A fault diagnosis method based on asymmetric stacked sparse autoencoders considering data characteristics

Abstract

The present invention discloses an asymmetric stacked sparse autoencoder fault diagnosis method considering data features, constructs a semi-supervised asymmetric stacked sparse autoencoder structure, adopts a convolutional pooling layer as an encoder, and a fully connected layer as a decoder to enhance the feature extraction capability of the network model; at the same time, a semi-supervised structure is adopted to introduce label information to improve the classification capability of the autoencoder; on the basis of the autoencoder, data feature constraints are considered, and the intra-class aggregation and inter-class separation are constrained by the nearest neighbor sample to improve the feature extraction performance of the network; variance distance and nearest neighbor sample distance are adopted to constrain the feature reconstruction direction of the feature detector to ensure the validity of the features extracted by the network model; finally, a dynamic weight is used to dynamically optimize the training process to ensure the convergence of the data feature constraints. The method of the present invention can effectively extract fault features, obtain good diagnostic accuracy, and realize reliable rolling bearing fault diagnosis.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent fault diagnosis of mechanical parts, and in particular relates to an asymmetric stacked sparse autoencoding fault diagnosis method considering data characteristics. Background Art

[0002] As a key component in rotating machinery, the working state and health of rolling bearings have attracted much attention. Since the working environment of high-speed rotating machinery is usually harsh and complex, rolling bearings are prone to failures when working for a long time under complex working conditions, which in turn affects the overall operation of the machinery. In order to avoid the significant impact of bearing failures on the safety and stability of mechanical operation, it is necessary to perform health monitoring and fault diagnosis on rolling bearings.

[0003] In recent years, fault diagnosis has entered the era of big data, and it is necessary to extract useful information from massive data. The autoencoder network is a typical unsupervised network that can easily reduce the dimension of data features. Therefore, this method has been widely used in bearing fault diagnosis research. However, due to its simple structure, the feature extraction performance of the traditional autoencoder network is limited when facing a variety of complex signals. In practical applications, it is still necessary to improve the traditional structure. At present, many improvement schemes have been produced based on the traditional autoencoder network to meet different application requirements, but there are still some problems that must be considered in practical applications.

[0004] First, due to the complexity and correlation of mechanical structures, fault signal features overlap, and how to use neural networks to efficiently and accurately learn the features of different faults is a key issue. Secondly, deep learning has a black box characteristic, and most intelligent diagnostic algorithms extract fault features without considering the original data feature structure of the sample data, and the direction of data feature extraction during training is uncontrollable. Third, how to avoid network overfitting. These problems have brought challenges to traditional intelligent fault diagnosis algorithms. To this end, the present invention proposes an asymmetric stacked sparse autoencoding fault diagnosis method that takes data features into account. Summary of the invention

[0005] Purpose of the invention: The present invention proposes an asymmetric stacked sparse autoencoding fault diagnosis method that takes data characteristics into consideration, which can constrain the network feature learning direction, improve the network feature extraction capability, improve the model diagnosis accuracy, and realize reliable rolling bearing fault diagnosis.

[0006] Technical solution: The present invention provides an asymmetric stacked sparse autoencoder fault diagnosis method considering data characteristics, comprising the following steps:

[0007] (1) Vibration signal acquisition: Place vibration acceleration sensors at key measuring points to measure the vibration signal data of the rolling bearing in normal state, inner ring fault, outer ring fault, and rolling element fault states;

[0008] (2) Data preprocessing: The collected vibration signals are converted into time-frequency data using fast Fourier transform. The converted frequency domain signal data of different fault types are then sorted into sample data sets and divided into training sets and test sets.

