CN112116029A

CN112116029A - Intelligent fault diagnosis method for gearbox with multi-scale structure and characteristic fusion

Info

Publication number: CN112116029A
Application number: CN202011055494.5A
Authority: CN
Inventors: 尚志武; 李万祥; 高茂生; 俞燕; 周士琦; 张宝仁; 刘飞; 庞海玉
Original assignee: Tianjin Polytechnic University
Current assignee: Tianjin Polytechnic University
Priority date: 2020-09-25
Filing date: 2020-09-25
Publication date: 2020-12-22

Abstract

The invention discloses a gearbox intelligent fault diagnosis method with a multi-scale structure and characteristic fusion function, which comprises the following steps of: the method comprises the steps of multi-scale depth feature extraction structure design, depth feature fusion strategy design based on information entropy and health state identification based on a depth confidence network. The method overcomes the defects of low diagnosis precision, poor stability and insufficient feature extraction of a single deep learning neural network model, enhances the feature extraction effect under noise interference, and ensures that the fused depth features have better robustness and representativeness, thereby realizing the high-precision diagnosis of the gear box.

Description

Intelligent fault diagnosis method for gearbox with multi-scale structure and characteristic fusion

Technical Field

The invention relates to an intelligent fault diagnosis method for a gearbox with multi-scale structure and characteristic fusion.

Background

In the industrial field, a gear box is widely used as a representative of rotary machines, contributing to the advancement of society. With the development of science and technology, rotary machines are developed towards automation, intellectualization and high speed, so that the gear box is operated in the environment of high temperature, high pressure and high load for a long time. Therefore, inevitable faults of the gearbox are forced, so that associated fault diagnosis and monitoring of the gearbox is necessary.

Traditional intelligent fault diagnosis methods, such as shallow learning algorithms based on BP neural networks, support vector machines, random deep forests, Bayesian classifiers and the like, make great research progress in the field of fault diagnosis. However, since the fault diagnosis method based on shallow learning has a simple structure, a complex nonlinear relationship cannot be accurately expressed, and such a method requires a feature extraction method that relies on a signal processing technique to a great extent. In addition, the method has low efficiency and generalization capability in fault diagnosis. In addition, intelligent fault diagnosis methods based on deep learning have also become a hot content of research. However, current studies on fault diagnosis of deep neural networks are mainly directed to performance studies of a single model. However, because the collected signals often contain a large amount of noise, the diagnosis of the gearbox by using the single deep neural network has the problems of low accuracy, poor stability and low generalization capability. The key point of these problems is that the depth feature information extracted from the original vibration signal by the single depth neural network is incomplete and contains noise features, so that the representativeness of the depth features is poor, the diagnostic result of the model is seriously affected, and the stability of the model is poor.

Therefore, in combination with the research form of gearbox fault diagnosis at home and abroad, a novel intelligent fault diagnosis method is urgently needed to be developed to realize high-precision diagnosis of the gearbox, improve the generalization capability of a network model and ensure the stability of a diagnosis result.

Disclosure of Invention

The invention aims to solve the problems, and designs an intelligent fault diagnosis method of a gearbox with a multi-scale structure and characteristic fusion, which comprises the following steps:

step 1: designing a multi-scale depth feature extraction structure;

step 2: designing a depth feature fusion strategy based on information entropy;

and step 3: and identifying the health state based on the deep belief network.

Further, in step 1, the multi-scale depth feature extraction structure design includes the specific steps of:

step 1.1: constructing a plurality of self-encoders with different properties;

step 1.2: and combining the self-encoders with different properties in parallel to generate the multi-scale depth feature extraction structure.

Further, in the step 2, the depth feature fusion strategy design based on the information entropy includes the specific steps:

step 2.1: obtaining an evaluation matrix A;

step 2.2: calculating fusion weight of each model based on the information entropy;

step 2.3: and calculating the depth fusion feature H.

