JP4605132B2 - Anomaly detection device and anomaly detection method - Google Patents

Description

  The present invention relates to an abnormality detection apparatus and an abnormality detection method for providing a competitive learning type neural network with a feature amount extracted from a target signal including a vibration component generated during operation of a device and monitoring whether the device being operated is normal or abnormal. Is.

  Conventionally, a sensory test has been performed in which an inspector listens to sound waves generated during operation of the device to inspect the presence or absence of the device. However, in this type of sensory test, the results may vary depending on individual differences in the judgment of the inspector and the physical condition of the inspector.

  Therefore, the threshold value is statistically determined by extracting the frequency component of the sound wave generated during the operation of the device, and obtaining the distribution of the maximum value of the extracted frequency component, and the maximum value of the frequency component as the threshold value. It has been proposed to separate non-defective products from defective products by comparison (see, for example, Patent Document 1).

  When this technology is adopted, there is no variation in the results during the inspection, but it is necessary to determine the frequency component to be noticed according to the type of defect, and the threshold for distinguishing it from non-defective products depending on the type of defect. It may be difficult to determine the value.

  On the other hand, Patent Document 1 describes a technique called peak value analysis that uses all frequency components instead of using only specific frequency components. In this technique, the peak value of the differential waveform generated by impact vibration is evaluated by differentiating the waveform of the sound wave. However, as described in Patent Document 1, even if this technique is used alone, the types of defects cannot be classified.

  By the way, conventionally, a technique for classifying a target signal including a vibration component by using a classification function of a neural network (neurocomputer) is known. This type of technology is also used in an abnormality detection apparatus that determines whether the operation of a device is normal or abnormal. For example, in an abnormality detection apparatus that employs this type of technology, the operation sound of the device or the vibration of the device is converted into an electrical signal by a sensor unit (transducer), and the output of the sensor unit is used as the target signal. Various techniques have been proposed for extracting feature quantities consisting of the above parameters and classifying the feature quantities using a neural network.

  Various configurations are known for neural networks. For example, it has been proposed to classify feature categories using a competitive learning type neural network (self-organizing map = SOM). The competitive learning type neural network is a neural network including two layers of an input layer and an output layer, and performs two operations of a learning mode and an inspection mode.

  In the learning mode, learning data is given without using a teacher signal. If a category is given to learning data, a category can be associated with a neuron in the output layer, and a cluster of neurons belonging to the same category can be formed. Accordingly, in the learning mode, a clustering map indicating a category can be associated with a cluster of neurons in the output layer.

  In the inspection mode, the feature quantity (input data) to be classified is given to the learned competitive learning neural network, and the category of the cluster to which the fired neuron belongs in the output layer of the competitive learning neural network is checked against the clustering map. By doing so, the category of the input data can be classified (for example, see Patent Document 2).

Therefore, if a competitive learning type neural network is trained using learning data having normal categories and learning data having categories for each abnormality, not only normality / abnormality but also abnormality type determination Is also possible. Further, since it is only necessary to determine the category of learning data, it is possible to determine the type of abnormality without selecting a specific frequency component. Further, in the configuration described in Patent Document 1, it is necessary to use both techniques in combination in order to compensate for the mutual disadvantages in the case of focusing on a specific frequency component and the case of using all frequency components. The technique described in No. 2 has an advantage that it is possible to discriminate between normal and abnormal and the type of abnormality only by providing learning data in which categories are appropriately set.
JP 2003-214944 A JP 2004-354111 A

  By the way, in order to classify feature quantities using a competitive learning type neural network, as described above, learning data for each category to be classified is required. For example, as a shipping test for equipment equipped with a motor, when testing whether the equipment operates normally even when the load applied to the motor is changed, learning data for each load magnitude is required. Become.

  On the other hand, in the shipping test, particularly important items are foreign matters generated at the time of manufacture (for example, cutting scraps generated by cutting, scraps generated by removing burrs, scraps generated by chipping of parts, and small intrusions inside. Abnormal noise caused by parts, etc., abnormal noise due to lack of lubricating oil, etc. If these specific abnormalities do not occur, it may be regarded as normal operation.

  That is, in the shipping test or the like, the type of abnormality to be inspected is specified, and the learning data for each abnormality is small in variation, but the variation in learning data may be large for normal categories. In such learning data, the variation of the learning data to which the normal category is assigned increases, so the number of neurons belonging to the normal category increases in the output layer of the competitive learning type neural network, and the neurons belonging to the abnormal category The problem arises that it becomes difficult to separate clearly.

