step, the reconstruction unit 135 may generate the second signal SG2 so as to reconstruct the first signal SG1 from a plurality of output signals output from the neural network 2. .

The determination unit 136 is configured to determine various information. For example, the determination unit 136 may determine whether there is an abnormality using the reference RC stored in the storage unit 12 .

As an example of the detection step, the detection unit 137 may detect an abnormality related to the first signal SG1 based on the determination result of the determination unit 136.

The transmission unit 138 is configured to output or transmit various information. The various information may be various signals SG, audio data relating to the abnormality, or display information. The display information may be visual information such as a screen, an image, an icon, text, or the like generated in a user-visible manner. It may be rendering information.

3. Configuration of Neural Network 2 In this section, the configuration of the neural network 2 in the information processing device 1 will be described.

FIG. 3 is a schematic diagram showing the configuration of the neural network 2. This neural network 2 includes an input layer 21 , a processing layer 22 and an output layer 23 .

A first signal SG1 is input to the input layer 21 as an input signal. This input layer 21 has input terminals 211 to 213 as an example, but it is only an example and is not limited to this.

The processing layer 22 is the main layer of the neural network 2, and multiple neurons are connected. Specifically, in FIG. 3 various neurons 31-37 are illustrated. Calculation results passing through the connecting line are transmitted from the side not marked with a white circle to the side marked with a white circle. For example, the neuron 31 receives an input from the input terminal 211, assigns weights to the

neurons

32 and 34, and transmits the computation results. The neuron 32 receives an input from the neuron 31, weights an output neuron 231, which will be described later, and transmits the operation result.

Particularly preferably, the processing layer 22 should include a reservoir 3 with recurrent coupling. As an example of the recurrent connection, the signal SG is input from the input terminal 212 to the neuron 34, then via the

neurons

37, 36 and 33, and is input again to the neuron 34 again. According to such an aspect, in the field of reservoir computing, which generally does not use time-series information, it is possible to realize reservoir computing using a time-series signal as an input.

The output layer 23 includes output neurons 231-233. That is, the output signal processed through the reservoir 3 is output through the output neurons 231 to 233, and these element signals are reconstructed by the reconstructing unit 135 to obtain the second signal SG2.

4. Information Processing Method In this section, the information processing method of the information processing apparatus 1 described above will be described. The information processing method according to the present embodiment includes each step of the information processing apparatus 1 as an example of the information processing system described above. FIG. 4 is an activity diagram showing the flow of information processing executed by the information processing apparatus 1. As shown in FIG. Each activity in this activity diagram will be described below.

In this embodiment, as an example, the information processing device 1 is a device related to a dedicated device such as a robot or a machine tool installed in a factory, and the sensor 51 can collect (detect) the operation sound of these devices. It is described as being a microphone. That is, the sensor 51 records an audio signal as a time-series signal, which is transmitted to the information processing apparatus 1 as the original signal SG0 via the communication section 11. FIG. In other words, the sensors 51 are assumed to correspond one-to-one to the information processing apparatus 1 (information processing system). According to such an aspect, it is possible to easily determine the location of the source of the abnormality, and improve usability.

First, the sensor 51 continuously collects the operation sound of the device as a time-series signal SG. Then, as an example of a receiving step, the receiving unit 131 receives the operation sound as an original signal SG0, which is a time-series signal (activity A101). In other words, the reception unit 131 receives the original signal SG0 (time-series signal) from the sensor 51 corresponding to the information processing system on a one-to-one basis. The original signal SG0 may be received in an uncompressed format such as wav or aiff, or may be received in various compressed formats such as aac, wmv, or mp3. The accepted original signal SG0 is stored in the temporary storage area of storage unit 12 .

Subsequently, the conversion unit 132 Fourier-transforms the original signal SG0 stored in the temporary storage area of the storage unit 12 using a dedicated program stored in advance in the storage unit 12 (activity A102). In other words, as an example of the conversion step, the conversion unit 132 converts the received original signal SG0 into a spectrogram decomposed for each frequency component.