[0009] (3) Build an asymmetric stacked sparse autoencoder model and initialize parameter settings: use two layers of convolutional layers in series as the encoding layer of the asymmetric stacked autoencoder, and three layers of fully connected layers as the decoding layer of the asymmetric stacked autoencoder; at the same time, connect a fully connected layer and a SoftMax classifier after the encoding layer to form a semi-supervised asymmetric stacked sparse autoencoder structure; introduce KL divergence to constrain the sparsity of the encoding layer;

[0010] (4) Introducing data feature loss into the asymmetric stacked sparse autoencoder model further constrains the direction of model data reconstruction, where the nearest neighbor samples are calculated using the K nearest neighbor algorithm;

[0011] (5) Inputting the training set into the model built in step (4) for model training, optimizing the total loss function by a dynamic weight method, and iterating the network weights according to the gradient descent method;

[0012] (6) Perform batch training according to the set number of iterations. Each training step extracts the same number of batches of data with different serial numbers. When the number of training steps meets the initial set number of iterations, stop training and save the model.

[0013] (7) Input the test set into the saved model to test the model performance, output the diagnostic results, and realize the fault diagnosis of rolling bearings.

[0014] Furthermore, the data preprocessing process in step (2) includes performing linear normalization processing on the frequency domain sample data set.

[0015] Furthermore, the step (3) comprises the following steps:

[0016] (31) The loss function of the semi-supervised asymmetric stacked sparse autoencoder structure consists of reconstruction error, cross entropy loss, and sparsity penalty term;

[0017] (32) The semi-supervised asymmetric stacked sparse autoencoder retains the unsupervised characteristics of the autoencoder and adaptively reconstructs samples to extract sample features. Its reconstruction error loss L AE It is expressed as follows:

[0018]

[0019] Among them, m is the number of samples, x i Input sample data for group i, Reconstruct sample data for the i-th group;

[0020] (33) Construct a sparse autoencoder. The asymmetric stacked sparse autoencoder model uses KL divergence to constrain the sparsity of the network to avoid network overfitting. The network sparse penalty term L KL for:

[0021]

[0022]

[0023] Where N is the number of neurons in the hidden layer, β is the sparse weight factor, is the average activation level of the jth neuron, and ρ is the expected activation level;

[0024] (34) The asymmetric stacked sparse autoencoder model introduces the true label y, constructs a semi-supervised structure, and uses the cross entropy function to measure the predicted label To enhance the classification performance of the network, when the number of samples is m, the cross entropy loss L sem for:

[0025]

[0026] Among them, y i is the true label of the i-th group of samples, is the predicted label of the i-th group of samples.

[0027] Furthermore, the data feature loss in step (4) includes data proximity aggregation constraint, data long-distance separation constraint and data nearest neighbor overall constraint; the data proximity aggregation constraint L dk for:

[0028]

[0029] Where h(x) refers to the feature layer of asymmetric stacked sparse autoencoders, KNN(x,j) means calculating the nearest neighbor sample of sample x in the data set that is closest to the jth nearest neighbor, and MMD(A,B) means calculating the maximum mean difference distance between sample A and sample B.

[0030] The data is remotely separated from the constraint L dks for:

[0031]

[0032] The data nearest neighbor overall constraint is divided into a distance constraint and a variance constraint; the distance constraint L gdand variance constraint L gv They are:

[0033] L gd =∑||h(x)-KNN(h(x),1)|| 2

[0034] L gv =||σ(h(x))-σ(KNN(h(x),1))|| 2

[0035] Among them, σ(x) represents the variance of the calculated sample x;

[0036] Therefore, the total loss of data features is L D for:

[0037]

[0038] Furthermore, the dynamic weighting method described in step (5) is as follows:

[0039] μ=u(L)=c·L

[0040] Among them, c is the weight factor, and μ is the dynamic weight of the loss term L, which changes with the size of the loss term L to ensure the convergence of the loss term during training;

[0041] λ1 and λ2 are weight factors, and the total loss function L of the asymmetric stacked sparse autoencoder model is sum It is expressed as follows:

[0042] L sum =λ1L AE +λ2L sem +L KL +u(L D )·L D

[0043] Among them, L AE is the reconstruction error loss, L KL is the network sparse penalty term, L sem is the cross entropy loss, L D is the total loss of data features.

