JP2019185580A - Device and method for detecting abnormality - Google Patents

Abstract

To provide a device and a method for detecting an abnormality which can detect an abnormality by automatically learning and acquiring a feature amount even if the device has only a relatively low computing capability.SOLUTION: An abnormality detector 1 includes: a one-dimensional signal acquisition unit 2 for acquiring a one-dimensional signal as a one-dimensional time-series signal from a monitoring target; a learning unit 5 for calculating a feature amount of the one-dimensional signal by making a one-dimensional CNN having a one-dimensional folding layer learned with the acquired one-dimensional signal as an input, and outputting setting information for constructing the one-dimensional CNN so as to classify the one-dimensional signal according to the feature amount and detect an abnormality in the monitoring target; a feature amount setting storage unit 61 and a classifier setting storage unit 62 for storing setting information; and a detection unit 6 for detecting an abnormality in the monitoring target by inputting an unknown one-dimensional signal into the one-dimensional CNN constructed on the basis of the setting information stored in the feature amount setting storage unit 61 and the classifier setting storage unit 62.SELECTED DRAWING: Figure 1

Description

  The present invention relates to an anomaly detection apparatus and method, and more particularly to an anomaly detection technique using a one-dimensional time series signal.

  In factory and plant production equipment, sensor data such as temperature, pressure, flow rate, control valve opening, vibration, torque, etc. measured by sensors placed in the equipment for safe operation and stable quality production. Control and monitor. These sensor data are one-dimensional time-series signals, and various methods for detecting such abnormal one-dimensional time-series signals have been proposed.

  There are various conventional methods for detecting an abnormality in a one-dimensional time-series signal (hereinafter referred to as “one-dimensional signal”). Generally, the following procedure is used in a conventional abnormality detection apparatus. .

  First, the abnormality detection device acquires a one-dimensional signal from the sensor. The abnormality detection apparatus performs preprocessing such as signal amplitude normalization and filtering on the acquired one-dimensional signal in order to increase the accuracy of abnormality detection. Thereafter, the abnormality detection device calculates a preset feature value from the preprocessed one-dimensional signal. A plurality of feature quantities may be calculated. Then, the abnormality detection device classifies whether the input one-dimensional data is normal or abnormal based on the feature amount.

  Here, the feature amount in the conventional abnormality detection device is an index value that is designed or selected based on the knowledge of the designer of the abnormality detection device, for example, the amplitude, frequency component, phase of these one-dimensional data, or these Index values calculated in combination have been used as feature quantities.

  In addition, another conventional abnormality detection device that classifies abnormalities and normality of a one-dimensional signal based on a feature amount includes a method of performing classification based on a threshold value determined in advance by a designer, a method using a statistical test, and the like. It is used. For example, Patent Document 1 discloses an anomaly detection technique in which only a classification process is automatically machine-learned by a neural network having only a fully connected layer based on a feature amount calculated by a designer by performing time-frequency analysis on a signal in advance. ing.

  On the other hand, in the research field of two-dimensional images in recent years, an analysis technique using a convolutional neural network (hereinafter referred to as “CNN”) using deep learning has been successful. It is well known that by using a CNN having a convolution layer and a fully connected layer, it is possible to automatically learn and acquire a feature quantity of a two-dimensional image and perform classification (for example, see Non-Patent Document 1). .

  When the CNN that performs the two-dimensional signal processing as described above is used to classify whether the one-dimensional data is normal or abnormal, the one-dimensional data to be classified is expanded to a two-dimensional signal and converted to the CNN. Machine learning is performed so that input signals can be classified. More specifically, one-dimensional data is analyzed in the same manner as a two-dimensional image by performing a short-time Fourier transform on a one-dimensional time series signal and converting it into a two-dimensional signal as a spectrogram.

  For example, Non-Patent Document 2 reports an example in which musical instruments are classified by converting one-dimensional data, which is a musical instrument sound, into a spectrogram and inputting it to a CNN having a convolutional layer and a fully connected layer.

  Here, FIG. 16 and FIG. 17 show an operation example of the abnormality detection apparatus that converts a conventional one-dimensional signal into a two-dimensional signal and uses it as an input signal of the CNN. The operation of the conventional abnormality detection apparatus is roughly divided into a learning process shown in FIG. 16 and a detection process shown in FIG. First, in the learning process of FIG. 16, CNN learning is performed using a large amount of one-dimensional signal groups (teacher data) that are classified in advance as normal and abnormal and labeled.

  More specifically, the conventional abnormality detection apparatus first acquires a one-dimensional signal group that is classified and labeled as normal and abnormal in advance from the monitoring target (step S500). Next, the acquired one-dimensional signal is subjected to preprocessing such as noise removal (step S501). Thereafter, the one-dimensional signal is converted into a two-dimensional signal by short-time Fourier transform (step S502). The signal converted into the two-dimensional signal is input to the CNN, and forward propagation processing is performed (step S503).

  Next, an error due to the loss function is derived with respect to the output of the CNN (step S504), and the CNN error back-propagation process is repeatedly performed until the error becomes sufficiently small (step S505: NO) (steps S506, S503, S504, S505). Then, the weight parameter determined by CNN learning is stored in a predetermined storage area (step S507).

  After learning by the above procedure, the conventional abnormality detection apparatus acquires an unknown one-dimensional signal that is not labeled as shown in FIG. 17 (step S510). Thereafter, preprocessing such as noise removal of the acquired one-dimensional signal is performed (step S511), and similarly, the one-dimensional signal is converted into a two-dimensional signal (step S512).

  Next, the two-dimensional signal is input to the learned CNN and forward propagation processing is performed (step S513). Thereafter, the abnormality detection device outputs a classification result indicating whether the one-dimensional signal expanded into a two-dimensional signal by the learned CNN is normal or abnormal (step S514).

Japanese Patent No. 6033140

Yutaka Matsuo "Artificial intelligence surpasses human beings or something beyond deep learning", published by Kadokawa EPUB, March 2015 Vincent Lostren and Carmine-Emanuel Cella: “DEEP CONVOLUTIONAL NETWORKS ON THE PITCH SPIRAL FOR MUSICAL INSTRUMENT RECOGNITION”, ISMIR 2016 Masato Mimura, “Front-end feature enhancement and robust speech recognition based on deep learning”, Journal of the Acoustical Society of Japan, Vol. 73, No. 1, pp. 47-54 (2017)

  However, in the conventional abnormality detection technique, since it is necessary to convert a one-dimensional signal into a two-dimensional signal every time a signal is input to the CNN, the calculation cost is high. For this reason, it has been difficult to implement and operate an abnormality detection device in equipment such as field equipment having a relatively low calculation capability.

  The present invention has been made to solve the above-described problems, and an abnormality detection apparatus and method capable of detecting an abnormality by automatically learning and acquiring a feature amount even in a device having a relatively low calculation capability. The purpose is to provide.

