A fault diagnosis method based on the integration of mechanism simulation and data drive

Abstract

The present invention discloses a fault diagnosis method based on the integration of mechanism simulation and data drive. Aiming at the small sample and imbalance problems in fault diagnosis, the fault mechanism is studied to construct a fault simulation formula to simulate fault data. At the same time, similar data is generated by using a generative adversarial network to expand and balance the sample size and match the convolutional neural network to achieve fault classification. The method includes the following steps: (1) selecting the driving end data in the bearing acceleration signal as experimental analysis data; (2) truncating each type of data as a sample during data preprocessing; (3) setting a training set and a test set and simulating an unbalanced data set; (4) studying the fault mechanism, constructing a simulation formula according to the fault mechanism, referring to the real test bench parameters and combining the time domain diagram and the envelope spectrum to determine the parameters in the simulation formula; (5) using a generative adversarial network to generate signals of three fault states and mix them with the original samples into a balanced data set; (6) generating sample quality evaluation.

Description

Technical Field

[0001] The invention relates to a fault diagnosis method, in particular to a fault diagnosis method combining fault mechanism simulation with a generative adversarial network and a convolutional neural network, and belongs to the technical field of bearing fault diagnosis. Background Art

[0002] Rolling bearings are important structural components and parts in large rotating machinery. They have complex working conditions and extremely harsh environments, which leads to frequent faults and more serious consequences. Therefore, the fault diagnosis of rolling bearings is of indispensable practical significance, but also has certain difficulties. Common fault diagnosis methods can be roughly divided into two categories. One category mainly adopts traditional machine learning methods, such as support vector machines (SVM), decision trees, and naive Bayes. The other category is intelligent diagnosis and processing methods using deep learning theoretical models, such as deep belief networks (DBNs), recurrent neural networks (RNNs), and convolutional neural networks (CNNs). Among them, convolutional neural networks have received a lot of attention because they can achieve data dimensionality reduction and have a powerful ability to automatically extract features.

[0003] Convolutional neural networks and other data-driven methods usually require a large amount of process operation sample data when training models. However, in actual production, fault signals are difficult to obtain, resulting in insufficient and unbalanced samples. Therefore, research on small and unbalanced samples has become increasingly important. Conventional diagnostic methods are mainly divided into two categories: the first type of method is to improve the sensitivity of the corresponding network to unbalanced samples from the perspective of improving network structure performance and improving algorithms, thereby improving the accuracy of fault classification. The second type of method focuses on the data level. Through certain data preprocessing methods, the unbalanced data set is enhanced, and then expanded to a balanced data set for classification. The commonly used preprocessing methods are mainly oversampling and downsampling. Generative adversarial networks mainly use generators to learn the distribution of data, and then generate data that is similar to but different from real data, thereby achieving data enhancement. Although the history of generative adversarial networks is relatively short, it has been used to solve imbalance problems. In addition, if the fault mechanism can be studied and the simulated signal of the fault can be obtained through simulation, data enhancement in the case of unbalanced samples or even missing data can be achieved, the imbalance problem can also be solved. The data-driven fault diagnosis method generally processes and analyzes the data to obtain the ability to mine data knowledge, but does not consider the information knowledge of the internal structure and mechanism. Therefore, it is relatively difficult to analyze and explain the fault, and the generalization ability of the model is not strong. Therefore, for the situation of unbalanced samples or even missing data, if the fault mechanism can be combined with data-driven, the fault data can be enhanced, and then the mapping between the fault mechanism and feature extraction can be realized, so as to achieve a higher fault diagnosis accuracy under unbalanced samples or even when the fault data is missing. This has very important practical significance, but it is also a relatively difficult problem, and there is no particularly obvious breakthrough at present.

