JP4875433B2 - Learning method of signal identification device - Google Patents

Description

  The present invention relates to a signal identification device learning method for appropriately learning a competitive learning type neural network in a signal identification device that classifies target signals including vibration components using a competitive learning type neural network.

  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 used in an abnormality monitoring apparatus that recognizes voice and determines whether an apparatus is operating normally or whether an apparatus has an abnormality. For example, in an abnormality monitoring apparatus that employs this type of technology, the operation sound of a device or the vibration of a device is converted into an electrical signal by a sensor unit (transducer), and the output of the sensor unit is used as a 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 a learned competitive learning neural network, and the category of the cluster to which the fired neuron belongs in the clustering map is checked against the clustering map. Can be classified (see, for example, Patent Document 1).
JP 2004-354111 A

  By the way, since the competitive learning type neural network does not use a teacher signal, the accuracy of category classification in the inspection mode is determined by the accuracy of the category of learning data input in the learning mode. That is, it is necessary to employ appropriate learning data to be used in the learning mode, but there is no appropriate standard for selecting learning data. Therefore, one neuron of the output layer in the competitive learning type neural network may fire with learning data of a plurality of categories. In other words, unless learning is performed using appropriate learning data, there arises a problem that neurons that cannot determine the category when fired are included in the output layer.

  The present invention has been made in view of the above-mentioned reasons, and its purpose is to improve the classification accuracy of the category of the signal identification device by optimizing the learning data used for learning of the competitive learning type neural network. It is to provide a learning method.

According to the first aspect of the present invention, a signal input unit that captures a target signal including a vibration component, a feature extraction unit that extracts a feature amount including a plurality of parameters for the target signal, and a feature amount extracted by the feature extraction unit is used as input data. It is expected to be classified into a specific category, a competitive learning type neural network that classifies input data according to categories learned in advance using learning data, an output unit that outputs a classification result by each competitive learning type neural network, and A learning method of a competitive learning type neural network in a signal identification device comprising a learning data selection unit for selecting data to be adopted as learning data from a data set consisting of a plurality of data, the output layer in the competitive learning type neural network Each neuron has a set of weighting factors. Weight vector is associated to, learning data selecting unit, when the neurons which belong to a plurality of categories in each neuron of the output layer in the competitive learning neural network which has learned by the data of the data set, the data set the average value of the discrepancy between the data and those the neurons ignite the neuron with obtaining for each category of, deleting other data leaving only the data category the average value is the minimum data set A data selection process is performed, and the data selection process is repeated until all the neurons in the output layer in the competitive learning type neural network are classified into a single category.

  According to this method, the learning data is sorted so that all neurons in the output layer in the competitive learning type neural network are each classified into a single category. Can be specified, and an ambiguous result in which the category cannot be specified can be avoided to improve the classification accuracy of the category. In addition, since only the neurons that are fired by learning data of multiple categories among the neurons in the output layer are targeted, the degree of belonging to the category is evaluated by the degree of divergence. Can be granted.

  According to a second aspect of the present invention, in the first aspect of the present invention, the magnitude of a difference vector between a weight vector of a focused neuron and input data to the competitive learning type neural network is used as the degree of divergence. .

  According to this method, the learning data is quantitatively evaluated by evaluating the difference between the weight vector between the data input as the degree of divergence and the firing neuron, and adopting the data of the category in which the difference is small as the learning data. Can be optimized.

  According to the method of the present invention, learning data is set so as to classify all neurons of the output layer in the competitive learning type neural network into a single category, and there is an advantage that the classification accuracy of the category can be improved. . Further, there is an advantage that an appropriate category can be assigned to each neuron by quantitative evaluation.

  In the embodiment described below, an example in which the technique of the present invention is applied to an abnormality monitoring apparatus that determines whether an operation of a device is normal or abnormal based on a feature amount of a target signal obtained by the operation of the device will be described. That is, an abnormality monitoring device is exemplified as the signal identification device. Moreover, although the equipment apparatus provided with motive power sources, such as a motor, is assumed as an apparatus, the kind of apparatus in particular is not ask | required.