Subsequently, the input processing unit 133 inputs the first signal SG1, which is a time-series signal, to the neural network 2 as an example of the input processing step. More preferably, the input processing unit 133 inputs the spectrogram to the neural network 2 as the first signal SG1 (activity A103). That is, the input first signal SG1 is input to the input terminals 211-213. In particular, it is preferable to input different frequency components to each of the input terminals 211 to 213 after classifying each frequency component included in the first signal SG1. According to such an aspect, compared to the case where the original signal SG0 is directly input to the neural network 2, the input signal SG can be analyzed more quantitatively including the frequency domain. As a result, more accurate abnormality detection can be realized.

Subsequently, in the reservoir 3 included in the processing layer 22, calculations including recurrent calculations are performed on the input first signal SG1 (activity A104). That is, the processing layer 22 is configured to hold the first signal SG1 input to the input layer 21 in the past for a finite time. This makes it possible to use the component of the first signal SG1 input before a certain time in order to output the second signal SG2 corresponding to the first signal SG1 corresponding to a certain time.

Subsequently, the output processing unit 134 outputs the signal SG processed through the processing layer 22 as a plurality of output signals to the output layer 23 (activity A105). After that, the components of the output signals output from the

neurons

32, 35 and 37 of the reservoir 3 are distributed to the

output neurons

231, 232 and 233 and input.

Subsequently, as an example of the reconstruction step, the reconstruction unit 135 generates the second signal SG2 so as to reconstruct the first signal SG1 from the plurality of output signals output from the neural network 2 (activity A106). More specifically, the reconstruction unit 135 generates the second signal SG2 based on the output signal and a learned model obtained by machine-learning weighting in the output layer 23 . According to such an aspect, it is possible to generate the second signal SG2 with high accuracy while reducing the processing load on machine learning.

It should be noted that the second signal SG2 output from the output layer 23 is preferably a signal generated so as to reconstruct the first signal SG1 at or before it was input. That is, preferably, the teacher signal for machine-learning the weighting in the output layer 23 is the first signal SG1 at or before the input to the input layer 21 . According to this aspect, the accuracy of the reconstruction itself improves compared to the case where the future prediction signal is often reconstructed in the neural network 2, and as a result, more accurate abnormality detection is realized. can do.

Subsequently, the determination unit 136 performs various determinations on the generated second signal SG2. Specifically, the determination unit 136 extracts the difference between the first signal SG1 and the second signal SG2 (activity A107). Next, the determination unit 136 refers to the reference standard RC stored in advance in the storage unit 12 or the like (activity A108). By comparing these differences with the reference standard RC, it is determined whether or not there is an abnormality (abnormality related to the first signal SG1) in the device installed in the factory. In other words, the detection unit 137 detects an abnormality in the first signal SG1 based on the error between the first signal SG1 and the second signal SG2 and the predetermined reference standard RC (activity A109). . When an anomaly is detected, visual information including a notice of the anomaly may be displayed on the display unit 14 . Furthermore, the transmission unit 138 may transmit a notification for notifying the user of the abnormality to another terminal (not shown) that the user can grasp. The information processing described above may be repeated until an abnormal operation is detected in activity A109.

It should be noted that the second signal SG2 is a signal generated to reconstruct the first signal SG1 at or before it was input, so when extracting the difference, the first signal at an appropriate time is used. is extracted from the signal SG1. For example, when the second signal SG2 is a signal generated to reconstruct the first signal SG1 one unit time earlier, the second signal SG2 and the first signal SG1 one unit time earlier are compared to extract these differences.

Also, the reference standard RC may be a lookup table that summarizes threshold information regarding the difference (error) between the first signal SG1 and the second signal SG2. The determination unit 136 can determine that there is an abnormality in the factory apparatus related to the information processing apparatus 1 when the difference is equal to or greater than the threshold value.

More preferably, the reference standard RC should have information on recruitment conditions. Then, it is preferable not to use the signal of the first signal SG1 that does not satisfy the adoption condition in the detection step by the detection unit 137 . According to such an aspect, it is possible to eliminate the signal SG which is undesirable in abnormality detection, and to realize more accurate abnormality detection. In particular, the signal SG at the beginning (within a predetermined period of time) tends to depend on the initial internal state of the reservoir 3, so it is preferable not to employ this.

In other words, the employment condition relates to the time after the input processing unit 133 is started. In other words, it is preferable not to use the first signal SG1 in the detection step by the detection unit 137 if it is within a predetermined period of time after the input processing unit 133 is started.