  The present invention has been made in view of the above reasons, and its purpose is to clearly separate normality and abnormality when it is sufficient to determine that a specific abnormal state during operation of the device is abnormal. Another object of the present invention is to provide an abnormality detection apparatus and an abnormality detection method that can reduce the time required for the learning mode by reducing the number of learning data.

The invention according to claim 1 is a signal input unit that captures a target signal including a vibration component that is generated during operation of the device, a feature extraction unit that extracts a feature amount including a plurality of parameters for the target signal, and learning data of a known category A competitive learning type neural network that uses as input data the feature quantity that has been learned in advance using the feature extraction unit, and a determination unit that determines whether the device in operation is normal or abnormal using the output of the competitive learning type neural network And using a feature amount obtained when the device is in a known abnormal state as learning data, the determination unit sets a weight vector for each neuron in the output layer of the competitive learning type neural network based on the learning data, a Gaussian function to determine the degree of membership of the input data to the firing neurons in the learning data of the abnormal state category Equipment when viewed vector and set in association with the Euclidean distance between the learning data, degree of membership of the input data that Gaussian function is obtained during operation of the device for the configured neurons is a threshold hereinafter the provisions The output value when the input data is input to the Gaussian function is used as the degree of attribution, and the Gaussian function retrains the learning data to the competitive learning type neural network after learning. The variance is determined by the Euclidean distance between the weight vector of each neuron that fires when input and the learning data, and the weight vector of the fired neuron is averaged .

In the invention of claim 2, in the invention of claim 1, prior SL threshold characterized in that it is set to three times the variance in the Gaussian function.

The invention according to claim 3 is a feature obtained when the device is in a known abnormal state after taking in a target signal including a vibration component generated during operation of the device and extracting a feature amount including a plurality of parameters for the target signal. the competitive learning neural network which has learned in advance using the amounts as learning data, given the feature amount obtained during the operation of equipment as input data, equipment is normal in operation using the output of the competitive learning neural network An anomaly detection method for determining whether an anomaly is detected by the determination unit, in which a weight vector is set for each neuron in the output layer of a competitive learning type neural network by learning data, and input to a neuron that fires with the learning data in the abnormal category in association a Gaussian function to determine the degree of membership of the data to the Euclidean distance between the heavy viewing vector and learning data Constant and, be those devices is determined to be operating normally when membership of the input data that Gaussian function is obtained during operation of the device for the configured neurons is a threshold hereinafter the provisions The output value when input data is input to the Gaussian function is used as the degree of attribution, and the Gaussian function is the weight vector of each neuron that fires when the learning data is re-input to the competitive learning type neural network after learning. The variance is determined by the Euclidean distance from the learning data, and the weight vectors of the fired neurons are averaged .

In the invention of claim 4, the invention smell of claim 3 Te, before Symbol threshold characterized in that it is set to three times the variance in the Gaussian distribution.

  According to the first and third aspects of the present invention, only the abnormal state is learned by using, as the learning data, the feature quantity obtained in the known abnormal state as the learning data. Can be reduced. As a result, the time required for the learning mode can be shortened. In addition, even if there is a large variation in feature values in the normal state as in a load test, if it is sufficient to determine that a specific abnormal state during operation of the device is abnormal, the neurons that fire in the abnormal state Only input data whose degree of belonging to the threshold is greater than or equal to the threshold is judged as abnormal, and input data whose degree of belonging is less than or equal to the threshold is judged as normal. Anomalies can be clearly separated.

According to the invention of claim 2 and 4, since the threshold value three times the variance in the Gaussian distribution, it is possible to automatically set a threshold value on a regular basis, regardless of the type of device for detecting an abnormality Judgment criteria can be formulated.

  The embodiment described below relates to an abnormality detection device that determines whether an operation of a device is normal or abnormal based on a feature amount of a target signal obtained during operation of the device. The device includes a motor alone and a motor as a drive source. An electric tool or machine tool is assumed. Further, it is assumed that the type of abnormality is specified, and it can be considered normal if no specific abnormality is detected.

  As shown in FIG. 1, the abnormality detection apparatus described in the present embodiment uses an unsupervised competitive learning type neural network (hereinafter simply referred to as “neural network”) 1. As shown in FIG. 2, the neural network 1 includes two layers of an input layer 11 and an output layer 12, and each neuron N <b> 2 of the output layer 12 is coupled to all the neurons N <b> 1 of the input layer 11. Have. Although the neural network 1 is assumed to be realized by executing an appropriate application program on a sequential processing type computer, a dedicated neurocomputer can also be used.