Further, in the step 3, the health status recognition based on the deep belief network specifically includes the steps of:

step 3.1: pre-training a restricted Boltzmann machine model;

step 3.2: and stacking the limited Boltzmann machine model to generate a depth confidence network.

Further, in step 1.1, a plurality of self-encoders with different properties are constructed, and the specific steps are as follows:

step 1.1.1: designing a self-encoder model with three neural network layers, wherein the neural network layers are an input layer, a hidden layer and an output layer respectively, the neuron numbers of the input layer and the output layer are the same, and the neuron number of the hidden layer is smaller than that of the input layer;

step 1.1.2: introducing a sparse regular term by utilizing the KL distance in an automatic encoder model, constructing a sparse automatic encoder, wherein a sparse regular term expression constructed by utilizing the KL distance is as follows:

wherein the content of the first and second substances,

and representing the average value of an input vector x of the input layer to the jth neuron of the hidden layer, wherein rho is a sparse response coefficient.

Step 1.1.3: on the basis of the self-encoder model, an input layer is changed into an input layer with random noise, so that a noise reduction self-encoder is constructed;

step 1.1.4: based on the self-encoder model, the output of the hidden layer is used for punishment relative to the input Jacobian matrix, so that the shrinkage self-encoder is constructed.

Further, in step 1.2, self-encoders with different properties are combined in parallel to generate a multi-scale depth feature extraction structure, which specifically comprises the following steps:

step 1.2.1: pre-training a plurality of self-encoder models, stacking the models in sequence according to a pre-training sequence to form a depth self-encoder, and constructing a depth sparse self-encoder, a depth denoising self-encoder and a depth contraction self-encoder in the same way;

step 1.2.2: and combining the generated depth self-encoder, the depth sparse self-encoder, the depth denoising self-encoder and the depth shrinkage self-encoder in a parallel connection manner to generate the multi-scale depth feature extraction structure.

Further, in the step 2.1, the evaluation matrix a is obtained, and the specific steps are as follows:

step 2.1.1: according to the training data and the training labels, pre-training a depth self-encoder, a depth sparse self-encoder, a depth denoising self-encoder and a depth shrinkage self-encoder of the multi-scale depth feature extraction structure;

step 2.1.2: obtaining the accuracy of each fault type of each model, and combining the accuracy into an evaluation matrix A, wherein the specific form is as follows:

wherein A is_ijThe accuracy corresponding to the ith fault of the jth model is shown, when j is 1, the depth self-encoder is represented, and when j is 2, the depth sparse self-encoder is represented; when j is 3, the depth denoising self-encoder is represented; when j is 4, it represents a depth-punctured self-encoder, and d represents the number of types of failure.

Further, in the step 2.2, the fusion weight of each model is calculated based on the information entropy, and the specific steps are as follows:

step 2.2.1: defining the information entropy of the jth model according to the evaluation matrix A as follows:

where j is 1, 2, 3, 4, d indicates the number of types of failure.

Step 2.2.2: on the basis of the information entropy, the fusion weight of the jth model is defined as:

wherein, w_jSatisfy 0 < w_j＜1，w₁+w₂+w₃+w₄＝1。

Further, in step 2.3, the depth fusion feature H has a specific expression as follows:

wherein H_jFor the depth feature learned for the jth model, j is 1, 2, 3, 4.

Further, in the step 3.1, the pre-training of the restricted boltzmann model includes the following specific steps:

step 3.1.1: designing neural networks of two neural network layers, namely a visible layer and a hidden layer, wherein neurons in the visible layer and the hidden layer are not interconnected, and only the neurons between the layers are symmetrically connected, and the number of neurons in the hidden layer is less than that of neurons in the visible layer, so that a restricted Boltzmann model is generated;

step 3.1.2: pre-training a restricted Boltzmann machine model by an unsupervised greedy layer-by-layer learning method, wherein in the process, a vector is generated by a display layer, and a value is transmitted to a hidden layer through the vector; in turn, the inputs to the display layers are randomly selected to attempt to reconstruct the input signal. And finally, the new neuron activation units of the presentation layer forward propagate and reconstruct the hidden layer activation function to obtain a hidden layer unit h.