In addition, the predetermined time not to be adopted (also understood as the number of times) is specifically, for example, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, 400, 410,420,430,440,450,460,470,480,490,500,510,520,530,540,550,560,570,580,590,600,610,620,630,640,650, 660,670,680,690,700,710,720,730,740,750,760,770,780,790,800,810,820,830,840,850,860,870,880,890,900, 910, 920, 930, 940, 950, 960, 970, 980, 990, 1000, and may be in the range between any two of the numbers exemplified herein.

By the way, in the activity diagram shown in FIG. 4, when the employment conditions related to the predetermined time, etc., are not satisfied, the process transitions to the signal processing of the next unit time before the process of comparing the difference with the threshold value. skip comparison). This is just an example, and the skip processing may be performed before the activities A101-103.

5. Theory and Experimental Results In this section, the theory using mathematical formulas and the experimental results of anomaly detection using neural network 2 are explained. FIG. 5 is a schematic diagram representing the time shift between the first signal SG1 and the second signal SG2.

A first signal SG1, which is an input time-series signal of normal data, is input to the reservoir 3 one after another. At the same time, the output layer 23 is presented with a first signal SG1 obtained by appropriately time-shifting this (represented by n_shift in FIG. 5). Then, the norm 2 of the difference between the second signal SG2, which is a temporary output signal during learning, and the first signal SG1, which is a teacher signal, at each time (root mean square error averaging with respect to k error: RMSE)) is calculated (see Equation 1).

This can be said to be an instantaneous error, and it is preferable to perform learning so that this value becomes small. For example, the load can be updated by the gradient method every time to obtain the optimum load W{̂opt}_out (see Equation 4). Alternatively, if the signal is in a statistically stationary state, the matrix X (see Equation 2/x(n) is the output signal from each neuron) defined by arranging the time-series signals obtained from the reservoir 3 and arranging the teacher signal Optimal weight W{^opt}_out (see Equation 4) may be obtained by the defined matrix S_out (see Equation 3).

where Xdagger is the pseudoinverse of X, N' is the number of x(n) (excluding rejected elements), and finverse is the inverse of f, working element by element. . Here, if n_shift≧0, the reservoir 3 learns to output the past first signal SG1 (input waveform) by making good use of the internal delay. called. On the other hand, if n_shift<0, the future input waveform is predicted. In this embodiment, the case of n_shift≧0 is adopted as a preferred mode.

An experiment was conducted in which normal data or abnormal data was input to a trained neural network with n_shift=1. The error E(n) at that time is shown in FIG. FIG. 6A shows the case of inputting normal data, and FIG. 6B shows the case of inputting abnormal data. In particular, the floor of the error E(n) is higher with the abnormal data in FIG. 6B than when the normal data in FIG. 6A is input. It is also confirmed that the variation is larger when abnormal data is used. This indicates that setting an appropriate threshold for E(n) enables anomaly detection by reconstruction of the input signal.

FIG. 7A shows the error E(n) when evaluating the normal first signal SG1 employing non-dominant combinations with weights w_outjj'<0.05, and FIG. 7B shows the error E(n) with weights w_outjj'≧0. 05 leading combination is used to evaluate the normal first signal SG1 error E(n). FIG. 8A shows the error E(n) when evaluating the abnormal first signal SG1 employing a non-dominant combination with weights w_outjj′<0.05 and FIG. 8B shows the error E(n) when weights w_outjj′≧0.05. 05 dominant combination is used to evaluate the abnormal first signal SG1, the error E(n).

When the abnormal first signal SG1 is evaluated, the error value is large regardless of the load w_outjj', but there are differences in the tendencies. Both error floors are significantly elevated when only minor connections are used. The value shows a similar response to both normal data and abnormal data, and it can be seen that the abnormality detection is not performed well. When only the principal couplings are used, the error floor rises slightly, but not significantly. However, the error floor for normal data is as high as for abnormal data, and here too, it is impossible to distinguish between normal and abnormal data. Moreover, the fluctuation is also increasing. In other words, it can be said that the neurons in the reservoir 3 include neurons that determine the majority and neurons that adjust the details.