[0004] In response to the above problems, this paper proposes a fault diagnosis method based on the fusion of mechanism simulation and data-driven. The generative adversarial network is used to directly generate a one-dimensional time domain signal instead of a two-dimensional image, which can reduce the loss of information in the signal, thereby completing the expansion of the data set and achieving data enhancement. In addition, it is proposed to construct a simulation formula to simulate the fault data by studying the fault mechanism, and at the same time, similar samples are generated through GAN to expand the data set. Finally, a convolutional neural network is composed of three convolutional layers and pooling layers for fault classification and identification, achieving a high fault classification accuracy in unbalanced data sets or even in scenarios without training fault data. Summary of the invention

[0005] The purpose of the present invention is to solve the problem that fault classification cannot be performed or the classification accuracy is low in scenarios with small bearing fault samples, unbalanced data sets or even missing fault data.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is a generative adversarial network sample generation method and a deep learning diagnosis method based on the integration of mechanism simulation and data drive. By studying the mechanism simulation fault formula to generate fault data, similar samples are generated through the generative adversarial network to enhance the data set. At the same time, the hidden information in the sample data is fully extracted through the convolutional neural network, thereby effectively improving the classification accuracy.

[0007] This method mainly includes three parts: generating similar samples by generative adversarial network, simulating fault signals based on fault mechanism, and realizing fault classification by convolutional neural network.

[0008] 1. Generate similar samples using Generative Adversarial Networks (GANs)

[0009] GAN consists of two parts: a generator G and a discriminator D. The steps of generating data in a generative adversarial network are as follows:

[0010] (1) The bearing signal under different states is truncated at certain intervals as a sample, i.e., the real sample x. Usually, the number of interval points needs to be greater than the number of sampling points when the bearing rotates one circle.

[0011] (2) A random noise sample of length 128 is input into the generator G, and the generator G outputs a generated sample G(z) that is similar to the real sample but not the same.

[0012] (3) The real sample x and the generated sample G(z) are input to D independently at the same time. D(x) is the output of D, that is, the probability that the input sample is a real sample. The label of the real sample x is set to 1, and the label of the generated sample G(z) is set to 0. When x is input, the discriminator D aims to make the value of D(x) equal to 1, that is, to judge it as true. When the input is the generated sample G(z), the discriminator hopes to judge it as false, that is, the output value D(G(z)) is equal to 0. However, at this time, the generator G hopes that its generated data can be indistinguishable from the real, so that the discriminator D(G(z)) = 1, that is, it is judged as true, thus forming an adversarial game. The loss function of GAN is shown in formula (1), and the optimization objective function is shown in formula (2), which is a minimax problem.

[0013]

[0014] Where: z is random noise, P z(z) is the prior distribution, z is the input of the generator G, G(z) is the generated sample, that is, the output of the generator, x is the real sample, and p data(x) is the distribution of real samples, D(G(z)) is the output of the discriminator, which is a probability value between [0,1].

[0015] (4) During the optimization training process, the two are trained alternately. When one is trained, the parameters of the other are fixed first. The optimization of the discriminator D is expected to correctly distinguish true and false samples. When the real sample x is input, D is optimized so that the value of D(x) changes from 0 to 1. When the generated sample G(z) is input, D is optimized so that the value of D(G(z)) changes from 1 to 0. Therefore, the optimization objective function of D is a maximization problem as shown in formula (3).

[0016]

[0017] The optimization of the generator G hopes to generate real-looking data G(z) to deceive D into believing it is real. Its optimization objective function is shown in formula (4), that is, the goal is to make the value of D(G(z)) change from 0 to 1, which becomes a minimization problem.

[0018]

[0019] The training is repeated many times until the discriminator cannot distinguish whether the input is a generated sample or a real sample. At this time, D(G(z)) = 0.5, and Nash equilibrium is reached. At this time, the training ends and samples that are very similar to but different from the real samples are obtained.

[0020] (5) Set the maximum number of network iterations to 25,000 and the threshold range of the discriminator D output value to 0.5-0.501. When the output probability D(G(z)) obtained in each iteration is within this range, the requirement is met and the generated sample is output.

[0021] 2. Simulate fault signals based on fault mechanisms

[0022] Failure mechanism refers to the law of mutual correlation or mapping between the signal characteristics of equipment failure status and the parameter characteristics of the failure system, obtained through system theory or a combination of a large number of field experiments and analysis. The vibration signal caused by bearing failure will continuously collide with the fault part with periodic rotation, thus showing periodic impact vibration, which generally shows exponentially decaying sinusoidal signal characteristics, and strong background noise or other interference components must also be considered.

