A rotating machinery fault diagnosis method based on prior wavelet group strategy fusion

Abstract

The application discloses a rotating machinery fault diagnosis method based on prior wavelet group strategy fusion, and steps are as follows: firstly, the characteristics of each wavelet are tested preliminarily by designing a wavelet convolution shrinkage network and cooperating with pre-training. The test result is used as priori to guide the selection of wavelet group strategy for different input signals. The selection of wavelet group strategy is based on training a wavelet group selection agent. The agent can select a suitable wavelet for each input signal sample to perform wavelet convolution, and then the convolution output result is fused. Finally, the classification network previously pre-trained is fine-tuned in combination with the agent. According to the characteristics of the diversity of the wavelet function, the application selects multiple wavelets to form a wavelet group, compared with the prior art, the application is more reasonable in feature selection, and the model has higher robustness, and the group strategy fusion can further improve the accuracy.

Description

TECHNICAL FIELD

[0001] The application relates to a rotating machinery fault diagnosis method based on prior wavelet group strategy fusion, and belongs to the technical field of fault detection in industrial Internet of Things. BACKGROUND

[0002] With the rise of artificial intelligence technology, intelligent fault diagnosis based on deep learning has become a hot spot in the field of fault diagnosis. The deep learning method has a deep structure and strong nonlinear feature extraction capability, and can directly realize fault feature extraction and pattern recognition in bearing vibration signals. However, in actual industrial scenes, the vibration signals collected from rotating machinery often contain a large amount of noise information, which makes the model unable to capture the fault pattern well in the training process. In fault diagnosis, the traditional method usually uses wavelet transform to process the vibration signal. First, the wavelet coefficients are generated by convolution of the wavelet basis and the vibration signal. Then, a suitable threshold is selected according to expert experience, the wavelet coefficients are soft-thresholded, the noise-related wavelet coefficients are filtered out, and then the remaining wavelet coefficients are used to reconstruct the signal, so as to effectively improve the data quality and improve the performance of the model. The equipment to be diagnosed in industrial production process may generate various impact signals due to changes in working conditions and external interference, so how to select a suitable wavelet for analysis has always been a difficult problem. In view of this problem, the application provides a rotating machinery fault diagnosis method based on prior wavelet group strategy fusion. The purpose is to improve the accuracy and improve the robustness to noise environment. SUMMARY

[0003] To solve the problems in the prior art, the application provides a rotating machinery fault diagnosis method based on prior wavelet group strategy fusion. According to the characteristics of the wavelet function, a plurality of wavelets are selected to form a wavelet group for strategy fusion, so as to further improve the accuracy.

[0004] The technical scheme mainly used in the application is as follows:

[0005] A rotating machinery fault diagnosis method based on prior wavelet group strategy fusion comprises the following steps:

[0006] Step 1:

[0007] Wavelet convolution shrinkage network pre-training: a wavelet convolution shrinkage network composed of a wavelet convolution layer, a threshold learning network, a soft threshold filter and a classifier is constructed, and pre-training is carried out.

[0008] In the pre-training stage, the learning rate a and the training period epoch are set, and the label smoothing strategy and the learning rate preheating strategy are added. Through the above configuration, the approximate performance of each wavelet can be quickly understood in the pre-training stage, and the generalization ability of the network parameters is maintained. After the sample is input into the wavelet convolutional shrinkage network, one wavelet is selected from the wavelet set containing'morlet', 'fbsp', 'gauss','mexhat','shan', and 'laplace' for convolution in turn, and the validation set accuracy of the wavelet convolution pre-training stage is taken as prior experience and brought into the training of the wavelet group strategy selection agent.

[0009] Step 2: Wavelet group strategy selection based on prior knowledge

[0010] In the formal wavelet group strategy selection stage, the wavelet group strategy selection process is converted into a Markov decision process MDP model ε = <S, A, R, T>, and the state space S, the action space A, the reward function R, the state transition function T, and the training target are designed to train an agent to select the wavelet group strategy.