  The operation of the neural network 1 has a learning mode and an inspection mode. After learning using appropriate learning data in the learning mode, a feature amount (input) generated from an actual target signal in the inspection mode. Data) category.

  The degree of connection (weighting coefficient) between the neuron N1 in the input layer 11 and the neuron N2 in the output layer 12 is variable. In the learning mode, the learning data is input to the neural network 1 to learn the neural network 1, and the input layer The weighting coefficients of the 11 neurons N1 and the neurons N2 of the output layer 12 are determined. In other words, each neuron N2 in the output layer 12 is associated with a weight vector whose element is a weight coefficient between each neuron N1 in the input layer 11. Therefore, the weight vector has the same number of elements as the neuron N1 of the input layer 11, and the number of parameters of the feature quantity input to the input layer 11 matches the number of elements of the weight vector.

  On the other hand, in the inspection mode, when the input data whose category is to be determined is given to the input layer 11 of the neural network 1, among the neurons N2 in the output layer 12, the neuron N2 having the smallest Euclidean distance between the weight vector and the input data is found. set a fire. If a category is associated with the neuron N2 of the output layer 12 in the learning mode, the category of the input data can be known from the category of the position of the fired neuron N2.

  The neuron N2 of the output layer 12 is associated one-to-one with each region of the two-dimensional clustering map 4 having, for example, 6 × 6 regions. Therefore, if the learning data category is associated with each region of the clustering map 4 in the learning mode, the category corresponding to the neuron N2 fired by the input data can be known from the clustering map 4. That is, the clustering map 4 can output the classification result by the neural network 1.

  When associating a category with each region of the clustering map 4 (substantially each neuron N2 of the output layer 12), the learned neural network 1 is operated in the opposite direction from the output layer 12 toward the input layer 11, and output. The data given to the input layer 11 is estimated for each neuron N2 of the layer 12, and the category of learning data having the closest Euclidean distance to the estimated data is used as the category of the neuron N2 in the output layer 12. In other words, as the category of each neuron N2 in the output layer 12, the category of learning data having the minimum Euclidean distance from the weight vector of each neuron N2 is used. Thereby, the category of learning data is reflected in the category of each neuron N2 of the output layer 12.

  In the present embodiment, categories classified by the neural network 1 are a specific abnormal state and a normal state, and one or more types of abnormal states are set as necessary. The category of the learning data is reflected in the category of each neuron N2 of the output layer 12, and when a large number (for example, 150) of learning data is given for one category, the category with high similarity is displayed on the clustering map 4. It is arranged at a close position. That is, the neuron N <b> 2 that fires the learning data having similar feature amounts is arranged at a close position on the clustering map 4.

  Accordingly, among the neurons N2 in the output layer 12, the neurons N2 that have fired corresponding to the learning data belonging to the same category gather at close positions on the clustering map 4 to form a cluster composed of the set of neurons N2. Note that the clustering map 4 having the original meaning is formed after the learning and the clusters are associated with each other. However, in the present embodiment, even before the learning, the clustering map 4 is called and is not particularly distinguished. The learning data given to the neural network 1 in the learning mode is stored in the learning data storage unit 6 and is read from the learning data storage unit 6 and given to the neural network 1 as necessary.

  By the way, the target signal to be classified by the neural network 1 is an electric signal obtained from the device X, for example, a microphone 2a for detecting the operation sound of the device X, and a vibration sensor 2b for detecting vibration generated during the operation of the device X. The output of the signal input unit 2 consisting of at least one of That is, the signal input unit 2 captures a target signal generated by the operation of the device X.

  The target signal, which is an electrical signal obtained by the signal input unit 2, is given to the feature extraction unit 3, and the feature amount of the target signal is extracted. In the present embodiment, the target signal given from the signal input unit 2 to the feature extraction unit 3 is a signal including a vibration component, and a plurality of parameters representing the vibration component of the target signal are input to the feature extraction unit 3. The feature quantity it has is extracted.

  The feature extraction unit 3 uses a timing signal (trigger signal) synchronized with the operation of the device X in order to extract a feature value from the target signal generated by the device X under the same conditions, or uses a waveform feature ( For example, the target signal is segmented by using a set of target signal start points and end points, and then divided into signals for each appropriate unit time, and feature quantities for each unit time To extract. Therefore, the feature extraction unit 3 includes a buffer that temporarily stores the target signal given from the signal input unit 2. Further, the feature extraction unit 3 performs preprocessing for reducing noise by limiting the frequency band as necessary. Furthermore, it also has a function of converting the target signal output from the signal input unit 2 into a digital signal.