[0023] In the construction of simulation models through mechanism research, there are mainly two categories:

[0024] 1) Constructing nonlinear dynamics model

[0025] 2) Construct bearing fault simulation formula

[0026] This paper adopts the second method to construct the bearing fault simulation formula. Therefore, the rolling bearing fault signal can be obtained by simulating the vibration response caused by the impact. The specific formula is as follows:

[0027] x(t)=s(t)+v(t) (5)

[0028] The simulation signal x(t) is composed of periodic shock s(t) and random noise v(t). x(t) is the bearing fault simulation signal. The fault form is a unilateral attenuation pulse. The expression for constructing the bearing fault simulation signal is as follows:

[0029]

[0030] Among them, A is the displacement coefficient, that is, the impact amplitude, f n is the natural frequency, ε is the damping coefficient, φ is the initial phase angle, T K represents the trigger time of the kth pulse, v(t) is random noise, and N is the number of sampling points. According to the actual rolling bearing fault diagnosis test bench parameters and combined with the time domain diagram and envelope spectrum diagram, the parameters of formula (6) are configured as follows:

[0031] Impact amplitude A=round(max(x)), x is the real signal, max is the maximum value function, and round is rounded;

[0032] The natural frequency of a general bearing failure is generally between 1KHz and 60KHz. The natural frequency f is adjusted by 1KHz each time. n , until max(abs(fft(x)))=max(abs(fft(y))) is satisfied, at which time f n That is what we need, where y is the simulation signal, fft is the fast Fourier transform, and abs is the absolute value function;

[0033] The initial phase angle φ = 0, and the size of v(t) can be set as long as the fault characteristic frequency in the envelope spectrum of the simulation signal is not submerged. z The calculation of the outer ring fault is as shown in formula (7):

[0034]

[0035] Among them, r is the actual speed of the rolling bearing, n is the number of balls, d is the rolling element diameter, D is the bearing pitch diameter, and α is the rolling element contact angle.

[0036] The value range of the damping coefficient ε is [0,1]. The optimal damping coefficient ε is selected by calculating the correlation between the simulated signal y and the real signal x. The correlation is evaluated by the Pearson correlation coefficient as shown in formula (8).

[0037]

[0038] Among them, cov is the covariance of the two, σ is the standard deviation, |ρ|∈[0,1], the larger the absolute value of the correlation coefficient, the stronger the correlation. The ε is initialized to 0, and each step is 0.001 until the value is 1. The ε with the largest correlation coefficient is the desired one.

[0039] 3. Convolutional Neural Network Model (CNN) for Fault Classification

[0040] The convolutional neural network model discussed in this article consists of the following eight layers: an input layer, three convolutional layers, and pooling layers connected alternately, and the last layer is the output layer. Each pooling layer adopts the maximum pooling method to reduce the influence of useless information and improve the calculation speed. The main training and steps of CNN are as follows:

[0041] (1) In the Keras framework of Python software, the proposed generative adversarial network sample generation method based on the fusion of mechanism simulation and data-driven is used to construct a training set as the input data of the convolutional neural network.

[0042] (2) Different labels are artificially defined as the output of the network according to different faults. During training, for normal state, rolling element, inner race, and outer race faults, each sample in each category is labeled with the numbers 0, 1, 2, and 3 at the end of the data, i.e., the true label. The output of the network is also the predicted label represented by numbers and their respective probabilities. They are compared and sorted according to the probability distribution values ​​of each label. The label with the highest probability is the predicted label, i.e., the predicted classification result.

[0043] (3) Set the number of samples for initial batch processing, the parameters of the convolution layer, the parameters of the pooling layer, etc.