This result shows that the reservoir 3 performs processing as an autoencoder, but unlike an autoencoder, for anomaly detection, information should be narrowed down with a small number of neurons rather than an autoencoder. It shows that it is necessary to fit small signals well. It can also be understood that since time-series signals have changes in the time direction unlike static signals, many neurons are required for their reconstruction.

[Second embodiment]
Next, the information processing device 1 according to the second embodiment will be described. It should be noted that descriptions of configurations and functions that are substantially the same as those of the information processing apparatus 1 according to the first embodiment will be omitted. FIG. 9 is a block diagram showing the hardware configuration of an information processing system according to the second embodiment. As illustrated, the information processing system may include components other than the information processing device 1 . Specifically, the information processing system according to the second embodiment includes an information processing device 1 and an external device 5 including a sensor 51 and the like, and these components are connected via a communication network 4. .

As described above, in the second embodiment, the information processing apparatus 1 is a server (cloud form), and performs the information processing shown in FIG. 4 in the form of SaaS (Software as a Service) or cloud computing, for example. can be adopted. In such a case, as an example of a receiving step, the receiving unit 131 may receive the original signal SG0 (time-series signal) from the plurality of sensors 51 or the external device 5 via the communication network 4 (activity A101). According to such an aspect, it is also possible to detect an abnormal operation based on a more general phenomenon. In particular, the information transmitted from the external device 5 and a plurality of sensors 51 can be collectively used, and can be applied to more advanced data utilization (so-called big data) and machine learning such as deep learning.

[others]
Regarding the information processing apparatus system described above, the following aspects may be adopted.

In each of the above-described embodiments, various functions have been described as the configuration of the information processing device 1, but a computer may be provided with a program for executing each step of the information processing device 1, which is an example of an information processing system.

In the above-described embodiment, the receiving unit 131 receives the original signal SG0, which is a time-series signal, from the sensor 51, the transforming unit 132 performs Fourier transform to generate a spectrogram, and the input processing unit 133 generates the first signal SG1, which is a spectrogram. is input to the neural network 2, the time-series signal received by the receiving unit 131 from the sensor 51 may be input to the neural network 2 as the first signal SG1 without conversion.

In the above-described embodiment, the reference standard RC is described as a lookup table, but the reference standard RC is not limited to a lookup table, and may be a database in which a plurality of pieces of information are mathematically related. A mathematical model may be used, or a trained model obtained by machine-learning the correlation of a plurality of pieces of information in advance may be used. Specifically, it may be a trained model that has learned the error between the first signal SG1 and the second signal SG2 and whether there is an abnormality in the operation.

In this embodiment, the reception unit 131, the conversion unit 132, the input processing unit 133, the output processing unit 134, the reconstruction unit 135, the determination unit 136, the detection unit 137, and the transmission unit 138 are controlled by the control unit 13 of the information processing apparatus 1. Although the description is given as a functional unit to be implemented, the information processing system may be implemented by implementing at least a part of this in another device.

In addition, it may be provided in each of the following aspects.

(1) An information processing system comprising a processor, the processor being configured to execute a program for executing the following steps, and in the input processing step, a first signal that is a time-series signal input to the network, and in the reconstruction step, generate a second signal so as to reconstruct the first signal from a plurality of output signals output from the neural network, and in the detection step, the first signal Detecting anomalies in the first signal based on the error between the signal and the second signal and a predetermined reference standard.

(2) In the information processing system described in (1) above, the neural network includes an input layer, a processing layer, and an output layer, and the processing layer includes the first layer input to the input layer in the past. The second signal is configured to hold a signal of 1 for a finite time, and the second signal output from the output layer is a signal generated so as to reconstruct the first signal at or before it was input. Become.

(3) In the information processing system described in (2) above, the processing layer includes a reservoir having recurrent coupling.

(4) In the information processing system described in (2) or (3) above, in the reconstruction step, the first 2 signals.

(5) In the information processing system described in (4) above, the machine learning teacher signal is the first signal at or before input to the input layer.

(6) In the information processing system according to any one of (1) to (5) above, further, the receiving step receives an original signal that is a time-series signal, and the converting step receives the received original signal. A signal is converted into a spectrogram decomposed for each frequency component, and in the input processing step, the spectrogram is input to the neural network as the first signal.