  In order to simplify the explanation, it is assumed here that a plurality of frequency components (power for each frequency band) are extracted from the vibration components of the target signal after segmentation, and each frequency component is used as a parameter for the feature amount. explain. For the extraction of frequency components, an FFT (Fast Fourier Transform) technique or a filter bank made up of a large number of bandpass filters is used. Which frequency component is used for the feature amount is appropriately selected according to the target device X and the abnormality to be extracted.

  The feature amount obtained for each unit time from the feature extraction unit 3 is stored in the learning data storage unit 6 when learning data is collected before the learning mode, and each time the feature amount is extracted in the inspection mode. The neural network 1 classifies the category of the input data using the feature amount as input data.

  Here, a set sharing data among the data stored in the learning data storage unit 6 is called a data set, and each data included in one data set has a specific abnormal state to be detected. Assume that categories are associated. The number of data constituting the data set is arbitrary within a range that can be stored in the learning data storage unit 6. Further, the category of abnormal states may be one or more types.

  Since only the data set in which the category of the specific abnormal state is associated is stored in the learning data storage unit 6, learning is performed with each data set stored in the learning data storage unit 6 using the neural network 1 as a learning mode. Then, the neural network 1 learns about a specific abnormal state. That is, the category for other abnormal states is not set in the clustering map 4 (substantially, the output layer 12 of the neural network 1), and the category in the normal state is not set in the clustering map 4.

In this embodiment, after setting a category in the clustering map 4 using learning data and before providing input data to the neural network 1, the determination unit 5 uses a Gaussian for each neuron N2 in the output layer 12 of the neural network 1. Assign functions. The following equation is used for the Gaussian function. In the following formula, a character with square brackets means a vector.
y = exp (−‖ [x] − [m] ‖ ² / 2σ ² )
However, [m]: average, σ: variance.

  For the average [m] in the above equation, the weight vector of the learned neural network 1 is used. [X] in the above equation is input data given to the neural network 1, and ‖ [x] − [m] ‖ is the Euclidean distance with respect to the average [m] (that is, the weight vector) of [x] of the input data. Further, y is an output value when input data [x] is input to the Gauss function of the above equation.

  According to the above equation, the closer the input data [x] is to the average [m], the larger the output value y. When the input data [x] matches the average [m], the output value y is 1 at the maximum. become. This output value y is used as the degree of attribution. That is, the range of the output value y is 0 to 1, and the greater the numerical value (the closer to 1), the higher the degree of attribution.

  On the other hand, in order to obtain the variance σ, first, learning data is re-input to the learned neural network 1, and the Euclidean distance between the weight vector [m] of each neuron N2 and each input data [x] (that is, learning data). Ask for. All the learning data of the same category is given to the neural network 1, a list of Euclidean distances is created, and the maximum value of the Euclidean distance obtained for each neuron N2 may be used for dispersion.

  As described above, as shown in FIG. 3, a Gaussian function can be set for each neuron N2 corresponding to a specific abnormal condition category. FIG. 3 is a conceptual diagram showing a state in which three neurons N2 belong to a specific abnormal state category, and a state in which a Gaussian function is set for each neuron N2. No Gaussian function is set for the neuron N2 outside the category. When a Gaussian function is set in the determination unit 5, the degree of attribution can be obtained by giving the Euclidean distance between the input data and the weight vector to the Gaussian function. Therefore, after setting the Gaussian function, input data [x] (feature value) obtained during operation of the device is input, the weight vector (that is, average [m]) of each neuron N2 of the neural network 1 and the input data [x ] Is substituted into a Gaussian function for each neuron N2, and an output value y of the Gaussian function is obtained for each category associated with the neuron N2. Since the output value y is the degree of attribution of input data to each category as described above, if a threshold value for the degree of attribution is set, it can be determined whether or not the input data belongs to the category. The threshold value may be set as the degree of attribution corresponding to three times the variance.

  In the output layer 12 of the neural network 1 (that is, the clustering map 4), the number of neurons N2 belonging to one category is not limited to one, and one category may include a plurality of neurons N2. For each category, the average value of the output values y of the Gaussian functions may be used for the degree of attribution, or the maximum value of the output values y within the category may be used for the degree of attribution.