  In order to solve the above-described problem, an abnormality detection device according to the present invention includes a signal acquisition unit that acquires one-dimensional data that is a one-dimensional time-series signal from a monitoring target, and the acquired one-dimensional data as an input. The neural network is constructed to learn a neural network, calculate a feature quantity of the one-dimensional data, classify the one-dimensional data based on the feature quantity, and detect an abnormality of the monitoring target. A learning unit that outputs setting information for the storage, a storage unit that stores the setting information, and the neural network that is constructed based on the setting information stored in the storage unit is newly acquired from the monitoring target A detecting unit that inputs unknown one-dimensional data that is a one-dimensional time series signal and detects an abnormality of the monitoring target, and the neural network , Characterized in that it comprises a one-dimensional convolution layer that performs convolution of a one-dimensional to the one-dimensional data.

  In the abnormality detection apparatus according to the present invention, the learning unit calculates feature values of the one-dimensional data, and sets feature value setting information for constructing the neural network based on the calculated feature values. And a classifier learning that classifies the one-dimensional data with the feature quantity setting information as input and outputs classification setting information for constructing the neural network based on the classification result. And the storage unit stores the feature amount setting information and the classification setting information, and the detection unit is based on the feature amount setting information stored in the storage unit, Based on the setting information of the classification stored in the storage unit, the feature amount calculated by the feature amount calculation unit based on the feature amount calculation unit that calculates the feature amount of the unknown one-dimensional data as input Or may be a classifying unit for detecting an abnormality performs classification of unknown said one-dimensional data.

  In the abnormality detection apparatus according to the present invention, the neural network may include at least one layer of the one-dimensional convolution layer and at least one layer of all coupling layers.

  In the abnormality detection apparatus according to the present invention, the neural network may further include a recursive neural network layer that receives the output of the one-dimensional convolutional layer as an input.

  In the anomaly detection device according to the present invention, the recursive neural network layer may be an LSTM layer.

  In addition, the abnormality detection method according to the present invention includes a signal acquisition step of acquiring one-dimensional data that is a one-dimensional time series signal from a monitoring target, and learning the neural network using the acquired one-dimensional data as an input, Learning that calculates feature values of the one-dimensional data and outputs setting information for constructing the neural network so as to classify the one-dimensional data based on the feature values and detect an abnormality of the monitoring target Step, and input to the neural network constructed based on the setting information stored in the storage unit, unknown one-dimensional data that is a one-dimensional time-series signal newly acquired from the monitoring target, A detection step of detecting an abnormality of the monitoring target, wherein the neural network performs a one-dimensional convolution on the one-dimensional data. Characterized in that it comprises a one-dimensional convolution layer for.

  In the abnormality detection method according to the present invention, the learning step calculates a feature amount of the one-dimensional data, and sets feature amount setting information for constructing the neural network based on the calculated feature amount. And a classifier learning that classifies the one-dimensional data using the feature quantity setting information as input and outputs classification setting information for constructing the neural network based on a classification result. A step of calculating a feature amount of the unknown one-dimensional data based on the setting information of the feature amount stored in the storage unit, and a step of storing the feature amount in the storage unit. Based on the stored setting information of the classification, the feature quantity calculated in the feature quantity calculation step is used as an input to classify the unknown one-dimensional data. It may comprise a classification step for detecting an abnormality.

  In the abnormality detection method according to the present invention, the neural network may further include a recursive neural network layer that receives the output of the one-dimensional convolutional layer as an input.

  In the abnormality detection method according to the present invention, the recursive neural network layer may be an LSTM layer.

  According to the present invention, since a one-dimensional CNN that performs processing with a one-dimensional signal is used, even if the device has a relatively low calculation capability, feature quantities are automatically learned and acquired to detect the presence or absence of an abnormality. can do.

FIG. 1 is a block diagram showing a functional configuration of the abnormality detection apparatus according to the first embodiment of the present invention. FIG. 2 is a block diagram showing a hardware configuration of the abnormality detection apparatus according to the first embodiment of the present invention. FIG. 3 is a diagram for explaining the configuration of the one-dimensional CNN used in the abnormality detection apparatus according to the first embodiment of the present invention. FIG. 4 is a diagram for explaining the calculation of the one-dimensional CNN used in the abnormality detection apparatus according to the first embodiment of the present invention. FIG. 5 is a flowchart for explaining the learning process of the abnormality detection device according to the first embodiment of the present invention. FIG. 6 is a flowchart for explaining detection processing of the abnormality detection device according to the first embodiment of the present invention. FIG. 7 is a block diagram showing a functional configuration of the abnormality detection apparatus according to the second embodiment of the present invention. FIG. 8 is a diagram for explaining the configuration of a one-dimensional CNN used in the abnormality detection apparatus according to the second embodiment of the present invention. FIG. 9 is a block diagram for explaining an LSTM layer according to the second embodiment of the present invention. FIG. 10 is a flowchart for explaining a learning process of the abnormality detection device according to the second embodiment of the present invention. FIG. 11 is a flowchart for explaining detection processing of the abnormality detection device according to the second embodiment of the present invention. FIG. 12 is a diagram showing a specific example of the one-dimensional signal according to the embodiment of the present invention. FIG. 13 is a diagram showing a specific example of the one-dimensional signal according to the embodiment of the present invention. FIG. 14 is a diagram for explaining the effect of the abnormality detection device according to the embodiment of the present invention. FIG. 15 is a diagram for explaining the effect of the abnormality detection device according to the embodiment of the present invention. FIG. 16 is a flowchart illustrating a learning process performed by a conventional abnormality detection device. FIG. 17 is a flowchart for explaining detection processing by a conventional abnormality detection device.

  Hereinafter, a preferred embodiment of the present invention will be described in detail with reference to FIGS. Constituent elements common to the drawings are given the same reference numerals.

[Principle of the Invention]
The abnormality detection apparatus according to the present invention inputs a one-dimensional signal (one-dimensional data) that is an object of abnormality detection directly into a neural network without performing a process of converting it into a two-dimensional signal, and further performs a one-dimensional process. Use a configuration that can.

  As described above, among neural networks, CNN is widely used for analyzing two-dimensional signals such as images. CNN is applicable not only to two-dimensional signals but also to analysis of one-dimensional signals. For this reason, in the abnormality detection device according to the present invention, attention is paid to CNN capable of inputting and processing a one-dimensional signal.

  In addition, as described above, conventionally, even if a signal acquired from a monitoring target is a one-dimensional signal, it is applied to a CNN that performs processing in two dimensions by extending the one-dimensional signal to a two-dimensional signal. It was. On the other hand, the CNN used in the abnormality detection device according to the present invention is characterized in that a one-dimensional signal acquired from a monitoring target is input as a one-dimensional signal and processed as a one-dimensional signal. Hereinafter, a CNN that performs one-dimensional processing as described above is referred to as “one-dimensional CNN”, as distinguished from CNN that performs two-dimensional processing of two-dimensional signals such as image data.

[First Embodiment]
FIG. 1 is a block diagram showing a functional configuration of an abnormality detection apparatus 1 according to the first embodiment of the present invention.