[0044] Beneficial effects: Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention proposes an asymmetric autoencoder structure, adopts a convolutional layer instead of a fully connected layer, and enhances the feature extraction capability of the network; the present invention fully considers the direction of network reconstruction of data features, provides data feature loss constraints, adopts nearest neighbor samples to enhance the aggregation within the data class and the separation outside the class, and at the same time applies variance distance to constrain the direction of network feature extraction; the present invention can constrain the network feature learning direction, improve the network feature extraction capability, improve the model diagnosis accuracy, and realize reliable rolling bearing fault diagnosis by constructing a semi-supervised asymmetric stacked sparse autoencoder structure. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is the overall flow chart of the present invention;

[0046] Figure 2 This is a schematic diagram of the asymmetric stacked autoencoder network model structure;

[0047] Figure 3 It is a complete structural schematic diagram of the method proposed by the present invention;

[0048] Figure 4 This is a comparison chart of the experimental accuracy of different methods on four sets of data sets;

[0049] Figure 5 This is a diagram showing the effect of feature extraction of the method proposed in the present invention on data set A. DETAILED DESCRIPTION

[0050] The present invention is further described in detail below with reference to the accompanying drawings.

[0051] like Figure 1 As shown, the present invention provides an asymmetric stacked sparse autoencoder fault diagnosis method considering data characteristics, and the specific steps are as follows:

[0052] Step 1: vibration signal acquisition. Vibration acceleration sensors are placed at key measuring points to measure the vibration signal data of the rolling bearing in normal state, inner ring fault, outer ring fault and rolling element fault state.

[0053] Step 2, data preprocessing, uses fast Fourier transform to convert the collected vibration signal into time-frequency; then organizes the converted frequency domain signal data of different fault types into sample data sets, and divides them into training sets and test sets according to the division ratio. The data preprocessing process includes linear normalization of the frequency domain sample data sets.

[0054] Step 3: Build an asymmetric stacked sparse autoencoder model and initialize the parameter settings. The structure diagram of the asymmetric stacked autoencoder model is as follows: Figure 2As shown in the figure, two layers of convolution-pooling layers are used as the encoding layer of the asymmetric stacked autoencoder, and three layers of fully connected layers are used as the decoding layer of the asymmetric stacked autoencoder to form a complete basic autoencoder structure. At the same time, a fully connected layer and a SoftMax classifier are connected after the encoding layer to form a semi-supervised asymmetric stacked sparse autoencoder structure. Finally, the KL divergence is introduced to constrain the sparsity of the encoding layer.

[0055] The loss function of the semi-supervised asymmetric stacked sparse autoencoder structure is mainly composed of reconstruction error, cross entropy loss and sparse penalty term, and its specific composition is as follows:

[0056] The semi-supervised asymmetric stacked sparse autoencoder retains the unsupervised characteristics of the autoencoder and adaptively reconstructs samples to extract sample features. Its reconstruction error loss L AE It is expressed as follows:

[0057]

[0058] Among them, m is the number of samples, x i Input sample data for group i, Reconstruct sample data for the i-th group.

[0059] Construct a sparse autoencoder. The asymmetric stacked sparse autoencoder model uses KL divergence to constrain the sparsity of the network to avoid overfitting of the network. The network sparse penalty term L KL It can be expressed as:

[0060]

[0061]

[0062] Where N is the number of neurons in the hidden layer, β is the sparse weight factor, is the average activation level of the jth neuron, and ρ is the expected activation level.

[0063] The asymmetric stacked sparse autoencoder model introduces the true label y, constructs a semi-supervised structure, and uses the cross entropy function to measure the predicted label To enhance the classification performance of the network, when the number of samples is m, the cross entropy loss L sem for:

[0064]

[0065] Among them, y i is the true label of the i-th group of samples, is the predicted label of the i-th group of samples.

[0066] Step 4: Introduce data feature loss into the model of step 3 to further constrain the direction of model data reconstruction, where the nearest neighbor samples are calculated using the K nearest neighbor algorithm.