[0044] (4) Adjustment of network parameters

[0045] After a training session, fine-tune the number of batch samples, the number of convolution kernels, and the size of the pooling layer according to the classification accuracy of the network output, and train again. When the classification accuracy of the network no longer improves, save the model and related parameters for testing.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] (1) Using a generative adversarial network to generate one-dimensional sample data instead of two-dimensional images reduces the loss of hidden information in the signal. At the same time, compared with the commonly used oversampling and downsampling sample production methods, it increases the diversity of samples and reduces the risk of overfitting.

[0048] (2) Study the fault mechanism. According to the fault mechanism, analyze the characteristics and components of the vibration signal exhibited by the bearing fault, and construct a simulation formula for the fault signal. Compared with building a nonlinear dynamic model, it has fewer parameters, faster calculation speed, and is more concise and convenient. At the same time, it can effectively replace the missing fault data, thereby realizing fault diagnosis without training fault data.

[0049] (3) The mechanism simulation is integrated with the data-driven generative adversarial network to realize the effective generation of data in small sample and unbalanced data sets or even in the absence of training data. At the same time, the mapping of fault mechanism and feature extraction is realized, thereby effectively improving the fault classification accuracy, which helps to solve the small sample and imbalance problems in the actual fault diagnosis engineering. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a flow chart of the fault diagnosis method based on mechanism simulation and data-driven fusion of the present invention.

[0051] Figure 2 This is the time domain waveform diagram of the bearing outer ring fault signal simulated by the mechanism in the present invention

[0052] Figure 3 This is the envelope spectrum of the bearing outer ring fault signal simulated by the mechanism in the present invention.

[0053] Figure 4 This is the model structure diagram of the generative adversarial network in the present invention.

[0054] Figure 5 This is a time domain waveform comparison diagram of the signal generated by the generative adversarial network in the present invention and the original signal

[0055] Figure 6 This is a comparison diagram of the envelope spectrum of the signal generated by the generative adversarial network in the present invention and the original signal

[0056] Figure 7 This is the model structure diagram of the convolutional neural network in the present invention.

[0057] Figure 8 is the fault classification confusion matrix when the proposed method is not applied

[0058] Fig. 9 is the fault classification confusion matrix when the proposed method is not applied DETAILED DESCRIPTION

[0059] In order to verify the effectiveness of this method, this paper uses the public data set of Case Western Reserve University (CWRU). The driving end data of bearing 6205-2RS JEM SKF with a speed of 1730r / min and a sampling frequency of 48k Hz is selected as the experimental analysis data. There are three types of bearing faults: rolling element, inner ring, outer ring faults, plus normal signals, a total of four types of data.

[0060] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0061] Figure 1 The following is a flow chart of the fault diagnosis method based on mechanism simulation and data-driven fusion of the present invention, and the process includes:

[0062] (1) The bearing drive end data was selected as the experimental analysis data, and during data preprocessing, each type of data was truncated at intervals of 416 data points as a sample;

[0063] (2) Set up training sets and test sets and simulate unbalanced data sets: In the training set, the normal state of the bearing, rolling element fault, and inner ring fault are all real data, and the outer ring fault is simulated by the mechanism simulation in this paper. This paper assumes that the number of samples after the data of each type is balanced is 100. Set up 100 samples of normal state in the training set, 50 samples of rolling element fault, 20 samples of inner ring fault, and 50 samples of outer ring fault simulated by mechanism simulation. In the test set, all four states are tested with real data, and each type has 40 samples;

[0064] (3) Bearing fault mechanism simulation formula simulation: This paper takes the acceleration simulation signal of the outer ring fault of the rolling bearing as an example for construction analysis. According to the parameter configuration method mentioned above, the parameters are set one by one, and the parameters are fine-tuned in combination with the time domain diagram and the envelope spectrum diagram. The various parameters in the simulation formula (6) are finally set to: A = 5, f n =45KHz, ε=0.004, φ=0rad, the theoretical value of the characteristic frequency of the outer ring fault is 103.4Hz. In order to simulate the bearing working condition, a Gaussian white noise component of v(t)=9db is added. The time domain waveform of the outer ring fault simulation signal is shown in Figure 2 As shown, the envelope spectrum analysis is performed on it. Figure 3 As shown in the figure, through the envelope spectrum, it can be found that the characteristic frequency of the outer race fault and its frequency multiple components are relatively obvious;