[Function block of abnormality detection device]
The abnormality detection device 1 includes a one-dimensional signal acquisition unit (signal acquisition unit) 2, a one-dimensional signal storage unit 3, a one-dimensional signal preprocessing unit 4, a learning unit 5, and a detection unit 6. The abnormality detection device 1 is roughly divided into a learning unit 5 and a detection unit 6 when their functions are roughly divided.

  The learning unit 5 receives the one-dimensional signal, learns the one-dimensional CNN, calculates the feature quantity of the one-dimensional signal, classifies the one-dimensional signal based on the feature quantity, and detects the abnormality to be monitored. Thus, setting information for constructing a one-dimensional CNN is output.

  The detection unit 6 inputs an unknown one-dimensional signal newly acquired from the monitoring target to the one-dimensional CNN constructed based on the setting information, and detects an abnormality of the monitoring target.

  The one-dimensional signal acquisition unit 2 acquires a one-dimensional time-series signal such as a vibration signal or a pressure signal from a monitoring target using electronic means such as network communication. The one-dimensional signal acquisition unit 2 acquires a one-dimensional signal for learning used when the learning unit 5 performs learning, that is, a one-dimensional signal labeled as abnormal or normal. In addition, the one-dimensional signal acquisition unit 2 acquires an unknown one-dimensional signal from which the detection unit 6 detects an abnormality from the monitoring target. Note that the one-dimensional signal acquisition unit 2 may acquire a time-series one-dimensional signal with multiple channels from a plurality of sensors.

  The one-dimensional signal storage unit 3 temporarily stores the one-dimensional signal acquired by the one-dimensional signal acquisition unit 2. More specifically, the one-dimensional signal storage unit 3 temporarily stores a labeled one-dimensional signal acquired for learning and an unlabeled unknown one-dimensional signal that is an abnormality detection target. To do. The one-dimensional signal storage unit 3 cuts out a one-dimensional signal and stores it.

  The one-dimensional signal preprocessing unit 4 performs processing such as amplitude normalization on the one-dimensional signal stored in the one-dimensional signal storage unit 3 and filtering for removing stationary noise. In addition, what is necessary is just to implement pre-processing of a one-dimensional signal as needed, and the structure which abbreviate | omits the one-dimensional signal pre-processing part 4 may be employ | adopted.

The learning unit 5 includes a feature amount learning unit 51 and a classifier learning unit 52.
The learning unit 5 learns a large number of one-dimensional signals acquired by the one-dimensional signal acquisition unit 2, and determines the calculation contents of the one-dimensional CNN so that the detection unit 6 can correctly classify unknown one-dimensional signals.

  The feature amount learning unit 51 calculates a feature amount of a one-dimensional signal, and outputs feature amount setting information for constructing a one-dimensional CNN based on the calculated feature amount. Therefore, the feature amount learning unit 51 has a function of learning a method for calculating a feature amount of a one-dimensional signal.

  More specifically, the feature amount learning unit 51 performs one-dimensional CNN learning based on a one-dimensional signal for learning. The feature quantity learning unit 51 includes a one-dimensional CNN convolution layer (one-dimensional convolution layer). Feature amount setting information, which is a method for calculating the feature amount of the one-dimensional signal obtained by the learning of the feature amount learning unit 51, is stored in a feature amount setting storage unit 61 of the detection unit 6 described later.

  The classifier learning unit 52 receives the feature amount setting information as input, classifies the one-dimensional signal, and outputs classifier setting information (classification setting information) for constructing the one-dimensional CNN based on the classification result. . More specifically, the classifier learning unit 52 receives the feature quantity of the one-dimensional signal calculated by the feature quantity learning unit 51 as an input, and classifies normality and abnormality in the one-dimensional signal, that is, the classifier is supervised. Learn by learning. The classifier learning unit 52 is configured by a fully connected layer in the one-dimensional CNN. The classifier setting information output by the classifier learning unit 52 is stored in a classifier setting storage unit 62 of the detection unit 6 described later.

The detection unit 6 includes a feature amount setting storage unit 61, a classifier setting storage unit 62, a feature amount calculation unit 63, a classification unit 64, and an output unit 65.
The detection unit 6 classifies whether the unknown one-dimensional signal is normal or abnormal according to the setting information of the calculation of the one-dimensional CNN learned by the learning unit 5, and detects the abnormality of the one-dimensional signal.

  The feature amount setting storage unit 61 stores feature amount setting information output from the feature amount learning unit 51. More specifically, the feature value setting storage unit 61 stores weight parameter values and the like in the convolutional layer of the one-dimensional CNN determined by learning.

  The classifier setting storage unit 62 stores classifier setting information output from the classifier learning unit 52. More specifically, the classifier setting storage unit 62 stores the values of parameters such as weights in the all connected layers of the one-dimensional CNN determined by learning.

  The feature amount calculation unit 63 calculates the feature amount of the unknown one-dimensional signal based on the feature amount setting information stored in the feature amount setting storage unit 61. More specifically, the feature amount calculation unit 63 reads the value of a parameter such as the weight of the learned one-dimensional CNN from the feature amount setting storage unit 61, performs the calculation of the convolution layer of the one-dimensional CNN, and performs an unknown one-dimensional The feature amount of the signal is calculated.

  Based on the classifier setting information stored in the classifier setting storage unit 62, the classifying unit 64 inputs the feature quantity calculated by the feature quantity calculation unit 63 and classifies an unknown one-dimensional signal to detect an abnormality. To detect. More specifically, the classification unit 64 reads the value of a parameter such as a learned one-dimensional CNN weight from the classifier setting storage unit 62 using the feature amount calculated by the feature amount calculation unit 63 as an input, and performs primary processing. The calculation of all coupling layers of the original CNN is performed to classify unknown one-dimensional signals.

  Here, the learning unit 5 and the detection unit 6 are configured by a one-dimensional CNN having a convolution layer and a fully connected layer having the same structure, and perform learning and detection, which are functions corresponding to each other. Details of the structure of the one-dimensional CNN used in this embodiment will be described later.

  The output unit 65 outputs the classification result of the one-dimensional signal by the classification unit 64. For example, the output unit 65 may display information indicating the presence / absence of an abnormality such as a classification result on a screen or the like, or may send a signal indicating the classification result to an external recording device, control device, or monitoring device.

[Hardware configuration of abnormality detection device]
Next, a configuration example of hardware for realizing the abnormality detection apparatus 1 having the above-described configuration will be described with reference to FIG.

  The anomaly detection apparatus 1 includes a computer having a CPU 103 and a main storage device 104 connected via a bus 101, a communication control device 105, a sensor 106, an external storage device 107, a display device 108, etc., and these It can be realized by a program that controls the hardware resources.

  The CPU 103 and the main storage device 104 constitute an arithmetic device 102. A program for the CPU 103 to perform various controls and calculations is stored in the main storage device 104 in advance.

  The communication control device 105 is a control device for connecting the abnormality detection device 1 and various external electronic devices via a communication network NW. The communication control apparatus 105 may transmit the one-dimensional signal classification result to an external device or the like via the communication network NW.