Further, in step 3.2, the limited boltzmann machine stack is used to generate a deep confidence network, which specifically includes: and connecting all hidden layers according to the pre-training sequence of the limited Boltzmann machine, thereby generating the deep confidence network model.

The method for intelligently diagnosing the fault of the gearbox by fusing the multi-scale structure and the characteristics by utilizing the technical method is used for extracting different fault characteristic information of an original vibration signal by constructing a multi-scale depth characteristic extraction structure, and ensures that the extracted characteristics contain all information of the running state of the gearbox to the maximum extent; then designing a depth feature fusion strategy based on the information entropy, and fusing the depth features extracted by different models to ensure that the fused features have stronger representativeness and less redundant information; and finally, generating a depth confidence network by utilizing a plurality of limited Boltzmann machine stacks, and realizing the health state identification of the gearbox on the fused depth characteristics by means of the strong characterization capability of the depth confidence network in characteristic identification, classification and nonlinear mapping. Therefore, the invention has strong practical value, can provide reliable gear box working state information for maintenance personnel, reduces economic loss caused by faults and meets the requirements of fault monitoring and diagnosis.

Drawings

FIG. 1 is a flow chart of a multi-scale structure and feature fused gearbox intelligent fault diagnosis method of the present invention;

FIG. 2 is a diagram of an experimental platform for acquiring vibration signals used in accordance with an embodiment of the present invention;

FIG. 3 is a diagram of gearbox fault locations for one example of the present invention;

FIG. 4 is a graph of a time domain signal and a frequency domain signal of a vibration signal used in the one embodiment of the present invention;

FIG. 5 is a block diagram of a standard self-encoder according to the present invention;

FIG. 6 is a schematic structural diagram of a noise reduction self-encoder according to the present invention;

FIG. 7 is a schematic diagram of a depth self-encoder generation according to the present invention;

FIG. 8 is a schematic diagram of a multi-scale depth feature extraction structure according to the present invention;

FIG. 9 is a schematic diagram of a deep belief network generation form according to the present invention;

FIG. 10 is a graph of the detailed diagnostic results of a comparative experiment according to one embodiment of the present invention;

FIG. 11 is a graph of multi-class confusion matrix results from a first experiment with other methods in accordance with an example of the present invention;

FIG. 12 is a diagram illustrating the result of the multi-class confusion matrix for the first experiment according to the proposed method in the embodiment of the present invention;

FIG. 13 is a graph of the visual comparison of the last layer of features extracted by different methods in one example of the invention.

Detailed Description

The invention is described in detail below with reference to the accompanying drawings, and as shown in fig. 1, an intelligent fault diagnosis method for a gearbox with a multi-scale structure and fused features comprises the following steps:

step 1: designing a multi-scale depth feature extraction structure;

step 2: designing a feature fusion strategy based on information entropy;

and step 3: and identifying the health state based on the deep belief network.

Next, an embodiment of the present invention and its effects will be specifically described by way of an example.

The experimental data acquisition of this example was derived from a laboratory drive train dynamic test rig, as shown in fig. 2. The collecting platform consists of a 3-phase motor with 3 horsepower, 1 2-stage planetary gear box, 1 2-stage fixed gear box and 1 programmable magnetic brake; the frequency of the motor is 30Hz, the sampling frequency of the acceleration sensor is 3kHz, one end of the acceleration is arranged in the vertical direction of the fixed shaft gear box, the other end of the acceleration is connected to the detailed parameters of the fault working condition of the gear box of the acquisition system, and the detailed parameters are shown in the table 1, and the position of the fault of the gear box is shown in the graph 3. In this example description, six conditions are considered, and time domain and frequency domain waveform diagrams of vibration signals of the six conditions are acquired and shown in fig. 4, where (a), (b), (c), (d), (e), and (f) are respectively a health condition signal, a gear hub crack signal, a gear tooth breakage signal, a composite fault 1 signal, a composite fault 2 signal, and a composite fault 3 signal. Each condition contained 300 samples, and the vibration signal for each sample consisted of 300 consecutive sampled data points. In order to verify the accuracy and reliability of the method, samples of each group of working conditions are randomly divided according to the proportion of 70% to 30%, and then a training set and a testing set are sequentially formed. That is, 210 samples are randomly selected to form a training set for each working condition, and the remaining 90 samples form a testing set for testing the diagnostic effect of the invention.