[0011] Step 3:

[0012] Rotating machinery fault diagnosis based on pre-trained network model reuse

[0013] The wavelet convolutional shrinkage network used in step 1 is used in the wavelet convolutional layer of the agent's wavelet group strategy selection result in step 2. In the network composed of the threshold learning network, the soft threshold filter, and the classifier, the parameters obtained in the pre-training stage in step 1 are reused, and the parameters are updated. When the loss converges, the training is terminated.

[0014] Preferably, the specific steps of step 1 are as follows:

[0015] Step 1-1: input the original vibration signal and the wavelet convolution kernel for convolution operation to generate a two-dimensional wavelet coefficient matrix, and the wavelet coefficient calculation formula is as follows:

[0016]

[0017] Where a is the translation coefficient, b is the scale coefficient, ψ(t) is the mother wavelet function, x(t) is the original vibration signal for wavelet convolution, and t is the time variable of the original vibration signal. After completing the continuous wavelet convolution and obtaining the two-dimensional wavelet coefficient matrix, the two-dimensional wavelet coefficient matrix is sequentially placed in the corresponding channel according to the decomposition frequency scale, i.e. one frequency scale corresponds to one channel;

[0018] Step 1-2: input the output of step 1-1 into the threshold learning network to learn the threshold. The essence of the threshold learning network is a small one-dimensional convolutional neural network, which specifically includes two one-dimensional convolutional layers, a global average pooling layer, and a fully connected layer;

[0019] Step 1-3: input the output of step 1-1 into the soft threshold filter, and perform soft thresholding according to the threshold output by step 1-2. The soft thresholding operation specifically sets the values in each channel that are less than the set threshold to zero according to the soft threshold function. The soft threshold function is calculated as follows:

[0020]

[0021] Where sign(x) is the sign function, which is used to ensure the positive and negative nature of the feature value; x is the soft threshold. λ is the threshold, which is output by the threshold learning network in step 1-2;

[0022] Step 1-4: input the output of the soft threshold filter into the classifier for fault classification, and the classifier structure adopts three one-dimensional convolutional layers and two fully connected layers.

[0023] Preferably, the specific steps of step 2 are as follows:

[0024] Step 2-1: After pre-training, the state space S is designed. The state is designed as the input vibration signal, so that the agent can select the appropriate wavelet for feature extraction according to the input signal, and obtain the final diagnosis result. The state space S is the set of all input vibration signals;

[0025] Step 2-2: Action space A design, action is designed as selecting appropriate wavelet group strategy, when the state input, the agent selects an action from the action space, according to the action selection to carry on the convolution to the wavelet convolution kernel corresponding, after the convolution, according to the prior experience to the convolution result carries on the fusion, obtains the final wavelet convolution output. The action space A is the set of all wavelet group strategies;

[0026] Step 2-3: design the reward function R, the reward function R is designed as follows: put the final convolution output into the network composed of threshold learning network, soft threshold filter and classifier to obtain the predicted label. Compare the predicted label with the label of the input sample, if the result is correct, obtain the reward, otherwise, be punished. At the same time, according to the predicted label and the actual label, update the historical experience of each wavelet convolution in the wavelet convolution set;

[0027] Step 2-4: design the state transition function T, the state transition is designed as randomly selecting a sample from the remaining samples, so the state transition function T is the set of all remaining samples.

[0028] After the above design is completed, the agent is trained, and the parameters of the network composed of the threshold learning network, the soft threshold filter and the classifier are kept fixed during the training process. When the average reward obtained by the agent tends to be stable, the training is terminated.

[0029] Beneficial effects: the application provides a rotating machinery fault diagnosis method based on prior wavelet group strategy fusion, which can perform fault diagnosis in a strong noise environment. According to the characteristics of the wavelet function, multiple wavelets are selected to form a wavelet group, which is more reasonable in feature selection compared with the prior art, and the model has higher robustness. The group strategy fusion can further improve the accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 The method flowchart of the application is shown in the figure;

[0031] Figure 2 The schematic diagram of the wavelet convolution shrink network is shown in the figure;

[0032] Figure 3 The schematic diagram of the threshold learning network is shown in the figure;

[0033] Figure 4 The schematic diagram of the agent performing action is shown in the figure. DETAILED DESCRIPTION

[0034] In order to enable the person skilled in the art to better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application are described clearly and completely below. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by the person skilled in the art without creative labor should be within the protection scope of the present application.