  The sensor 106 includes a thermometer, a pressure gauge, a flow meter, and the like that measure one-dimensional signals such as temperature, pressure, flow rate, opening degree of the control valve, vibration, and torque. The one-dimensional signal measured by the sensor 106 is acquired by the one-dimensional signal acquisition unit 2 described with reference to FIG. A configuration including the sensor 106 in the one-dimensional signal acquisition unit 2 may be employed.

  The external storage device 107 includes a readable / writable storage medium and a drive device for reading / writing various information such as programs and data from / to the storage medium. For the external storage device 107, a semiconductor memory such as a hard disk or a flash memory can be used as a storage medium. The external storage device 107 is a one-dimensional signal storage unit 107a, a feature amount setting storage unit 107b, a classifier setting storage unit 107c, a program storage unit 107d, and other storage devices (not shown), for example, stored in the external storage device 107. It is possible to have a storage device or the like for backing up stored programs and data.

  The one-dimensional signal storage unit 107a, the feature amount setting storage unit 107b, and the classifier setting storage unit 107c are respectively included in the one-dimensional signal storage unit 3, the feature amount setting storage unit 61, and the classifier setting storage unit 62 described in FIG. Correspond.

  The program storage unit 107d executes arithmetic processing necessary when the abnormality detection device 1 detects an abnormality in the one-dimensional signal, such as preprocessing, learning processing, and abnormality detection processing of the one-dimensional signal in the present embodiment. The various programs are stored.

  The display device 108 constitutes the display screen of the output unit 65 described with reference to FIG. The display device 108 is realized by a liquid crystal display or the like.

[One-dimensional CNN configuration example]
Next, a configuration example of the one-dimensional CNN used for the learning unit 5 and the detection unit 6 will be described with reference to FIG.
As shown in FIG. 3, the feature quantity learning unit 51 of the learning unit 5 includes a neural network having at least one convolutional layer. More preferably, the feature amount learning unit 51 is configured by a neural network having only a plurality of convolutional layers as shown in FIG. Since the convolutional layer is excellent in extracting signal features, as shown in FIG. 3, the convolutional layer may be arranged in the front row of the network using a one-dimensional signal as an input.

  The classifier learning unit 52 of the learning unit 5 includes at least one layer of neural network. More preferably, the classifier learning unit 52 includes a neural network having only one or more fully connected layers as shown in FIG.

  As described above, the learning unit 5 composed of the one-dimensional CNN having the convolutional layer and the fully connected layer uses a one-dimensional learning signal that is prepared in advance and is labeled with normality and abnormality classified. Then, supervised learning is performed, and the values of parameters such as weights are updated for the entire one-dimensional CNN so that the classification error is minimized.

Here, the principle of the one-dimensional CNN used in this embodiment will be described with reference to FIG. 4 in the case where the number of filters in the convolution layer is 1 and the stride of the filter is 1.
As shown in FIG. 4, the forward propagation process for one convolutional layer in the one-dimensional CNN is expressed by the following equation (1).

Here, the input signal X = [x ₀ , x ₁ ,...] Of the convolutional layer of the one-dimensional CNN, the output signal H = [h ₀ , h ₁ ,...], The filter [w ₀ , w ₁ ,. ..], The filter size n, and the bias b, and the above equation (1) shows the calculation for the i-th element of the input signal.

  When the filter stride is 1, the calculation is sequentially performed for the (i + 1) th element. When the convolutional layer is composed of multiple layers in the one-dimensional CNN, the output signal H of the above equation (1) may be calculated again as the input signal of the next convolutional layer. Thus, it can be seen that the calculation content of the convolution layer in the one-dimensional CNN is centered on the convolution calculation.

  The convolution operation has the property of being the product of the frequency component of the input signal and the frequency component of the filter when considered in the frequency space. That is, the effect of enhancing the input signal with the frequency component of the convolution filter can be obtained.

The output signal H of the convolutional layer in the one-dimensional CNN is classified in the total coupling layer. The forward propagation process for one layer of all coupling layers is expressed by the following equation (2), for example.
Y = HW + B (2)
Here, the output vector Y, the weight matrix W, and the bias vector B.

Each element of the output vector Y obtained by the above equation (2) is subjected to threshold determination by an activation function. When all the coupling layers are formed in multiple layers, the output vector Y in the above equation (2) may be calculated again as the input signal H. However, in the one-dimensional CNN, the dimension of the output vector Y in the last coupled layer in all the columns needs to be the number of classes to be classified. In this embodiment, since the one-dimensional signal is classified as normal or abnormal, the number of classes is 2 (Y = [y ₀ , y ₁ ]).

  In the classification result Y, an error from the teacher label is obtained by the loss function, and the filter and bias of the convolution layer and the weight matrix and bias vector of the all coupling layer are updated so as to reduce the error by the error back propagation method. By continuing to update the one-dimensional CNN repeatedly a plurality of times so that the classification accuracy becomes high, the effect of learning progresses so as to automatically acquire a filter that emphasizes a characteristic frequency component in the one-dimensional signal to be classified. is there.

  Examples of the activation function used by the classifier learning unit 52 of the learning unit 5 include a sigmoid function. The activation function is not limited to a sigmoid function, and for example, a rectified linear unit (ReLU), a hyperbolic tangent, a soft sign, a soft plus, a linear or higher order polynomial, and the like may be used.

  As the loss function for calculating the classification error, for example, softmax cross entropy may be used. If two-class classification is performed such as normal or abnormal, a square sum error may be used as a loss function other than softmax cross entropy.

By using the one-dimensional CNN as described above, the abnormality detection apparatus 1 can automatically learn and acquire a feature amount from a one-dimensional signal.
For the implementation of the one-dimensional CNN, a plurality of open source software frameworks can be used, and these software frameworks may be used for the implementation.

  As described above, the feature amount calculation unit 63 and the classification unit 64 of the detection unit 6 are one-dimensional CNNs having the same structure as the learning unit 5, and weights of learned one-dimensional CNNs obtained by learning by the learning unit 5. One-dimensional CNN forward propagation processing is performed using the parameter value.

[Operation of abnormality detection device]
Next, the operation of the abnormality detection apparatus 1 having the above-described configuration will be described using the flowcharts of FIGS. The operation of the abnormality detection apparatus 1 includes a learning process and a detection process.

[Learning process]
FIG. 5 is a flowchart of the learning process in the abnormality detection device 1.
First, the one-dimensional signal acquisition unit 2 acquires a one-dimensional signal group that has been classified as normal and abnormal in advance and labeled as teacher data from the monitoring target (step S100). The acquired one-dimensional signal group of teacher data is temporarily stored in the one-dimensional signal storage unit 3.

  Next, the one-dimensional signal preprocessing unit 4 performs preprocessing such as amplitude normalization of the acquired one-dimensional signal and removal of stationary noise by filtering (step S101). The preprocessed one-dimensional signal for learning is input to the learning unit 5, and forward propagation processing in the one-dimensional CNN is performed (step S102).

