CN115358274A

CN115358274A - DCGAN-CNN-based fault classification method

Info

Publication number: CN115358274A
Application number: CN202211046543.8A
Authority: CN
Inventors: 黑新宏; 张宽; 费蓉; 姬文江; 谢国; 高苗
Original assignee: Xian University of Technology
Current assignee: Xian University of Technology
Priority date: 2022-08-30
Filing date: 2022-08-30
Publication date: 2022-11-18

Abstract

The invention discloses a DCGAN-CNN-based fault classification method, which comprises the steps of firstly, extracting time-frequency characteristic information from a fault signal by using wavelet transform to form a new characteristic fault data set; then constructing a generator and a discriminator based on a convolutional network; inputting the characteristic fault data set into a DCGAN network model for training, and generating a synthesized fault data sample after training is finished; adding the synthesized fault data sample into a characteristic fault data set, and dividing a training set and a testing set; constructing a classifier based on the CNN, and training a network by using a training set to obtain a CNN classification model; and testing the classification model by using the test set, and performing model evaluation through accuracy, precision, recall rate and F1-Score evaluation indexes. The invention enhances the data by combining the wavelet transformation with the DCGAN to improve the fault diagnosis accuracy under the condition of data imbalance.

Description

DCGAN-CNN-based fault classification method

Technical Field

The invention belongs to the technical field of fault diagnosis, and particularly relates to a fault classification method based on DCGAN-CNN.

Background

The modern industrial process system has irreplaceable effect in promoting social progress and development. Along with the improvement of the science and technology level, industrial equipment is increasingly complicated, informationized and intelligentized, and a fault occurs in equipment, so that other parts can be influenced to a certain extent, and therefore, the safety and the reliability of machinery can be effectively improved by accurately diagnosing the fault. In recent years, the intelligent fault diagnosis technology based on deep learning develops rapidly, and reliable guarantee is provided for normal operation of industrial machines. Compared with the traditional method for manually selecting the features, the fault diagnosis framework based on deep learning has strong feature learning capability, and can automatically select the distinguishing features beneficial to accurate classification. Due to the particularity of deep learning training, the amount of data required is large, but in practice training samples between different machine operating states are often unbalanced. For a mechanical system in operation, which works under normal conditions most of the time, sufficient normal samples are collected, but the frequency of mechanical failure is low, the number of collected samples is limited, and therefore an imbalance exists between normal and failed samples.

For fault sample classification of unbalanced data, generating a countermeasure network (GAN) can solve the data imbalance problem by generating a small number of class samples. As a generation model, GAN and its derivative models are commonly used for generating samples for data enhancement and data preprocessing methods in deep learning, and have wide application scenarios in the fields of image processing, biomedicine, network, information security, and the like. The invention adopts wavelet transformation to extract the fault signal characteristics, generates a countermeasure network (DCGAN) based on the variant-depth convolution generated countermeasure network to solve the problem of data imbalance, and constructs a classifier by combining with the CNN to realize the fault classification task.

Disclosure of Invention

The invention aims to provide a DCGAN-CNN-based fault classification method, which can improve the fault diagnosis accuracy under data imbalance.

The technical scheme adopted by the invention is that a DCGAN-CNN fault classification method is implemented according to the following steps:

step 1, extracting time-frequency characteristic information from a fault signal by using wavelet transform to form a new characteristic fault data set;

step 2, constructing a DCGAN network model: constructing a generator and a discriminator based on a convolutional network;

step 3, inputting the characteristic fault data set in the step 1 into a DCGAN network model for training, and generating a synthesized fault data sample after training is finished;

step 4, adding the synthesized fault data sample obtained in the step 3 into the characteristic fault data set obtained in the step 1, and dividing a training set and a testing set;

step 5, constructing a classifier based on the CNN, and training the network by using a training set to obtain a CNN classification model;

and 6, testing the classification model obtained in the step 5 by using the test set in the step 4, and performing model evaluation according to the accuracy, precision, recall rate and F1-Score evaluation index.