[0065] (4) GAN-generated samples are expanded to a balanced dataset: The GAN model structure diagram in this article is as follows: Figure 4As shown. The generator and discriminator are both composed of multi-layer perceptrons, that is, each layer is composed of a fully connected layer. The generator consists of a total of four layers of networks. The input layer is a fully connected layer composed of 128 neurons, and the input data is a random noise of length 128. There are two hidden layers, both of which are fully connected layers composed of 128 neurons. The output layer is a fully connected layer composed of 416 neurons, and the output is a generated sample of length 416. After each layer, an activation function is used for nonlinear mapping to the next layer. This paper uses the RELU function. The input layer in the discriminator is the same as the output layer in the generator. The input is a real sample and a generated sample of length 416. The middle layer is a fully connected layer with 128 neurons. The output layer consists of one neuron, and the output is a probability value between 0 and 1. When the output value is 0.5 or closer to 0.5, it means that the generated sample is more similar to the real sample but still different. Set the maximum number of network iterations to 25,000, and the output value threshold range to 0.5-0.501. When the output probability obtained in each iteration is within this range, the required output generation sample is met;

[0066] (5) Generated sample quality evaluation: Taking the bearing outer ring fault signal as an example, the time domain waveform comparison diagram and envelope spectrum diagram of the original signal and the generated signal are respectively made, as shown in the figure below: Figure 5 and Figure 6 As shown. Figure 5 The time domain waveform comparison shows that the signals are very similar but still different. Figure 6 The envelope spectrum comparison shows that the characteristic frequency and its frequency multiple components of the generated outer race fault signal are quite obvious;

[0067] (6) Mixing the original real samples with the generated samples: Generate each type of sample in the training set using the method in step (4), and mix them with the original real samples until there are 100 samples to achieve a balanced data set;

[0068] (7) Use CNN to conduct classification experiments: The model structure diagram of CNN is as follows: Figure 7As shown. The input is a sample consisting of a vibration signal and a label. The first convolution layer is equipped with 32 3*1 convolution kernels, and the step size of the first pooling layer is 3*1. The settings of the convolution layer and the pooling layer in the second and third layers are the same, with 128 5*1 convolution kernels and a step size of 3*1 in the pooling layer. The fully connected layer consists of 256 nodes, and the output layer is the sample label. The network iterates about eighty times until the training accuracy no longer changes, and the model is saved for testing. The test data and training data are not repeated. All the test data, a total of 160 samples, are input into the saved model, and the corresponding 160 predicted labels are output. The predicted labels and the true labels are saved to draw the classification confusion matrix. Using the above steps, for the unbalanced data set, the method proposed in this article is used to expand it to the balanced data set to conduct classification test experiments, and confusion matrices are drawn to verify the effectiveness of the method proposed in this article. As shown Figure 8 It is the confusion matrix of the original unbalanced dataset when the method in this paper is not applied. Fig. 9 This is the confusion matrix when the proposed method is expanded to a balanced data set. Comparing the two figures, it can be seen that the accuracy of the inner circle fault, i.e., label 2, is increased from 62% to 83% after expansion, and the accuracy of the outer circle fault simulation signal, i.e., label 3, is increased from 87% to 94% after expansion. This fully proves that the fault diagnosis method based on the fusion of mechanism simulation and data-driven can generate effective simulation signals, realize data expansion in small sample and unbalanced scenarios, thereby effectively improving the fault classification accuracy, which helps to solve the small sample and unbalanced problems in the actual fault diagnosis engineering.