  More specifically, the feature value learning unit 51 of the learning unit 5 configured by the convolutional layer of the one-dimensional CNN calculates the feature value of the one-dimensional signal using the above-described equation (1). Further, the classifier learning unit 52 configured by a fully connected layer of one-dimensional CNN performs classification of the one-dimensional signal using the above-described equation (2) with the output signal of the feature amount learning unit 51 as an input.

  Next, the learning unit 5 obtains an error between the output value obtained by the one-dimensional CNN forward propagation process and the teacher label as the target value (step S103). After that, when the error between the output value from the one-dimensional CNN forward propagation process and the teacher label is not sufficiently small (step S104: NO), the learning unit 5 performs the one-dimensional CNN error back-propagation process (step S104). S105).

  More specifically, in the error back-propagation process, the convolutional layer that constitutes the feature quantity learning unit 51 and the classifier learning unit 52 are configured so that the error between the output value and the teacher label due to the one-dimensional CNN forward propagation process is reduced. The values of parameters such as the weights of all the connected layers that are configured are updated.

  Next, the learning unit 5 uses a one-dimensional CNN forward propagation process (step S102) and an error using the updated values of parameters such as the weights of the one-dimensional CNNs constituting the feature amount learning unit 51 and the classifier learning unit 52. Is derived (step S103).

  Thereafter, when the error between the output value from the one-dimensional CNN forward propagation process and the teacher label is sufficiently small (step S104: YES), the parameter values such as the updated weight are stored (step S104). S106). More specifically, the final update value of the weight parameter of the convolution layer constituting the feature amount learning unit 51 is stored in the feature amount setting storage unit 61. The final updated values of the weight parameters of all the connection layers that constitute the classifier learning unit 52 are stored in the classifier setting storage unit 62.

[Detection processing]
Next, the detection process by the detection part 6 is demonstrated using the flowchart of FIG.
The detection unit 6 reads the learned one-dimensional CNN obtained by the learning process by the learning unit 5 and classifies whether the unknown one-dimensional signal is normal or abnormal.
First, the one-dimensional signal acquisition unit 2 acquires an unknown one-dimensional signal that is not labeled from the monitoring target (step S110). The acquired one-dimensional signal is temporarily stored in the one-dimensional signal storage unit 3.

  Next, the one-dimensional signal pre-processing unit 4 performs pre-processing such as amplitude normalization of the one-dimensional signal and removal of stationary noise by filtering in the same manner as the pre-processing in the learning process (step S111). Thereafter, the one-dimensional signal is input to the detection unit 6, and the detection unit 6 performs the forward propagation process of the learned one-dimensional CNN (step S112).

  More specifically, the feature amount calculation unit 63 reads the values of parameters such as weights in the learned convolutional layer stored in the feature amount setting storage unit 61. The feature amount calculation unit 63 performs a convolution layer operation using an unknown one-dimensional signal as an input, using the above-described equation (1). Further, the classification unit 64 reads the value of the weight parameter in the learned all connected layers stored in the classifier setting storage unit 62. The classification unit 64 performs an operation on all connected layers using the output value of the feature amount calculation unit 63 as an input using the above-described equation (2).

  Thereafter, the output unit 65 outputs the detection result of the presence or absence of abnormality in the one-dimensional signal based on the calculation result by the classification unit 64 (step S113). More specifically, the calculation result by the classification unit 64 is a classification result indicating that the unknown one-dimensional signal is normal or abnormal. The classification unit 64 detects an abnormality of the one-dimensional signal from the classification result.

  As described above, according to the first embodiment, the abnormality detection apparatus 1 can reduce the calculation cost for converting a one-dimensional signal into a two-dimensional signal by using the one-dimensional CNN. In addition, the calculation cost is further reduced by changing the CNN processing from two-dimensional signal processing to one-dimensional signal processing. As a result, even a device such as a field device having a relatively low calculation ability can automatically learn and acquire a feature quantity of a one-dimensional signal and detect an abnormality.

[Second Embodiment]
Next, a second embodiment of the present invention will be described. In the following description, the same components as those in the first embodiment described above are denoted by the same reference numerals, and the description thereof is omitted.

  In the first embodiment, the case where the feature amount learning unit 51 and the feature amount calculation unit 63 are configured by only one or more convolution layers has been described. On the other hand, in the second embodiment, the feature quantity learning unit 51A has a neural network including one or more convolution layers and one or more recursive neural network layers.

[Outline of Second Embodiment]
As is well known, since the CNN learns with a local band filter, it is excellent in learning the feature quantity in the local band of the input signal. On the other hand, CNN has a property that classification performance deteriorates when wideband noise such as white noise is included in an input signal (see Non-Patent Document 2).

  For this reason, attention is paid to a recursive neural network as a neural network having a characteristic capable of learning a robust feature amount even when broadband noise is included in an input signal. The recurrent neural network has a feature that it has a feedback path inside the network and can hold past signal values.

  Further, among recursive neural networks, in particular, Long Short-Term Memory (hereinafter referred to as “LSTM”) is known to be able to store past signal values for a long period of time (see Non-Patent Document 3). For this reason, LSTM is considered to be able to learn more robust feature values when broadband noise is included in the input signal.

  As described above, in the present embodiment, by using a neural network having a structure in which LSTM is combined with one-dimensional CNN, even when broadband noise is included in the input signal, feature quantities are automatically learned and acquired from the one-dimensional signal. An abnormality detection device 1A that can perform the above is realized.

[Function block of abnormality detection device]
FIG. 7 is a block diagram showing a functional configuration of the abnormality detection apparatus 1A according to the present embodiment. Hereinafter, a description will be given focusing on a configuration different from the first embodiment.

The learning unit 5A includes a feature amount learning unit 51A and a classifier learning unit 52.
The feature amount learning unit 51A includes an LSTM learning unit 510. As described above, the feature amount learning unit 51A includes the convolution layer and the LSTM layer. The feature amount learning unit 51A performs a calculation process on the convolutional layer of the one-dimensional CNN to learn a method for calculating the feature amount of the one-dimensional signal, and outputs feature amount setting information.

  The LSTM learning unit 510 receives the output signal from the feature amount learning unit 51A as input, learns the calculation method of the feature amount of the one-dimensional signal by performing arithmetic processing of the LSTM layer included in the one-dimensional CNN, and sets the feature amount Output information. Details of the LSTM layer constituting the LSTM learning unit 510 will be described later.

  As described above, the feature amount learning unit 51A configured by the convolution layer is excellent in learning the feature amount of a periodic signal in a certain narrow frequency region. On the other hand, the LSTM learning unit 510 configured by the LSTM layer is excellent by learning the feature amount of the non-stationary signal in a wider frequency region.

  In this way, the feature quantity learning unit 51A and the LSTM learning unit 510 learn different feature quantities in the one-dimensional signal, and by combining these, depending on the characteristics of the one-dimensional signal to be classified, they complement each other and perform classification. Accuracy may be improved.

The detection unit 6A includes a feature amount setting storage unit 61A, a classifier setting storage unit 62, a feature amount calculation unit 63A, a classification unit 64, and an output unit 65.
The feature amount setting storage unit 61A includes an LSTM feature amount setting storage unit 610.
The feature amount calculation unit 63A includes an LSTM feature amount calculation unit 630.