Table 1: details of six gearbox operating conditions

Operating condition of gear box	Training/testing set capacity	Label (R)
			Health care	210/90	1
Wheel hub crack	210/90	2
			Broken gear of gear	210/90	3
Composite fault 1	210/90	4
			Composite fault 2	210/90	5
Composite fault 3	210/90	6

Description of the drawings: the compound fault 1 is gear hub crack, gear abrasion and bearing rolling element abrasion; the compound failure 2 is the gear wheel hub crack, gear broken tooth and bearing rolling body abrasion; the compound fault 3 is gear abrasion, gear breakage and bearing rolling body abrasion.

Further, the multi-scale depth feature extraction structure design described in step 1 specifically includes the steps of:

step 1.1: constructing a plurality of self-encoders with different properties;

design of self-encoder with three neural network layersThe model, as shown in fig. 5, the neural network layer is an input layer, a hidden layer and an output layer, wherein the neuron numbers of the input layer and the output layer are the same, and the neuron number of the hidden layer is smaller than that of the input layer. Suppose an input vector x ═ x₁，x₂，…，x_n]Then the output vector h of the hidden layer is:

h＝sigm(Wx+b)

the decoding process reconstructs the input vector using h as:

wherein, W represents a weight matrix from an input layer to a hidden layer, W 'represents a weight matrix from a hidden layer to an input layer, b and b' represent bias vectors of an encoding layer and a decoding layer, respectively, and sigm () is a nonlinear activation function, which is specifically expressed as:

sigm(t)＝1/(1+e^-t)

the purpose of extracting the depth features is achieved by minimizing the error of the input vector and the reconstructed input vector, and the loss function formula for measuring the error is as follows:

wherein the content of the first and second substances,

is a loss function for measuring the input vector x and reconstructing the input vector

Of different size, x_iAnd

the ith dimension elements of the input vector and the output vector, respectively.

Introducing a sparse regular term by using the KL distance in the self-encoder model to construct a sparse self-encoder (SAE), wherein a sparse regular term expression constructed by using the KL distance is as follows:

wherein the content of the first and second substances,

At this time, the loss function expression of the sparse self-encoder is:

wherein, beta is a coefficient of sparse penalty constraint degree, and p represents the number of hidden layer neurons.

And thirdly, on the basis of the self-encoder model, changing an input layer into an input layer with random noise so as to construct a noise reduction self-encoder (DAE), wherein the structure of the noise reduction self-encoder is shown in FIG. 6. The noise-reducing self-encoder adopted in the invention adds Gaussian noise into an input vector x to construct a damaged input vector

At this time, the loss function of the noise-reducing self-encoder is:

wherein the content of the first and second substances,

representing the reconstructed input vector, sigma is the damage degree of the noise parameter representing the original input vector, and I is an identity matrix.

Based on the self-encoder model, punishment is carried out by utilizing the Jacobian matrix of the output of the hidden layer relative to the input, so as to construct the shrinkage self-encoder, wherein the loss function is as follows:

wherein, J_b(x) A Jacobian matrix of the hidden layer output vector h relative to the input vector x can contain information of data in all directions, and the disturbance of the training sample in all directions can be restrained by the regularity constraint of the Jacobian matrix; | J_h(x)||²Is a shrinkage penalty term, and lambda is a shrinkage regularization constraint coefficient.