  As shown in FIG. 1, the signal identification apparatus described in this 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, the category of the feature quantity (input data) generated from the actual target signal in the inspection mode is selected. Classify.

  That is, the degree of connection (weighting coefficient) between the neuron N1 of the input layer 11 and the neuron N2 of the output layer 12 is variable, and in the learning mode, the learning data is input to the neural network 1 to learn the neural network 1, The weighting coefficient for each neuron N1 in the input layer 11 and each neuron N2 in the output layer 12 is 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 with each region of the two-dimensional clustering map 5 having, for example, 6 × 6 regions on a one-to-one basis. Therefore, if the learning data category is associated with each area of the clustering map 5 in the learning mode, the category corresponding to the neuron N2 fired by the input data can be known from the clustering map 5. It functions as an output unit that outputs the classification result by the neural network 1.

  Categories are associated with each region of the clustering map 5 (substantially, each neuron N2 of the output layer 12) in the procedure described below. Further, in this embodiment, for simplicity of explanation, the categories to be classified in the neural network 1 are two types, normal and abnormal, and in the learning mode, learning data of two types of categories, normal and abnormal, are stored. input. The category of each learning neuron N2 in the output layer 12 reflects the category of the learning data. When a large number (for example, 150) of learning data is given, the category having a high attribute similarity is close to the clustering map 5. Placed in.

  Therefore, the neurons N2 fired corresponding to the learning data belonging to the same category among the neurons N2 in the output layer 12 gather at close positions on the clustering map 5 to form a cluster composed of the set of neurons N2. The learning data given to the neural networks 1a and 1b 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 sensor unit 2 consisting of at least one of the above is used. The configuration of the sensor 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. Since the sensor unit 2 captures a target signal generated by the operation of the device X, it functions as a signal input unit.

  The target signal, which is an electrical signal obtained by the sensor 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 sensor unit 2 to the feature extraction unit 3 is a signal including a vibration component, and is input to the feature extraction unit 3 so as to represent a plurality of types of feature amounts representing the vibration component of the target signal. 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 a target signal given from the sensor 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 a target signal output from the sensor unit 2 into a digital signal.

  In order to simplify the description, 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 feature amount. . 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 quantity obtained from the feature extraction unit 3 per unit time is given to the neural network 1 every time the feature quantity is extracted. The feature amount is also stored in the learning data storage unit 6 for use as learning data. The learning data storage unit 6 is a FIFO (first-in first-out), and holds, for example, 150 feature values as learning data at the maximum.

  Below, the learning data selection part 4 which is the characteristic of this invention is demonstrated. As described above, a large number of pieces of learning data can be stored in the learning data storage unit 6, but the category associated with each piece of learning data is not always appropriate. Therefore, the learning data storage unit 6 evaluates the result when the data stored in the learning data storage unit 6 is applied to the neural network 1, and learning is performed from the data stored in the learning data storage unit 6 by the evaluation. Only appropriate learning data remains in the learning data storage unit 6 by selecting appropriate data.

  Here, a set of data stored in the learning data storage unit 6 is called a data set, and each data constituting the data set is classified into a specific category (that is, a category of normal and abnormal). Suppose that is expected. That is, it is assumed that a category is associated with each data. Although the data set is considered as the whole data stored in the learning data storage unit 6, a plurality of data sets may be stored in the learning data storage unit 6. Further, the number of data constituting the data set is arbitrary within a range that can be stored in the learning data storage unit 6.

  When one of the data sets stored in the learning data storage unit 6 is given to the neural network 1 in the learning mode, the neuron N2 in the output layer 12 is fired for each data. The learning data selection unit 4 stores the fired neuron N2 and the input data (information for identifying the data and the category of the data) as a set. After learning, the category of data fired for each neuron N2 of the output layer 12 is examined. If there is only one category of data fired for the neuron N2, that category is set as the category of the neuron N2.

  On the other hand, there may be a neuron N2 that has fired with data of different categories in the neuron N2. Since the neuron N2 fired with data of a plurality of categories cannot determine the category, the category is set to “gray”.