The feature amount setting storage unit 61A stores feature amount setting information of a one-dimensional signal obtained by learning of the convolutional layer by the feature amount learning unit 51A.
The LSTM feature value setting storage unit 610 stores feature value setting information of a one-dimensional signal obtained by learning the LSTM layer by the LSTM learning unit 510.

The feature amount calculation unit 63A refers to the feature amount setting information of the one-dimensional signal stored in the feature amount setting storage unit 61A, performs a calculation process on the convolution layer of the one-dimensional CNN, and performs an unknown one-dimensional signal The feature amount is calculated.
The LSTM feature value calculation unit 630 refers to the setting information of the feature value of the one-dimensional signal stored in the LSTM feature value setting storage unit 610, performs an LSTM layer calculation process in the one-dimensional CNN, and performs the processing of the one-dimensional signal. The feature amount is calculated.

  More specifically, the LSTM feature quantity calculation unit 630 reads a parameter value such as the weight of the LSTM layer, inputs the feature quantity output from the feature quantity calculation unit 63A, performs an operation on the LSTM layer, and performs an unknown operation. The feature amount of the one-dimensional signal is calculated.

  The classification unit 64 receives the one-dimensional signal feature value calculated by the LSTM feature value calculation unit 630 as an input, and refers to the information on the method for classifying the one-dimensional signal stored in the classifier setting storage unit 62 and is unknown. Anomalies are detected by classifying one-dimensional signals.

[One-dimensional CNN configuration example]
Next, a configuration example of a one-dimensional CNN having an LSTM layer used in the learning unit 5A and the detection unit 6A described above will be described with reference to FIG.

  The feature amount learning unit 51A of the learning unit 5A has one or more convolution layers, and has the LSTM layer of the LSTM learning unit 510 in the rear row of the convolution layer. More preferably, as shown in FIG. 8, the LSTM layer may be arranged in the column next to the last column of the convolution layer, and the output of the convolution layer may be used as the input of the LSTM layer. The feature amount calculation unit 63A and the LSTM feature amount calculation unit 630 of the detection unit 6A are configured by the same convolution layer and LSTM layer as the feature amount learning unit 51A and the LSTM learning unit 510 of the learning unit 5A.

  As in the first embodiment, the classifier learning unit 52 of the learning unit 5A is configured by one or more fully connected layers. The classification unit 64 of the detection unit 6A includes one or more fully connected layers that are the same as the classifier learning unit 52 of the learning unit 5A.

[Operation processing of LSTM layer]
Next, the arithmetic processing of the LSTM layer which comprises the LSTM learning part 510 and the LSTM feature-value calculation part 630 is demonstrated using the block diagram of FIG.

FIG. 9 shows the contents of forward propagation processing for one LSTM layer. In FIG. 9, “+” indicates addition, “×” indicates multiplication, and “S” indicates a sigmoid function. The broken-line arrows represent first-order lags, and C1 and C2 are memory cells, respectively. The input signal x _t and the output signal h _t are values at time t, and the parameters starting from w and R are weights.

In the LSTM layer, there is a recursive coupling in the memory cell C1 portion. The LSTM layer has three gates G1, G2, and G3. Input signal x _t is also inputted to these three gates G1, G2, G3, used for opening and closing the respective gates.

  The three gates G1, G2, and G3 are used for controlling how much information is passed. When the gate is closed (value approaches 0), it becomes difficult for information to pass. When the gate is open, the sigmoid function is close to 1.

  In particular, when the gate G2 is open, the state of the memory cell C1 one time before (t-1) affects the state at the time t. Therefore, the gate G2 determines the degree of consideration for the influence of the immediately preceding state.

  The LSTM layer can learn long-term dependency by learning weight parameters in the gates G1, G2, and G3. When the LSTM layer is composed of multiple layers, the output h in FIG. 9 may be calculated again as the input signal x.

  Since the implementation of the above LSTM and recursive neural network can use a plurality of open source software frameworks, these software frameworks may be used.

[Operation of abnormality detection device]
Next, the operation of the abnormality detection apparatus 1A having the above-described configuration will be described using the flowcharts of FIGS. In the following, a description will be given focusing on processing different from that of the first embodiment.

[Learning process]
First, the learning process by the learning unit 5A of the abnormality detection device 1A will be described.
As shown in FIG. 10, the one-dimensional signal acquisition unit 2 acquires a labeled one-dimensional signal group from the monitoring target as teacher data (step S100). Thereafter, the one-dimensional signal preprocessing unit 4 performs preprocessing of the acquired one-dimensional signal (step S101).

  Next, the learning unit 5A performs a one-dimensional CNN forward propagation process on the input one-dimensional signal (step S102A). More specifically, in the one-dimensional CNN convolutional layer included in the feature amount learning unit 51A, the feature amount in the one-dimensional signal for learning is calculated using the above-described equation (1).

  Further, with the output signal of the convolution layer included in the feature amount learning unit 51A as an input, the LSTM learning unit 510 calculates the feature amount in the one-dimensional signal for learning in the LSTM layer shown in FIG.

  Furthermore, the classifier learning unit 52 configured by a fully connected layer receives the output signal of the LSTM learning unit 510 and outputs the classification result of the learning one-dimensional signal using the above-described equation (2).

  Next, the learning unit 5A obtains an error between the output value by the forward propagation process of the one-dimensional CNN having the LSTM layer and the teacher label that is the target value (step S103). Thereafter, when the error between the output value obtained by the forward propagation process of the one-dimensional CNN and the teacher label is not sufficiently small (step S104: NO), the learning unit 5A performs the error back-propagation process of the one-dimensional CNN (step S104). S105A).

  More specifically, in the error back-propagation process, the convolution layer and the LSTM learning unit 510 that configure the feature amount learning unit 51A are configured so that the error between the output value and the teacher label by the one-dimensional CNN forward propagation process is reduced. The values of the weight parameters of the LSTM layer and all connected layers constituting the classifier learning unit 52 are updated.

  Next, the learning unit 5A uses the value of the parameter such as the weight of the one-dimensional CNN having the LSTM layer updated by the feature amount learning unit 51A, the LSTM learning unit 510, and the classifier learning unit 52, to determine the one-dimensional CNN. Forward propagation processing (step S102A) and error derivation (step S103) are performed.

  Thereafter, when the error between the output value from the one-dimensional CNN forward propagation process and the teacher label is sufficiently small (step S104: YES), the parameter values such as the updated weight are stored (step S104). S106A).

  More specifically, final update values of parameters such as the weight of the convolution layer included in the feature amount learning unit 51A are stored in the feature amount setting storage unit 61. The final update values of parameters such as the weights of the LSTM layers that constitute the LSTM learning unit 510 are stored in the LSTM feature amount setting storage unit 610. The final updated values of parameters such as the weights of all connected layers constituting the classifier learning unit 52 are stored in the classifier setting storage unit 62.