[0067] Data feature loss mainly includes data proximity aggregation constraints, data long-distance separation constraints, and data nearest neighbor overall constraints. The specific process is as follows:

[0068] Data proximity aggregation constraint: The K nearest neighbor algorithm is used to calculate the three groups of nearest neighbor samples closest to the sample. During the training process, the distance between the sample and the three groups of nearest neighbor samples closest to its initial distance is minimized as much as possible. The maximum mean difference (MMD) is used to measure the distance between samples. Data proximity aggregation constraint L dk It can be expressed as follows:

[0069]

[0070] Where h(x) refers to the feature layer of the asymmetric stacked sparse autoencoder, KNN(x,j) means calculating the nearest neighbor sample of sample x in the data set that is closest to the jth neighbor, and MMD(A,B) means calculating the maximum average difference distance between sample A and sample B.

[0071] The data long-distance separation constraint uses the K nearest neighbor algorithm to calculate the three groups of nearest neighbor samples farthest from the sample. During the training process, try to ensure that the distance between the sample and its three groups of nearest neighbor samples farthest from the initial distance is the largest. The maximum average difference is also used to measure the distance between samples. When the total number of samples is m, the data long-distance separation constraint L dks It can be expressed as follows:

[0072]

[0073] The overall constraints of the nearest neighbor of the data are mainly divided into distance constraints and variance constraints. For distance constraints, the K nearest neighbor algorithm is used to calculate a group of nearest neighbor samples that are closest to the sample, and the distance between the two is kept to a minimum during the training process. The distance constraint L gd It can be expressed as:

[0074] L gd =∑||h(x)-KNN(h(x),1)|| 2

[0075] Variance constraint, calculate the variance of the nearest neighbor sample data and the variance distance of the feature layer of the nearest neighbor sample extracted by the encoding layer, and ensure that the difference in variance is minimized during the training process. σ(x) represents the variance of the calculated sample x, then the variance constraint L gv It can be expressed as:

[0076] L gv=||σ(h(x))-σ(KNN(h(x),1))|| 2

[0077] The total loss of data features L D It can be expressed as follows:

[0078]

[0079] Step 5: Input the training set into the model built in step 4 to start training. At the same time, the total loss function is optimized by dynamic weights, and the network weights are iterated according to the gradient descent method. The complete model structure of this method is as follows: Figure 3 As shown in the figure, an asymmetric stacked sparse autoencoder structure is constructed to enhance the feature extraction performance, and data feature loss is introduced to constrain the feature reconstruction direction. At the same time, the dynamic weight method and semi-supervised mode are used to improve the network convergence and classification performance.

[0080] The dynamic weighting method is as follows:

[0081] μ=u(L)=c·L

[0082] Among them, c is the weight factor, and μ is the dynamic weight of the loss term L, which changes with the size of the loss term L to ensure the convergence of the loss term during training.

[0083] λ1 and λ2 are weight factors, and the total loss function L of the proposed model is sum It is expressed as follows:

[0084] L sum =λ1L AE +λ2L sem +L KL +u(L D )·L D .

[0085] Step 6: Perform batch training according to the set number of iterations. Extract batch data of the same quantity but different serial numbers for each training. When the number of training steps meets the initially set number of iterations, stop training and save the model.

[0086] Step 7: Input the test set into the saved model to perform model performance testing, output the diagnosis results, and implement rolling bearing fault diagnosis.

[0087] The present invention selects the rolling bearing public data set (CWRU) of the bearing test bench of Case Western Reserve University in the United States for testing and verification. The experimental bearing is a 6205-RS JEM SKF deep groove ball bearing, and the drive end data at a sampling frequency of 48KHz is selected. Three types of faults are mainly selected, namely inner ring fault, outer ring fault and rolling element fault, and each fault is divided into three fault degrees. A total of 4 loads were tested, and the load conditions were 0HP, 1HP, 2HP and 3HP respectively. Each load state contains a normal state data set with a total of ten data categories. Data acquisition adopts a data enhancement method, that is, overlapping sampling with a small sample length, so as to increase the number of samples. There are 600 groups of samples for each fault category, and the length of each sample group is 600. The ratio of training set to test set is 1:1. The detailed division of data is shown in Table 1.