The learning data selection unit 4 obtains the degree of divergence between the weight vector associated with the neuron N2 and the input data for the “gray” neuron N2. The degree of divergence is a value reflecting the magnitude of the difference vector between the input data (vector) and the weight vector associated with the neuron N2 fired by the input data, and the input data is [X], and the fired neuron If the weight vector of N2 is [Wwin] ([a] means that a is a vector), the divergence degree Y is defined by the following equation.
Y = ([X] / X- [Wwin] / Wwin) ^T ([X] / X- [Wwin] / Wwin)
Here, T represents transposition, and X and Wwin without a square bracket indicate the norm of each vector. The divergence degree Y is normalized by dividing the elements of each vector by the norm.

  The neuron N2 whose category is “gray” is a neuron N2 that is fired by data of two types of categories, normal and abnormal. Therefore, the average value of the divergence degree Y is obtained for each category of data of normal and abnormal, and the category having the smaller average value of the divergence degree Y is adopted as the category of the neuron N2. At this time, data belonging to other categories is deleted from the data set.

  As described above, for the neuron N2 whose category is “gray”, the data that fired the neuron N2 is divided for each category, the average value of the divergence degree Y is obtained for each category, and the average value is smaller (3 The process of leaving only the data of the smallest (if there are more than categories) in the data set and deleting the other data that fired the “gray” neuron N2 will be referred to as a data selection process. The data selection process is performed for all the neurons “Gray” N2, and the process is repeated until all the neurons N2 of the output layer 12 in the neural network 1 are classified into a single category.

  That is, as shown in FIG. 3, the output layer 12 has 6 × 6 neurons N2 (equivalent to the clustering map 5), and normal (outlined part) and abnormal (blacked out part) are learned by the data set. It is assumed that it is divided into “gray” (shaded area). Now, assume that the data of the data set is represented in the format of (identification number-category), the identification number identifies the data with three digits, and the category has normal (OK) and abnormal (NG). For example, it is expressed in a format of 002-OK.

  Data in which the neuron N2 in the area where the category is normal (002-OK, 009-OK, 017-OK) is normal, and the data in which the neuron N2 in the area where the category is abnormal is fired Any of the categories (041-NG, 059-NG, 072-NG) becomes abnormal. However, data obtained by firing the “gray” neuron N2 includes both normal and abnormal data.

  Now, in the example of FIG. 3, the data of firing the neurons N2 in the fifth row from the top and the second column from the left are 001-OK, 005-OK, 010-OK, 037-OK, 051-NG, 070-. NG, and data obtained by firing the neuron N2 in the third row from the top and the fourth column from the left are 003-OK, 008-OK, 015-OK, 048-NG, 057-NG, 063-NG. Suppose.

  In this case, the divergence degree Y from the data that fired the neuron N2 is obtained by the above-described calculation, and the average value of the divergence degrees Y is obtained for each category of OK and NG. For example, if the average values of the divergence degrees Y obtained for the normal and abnormal category data for the neuron N2 in the fifth row from the top and the second column from the left are 0.83 and 0.25, respectively, Since .25 is smaller, data having normal categories (001-OK, 005-OK, 010-OK, 037-OK) are deleted from the data that fired this neuron N2. Similarly, the average values of the divergence degrees Y obtained for the normal and abnormal category data for the neuron N2 in the third row from the top and the fourth column from the left are 0.40 and 1.13, respectively. Then, data having an abnormal category (048-NG, 057-NG, 063-NG) is deleted.

  In this embodiment, since there are two types of categories, a single category is assigned to all neurons N2 in the output layer 12 if data selection processing is normally performed once for each neuron N2 of “gray”. Can be granted. Even in the case of classification into three or more categories, convergence is achieved if the data selection process is performed a finite number of times. In this way, it is possible to optimize the learning data so that one type of category can be associated with each neuron N2 of the output layer 12.