Firstly, pre-training a plurality of standard self-encoder models, and stacking the models in sequence according to the pre-training sequence to form a depth self-encoder, wherein the specific form is shown in FIG. 7; a depth sparse self-encoder, a depth denoising self-encoder and a depth shrinkage self-encoder are constructed in the same way. The training mode adopts a back propagation algorithm and a gradient descent algorithm to update the parameters of the models of the standard self-encoder, the sparse self-encoder, the noise reduction self-encoder and the contraction self-encoder, and the parameter updating formula is as follows:

wherein α represents a learning rate; k represents the number of iterations of the self-encoder training.

And secondly, combining the generated depth autoencoder, the depth sparse autoencoder, the depth denoising autoencoder and the depth shrinkage autoencoder in a parallel connection mode to generate a multi-scale depth feature extraction structure, wherein the specific form is shown in FIG. 8.

Further, the characteristic fusion strategy design based on the information entropy in step 2 specifically comprises the following steps:

step 2.1: obtaining an evaluation matrix A;

according to training data and training labels, a depth self-encoder, a depth sparse self-encoder, a depth denoising self-encoder and a depth shrinkage self-encoder of the multi-scale depth feature extraction structure are pre-trained.

Obtaining the accuracy of each fault type of each model, and combining the accuracy into an evaluation matrix A, wherein the specific form is as follows:

defining the information entropy of the jth model according to the evaluation matrix A as follows:

where j is 1, 2, 3, 4, d indicates the number of types of failure.

Secondly, on the basis of the information entropy, the fusion weight of the jth model is defined as:

wherein, w_jSatisfy 0 < w_j＜1，w₁+w₂+w₃+w₄＝1。

Step 2.3: and calculating the depth fusion feature H, wherein the specific expression is as follows:

wherein H_jFor the depth feature learned for the jth model, j is 1, 2, 3, 4.

Further, the health status recognition based on the deep belief network described in step 3 specifically includes the steps of:

step 3.1: pre-training a restricted Boltzmann machine model;

firstly, designing neural networks of two neural network layers, namely a visible layer and a hidden layer, wherein neurons in the visible layer and the hidden layer are not interconnected, only the neurons between the layers are symmetrically connected, and the number of the neurons in the hidden layer is less than that of the neurons in the visible layer, so that a restricted Boltzmann model is generated.

Pre-training a restricted Boltzmann machine model by an unsupervised greedy layer-by-layer learning method, wherein in the process, a vector is generated by a display layer, and a value is transmitted to a hidden layer through the vector; in turn, the inputs to the display layers are randomly selected to attempt to reconstruct the input signal. And finally, the new neuron activation units of the presentation layer forward propagate and reconstruct the hidden layer activation function to obtain a hidden layer unit h. The specific parameter updating mode of the limited Boltzmann machine is as follows:

assuming x is the input of the explicit layer, the probability of hidden layer neuron being turned on is calculated as:

the superscript in the formula is used for distinguishing different limited Boltzmann machines, and the subscript is used for distinguishing different neurons. Then, a Gibbs sampling method is adopted to extract a sample h from the calculated probability distribution⁽⁰⁾～p(h⁽⁰⁾|v⁽⁰⁾) By using h⁽⁰⁾Reconstructing the input of the display layer, wherein the specific formula is as follows:

likewise, a sample v of the layer is taken⁽¹⁾～p(v⁽¹⁾|h⁽⁰⁾) And calculating the probability that the hidden layer neuron is opened by using the apparent layer neuron again as follows:

the weights and offsets of the restricted boltzmann machine are then updated according to the following equations:

W←W+λ(p(h⁽⁰⁾＝1|v⁽⁰⁾)v^(0)T-p(h⁽¹⁾＝1|v⁽¹⁾)v^(1)T)

b←b+λ(v⁽⁰⁾-v⁽¹⁾)

c←c+λ(h⁽⁰⁾-h⁽¹⁾)

wherein W is the weight of the restricted Boltzmann machine, b is the bias of the explicit layer, and c is the bias of the implicit layer.