[0035] Embodiment 1:

[0036] The rotating machinery fault diagnosis method based on prior wavelet group strategy fusion in the embodiment comprises the following steps: Figure 1

[0037] Step 1:

[0038] Wavelet convolution shrink network pre-training: a wavelet convolution shrink network composed of a wavelet convolution layer, a threshold learning network, a soft threshold filter and a classifier is constructed, as shown in the figure. Figure 2 ​The pre-training phase is set with learning rate a and training period epoch, while label smoothing strategy and learning rate preheating strategy are added. Through the above configuration, the approximate performance of each wavelet can be quickly understood in the pre-training phase, and the generalization ability of the network parameters is maintained. After the sample is input into the wavelet convolutional shrinkage network, one wavelet is selected from the wavelet set containing'morlet', 'fbsp', 'gauss','mexhat','shan', 'laplace' for convolution in turn, and the validation set accuracy of the wavelet convolution pre-training phase is taken as prior experience and brought into the training of the wavelet group strategy selection agent.

[0039] Step 2:

[0040] Wavelet group strategy selection based on prior knowledge: In the formal wavelet group strategy selection phase, the wavelet group strategy selection process is converted into a Markov decision process MDP model ε = <S, A, R, T>, and the state space S, the action space A, the reward function R, the state transition function T and the training target are designed to train an agent to select the wavelet group strategy.

[0041] Step 3: Rotary machinery fault diagnosis based on pre-trained network model reuse: The wavelet convolutional shrinkage network used in step 1 pre-training phase is used, and the wavelet group strategy selection result of the agent in step 2 is used in the wavelet convolutional layer. In the network composed of threshold learning network, soft threshold filter and classifier, the parameters obtained in the pre-training phase in step 1 are reused, and the parameters are updated. When the loss converges, the training is terminated.

[0042] Preferably, the specific steps of step 1 are as follows:

[0043] Step 1-1: input the original vibration signal and the wavelet convolution kernel for convolution operation to generate a two-dimensional wavelet coefficient matrix of MxN, M represents the number of frequency scales of decomposition, and N is the length of the vibration signal. The wavelet coefficient calculation formula is as follows:

[0044]

[0045] Where a is the translation coefficient, b is the scale coefficient, ψ(t) is the mother wavelet function, x(t) is the original vibration signal for wavelet convolution, and t is the time variable of the original vibration signal. After completing the continuous wavelet transform and obtaining the MxN two-dimensional wavelet coefficient matrix, the two-dimensional wavelet coefficient matrix is decomposed into M one-dimensional feature values of length N according to the frequency scale decomposition, and is put into M channels, that is, one frequency scale corresponds to one channel;

[0046] Step 1-2: Input the output of Step 1-1 into the threshold learning network to learn the threshold. The threshold learning network is essentially a small one-dimensional convolutional neural network, specifically consisting of two one-dimensional convolutional layers, a global average pooling layer, and a fully connected layer. For example... Figure 3 As shown, the input to the threshold learning network is the feature values ​​of M channels output by the wavelet convolutional layer, with each channel's feature value having a length of N, resulting in a total input of an M×N matrix. A one-dimensional convolutional layer 1 takes M input channels and outputs M1 channels, producing an M1×N1 feature map. This is then passed through a one-dimensional convolutional layer 2 to obtain an M2×N2 feature map. Global average pooling is then applied to each channel to reduce the feature map to M2×1, which serves as the input to the fully connected layer. The fully connected layer outputs M values ​​as weights. Simultaneously, the input feature values ​​are first absolute-valued and then subjected to global average pooling to obtain the absolute average of the M channel features. Finally, the M weights and the M absolute averages are multiplied by a dot to obtain the final M thresholds.