The present invention is also characterized in that,

the step 1 is implemented according to the following steps:

step 1.1, selecting a wavelet function and a scale a, and aligning the wavelet function with a starting point of a signal to be analyzed;

step 1.2, starting from the signal initial position, calculating the approximation degree of the signal to be analyzed and the wavelet function at the moment, namely calculating to obtain the wavelet coefficient by using a formula (1), wherein the larger the wavelet coefficient is, the closer the waveform of the signal and the selected wavelet function at the moment is represented;

step 1.3, shifting the wavelet function to the right for a unit time along a time axis, namely scanning along the time axis, and repeating the step 1.2 to obtain the wavelet coefficient at the moment until the whole length of the signal to be analyzed is scanned;

step 1.4, changing the scale a, and repeating the steps 1.2-1.3 to complete the scanning of the frequency axis;

and respectively completing the time domain and frequency domain feature analysis of the fault signal based on the step 1.3 and the step 1.4.

The DCGAN network model in step 2 consists of two modules: the generator network G and the discriminator network D, the task of the generator is to receive the noise z distributed randomly, make the sample G (z) synthesized accord with the true sample distribution; the task of the discriminator is to receive the data G (z) of the generator and the real sample data x and to discriminate between true and false of the received data.

Step 2 the network construction process is specifically implemented according to the following steps:

step 2.1, constructing a generator network G: the input of the generator network G is random noise z, the random noise z is converted into the size of a real sample through three-layer deconvolution operation, batch normalization is carried out among the three-layer deconvolution operation, and the output of the generator is an image with the same size as the data set in the step 1;

step 2.2, constructing a discriminator network D: the input of the discriminator network D is the real sample data x in the step 1 and the synthetic sample data G (z) output by the generator G in the step 2, the discriminator network D passes the input data through three convolution layers, and finally the output layer calculates the probability of whether the input image is real by using a sigmoid activation function.

Step 3 is specifically implemented according to the following steps:

the loss function of DCGAN is given by the following formula (2):

wherein G, D represent the generator and the discriminator, x-p _data (x) Represents the distribution of real data, z-p _z (z) represents a random noise distribution, D (x) represents an output result of the real data passing through the discriminator, and D (G (z)) represents an output result of the generator synthetic data G (z) passing through the discriminator;

optimizing the DCGAN network model based on the formula (2), specifically as follows:

step 3.1, updating the parameters of the generator network by minimizing the loss of the formula (2), so that the output of the generator network is more similar to the real data;

step 3.2, updating network parameters of the discriminator by maximizing the loss result of the formula (2), so that the discriminator network can more accurately distinguish real data from synthetic data;

the countermeasure training is carried out through the step 3.1 and the step 3.2, the DCGAN network parameters are updated, the generator and the discriminator reach a relative balance state, the trained DCGAN can be used for generating data, and the problem of unbalanced original data is solved.

Step 5 is specifically implemented according to the following steps:

and (3) adopting the CNN as a classifier, taking the data in the step (4) as input, forming a CNN network by using 3 convolution kernels of 3 × 3 convolution layers, two 2 × 2 maximum pooling layers, a full connection layer and a softmax output layer, and training a CNN network model by using a training set to obtain a CNN classification model.

The invention has the advantages that the DCGAN-CNN-based fault classification method improves the problems of gradient disappearance, mode collapse and the like of a basic GAN model, simultaneously uses wavelet transformation, better utilizes time domain and frequency domain information of fault signals, extracts time-frequency image characteristics to facilitate DCGAN processing, effectively expands fault data and improves classification accuracy.

Drawings

FIG. 1 is an overall flow chart of a DCGAN-CNN-based fault classification method according to the present invention.