[Detection processing]
Next, detection processing by the detection unit 6A will be described with reference to the flowchart of FIG.
Using the one-dimensional CNN having the learned LSTM layer obtained by the learning process by the learning unit 5A, the detection unit 6A classifies whether the unknown one-dimensional signal is normal or abnormal.
First, the one-dimensional signal acquisition unit 2 acquires an unknown one-dimensional signal that is not labeled (step S110).

  Next, the one-dimensional signal preprocessing unit 4 performs preprocessing of the acquired one-dimensional signal (step S111). Thereafter, the one-dimensional signal is input to the detection unit 6A, and the detection unit 6A performs forward propagation processing of the one-dimensional CNN having the learned LSTM layer (step S112A).

  More specifically, the feature amount calculation unit 63A reads the values of parameters such as weights in the learned convolutional layer stored in the feature amount setting storage unit 61A, and uses the above-described equation (1) to perform one-dimensional A convolutional layer is calculated using the signal as an input to calculate the feature quantity of the one-dimensional signal.

  Further, the LSTM feature value calculation unit 630 reads the values of parameters such as weights in the learned LSTM layer stored in the LSTM feature value setting storage unit 610, and uses the feature values calculated by the feature value calculation unit 63A. An LSTM layer calculation is performed as an input, and a feature value of the one-dimensional signal is calculated.

  Further, the classification unit 64 reads the weight parameter values in the learned all connection layers stored in the classifier setting storage unit 62, and uses the above-described equation (2) to output the feature amount calculation unit 63. Performs calculation for all connected layers using the value as input.

  The calculation result by the classification unit 64 is output by the output unit 65 (step S113). The calculation result by the classification unit 64 indicates whether the unknown one-dimensional signal is normal or abnormal.

  As described above, according to the second embodiment, since the one-dimensional CNN having the LSTM layer is used for the learning unit 5A and the detection unit 6A, the feature amount in the local band of the one-dimensional signal is convolved. In the LSTM layer, non-stationary signal feature amounts in a wider frequency range are learned. Therefore, the abnormality detection device 1A can automatically detect and acquire a robust feature amount and detect an abnormality even when a one-dimensional signal includes broadband noise.

  In addition, since the LSTM layer included in the learning unit 5A and the detection unit 6A can directly handle a one-dimensional signal, it automatically learns a more robust feature amount while reducing the calculation cost in the abnormality detection device 1A. Can be earned.

[Example]
Next, abnormality detection inside the control valve using the abnormality detection devices 1 and 1A according to the first and second embodiments described above will be described.

  In an oil plant or a chemical plant, a control valve is used as an operation end for controlling the flow rate of a fluid. Depending on the conditions of temperature, pressure, and flow rate of the fluid flowing inside the control valve, the generation and disappearance of bubbles called cavitation may occur. Cavitation is a factor that damages the inside of the control valve. In this embodiment, the purpose of the present embodiment is to detect the occurrence of cavitation as an abnormality in the control valve to be monitored.

  In the present embodiment, a vibration sensor is used as the one-dimensional signal acquisition unit 2 to acquire vibration when a fluid flows from the outside of the control valve through the inside of the control valve to obtain a one-dimensional signal of the input signal. The vibration signal acquired by the vibration sensor includes only a fluid sound when normal, but includes a cavitation sound in addition to the fluid sound when abnormal.

  12 and 13 are diagrams illustrating an example of a vibration signal that is a target of abnormality detection in the present embodiment. FIG. 12A shows a vibration signal in a normal state, and FIG. 12B shows a vibration signal in an abnormal state. As can be seen from FIGS. 12 (a) and 12 (b), the difference in vibration signal between normal and abnormal conditions including cavitation sound is relatively small.

  FIG. 13A shows the distribution of the spectrum of the vibration signal at the normal time, and FIG. 13B shows the distribution of the spectrum of the vibration signal at the time of the abnormality. As can be seen from FIGS. 13A and 13B, the frequency bands of the vibration signals at the normal time and at the time of abnormality including the cavitation sound are both wide and overlapping.

Next, in the present embodiment, the one-dimensional signal storage unit 3 samples an analog signal from the vibration sensor at a sampling frequency of 96 kHz, and stores a signal divided every 10,000 samples.
The one-dimensional signal preprocessing unit 4 normalizes the effective value of the vibration signal so that the amplitude scale of the vibration signal does not change significantly.

Next, a specific configuration of the one-dimensional CNN of the learning units 5 and 5A and the detection units 6 and 6A will be described.
As a specific example of the learning unit 5 and the detection unit 6 according to the first embodiment, the one-dimensional CNN shown in FIG. 3 having three convolution layers and three total coupling layers is used.
As a specific example of the learning unit 5A and the detection unit 6A according to the second embodiment, the LSTM layer shown in FIG. 8 has three convolution layers, one LSTM layer, and one total coupling layer. A one-dimensional CNN was used.

Further, as a setting common to the specific examples of the first and second embodiments, the number of filters of the convolutional layer of the one-dimensional CNN is 1, the filter size is 100, the stride is 2, and the activation function of the all coupling layers is the sigmoid. It was a function.
As teacher data, 1,500 sets of normal vibration signals and 1,100 sets of abnormal vibration signals were prepared. As the loss function, softmax cross entropy was used.

  Under the above conditions, supervised learning was performed in each of the abnormality detection device 1 according to the first embodiment and the abnormality detection device 1A according to the second embodiment. Thereafter, the classification correct answer rate was evaluated using 100 sets of vibration signals at normal time and 100 sets of vibration signals at abnormal time as verification data.

  FIG. 14 is a diagram illustrating a classification correct answer rate by the abnormality detection device 1 according to the first embodiment. Moreover, FIG. 15 is a figure which shows the classification | category correct answer rate by 1 A of abnormality detection apparatuses which concern on 2nd Embodiment. 14 and 15, the horizontal axis represents the number of learnings, and the vertical axis represents the classification correct answer rate. The broken line indicates the classification correct answer rate of the learning data, and the solid line indicates the classification correct answer rate of the verification data.

  As shown in FIGS. 14 and 15, the abnormality detection devices 1 and 1A according to the first and second embodiments gradually learn the feature amount of the vibration signal that is an input signal by performing learning with repeated learning data. It can be seen that the classification correct answer rate has improved.

  As shown in FIG. 14, in the anomaly detection apparatus 1 using a one-dimensional CNN having a convolution layer and a fully connected layer, the classification data correctness rate of verification data has achieved about 90%. From this, it can be said that the abnormality detection device 1 according to the first embodiment has sufficiently achieved the object of abnormality detection.

  As shown in FIG. 15, in the abnormality detection apparatus 1A using the one-dimensional CNN having the LSTM layer, the classification data correct rate of the verification data is about 100%, and the object of abnormality detection can be sufficiently achieved. In particular, when the signal difference between the normal time and the abnormal time is smaller, such as the cavitation vibration signal in the control valve that is the target of abnormality detection in the present embodiment, and the frequency band also overlaps in a wide band, It is considered that the abnormality detection apparatus 1A using a one-dimensional CNN having an LSTM layer is more effective.