Generating a depth confidence network by a limited Boltzmann machine stack, which specifically comprises the following steps: and connecting the hidden layers according to the pre-training sequence of the limited Boltzmann machine, thereby generating a deep confidence network model, wherein the structure of the deep confidence network model is shown in FIG. 9.

Secondly, fine tuning the deep confidence network model by adopting a back propagation algorithm and a gradient descent algorithm, wherein the specific process is as follows: firstly, inputting an input signal into a deep belief network, and enabling the network to forward propagate and learn to obtain a prediction tag; then calculating an error value of the predicted label and the real label, and reversely propagating the error value to each layer of neural network by using a reverse propagation algorithm; and finally, updating the weight and the bias in the network by using a gradient descent algorithm.

The steps of the intelligent gearbox fault diagnosis method based on the multi-scale depth feature extraction structure and the feature fusion are explained. In order to further verify the effectiveness and the advancement of the technical method, the fault diagnosis is carried out on the gearbox by adopting a depth standard self-encoder, a depth sparse self-encoder, a depth denoising self-encoder, a depth contraction self-encoder and a Convolutional Neural Network (CNN) of a depth learning method besides adopting a BP neural network (BPNN), a Softmax classifier, a Support Vector Machine (SVM) and a Random Forest (RF) fault diagnosis method of a shallow learning method to verify the same data set. The following points need to be further explained:

firstly, the method only needs to carry out data segmentation processing on the acquired vibration signals, and no feature extraction technology participates in the process;

the input of a depth standard self-encoder, a depth sparse self-encoder, a depth denoising self-encoder and a depth shrinkage self-encoder and the input of the method belong to the same data set, and the input dimension of the CNN is a sample of 400;

only one form of input is provided for BPNN, SVM, Softmax classifier and RF; namely, the 28 features extracted by the signal processing technology comprise 10 time domain features, 10 frequency domain features and 8 time-frequency domain features.

In addition, ten experiments were performed on the data set of the fixed axis gearbox in order to ensure the reliability of the experimental results of this method. The average test accuracy and the average standard deviation of the invention and other fault diagnosis methods are shown in table 2. From the analysis in table 2, the present invention has higher test accuracy (94.31%) and lower standard deviation (0.3187) compared to other methods. Compared with the BPNN, SVM, Softmax classifier and RF of shallow learning, the average accuracy of the test samples of the present invention is higher than 84.07% of BPNN, 89.39% of SVM, 90.40% of RF and 83.14% of Softmax classifier. The invention can directly extract the fault characteristics from the vibration signals to carry out fault diagnosis, and avoids the complicated process of manually participating in characteristic extraction. Compared with the standard depth learning model, the accuracy of the test sample is also higher than 88.19% of the depth standard self-encoder, 90.13% of the depth noise reduction self-encoder, 90.69% of the depth sparse self-encoder, 90.94% of the depth contraction self-encoder and 90.74% of the CNN. Moreover, the standard deviation of the diagnostic result of the test sample of the present invention is 0.3187, which is also much lower than 1.3593, 1.2359, 1.2823, 1.0287, 0.6171, 1.4211, 0.8097, 1.3824, 1.7112 of other methods. Therefore, when fault diagnosis is carried out on the vibration signal of the gearbox, the stability of fault identification and the capability of extracting the depth feature are improved.