Detailed Description

The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

The invention relates to a DCGAN-CNN fault classification based method, a flow chart is shown in figure 1, and the method is implemented according to the following steps:

the step 1 is implemented according to the following steps:

wavelet transform is a time domain-frequency domain analysis method of signals, and has the capability of representing local characteristics of the signals in both time domain and frequency domain. The meaning is that after the basic wavelet function is shifted by tau, the basic wavelet function and the signal to be analyzed x (t) are subjected to inner product under different scales a, namely

Wherein WT (a, τ) represents wavelet coefficients; a is more than 0 and is called as a scale factor, the expansion and contraction of the wavelet function are controlled, and the frequency of the corresponding signal is controlled; and tau represents displacement, controls the translation of the wavelet function and corresponds to the time information of the signal.

step 1.2, starting from the initial position of the signal, calculating the approximation degree of the signal to be analyzed and the wavelet function at the moment, namely calculating to obtain the wavelet coefficient by using a formula (1), wherein the larger the wavelet coefficient is, the closer the waveform of the signal and the selected wavelet function at the moment is represented;

the DCGAN network model in step 2 consists of two modules: the generator network G and the discriminator network D, the task of the generator is to receive the noise z distributed randomly, make the sample G (z) synthesized accord with the true sample distribution; the task of the discriminator is to receive the data G (z) of the generator and the real sample data x and to discriminate the authenticity of the received data.

Step 2, the network construction process is implemented according to the following steps:

step 2.1, constructing a generator network G: the input of the generator network G is random noise z, the random noise z is converted into the size of a real sample through three layers of deconvolution operations, batch normalization is carried out among the three layers of deconvolution operations, and the output of the generator is an image with the same size as the data set in the step 1; the method is beneficial to processing the training problem caused by poor initialization, and meanwhile, the model training is accelerated, so that the training stability is improved;

step 3 is specifically implemented according to the following steps:

the loss function of DCGAN is given by the following formula (2):

step 3.1, updating generator network parameters through minimizing the loss of the formula (2), so that the output of the generator network is more similar to real data;

the countermeasure training is carried out through the step 3.1 and the step 3.2, the DCGAN network parameters are updated, so that the generator and the discriminator reach a relative balance state, the trained DCGAN can be used for generating data, and the problem of original data imbalance is solved.

step 5 is specifically implemented according to the following steps:

Claims

1. A DCGAN-CNN fault classification based method is characterized by being implemented according to the following steps:

2. The method for fault classification of DCGAN-CNN according to claim 1, wherein the step 1 is specifically implemented according to the following steps:

3. The method according to claim 2, wherein the DCGAN network model in step 2 comprises two modules: the generator network G and the discriminator network D, the task of the generator is to receive the noise z distributed randomly, make the sample G (z) synthesized accord with the true sample distribution; the task of the discriminator is to receive the data G (z) of the generator and the real sample data x and to discriminate the authenticity of the received data.

4. The method for fault classification based on DCGAN-CNN according to claim 3, wherein the step 2 network construction process is specifically implemented according to the following steps:

step 2.2, constructing a discriminator network D: and (3) inputting the real sample data x in the step (1) and the synthetic sample data G (z) output by the generator G in the step (2) by the discriminator network D, passing the input data through three convolution layers by the discriminator network D, and finally calculating the probability of whether the input image is real or not by using a sigmoid activation function by the output layer.

5. The method for fault classification of DCGAN-CNN according to claim 4, wherein the step 3 is specifically implemented according to the following steps:

the loss function of DCGAN is given by the following formula (2):

wherein G and D represent a generator and a discriminator, respectively, and x to p _data (x) Represents the distribution of real data, z-p _z (z) represents a random noise distribution, D (x) represents an output result of the real data passing through the discriminator, and D (G (z)) represents an output result of the generator synthetic data G (z) passing through the discriminator;

based on the above formula (2), the DCGAN network model is optimized, specifically as follows:

6. The method for fault classification of DCGAN-CNN according to claim 5, wherein the step 5 is specifically implemented according to the following steps:

and (3) adopting the CNN as a classifier, taking the data in the step (4) as input, forming a CNN network by using a convolution layer with 3 convolution kernels of 3 x 3, two maximum pooling layers of 2 x 2, a full connection layer and a softmax output layer, and training a CNN network model by using a training set to obtain a CNN classification model.