  The embodiments of the abnormality detection apparatus and the abnormality detection method of the present invention have been described above. However, the present invention is not limited to the described embodiments, and those skilled in the art can assume the scope of the invention described in the claims. Various possible modifications can be made.

  Note that the learning units 5 and 5A and the detection units 6 and 6A of the abnormality detection apparatus 1 and 1A described above may be implemented by the same computer or may be implemented independently. In particular, when the learning units 5 and 5A and the detection units 6 and 6A are independently mounted, abnormality detection is performed by automatically learning and acquiring feature amounts even in field devices with lower calculation capabilities. Can do.

  More specifically, when the learning units 5 and 5A and the detection units 6 and 6A are implemented independently, the learning units 5 and 5A that learn a large amount of teacher data use a computer with higher calculation ability. Good. On the other hand, a computer with lower calculation capability can be used for the detection units 6 and 6A centering on the forward propagation processing of a neural network having a relatively low calculation cost consisting of sum-of-product calculation and threshold determination by an activation function. . As described above, when the learning units 5 and 5A and the detection units 6 and 6A are respectively configured independently, the abolition of the one-dimensional signal to the two-dimensional signal and the one-dimensional processing of the signal in the one-dimensional CNN The anomaly detection devices 1 and 1A can be mounted on field equipment having a lower calculation capability.

  In the described embodiment, the case where the one-dimensional signal acquisition unit 2 and the one-dimensional signal storage unit 3 have independent configurations has been described. However, the one-dimensional signal acquisition unit 2 and the one-dimensional signal storage unit 3 may be implemented as one integrated configuration.

  The one-dimensional signal acquisition unit 2 and the one-dimensional signal storage unit 3 may be shared by the learning units 5 and 5A and the detection units 6 and 6A, and the learning units 5 and 5A and the detection unit 6 6A, a dedicated one-dimensional signal acquisition unit 2 and a one-dimensional signal storage unit 3 may be prepared.

  Further, in the embodiment described above, the case where the feature amount setting storage unit 61 and the classifier setting storage unit 62 are provided in different areas has been described, but these may be configured as one storage unit. Good.

  In the embodiment described above, the classification unit 64 has described the case of performing two-class classification of whether the one-dimensional signal is normal or abnormal. However, the number of classification classes is not limited to two, and multi-class classification is performed. Also good.

  DESCRIPTION OF SYMBOLS 1, 1A ... Abnormality detection apparatus, 2 ... One-dimensional signal acquisition part, 3, 107a ... One-dimensional signal storage part, 4 ... One-dimensional signal pre-processing part, 5, 5A ... Learning part, 6, 6A ... Detection part, 51 , 51A ... feature amount learning unit, 52 ... classifier learning unit, 61, 61A, 107b ... feature amount setting storage unit, 62, 107c ... classifier setting storage unit, 63, 63A ... feature amount calculation unit, 64 ... classification unit , 65 ... output unit, 101 ... bus, 102 ... arithmetic device, 103 ... CPU, 104 ... main storage device, 105 ... communication control device, 106 ... sensor, 107 ... external storage device, 107d ... program storage unit, 108 ... display Device, NW ... communication network, 510 ... LSTM learning unit, 610 ... LSTM feature value setting storage unit, 630 ... LSTM feature value calculation unit.

Claims (9)

A signal acquisition unit for acquiring one-dimensional data, which is a one-dimensional time series signal, from a monitoring target;
Using the acquired one-dimensional data as input, learn a neural network, calculate the feature quantity of the one-dimensional data, classify the one-dimensional data based on the feature quantity, and detect an abnormality of the monitoring target A learning unit that outputs setting information for constructing the neural network;
A storage unit for storing the setting information;
An unknown one-dimensional data which is a one-dimensional time series signal newly acquired from the monitoring target is input to the neural network constructed based on the setting information stored in the storage unit, and the monitoring is performed. A detection unit for detecting a target abnormality;
With
The neural network includes a one-dimensional convolutional layer that performs one-dimensional convolution on the one-dimensional data. The abnormality detection device according to claim 1,
The learning unit
A feature amount learning unit that calculates a feature amount of the one-dimensional data and outputs setting information of a feature amount for constructing the neural network based on the calculated feature amount;
A classifier learning unit that classifies the one-dimensional data using the feature amount setting information as input, and outputs classification setting information for constructing the neural network based on a classification result;
With
The storage unit stores setting information of the feature amount and setting information of the classification,
The detector is
A feature amount calculation unit that calculates a feature amount of the unknown one-dimensional data based on the setting information of the feature amount stored in the storage unit;
A classification unit configured to classify the unknown one-dimensional data and detect an abnormality by using the feature amount calculated by the feature amount calculation unit based on the classification setting information stored in the storage unit; An abnormality detection device characterized by the above. In the abnormality detection device according to claim 1 or 2,
The anomaly detection apparatus, wherein the neural network includes at least one layer of the one-dimensional convolution layer and at least one full coupling layer. In the abnormality detection device according to any one of claims 1 to 3,
The abnormality detection apparatus according to claim 1, wherein the neural network further includes a recursive neural network layer that receives an output of the one-dimensional convolutional layer. In the abnormality detection device according to claim 4,
The abnormality detection device, wherein the recursive neural network layer is an LSTM layer. A signal acquisition step of acquiring one-dimensional data that is a one-dimensional time-series signal from a monitoring target;
Using the acquired one-dimensional data as an input, the neural network is trained to calculate a feature quantity of the one-dimensional data, and the one-dimensional data is classified based on the feature quantity to detect an abnormality of the monitoring target. Learning step for outputting setting information for constructing the neural network,
An unknown one-dimensional data which is a one-dimensional time-series signal newly acquired from the monitoring target is input to the neural network constructed based on the setting information stored in the storage unit, and the monitoring target A detection step for detecting an abnormality of
With
The abnormality detection method, wherein the neural network includes a one-dimensional convolution layer that performs one-dimensional convolution on the one-dimensional data. In the abnormality detection method according to claim 6,
The learning step includes
A feature amount learning step of calculating a feature amount of the one-dimensional data and outputting setting information of a feature amount for constructing the neural network based on the calculated feature amount;
A classifier learning step for classifying the one-dimensional data using the feature amount setting information as input, and outputting classification setting information for constructing the neural network based on a classification result;
With
The detecting step includes
A feature amount calculating step of calculating a feature amount of the unknown one-dimensional data based on the setting information of the feature amount stored in the storage unit;
Based on the setting information of the classification stored in the storage unit, with the feature quantity calculated in the feature quantity calculation step as an input, a classification step of classifying unknown one-dimensional data and detecting an abnormality,
An abnormality detection method comprising: In the abnormality detection method according to claim 6 or 7,
The abnormality detection method, wherein the neural network further includes a recursive neural network layer that receives an output of the one-dimensional convolution layer. The abnormality detection method according to claim 8,
The abnormality detection method, wherein the recurrent neural network layer is an LSTM layer.