[0088] Table 1. Western Reserve University dataset partition table

[0089]

[0090] The parameter setting of the proposed asymmetric stacked sparse autoencoder model considering data characteristics mainly includes the parameters of the encoding layer, decoding layer and classification layer. The specific parameters are shown in Table 2, where the pooling layer adopts maximum pooling.

[0091] Table 2 Parameter settings of asymmetric stacked sparse autoencoder model considering data characteristics

[0092]

[0093]

[0094] Except for the model structure parameters, other training parameter settings are shown in Table 3.

[0095] Table 3 Model training parameter settings

[0096]

[0097] The present invention and other different methods are tested on four sets of data sets, and the experimental results are shown in Table 4. In order to reduce the influence of random factors, the data sets in the table are the average results of 10 experiments. It can be seen that although all the comparison methods have achieved an average accuracy greater than 85% on the four sets of data sets, the proposed method can achieve an average accuracy greater than 99% on the four sets of data sets, indicating that the proposed method has good stability and reliability.

[0098] Table 4 Experimental results of different methods on the dataset

[0099]

[0100] Furthermore, the average results of different methods on four sets of data sets are visualized, as shown in Figure 4 As shown in the figure. It can be seen that the diagnostic accuracy of BNAE and LGSSAE is very similar, but the stability of BNAE is stronger than that of LGSSAE. The highest diagnostic accuracy of FDSAE ​​is similar to that of DisAE, but its stability and reliability are far inferior to DisAE. Compared with the above four methods, the proposed method shows similar accuracy in all four data sets, and its stability is significantly greater than that of the other four methods, indicating that the proposed method has strong robustness.

[0101] In order to explore the performance of this method in feature extraction, the t-SNE graph is used to visualize the feature layer. The features extracted by the proposed method in dataset A are as follows Figure 5 shown. Figure 5 The labels in the legend correspond strictly to the fault descriptions in Table 1, and different fault data are marked with different icons and colors. Among them, C0-C2 represents three different degrees of inner ring fault, C3-C5 represents three different degrees of outer ring fault, C6-C8 represents three different degrees of rolling element fault, and C9 represents the normal condition of the bearing. As you can see, Figure 5 The same markers are well aggregated together, while the distances between different markers are obvious, reflecting good classification ability. Therefore, for multi-class sample data, this method can well aggregate the same data and separate the out-of-class data, showing reliable feature extraction performance.

[0102] In summary, the present invention constructs a semi-supervised asymmetric stacked sparse autoencoder model, selects convolutional layer as encoder, and fully connected layer as decoder to enhance the feature extraction capability of encoder. Label information is introduced to enhance the classification capability of the network. Considering the intra-class distribution, the nearest neighbor samples are used to measure the class aggregation and separation, and the neighboring features and data features of the samples are constrained at the same time to enhance the feature separation performance of the network. Finally, a dynamic weight is used to dynamically optimize the training process to ensure the training balance direction and convergence. It has been verified by the implementation case that the method of the present invention can realize reliable rolling bearing fault diagnosis.