(7) In the information processing system according to any one of (1) to (6) above, further, in the receiving step, a time-series signal is received from a sensor corresponding to the information processing system on a one-to-one basis. .

(8) In the information processing system according to any one of (1) to (6) above, further, in the receiving step, time-series signals are received from a plurality of sensors or external devices via a communication network. .

(9) In the information processing system according to any one of (1) to (8) above, the reference standard has information about employment conditions, and the first signal does not satisfy the employment conditions. A signal is not used in said detecting step.

(10) In the information processing system described in (9) above, the employment condition relates to the time after the input processing step is started, and the input processing step is started in the first signal. signals that are within a predetermined time from are not used in the detection step.

(11) An information processing method, comprising the steps of the information processing system according to any one of (1) to (10) above.

(12) A program that causes a computer to execute each step of the information processing system according to any one of (1) to (10) above.
Of course, this is not the only case.

Finally, although various embodiments of the present invention have been described, these are presented as examples and are not intended to limit the scope of the invention. The novel embodiment can be embodied in various other forms, and various omissions, replacements, and modifications can be made without departing from the scope of the invention. The embodiment and its modifications are included in the scope and gist of the invention, and are included in the scope of the invention described in the claims and equivalents thereof.

1: information processing device 10: communication bus 11: communication unit 12: storage unit 13: control unit 131: reception unit 132: conversion unit 133: input processing unit 134: output processing unit 135: reconstruction unit 136: determination unit 137: Detector 138 : Transmitter 14 : Display 15 : Input 2 : Neural network 21 : Input layer 211 : Input terminal 212 : Input terminal 213 : Input terminal 22 : Processing layer 23 : Output layer 231 : Output neuron 232 : Output neuron 233: output neuron 3: reservoir 31: neuron 32: neuron 33: neuron 34: neuron 35: neuron 36: neuron 37: neuron 4: communication network 5: external device 51: sensor RC: reference standard SG: signal SG0: original signal SG1: first signal SG2: second signal

Claims

An information processing system,
with a processor
The processor is configured to execute a program that causes the following steps to be performed:
In the input processing step, the first signal, which is a time-series signal, is input to the neural network,
In the reconstruction step, generating a second signal so as to reconstruct the first signal from a plurality of output signals output from the neural network;
The detection step detects an abnormality in the first signal based on an error between the first signal and the second signal and a predetermined reference standard.
In the information processing system according to claim 1,
The neural network includes an input layer, a processing layer, and an output layer,
the processing layer is configured to hold the first signal input to the input layer in the past for a finite time;
The second signal output from the output layer is a signal generated so as to reconstruct the first signal at or before the input.
In the information processing system according to claim 2,
The processing layer includes a reservoir with recurrent coupling.
In the information processing system according to claim 2 or claim 3,
In the reconstruction step, the second signal is generated based on the output signal and a learned model obtained by machine-learning the weighting in the output layer.
In the information processing system according to claim 4,
The machine learning teacher signal is the first signal at or before input to the input layer.
In the information processing system according to any one of claims 1 to 5,
Furthermore, in the receiving step, the original signal, which is a time-series signal, is received,
Furthermore, in the conversion step, the received original signal is converted into a spectrogram decomposed for each frequency component,
In the input processing step, the spectrogram is input to the neural network as the first signal.
In the information processing system according to any one of claims 1 to 6,
Furthermore, in the receiving step, time-series signals are received from sensors corresponding to the information processing system on a one-to-one basis.
In the information processing system according to any one of claims 1 to 6,
Furthermore, in the receiving step, time-series signals are received from a plurality of sensors or external devices via a communication network.
In the information processing system according to any one of claims 1 to 8,
The reference standard has information on recruitment conditions,
Signals that do not satisfy the adoption condition among the first signals are not used in the detection step.
In the information processing system according to claim 9,
The employment condition relates to the time from the start of the input processing step,
A signal of the first signal that is within a predetermined period of time after starting the input processing step is not used in the detecting step.
An information processing method,
A method comprising the steps of an information processing system according to any one of claims 1-10.
a program,
A computer that executes each step of the information processing system according to any one of claims 1 to 10.