Table 2: diagnosis results of different methods of 10 experiments

Comparison method	Dimension of input	Average test accuracy (%)	Standard deviation of mean test accuracy
				Method
1	300	94.31	0.3187
				Method 2	300	88.19	1.3593
Method 3	300	90.13	1.2359
				Method 4	300	90.69	1.2823
Method 5	300	90.94	1.0287
				Method 6	300	90.74	0.6171
Method 7	28	84.07	1.4211
				Method 8	28	89.39	0.8097
Method 9	28	90.40	1.3824
				Method 10	28	83.14	1.7112

Description of the drawings: method 1-invention; method 2-depth standard auto-encoder; method 3-deep denoising autoencoder; method 4-depth sparse autoencoder; method 5-depth-shrink self-encoder; method 6-convolutional neural network; method 7-BP neural network; method 8 — support vector machine; method 9-random forest; method 10-Softmax classifier.

The detailed diagnosis results of the ten tests on the test sample by different methods are shown in fig. 10 and are presented in an intuitive form. The ten-time test accuracy of the method is 94.07%, 93.86%, 94.26%, 94.44%, 94.81%, 94.26%, 94.07%, 94.81% and 94.07% respectively. As seen in fig. 10, the test accuracy of the present invention is significantly higher than other fault diagnosis methods. Table 3 shows the main parameters required by the present invention, and the four depth autocoders of the multi-channel depth feature extraction structure have the structure of 300-200-80, and the feature learning between them is independent and not interfered. The model structure of the DBN classifier is 80-40-40-40-6, and the DBN classifier comprises an input layer, three hidden layers and an output layer.

Table 3: the main parameters of the method

Description of example parameters of the invention	Value taking
		Number of hidden layers per depth autoencoder	3
Number of neurons in input layer	300
		Number of first hidden layer neurons	200
Number of second hidden layer neurons	100
		Number of third hidden layer neurons	80
Learning rate for depth autoencoder	0.025
		Number of iterations of depth autoencoder	500
Sparse punishment of sparse autoencoderPenalty constraint coefficient	0.2
		De-noising noise destruction coefficients from an encoder	0.15
Shrinking regularization constraint coefficients for a compressed self-encoder	0.1
		Learning rate for deep belief networks	0.02
Momentum parameters for deep belief networks	0.15

The main parameters of the other comparative methods are described below:

method 2 (standard depth autoencoder): the structure is selected to be 300-200-100-80, the learning rate is 0.01, and the number of pre-training iterations of each standard self-encoder is 500.

Method 3 (depth noise reduction self-coding): the structure is selected from 300-200-80, the learning rate and the noise loss coefficient are respectively 0.017 and 0.1, and the number of pre-training iterations of each noise reduction self-encoder is 500.

Method 4 (depth sparse self-editor): the structure is selected from 300-200-80, the learning rate, the parameter of the sparse penalty term constraint degree and the sparsity parameter are respectively 0.016, 0.1 and 0.15, and the pre-training iteration number of each sparse self-coding is 500.

Method 5 (depth-shrinking auto-encoder): the structure is selected from 300-200-100-80, the learning rate and the regularization coefficient are 0.025 and 0.05, respectively, and the number of pre-training iterations of each systolic self-encoder is 500.

Method 6 (convolutional neural network): the convolutional neural network has two convolutional layers, two pooling layers, and a fully-connected layer. The size of the input layers is 20 × 20, and the convolution kernels of the first convolution layer and the second convolution layer are 3 × 3 and 4 × 4 respectively; the step size for both pooling layers was set to 2, and the learning rate and number of iterations were 0.01 and 500, respectively.

Method 7 (BPNN): the structure is selected to be 28-40-6, the learning rate is 0.15, and the iteration number is 1000.

Method 8 (SVM): the kernel function is of a gaussian type and the penalty factor of the loss function is 0.54.

Method 9 (RF): the number of trees is 400, the maximum depth of the tree is 70, the minimum number of samples required to split an internal node is 70, and the minimum number of samples required to split a leaf node is 80.

Ninthly method 10(Softmax classifier): the learning rate was 0.25 and the number of iterations was 1000.