[0047] Steps 1-3: Input the output of Step 1-1 into the soft threshold filter, and perform soft thresholding based on the threshold output from Step 1-2. Specifically, the soft thresholding operation sets the values ​​in each channel that are less than a set threshold to zero according to the soft threshold function. The soft threshold function is calculated as follows:

[0048]

[0049] Where sign(x) is the sign function, used to ensure the positive or negative nature of the feature value; x is the soft threshold, and λ is the threshold, output by the threshold learning network in steps 1-2;

[0050] Steps 1-4: Input the output of the soft threshold filter into the classifier for fault classification. The classifier structure uses three one-dimensional convolutional layers and two fully connected layers.

[0051] Preferably, the specific steps of step 2 are as follows:

[0052] Step 2-1: After pre-training, design the state space S. The state is designed as the input vibration signal, allowing the agent to select appropriate wavelets for feature extraction based on the input signal, thus obtaining the final diagnostic result. The state can be represented as s = {X}, where X = {x1, x2, x3, ..., x...}. n} refers to the input vibration signal, x i Let n be the sampled data points, and n represent the signal length. The state space S is the set of all possible input vibration signals.

[0053] Step 2-2: Design the action space A. The action is designed to select an appropriate wavelet group strategy. Upon receiving the state input, the agent selects an action from the action space and performs convolution using the corresponding wavelet convolution kernel. After convolution, the results are fused according to prior experience to obtain the final wavelet convolution output. The action can be represented as a = {model1, model2, model3, ..., model...} m}, where model i ={0 or 1} represents the selection state of the i-th wavelet, when model i =1 indicates that the i-th wavelet is selected for convolution in this round, otherwise the i-th wavelet is not selected. The action space A is the set of all wavelet group policies.

[0054] by Figure 4 For example, after inputting state s, the resulting action a = {1, 0, 1, 0, 1} is selected. Then, the 1st, 3rd, and 5th wavelets from the wavelet pool are chosen for convolution. Let the convolution be F. i (·), after inputting sample X, the convolution output result Result=F i (X). Based on action a, a set of outputs can be obtained at this time: Outputs = {F1(X), F3(X), F5(X)}. Meanwhile, through pre-training, a general understanding of the characteristics of each wavelet has been obtained. Here, the accuracy of the model on the validation set during the pre-training stage is used as historical experience for wavelet selection by the subsequent agent, providing a basis for the final weighted fusion of convolutional outputs. The historical experience of the wavelet pool can be expressed as E = {e1, e2, e3, ..., e...} m}, e i Let be the historical experience of the i-th wavelet in the wavelet pool, and m be the number of wavelet types in the wavelet pool. Finally, this set of results is weighted and averaged according to historical experience to obtain the final convolution result, i.e.

[0055] Steps 2-3: Design the reward function R as follows: The final convolutional output is fed into a network consisting of a threshold learning network, a soft thresholding filter, and a classifier to obtain the predicted label. The predicted label is compared with the label of the input sample. If the result is correct, a reward is given; otherwise, a penalty is applied. Let the predicted label be y′, the input sample label be y, the reward value be a, and the penalty value be p. Therefore, the reward function is defined as follows:

[0056]

[0057] At the same time, the historical experience of the model pool is updated based on the predicted and actual labels. This historical experience is stored in fractional form. e righte represents the number of historical correct results of the model n e represents the number of historical diagnoses of the model. Let e i ′ be the updated historical experience, then:

[0058] Step 2-4: Perform state transition function T design, the state transition is designed to randomly select a sample from the remaining samples, so the state transition function T is the set of all remaining samples. It can be represented as s'∈{X1,X2,…,X j}, {X1,X2,…,X j} indicates the set of all remaining samples.

[0059] After the above design is completed, the agent is trained, and the parameters of the network composed of the threshold learning network, the soft threshold filter and the classifier are kept fixed during the training process. When the average reward obtained by the agent tends to be stable, the training is terminated.