Claims

1. A fault diagnosis method based on the integration of mechanism simulation and data drive, characterized in that: Training a generative adversarial network to generate one-dimensional sample data, the steps are as follows: (1) The bearing acceleration signal under different states is truncated at a certain interval as a sample, i.e., the real sample x. The number of interval points needs to be greater than the number of sampling points for one rotation of the bearing; (2) A random noise sample of length 128 is input into the generator G, and the generator G outputs a generated sample G(z) that is similar to the real sample but not the same; (3) The real sample x and the generated sample G(z) are input to the discriminator D independently at the same time. D(x) is the output of the discriminator D, that is, the probability that the input sample is a real sample. The label of the real sample x is set to 1, and the label of the generated sample G(z) is set to 0. When x is input, the discriminator D aims to make the value of D(x) equal to 1, that is, it is judged to be true. When the input is the generated sample G(z), the discriminator hopes to judge it as false, that is, the output value D(G(z)) is equal to 0, that is, it is judged to be false. However, at this time, the generator G hopes that its generated data can be indistinguishable from the real so that the discriminator D(G(z)) = 1, that is, it is judged to be true, thus forming an adversarial game. The loss function of GAN is shown in formula (1), and the optimization objective function is shown in formula (2), which is a minimax problem. Where: z is random noise and is also the input of the generator G, P z(z) is the prior distribution, G(z) is the output of the generator, i.e., the generated sample, x is the real sample, and p data(x) is the distribution of real samples, D(G(z)) is the output of the discriminator, which is a probability value between [0,1]; (4) In the process of optimization training, the two are trained alternately. When one is trained, the parameters of the other are fixed first. The optimization of the discriminator D is expected to correctly distinguish true and false samples. When the real sample x is input, D is optimized so that the value of D(x) changes from 0 to 1; when the generated sample G(z) is input, D is optimized so that the value of D(G(z)) changes from 1 to 0. Therefore, the optimization objective function of D is a maximization problem as shown in formula (3). The optimization of the generator G is to generate fake data G(z) that is indistinguishable from real data, thereby deceiving D and making D believe it. Its optimization objective function is shown in formula (4), that is, the goal is to make the value of D(G(z)) change from 0 to 1, thus evolving into a minimization problem; The training is repeated many times until the discriminator cannot distinguish whether the input is a generated sample or a real sample. At this time, D(G(z)) = 0.5, and Nash equilibrium is reached. At this time, the training ends and samples that are very similar to but different from the real samples are obtained. (5) Set the maximum number of network iterations to 25,000, and the threshold range of the discriminator D output value to 0.5-0.

501. When the output probability D(G(z)) obtained in each iteration is within this range, the requirement is met and the generated sample is output; The expression for constructing the bearing fault simulation signal is as follows: Among them, A is the displacement coefficient, that is, the impact amplitude, f n is the natural frequency, ε is the damping coefficient, φ is the initial phase angle, T K represents the trigger time of the kth pulse, v(t) is random noise, and N is the number of sampling points; the parameters of formula (5) are configured according to the actual rolling bearing fault diagnosis test bench parameters and combined with the time domain diagram and envelope spectrum diagram, and the method is as follows: Impact amplitude A=round(max(x)), x is the real signal, max is the maximum value function, and round is rounded; The natural frequency of the bearing failure is [1KHz-60KHz], and the natural frequency f is adjusted by 1KHz each time. n , until max(abs(fft(x)))=max(abs(fft(y))) is satisfied, at which time f n That is what we need, where y is the simulation signal, fft is the fast Fourier transform, and abs is the absolute value function; The initial phase angle φ = 0, and the size of v(t) can be set as long as the fault characteristic frequency in the envelope spectrum of the simulation signal is not submerged. z The calculation of the outer ring fault is shown in formula (6): Among them, r is the actual speed of the rolling bearing, n is the number of balls, d is the rolling element diameter, D is the bearing pitch diameter, and α is the rolling element contact angle; The value range of the damping coefficient ε is [0,1]. The optimal damping coefficient ε is selected by calculating the correlation between the simulation signal y and the real signal x. The correlation is evaluated by the Pearson correlation coefficient as shown in formula (7); Among them, cov is the covariance of the two, σ is the standard deviation, |ρ|∈[0,1], the larger the absolute value of the correlation coefficient, the stronger the correlation; the ε initialization value is 0, and each step is 0.001 until the value reaches 1. The ε with the largest correlation coefficient is the desired one.

Images