  By the above-described operation, for example, as shown in FIG. 4A, in the first data registered in the learning data storage unit 6, the region classified as “gray” (hatched portion) in the neuron N2 of the output layer 12 Even if is large, the learning data can be corrected so that the “gray” region disappears as shown in FIG. 4C while the above-described operation is repeated.

  The operation of the learning data selection unit 4 is summarized in FIG. In FIG. 5, steps S <b> 1 to S <b> 4 are preprocessing until the learning data selection unit 4 selects learning data. That is, a signal for use in learning from the device X or the like (not only the device X but also a signal generated by simulation) can be collected (S1), and the feature extraction unit 3 extracts the feature amount of the signal. (S2). The feature amount is stored in the learning data storage unit 6, and the neural network 1 is learned using the data set stored in the learning data storage unit 6 (S3). Further, a category associated with each neuron N2 by learning of the neural network 1 is given to the clustering map 5 (S4).

  By the way, while learning the neural network 1 in the learning mode, the learning data selection unit 4 stores the neuron N2 fired in the output layer 12 and the data that fired the neuron N2 in association with each other. A correspondence list between N2 and category is created (S5), and if there is no neuron N2 that has fired with data of multiple categories in the list (S6), a single category is assigned to all neurons N2 respectively. Therefore, the process by the learning data selection part 4 is complete | finished (S7).

  On the other hand, if there are “gray” neurons N2 fired with data of a plurality of categories in the list, one “gray” neuron is selected one by one (S8), and the data that fired the neuron N2 is displayed for each category. And an average value of the divergence degrees Y for each category is obtained (S9). Furthermore, only the data of the category with the smallest average divergence is left in the data set, and other data that fired the same neuron N2 is deleted from the data set (S10). That is, steps S7 to S9 are data selection processing.

  If there is another “gray” neuron N2 in the list, the data selection process is repeated, and after all the “gray” neurons N2 in the output layer 12 have been subjected to the data selection process (S11, S12), The data remaining in the learning data storage unit 6 is used as learning data, and the process returns to step S3 to learn the neural network 1 again.

  Thereafter, the same processing is repeated until there is no neuron N2 whose category is “gray”. In the above-described operation, a large number of data is stored in the learning data storage unit 6 in advance, and the data selection process is repeated so that the “gray” neuron N2 disappears. Instead, it is also possible to store a relatively small amount of data in the learning data storage unit 6 and add the data stored in the learning data storage unit 6 in a timely manner.

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 operation | movement explanatory drawing same as the above. It is operation | movement explanatory drawing same as the above. It is operation | movement explanatory drawing same as the above.

Explanation of symbols

1 Neural network (competitive learning type neural network)
2 Sensor unit 2a Microphone 2b Vibration sensor 3 Feature extraction unit 4 Learning data selection unit 5 Clustering map 6 Learning data storage unit 11 Input layer 12 Output layer N1, N2 neuron

Claims (2)

A signal input unit that captures a target signal including a vibration component, a feature extraction unit that extracts a feature amount including a plurality of parameters for the target signal, and learning in advance using learning data using the feature amount extracted by the feature extraction unit as input data Competitive learning type neural network that classifies input data according to the selected category, output unit that outputs the classification result by each competitive learning type neural network, and data that is expected to be classified into a specific category A learning method of a competitive learning type neural network in a signal identification device including a learning data selection unit that selects data to be adopted as learning data from a set, wherein each neuron of the output layer in the competitive learning type neural network has a weight A weight vector whose elements are coefficient pairs Is associated, the learning data selecting unit, when the neurons belonging in each neuron of the output layer in the competitive learning neural network which has learned by the data of the data set into a plurality of categories exist, the neurons of the data set the average value of the discrepancy between the data obtained by firing the person said neuron with obtaining for each category, the data sorting process of the average value is delete other data leaving only the data set categories of data is minimal A learning method for a signal identification device, characterized in that the data selection processing is repeated until all neurons in the output layer in the competitive learning type neural network are classified into a single category.   2. The signal identification device learning method according to claim 1, wherein a magnitude of a difference vector between a weight vector of a focused neuron and input data to the competitive learning type neural network is used as the divergence degree.

Images