[0060] The above only describes the preferred embodiments of the present application. It should be noted that, for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. A method for fault diagnosis of rotating machinery based on prior wavelet group strategy fusion, characterized in that, Includes the following steps: Step 1: Pre-training the wavelet convolution shrinking network A wavelet convolution shrinking network consisting of wavelet convolutional layers, a threshold learning network, a soft threshold filter, and a classifier is constructed and pre-trained. During the pre-training stage, after the samples are input into the wavelet convolution shrinking network, a wavelet is selected sequentially from the wavelet set containing 'morlet', 'fbsp', 'gauss', 'mexhat', 'shan', and 'laplace' for convolution. The validation set accuracy of the wavelet convolution pre-training stage is used as prior experience and incorporated into the training of the wavelet group policy selection agent. Step 2: Wavelet group strategy selection based on prior knowledge In the formal wavelet group policy selection phase, the wavelet group policy selection process is transformed into a Markov decision process (MDP) model ε=<S,A,R,T> We designed a state space S, an action space A, a reward function R, a state transition function T, and a training objective, and trained an agent to select wavelet group policies. Step 2-1: After pre-training, design the state space S; design the state as the input vibration signal, so that the agent selects an appropriate wavelet to extract features based on the input signal, and obtains the final diagnosis result; the state space S is the set of all input vibration signals; Step 2-2: Design the action space A; design the action to select a suitable wavelet group strategy. When the state is input, the agent selects an action from the action space, selects the corresponding wavelet convolution kernel according to the action, and performs convolution. After convolution, the convolution results are fused according to prior experience to obtain the final wavelet convolution output; the action space A is the set of all wavelet group strategies. Steps 2-3: Design the reward function R; The reward function R is designed as follows: The final convolution output is fed into a network consisting of a threshold learning network, a soft thresholding filter, and a classifier to obtain the predicted label; The predicted label is compared with the label of the input sample. If the result is correct, a reward is obtained; otherwise, a penalty is imposed; At the same time, the historical experience of each wavelet convolution in the wavelet convolution set is updated according to the predicted label and the actual label. Step 2-4: Design the state transition function T. The state transition is designed to randomly select a sample from the remaining samples, so the state transition function T is the set of all remaining samples. After completing the above design, the agent is trained. During the training process, the parameters of the network consisting of the threshold learning network, the soft thresholding filter and the classifier are kept fixed. The training is terminated when the average reward obtained by the agent tends to stabilize. Step 3: Rotating machinery fault diagnosis based on the reuse of pre-trained network models The wavelet convolution shrinking network used in the pre-training phase of step 1 is used, and the wavelet group policy selection results of the agent in step 2 are used in its wavelet convolution layer; in the network composed of threshold learning network, soft thresholding filter and classifier, the parameters obtained in the pre-training phase of step 1 are reused and the parameters are updated; training is terminated when the loss converges.

2. The rotating machinery fault diagnosis method based on prior wavelet group strategy fusion according to claim 1, characterized in that, The specific steps of step 1 are as follows: Step 1-1: Input the original vibration signal and perform convolution operation with the wavelet convolution kernel to generate a two-dimensional wavelet coefficient matrix. The wavelet coefficient calculation formula is as follows: Where a is the translation coefficient, b is the scaling coefficient, ψ(t) is the wavelet mother function, x(t) is the original vibration signal for wavelet convolution, and t is the time variable of the original vibration signal; after completing continuous wavelet convolution and obtaining the two-dimensional wavelet coefficient matrix, the two-dimensional wavelet coefficient matrix is ​​put into the corresponding channel according to the decomposed frequency scale, that is, one frequency scale corresponds to one channel. Step 1-2: Input the output of Step 1-1 into the threshold learning network to learn the threshold; the threshold learning network consists of two one-dimensional convolutional layers, one global average pooling layer, and one fully connected layer; Steps 1-3: Input the output of Step 1-1 into the soft threshold filter, and perform soft thresholding based on the threshold output from Step 1-2. Soft thresholding involves setting values ​​in each channel that are less than a set threshold to zero according to the soft threshold function. The soft threshold function is calculated as follows: Where sign(x) is the sign function, used to ensure the positive or negative nature of the feature value; x is the soft threshold, and λ is the threshold, output by the threshold learning network in steps 1-2; Steps 1-4: Input the output of the soft threshold filter into the classifier for fault classification. The classifier structure uses three one-dimensional convolutional layers and two fully connected layers.

Images