  In this embodiment, since the category is associated with a specific abnormal state, it is estimated that the degree of belonging to the category set in the clustering map 4 is equal to or less than the threshold is another abnormal state or a normal state. Is done. However, since this embodiment assumes that the possibility of another abnormal state is extremely low, when the degree of belonging to the category set in the clustering map 4 is equal to or less than the threshold value, it is substantially regarded as a normal state. be able to. That is, the determination unit 5 calculates the degree of belonging to each category by the above-described calculation, and outputs a determination result of a normal state when all the calculated degrees of belonging are less than or equal to the threshold value.

  Therefore, if there is no abnormality such as abnormal noise due to mixing of waste generated during manufacturing or abnormal noise due to lack of lubricating oil, it can be determined to be normal, and abnormalities can be detected when a load is applied during shipping inspection. If it is not normal, it is possible to determine whether there is an abnormality regardless of variations in the characteristic amount of the target signal that occurs even during normal operation. In addition, it is not necessary to prepare a large number of categories of learning data as in the case of performing a test with a load, and it is only necessary to collect learning data related to a specific abnormal state. Since it is relatively short and learning data is small, the time required for learning in the learning mode is also shortened.

  The configuration of the signal input unit 2 is appropriately selected according to the type of the device X, and various sensors such as a TV camera and an odor sensor can be used alone or in combination in addition to the microphone 2a and the vibration sensor 2b. Alternatively, a signal generated by the device X can be taken out and used as a target signal.

It is a block diagram which shows embodiment of this invention. It is a schematic block diagram of the neural network used for the same as the above. It is a conceptual diagram of the Gauss function in the same as the above.

Explanation of symbols

1 Neural network (competitive learning type neural network)
2 Signal Input Unit 2a Microphone 2b Vibration Sensor 3 Feature Extraction Unit 4 Clustering Map 5 Determination Unit 6 Learning Data Storage Unit 11 Input Layer 12 Output Layer N1 Neuron N2 Neuron X Device

Claims (4)

A signal input unit that captures a target signal including a vibration component generated during operation of the device, a feature extraction unit that extracts a feature amount including a plurality of parameters for the target signal, and a feature that has been learned in advance using learning data of a known category Comprising a competitive learning type neural network that uses the feature value extracted by the extraction unit as input data, and a determination unit that determines whether the device in operation is normal or abnormal using the output of the competitive learning type neural network, as learning data Using the feature value obtained when the device is in a known abnormal state, the determination unit sets the weight vector for each neuron in the output layer of the competitive learning type neural network using the learning data, and the learning data for the abnormal state category learning in a Gaussian function to determine the degree of membership of the input data to the firing neurons to a heavy real vector Set in association with the Euclidean distance between the over data, equipment normally when membership of the input data that Gaussian function is obtained during operation of the device for the configured neurons is a threshold hereinafter the provisions When the input data is input to the Gaussian function, the Gaussian function re-inputs the training data to the competitive neural network after learning. An anomaly detection apparatus characterized by determining a variance based on a Euclidean distance between a weight vector of each neuron that fires and the learning data, and averaging the weight vectors of fired neurons . Before SL threshold abnormality detecting device according to claim 1, wherein a is set to three times the variance in the Gaussian function. After capturing the target signal including the vibration component generated during the operation of the device and extracting the feature amount consisting of multiple parameters for the target signal, the feature amount obtained when the device is in a known abnormal state is used as learning data the learned in advance competitive learning neural network was, given the feature amount obtained during the operation of equipment as input data, determining by the determination unit equipment is normal or abnormal in operation using the output of the competitive learning neural network A Gaussian that determines the degree of attribution of input data to neurons fired by learning data in the abnormal category category by setting a weight vector for each neuron in the output layer of the competitive learning type neural network. set in association functions in the Euclidean distance between the heavy viewing vector and learning data, the Gaussian function Be one device is determined to be operating normally when the set degree of membership of the input data obtained during the operation of the equipment to neurons is a threshold or less under provisions, Gauss as degree of membership Using the output value when input data is input to the function, the Gaussian function is the Euclidean distance between the weight vector of each neuron that fires when the learning data is re-input to the competitive learning neural network after learning and the learning data. An anomaly detection method characterized by determining variance by means of averaging the weighted vectors of fired neurons . Before SL threshold abnormality detecting method according to claim 3, wherein a is set to three times the variance in the Gaussian function.

Images