[0103] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for fault diagnosis of asymmetric stacked sparse autoencoders considering data characteristics, characterized in that: The following steps are involved: (1) Vibration signal acquisition: Place vibration acceleration sensors at key measuring points to measure the vibration signal data of the rolling bearing in normal state, inner ring fault, outer ring fault, and rolling element fault states; (2) Data preprocessing: The collected vibration signals are converted into time-frequency data using fast Fourier transform. The converted frequency domain signal data of different fault types are then sorted into sample data sets and divided into training sets and test sets. (3) Build an asymmetric stacked sparse autoencoder model and initialize parameter settings: use two layers of convolutional layers in series as the encoding layer of the asymmetric stacked autoencoder, and three layers of fully connected layers as the decoding layer of the asymmetric stacked autoencoder; at the same time, connect a fully connected layer and a SoftMax classifier after the encoding layer to form a semi-supervised asymmetric stacked sparse autoencoder structure; introduce KL divergence to constrain the sparsity of the encoding layer; (4) Introducing data feature loss into the asymmetric stacked sparse autoencoder model further constrains the direction of model data reconstruction, where the nearest neighbor samples are calculated using the K nearest neighbor algorithm; (5) Inputting the training set into the model built in step (4) for model training, optimizing the total loss function by a dynamic weight method, and iterating the network weights according to the gradient descent method; (6) Perform batch training according to the set number of iterations. Each training step extracts the same number of batches of data with different serial numbers. When the number of training steps meets the initial set number of iterations, stop training and save the model. (7) Input the test set into the saved model to perform model performance testing, output the diagnosis results, and realize rolling bearing fault diagnosis; The data feature loss in step (4) includes data proximity aggregation constraint, data long-distance separation constraint and data nearest neighbor overall constraint; the data proximity aggregation constraint L dk for: Where h(x) refers to the feature layer of asymmetric stacked sparse autoencoders, KNN(x,j) means calculating the nearest neighbor sample of sample x in the data set that is closest to the jth nearest neighbor, and MMD(A,B) means calculating the maximum mean difference distance between sample A and sample B. The data is remotely separated from the constraint L dks for: The data nearest neighbor overall constraint is divided into a distance constraint and a variance constraint; the distance constraint L gd and variance constraint L gv They are: L gd =∑||h(x)-KNN(h(x),1)|| 2 L gv =||σ(h(x))-σ(KNN(h(x),1))|| 2 Among them, σ(x) represents the variance of the calculated sample x; Therefore, the total loss of data features is L D for:

2. According to the asymmetric stacked sparse autoencoder fault diagnosis method considering data characteristics according to claim 1, it is characterized in that: The data preprocessing process in step (2) includes performing linear normalization processing on the frequency domain sample data set.

3. The asymmetric stacked sparse autoencoder fault diagnosis method considering data characteristics according to claim 1, characterized in that: The step (3) comprises the following steps: (31) The loss function of the semi-supervised asymmetric stacked sparse autoencoder structure consists of reconstruction error, cross entropy loss, and sparsity penalty term; (32) The semi-supervised asymmetric stacked sparse autoencoder retains the unsupervised characteristics of the autoencoder and adaptively reconstructs samples to extract sample features. Its reconstruction error loss L AE It is expressed as follows: Among them, m is the number of samples, x i Input sample data for group i, Reconstruct sample data for the i-th group; (33) Construct a sparse autoencoder. The asymmetric stacked sparse autoencoder model uses KL divergence to constrain the sparsity of the network to avoid network overfitting. The network sparse penalty term L KL for: Where N is the number of neurons in the hidden layer, β is the sparse weight factor, is the average activation level of the jth neuron, and ρ is the expected activation level; (34) The asymmetric stacked sparse autoencoder model introduces the true label y, constructs a semi-supervised structure, and uses the cross entropy function to measure the predicted label To enhance the classification performance of the network, when the number of samples is m, the cross entropy loss L sem for: Among them, y i is the true label of the i-th group of samples, is the predicted label of the i-th group of samples.

4. The asymmetric stacked sparse autoencoder fault diagnosis method considering data characteristics according to claim 1, characterized in that: The dynamic weight method described in step (5) is as follows: μ=u(L)=c·L Among them, c is the weight factor, and μ is the dynamic weight of the loss term L, which changes with the size of the loss term L to ensure the convergence of the loss term during training; λ1 and λ2 are weight factors, and the total loss function L of the asymmetric stacked sparse autoencoder model is sum It is expressed as follows: L sum =λ1L AE +λ2L sem +L KL +u(L D )·L D Among them, L AE is the reconstruction error loss, L KL is the network sparse penalty term, L sem is the cross entropy loss, L D is the total loss of data features.

Images