The multi-class confusion matrix is used as a method for measuring the performance of the deep learning model. FIGS. 11 and 12 show the multi-class confusion matrix of the present invention in other self-encoders for the test set in the first experiment. In fig. 11, the horizontal axis represents a prediction tag of a failure, the vertical axis represents a true tag of a failure, and diagonal elements represent probabilities that predicted values are equal to true values; fig. 11(a), (b), (c), and (d) are multi-class confusion matrices of depth-normalized self-coding, depth-denoised self-encoder, depth-sparse self-encoder, and depth-shrunk self-encoder, respectively. The color bars on the right correspond to the values of the multi-class confusion matrix. The accuracy of the experimental prediction label and the actual label is visually expressed through the multi-class confusion matrix. Compared with fig. 11, fig. 12 shows that the prediction accuracy of the invention for tag 1 is significantly improved, the prediction accuracy of

Claims

1. A gearbox intelligent fault diagnosis method with multi-scale structure and feature fusion is characterized by comprising the following steps:

step 1: designing a multi-scale depth feature extraction structure;

step 2: designing a depth feature fusion strategy based on information entropy;

and step 3: and identifying the health state based on the deep belief network.

2. The intelligent fault diagnosis method for the gearbox with the fusion of the multi-scale structure and the features as claimed in claim 1, wherein in the step 1, the design of the multi-scale depth feature extraction structure comprises the following specific steps:

step 1.1: constructing a plurality of self-encoders with different properties;

3. The intelligent fault diagnosis method for the gearbox with the multi-scale structure and the characteristic fusion function according to claim 1, wherein in the step 2, a depth characteristic fusion strategy design based on information entropy is specifically carried out by the following steps:

step 2.1: obtaining an evaluation matrix A;

step 2.3: and calculating the depth fusion feature H.

4. The intelligent gearbox fault diagnosis method based on multi-scale structure and feature fusion as claimed in claim 1, wherein in the step 3, the health state identification based on the deep belief network comprises the following specific steps:

step 3.1: pre-training a restricted Boltzmann machine model;

5. The intelligent fault diagnosis method for the gearbox with the multi-scale structure and the characteristic fusion function according to claim 2, wherein in the step 1.1, a plurality of self-encoders with different properties are constructed, and the specific steps are as follows:

wherein the content of the first and second substances,

6. The intelligent fault diagnosis method for the gearbox with the fused multi-scale structure and the features as claimed in claim 2, wherein in step 1.2, self-encoder models with different properties are combined in parallel to generate the multi-scale depth feature extraction structure, and the specific steps are as follows:

7. The intelligent fault diagnosis method for the gearbox with the multi-scale structure and the characteristic fusion function according to claim 3, wherein in the step 2.1, an evaluation matrix A is obtained, and the specific steps are as follows:

wherein A is_ijThe accuracy corresponding to the ith fault of the jth model is shown, when j is 1, the depth self-encoder is represented, and when j is 2, the depth sparse self-encoder is represented; when j is 3, the number of the adjacent groups is 3,representing a depth denoise autocoder; when j is 4, it represents a depth-punctured self-encoder, and d represents the number of types of failure.

8. The intelligent gearbox fault diagnosis method based on multi-scale structure and feature fusion as claimed in claim 3, wherein in the step 2.2, fusion weights of each model are calculated based on information entropy, and the specific steps are as follows:

step 2.2.1: defining the information entropy of the jth model as

Where j is 1, 2, 3, 4, d indicates the number of types of failure.

wherein, w_jSatisfy 0 < w_j＜1，w₁+w₂+w₃+w₄＝1。

9. The intelligent fault diagnosis method for the gearbox with the multi-scale structure and the characteristic fusion function according to claim 3, wherein in the step 2.3, a fusion characteristic matrix H is calculated, and the specific expression is as follows:

wherein H_jFor the depth feature learned for the jth model, j is 1, 2, 3, 4.

10. The intelligent fault diagnosis method for the gearbox with the multi-scale structure and the characteristic fused according to claim 4, wherein in the step 3.1, a restricted Boltzmann machine model is pre-trained, and the specific steps are as follows: