A composite fault diagnosis method based on feature enhancement and double-layer random architecture

Abstract

This invention discloses a composite fault diagnosis method based on feature enhancement and a two-layer stochastic architecture: A bearing vibration signal is acquired, and a shallow feature matrix is ​​generated through local mean decomposition and temporal statistical feature extraction; a Transformer network is constructed for multi-label classification pre-training to extract a deep feature matrix containing global dependencies; physical feature templates for cloud models of each category are constructed based on the deep features, and the similarity matrix between the test sample and the template is calculated; a counting model and a label model are constructed separately and trained using a stochastic incremental learning algorithm with global constraints; based on the number of faults output by the counting model, the probability matrix of the label model and the similarity matrix of the cloud model are fused for decision-making, and the diagnostic result is output. This invention solves the problem of insufficient model feature extraction capability, overcomes the label dimension explosion problem faced when multiple concurrent faults occur, and compensates for the low classification accuracy of a single model, achieving efficient and accurate diagnosis of single and composite bearing faults.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis technology, and in particular to a method for diagnosing composite bearing faults based on feature enhancement and a two-layer random architecture. Background Technology

[0002] Rolling bearings are core components of rotating machinery, and their health directly affects the safety and stable operation of the entire electromechanical system. Under complex and variable industrial conditions, bearings are prone to concurrent damage in multiple areas, and their early, weak characteristics are often obscured by strong background noise and intense nonlinear interference signals. Therefore, conducting efficient and accurate research on complex fault diagnosis is crucial. In recent years, deep learning models have been widely applied to bearing fault diagnosis, capable of automatically extracting deep features containing high-order discriminative information. However, traditional deep networks typically use gradient-based backpropagation algorithms to update parameters, which suffers from long training times and susceptibility to getting trapped in local optima. To address these issues, incremental random weighted networks, as a novel type of feedforward neural network, have been introduced. However, traditional incremental networks still face significant challenges when processing complex, high-dimensional industrial data: they are essentially shallow architectures, heavily reliant on the quality of input feature representations, and lack the ability to extract deep features for strongly coupled operating conditions.

[0003] On the other hand, when dealing with multi-label identification of complex faults and the dynamic evolution of network structures, existing diagnostic models generally suffer from a triple defect: rigid classification architecture, simplistic decision-making mechanism, and redundant nodes. At the classification and diagnosis level, traditional networks often employ a single-layer closed-set classification architecture. Their mutually exclusive logic forcibly defines complex faults as independent new categories, which not only easily leads to the dimensionality curse caused by "label combination explosion" but also severs the physical essence of concurrent faults being formed by the superposition of multiple single damages. Furthermore, in the final diagnostic decision-making stage, traditional methods often heavily rely on the classification output of a single network model, making them prone to misjudgment when faced with feature distribution shifts caused by extreme variable conditions. At the network structure evolution level, the construction process of conventional incremental random networks lacks effective supervision based on a global perspective, and the introduction of candidate nodes often manifests as blind random mapping. This directional selection mechanism, lacking global inequality error constraints, easily leads to the system accepting a large number of redundant nodes with limited contribution to reducing global residuals, causing the generated network structure to continuously increase in size and weakening its compactness and lightweight advantages. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a composite fault diagnosis method based on feature enhancement and a two-layer random architecture, which solves a series of technical problems in existing technologies, such as insufficient deep feature extraction capabilities, the tendency of concurrent faults to cause label dimension explosion, poor robustness of single model decision-making, and bloated structure and lack of global supervision during incremental network evolution.

[0005] This invention provides a composite fault diagnosis method based on feature enhancement and a two-layer random architecture, comprising the following steps:

[0006] Step 1: Acquire bearing vibration signals and extract shallow feature matrices using local mean decomposition and time-domain statistical methods;

[0007] Step 2: Construct a Transformer network based on self-attention mechanism, perform multi-label classification pre-training with shallow feature matrix as input, and extract deep feature matrix containing global dependencies after solidifying network parameters;

[0008] Step 3: Construct cloud model physical feature templates for each type of fault based on deep features, and calculate the similarity matrix between the sample to be tested and each cloud model physical feature template;

[0009] Step 4: Construct a counting model for predicting the number of faults and a labeling model for predicting specific fault categories, respectively. Use the deep feature matrix as input and train the models using a stochastic incremental learning algorithm with global constraints.

[0010] Step 5: Based on the number of faults output by the counting model, the output probability matrix of the label model and the similarity matrix are fused to obtain a weighted fusion confidence matrix, and the final multi-label fault diagnosis result is output according to the weighted fusion confidence matrix.

[0011] Furthermore, the specific process of step 1 is as follows:

[0012] The bearing vibration signal is sliced ​​by a sliding window, and multiple product function components are obtained by extreme point envelope calculation. The 10-dimensional time-domain statistical features of the first k PF components are calculated respectively. The k×10-dimensional time-domain statistical features extracted from each sample are normalized to generate a shallow feature matrix. The time-domain statistical features include: mean, standard deviation, kurtosis, peak value, root mean square, impulse factor, waveform factor, margin factor, skewness, and peak value factor.

[0013] Further, step 2 includes the following steps:

[0014] Step 21: Map the shallow feature matrix to a high-dimensional feature space through a fully connected embedding layer. The specific formula is as follows:

[0015] ;

[0016] In the formula, For embedding feature sequences; and Assign its weights and biases;

[0017] Step 22: Input the embedded feature sequence into the multi-head self-attention layer to extract coupling features, and then normalize the distribution through the first normalization layer to obtain the initial normalized features. The specific formula is as follows:

[0018] ;

[0019] In the formula, These are the initial normalized features; This is a coupling characteristic; and These are the mean and variance, respectively. and These are the learnable parameters;

[0020] Step 23: The initial normalized features are transformed into higher-order nonlinear features through a two-layer fully connected feedforward network with ReLU activation, and then the normalized deep feature matrix is ​​output through the second normalization layer.

[0021] Step 24: Use the multi-label one-hot matrix of composite faults as the supervision signal to perform network regression pre-training. During the training process, after the second normalization layer, a fully connected output layer and a Sigmoid activation layer are cascaded to decouple and map the predicted probabilities of each independent fault category.

[0022] After training, the network parameters are fixed and the output fully connected layer and Sigmoid activation layer are stripped. The output of the second normalized layer is directly forward propagated as the deep feature matrix.

[0023] Furthermore, in step 3, a cloud model feature template is generated based on deep features, wherein the expected value in the cloud model feature template is calculated using the following formula:

[0024] ;

[0025] The formula for calculating entropy in the cloud model feature template is:

[0026] ;

[0027] The formula for calculating hyperentropy in the cloud model feature template is:

[0028] ;

[0029] In the formula, For the first The number of class samples; For the next category Deep feature vectors of each sample; To calculate the variance by column; It is a preset, extremely small positive number;

[0030] The dimensional membership degree calculated using the cloud model feature template is:

[0031] ;

[0032] In the formula, x represents the sample;

[0033] Calculate the similarity matrix between the sample to be tested and the physical feature templates of each cloud model, specifically as follows:

[0034] First, calculate the similarity vector based on the membership dimension. The formula is:

[0035] ,

[0036] in, ;

[0037] Then, obtain the similarity matrix based on the similarity vector.

[0038] Furthermore, the stochastic incremental learning construction method with global constraints in step 4 is specifically as follows:

[0039] We begin by building a small network. Several candidate nodes are randomly generated within a range of random parameters. The generated nodes are then globally constrained using the Greville iteration method. Candidate nodes that satisfy the constraints are sorted, and the node with the largest constraint value is selected as the new node. The output weights and residuals after adding the new node are as follows:

[0040] ;

[0041] ;

[0042] ;

[0043] ;

[0044] ;

[0045] ,

[0046] in, ;

[0047] ;

[0048] In the formula, These are the initial weights; The initial residual; For regularization parameters; For having The output of the hidden layer of each node; For having The output of the hidden layer of each node; These are the extracted deep features; Output the target.

[0049] Furthermore, the global constraint expression is:

[0050] ,

[0051] in, ;

[0052] In the formula, , , For learning parameters;

[0053] Furthermore, in step 4, when there are no candidate nodes in the candidate pool that meet the constraints, the learning parameters are scaled, and the generation range of the random parameters is expanded. The scaling formula for the learning parameters is:

[0054] ;

[0055] In the formula, for Random numbers within.

[0056] Furthermore, in step 5, the weighted fusion confidence matrix is:

[0057] ;

[0058] In the formula, It is a probability matrix; This is a similarity matrix; These are weight hyperparameters;

[0059] The fusion decision is:

[0060] ;

[0061] In the formula, These are the predicted values ​​from the counting model; X represents the label dimension; X represents the deep feature vector.

[0062] The beneficial effects of this invention are:

[0063] This invention first utilizes the powerful multi-head self-attention mechanism of the Transformer network to overcome the shallow representation bottleneck of traditional incremental networks, adaptively capturing the nonlinear cross-coupling weights between various indicators, and significantly improving the ability to extract deep features from complex high-dimensional industrial data.

[0064] This invention innovatively constructs a "counting-location" dual-layer diagnostic structure, which completely breaks the rigid architecture of traditional single-layer closed-set classification. It flexibly and accurately identifies and discovers multiple parts and phenomena, just like "building blocks". It perfectly matches the physical nature of multiple faults being superimposed by a single damage, and effectively avoids the problem of "label combination explosion".

[0065] This invention introduces a segmented weighted multidimensional joint decision-making mechanism, which uses the similarity between the predicted score of the label model and the spatial overlap of the cloud model to complement information, completely overcoming the defect that a single network is prone to misjudgment under extreme variable conditions, and greatly enhancing the robustness of system diagnosis.

[0066] This invention employs a stochastic incremental learning algorithm with global constraints for model training.

[0067] In summary, this invention successfully constructs an intelligent diagnostic paradigm that takes into account deep feature extraction, accurate separation of concurrent faults, and complementary multi-dimensional decision-making, ultimately providing an effective solution for health monitoring of rotating machinery under complex working conditions. Attached Figure Description

[0068] The features and advantages of the invention will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the invention in any way. In the drawings:

[0069] Figure 1 This is a flowchart of a specific embodiment of the present invention;

[0070] Figure 2 This is a structural diagram of the bearing fault diagnosis model in a specific embodiment of the present invention;

[0071] Figure 3 This is a confusion matrix diagram of the bearing fault diagnosis model in a specific embodiment of the present invention. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] The present invention will be further illustrated below with reference to specific embodiments. Those skilled in the art should understand that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Modifications to the present invention in various equivalent forms all fall within the scope defined by the appended claims.

[0074] like Figure 1 , Figure 2 As shown, this invention provides a composite fault diagnosis method based on feature enhancement and a two-layer random architecture, comprising the following steps:

[0075] Step 1: Select the bearing fault dataset, which includes data for four types of faults: normal state, inner ring fault, rolling element fault, outer ring fault, and compound fault (inner ring + outer ring) under two different speed and load combinations: 20Hz-0Nm and 30Hz-20Nm.

[0076] Local mean decomposition was employed, selecting the first three Power Factor (PF) components. Ten-dimensional time-domain statistical features were calculated for each component, including: mean, standard deviation, kurtosis, peak value, root mean square (RMS), impulse factor, waveform factor, margin factor, skewness, and kurtosis factor. The extracted values ​​for each sample were then... The statistical features are normalized to the [0,1] interval to generate a shallow feature matrix. .

[0077] Step 2: Construct a Transformer network based on a self-attention mechanism. Use the shallow feature matrix as input for multi-label classification pre-training. After fixing the network parameters, extract the deep feature matrix containing global dependencies. The specific process includes:

[0078] Step 21: The feature space is mapped to a high-dimensional feature space through a fully connected embedding layer. The specific formula is as follows:

[0079] ;

[0080] In the formula, For embedding feature sequences; and Assign its weights and biases;

[0081] Step 22: Input multi-head self-attention layer extracts coupled features Then, the distribution is standardized by connecting to the first normalization layer to obtain the initial normalized features:

[0082] ;

[0083] In the formula, These are the initial normalized features; and The mean and variance are given. and These are learnable parameters;

[0084] Step 23: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the A high-order nonlinear transformation of features is performed through a two-layer fully connected feedforward network with ReLU activation, followed by a normalized deep feature matrix output through a second normalization layer. :

[0085] ,

[0086] in, ;

[0087] In the formula, Recombination characteristics; , , , These are the parameters of the perceptron; and These are the parameters for the second normalized layer;

[0088] Step 24: Use the multi-label one-hot matrix of the composite fault as the supervision signal for network regression pre-training. During training, after the second normalization layer, a fully connected output layer and a sigmoid activation layer are cascaded to decouple and map the predicted probabilities of each independent fault category. :

[0089] ;

[0090] After training, the network parameters are fixed and the output fully connected layer and Sigmoid activation layer are stripped. The output of the second normalized layer is directly forward propagated as the deep feature matrix.

[0091] Step 3: Construct cloud model physical feature templates for each type of fault based on deep features, and calculate the similarity matrix between the test sample and each cloud model physical feature template, specifically:

[0092] Based on the deep feature sample set The three core digital features of the cloud model are calculated according to category: expectation. ,entropy and hyperentropy ,

[0093] expect The formula reflecting the distribution center of this type of fault in the deep feature space is as follows:

[0094] ;

[0095] entropy The degree of dispersion and ambiguity of the deep feature distribution is reflected by an unbiased estimate of the first-order absolute central moment.

[0096] ;

[0097] hyperentropy Reflecting entropy The uncertainty and dispersion are calculated by introducing a lower bound minimum constraint to prevent negative numbers within the square root caused by variance drift:

[0098] ;

[0099] In the formula, For the first The number of class samples; For the next category Deep feature vectors of each sample; To calculate the variance by column; Using a preset minimum normal number, a comprehensive cloud model feature template containing different health states and single / compound faults is constructed.

[0100] Using the constructed physical feature templates of various cloud models, the dimensionality membership degree of the test sample features in the cloud space of each fault category is calculated. The specific formula is as follows:

[0101] ;

[0102] In the formula, x represents the sample;

[0103] Calculate the current sample to be diagnosed With the Overall overlap similarity of cloud-like model templates Its value is the arithmetic mean of the membership degrees of all feature dimensions of the sample:

[0104] ;

[0105] Regarding the first After calculating the overlap similarity of each sample to be diagnosed for all categories, the similarity matrix is ​​normalized row by row to stabilize its values ​​within the [0,1] interval, in order to align with the similarity vector output by the subsequent labeling model. The formula is as follows:

[0106] ;

[0107] Then, obtain the similarity matrix based on the similarity vector.

[0108] Step 4: Construct a counting model to predict the number of faults and a labeling model to predict specific fault categories, respectively. Using the deep feature matrix as input, train the models using a stochastic incremental learning algorithm with global constraints. The specific method for constructing the weights and residuals in the stochastic incremental learning algorithm is as follows:

[0109] Initialize the current weights of the stochastic incremental learning algorithm. and residual Specifically:

[0110] ;

[0111] ;

[0112] First within the range of random parameters Internal random generation The generated nodes are selected as candidate nodes, and then the Greville iteration method is used to perform global constraint screening on the generated nodes. Nodes that meet the constraints are added to the candidate pool, and the node with the largest constraint value is selected as the new node. The output weights and residuals after adding the new node are as follows:

[0113] ;

[0114] ;

[0115] ;

[0116] ;

[0117] ;

[0118] ;

[0119] In the formula, These are the initial weights; The initial residual; For regularization parameters; For having The output of the hidden layer of each node; For having The output of the hidden layer of each node; These are the extracted deep features; This is for output data.

[0120] The global constraint expression is constructed as follows:

[0121] ,

[0122] in, ;

[0123] In the formula, , , For learning parameters;

[0124] When there are no candidate nodes in the candidate pool that meet the constraints, the learning parameters are... Scaling is applied, while simultaneously expanding the range of random parameter generation. Learning parameters The scaling formula is:

[0125] ;

[0126] In the formula, for Random numbers within the model. When the model's accuracy reaches the preset error tolerance... Or the maximum number of nodes in the model Then, incremental learning is stopped, thereby obtaining a bearing fault diagnosis model.

[0127] Step 5: Based on the number of faults output by the counting model, fuse the output probability matrix of the label model with the similarity matrix to obtain a weighted fusion confidence matrix. Then, output the final multi-label fault diagnosis result based on the weighted fusion confidence matrix. The weighted fusion confidence matrix is ​​as follows:

[0128] ;

[0129] In the formula, It is a probability matrix; This is a similarity matrix; The weighted hyperparameters; combined with the number of faults predicted by the counting model. The following decision is made:

[0130] ;

[0131] In the formula, These are the predicted values ​​from the counting model; For the label dimension.

[0132] Experimental results are as follows Figure 3 As shown, the accuracy on the test set reaches 97.04%. The confusion matrix reveals that the main diagonal is darker while the rest is lighter, visually demonstrating the high accuracy of this method for fault identification. This fully proves that the method of this invention has high prediction accuracy and can provide an effective solution for the diagnosis of composite bearing faults.

[0133] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A composite fault diagnosis method based on feature enhancement and a two-layer random architecture, characterized in that, The steps include the following: Step 1: Acquire bearing vibration signals and extract shallow feature matrices using local mean decomposition and time-domain statistical methods; Step 2: Construct a Transformer network based on self-attention mechanism, perform multi-label classification pre-training with shallow feature matrix as input, and extract deep feature matrix containing global dependencies after solidifying network parameters; Step 3: Construct cloud model physical feature templates for each type of fault based on deep features, and calculate the similarity matrix between the sample to be tested and each cloud model physical feature template; Step 4: Construct a counting model for predicting the number of faults and a labeling model for predicting specific fault categories, respectively. Use the deep feature matrix as input and train the models using a stochastic incremental learning algorithm with global constraints. Step 5: Based on the number of faults output by the counting model, the output probability matrix of the label model and the similarity matrix are fused to obtain a weighted fusion confidence matrix, and the final multi-label fault diagnosis result is output according to the weighted fusion confidence matrix.

2. The composite fault diagnosis method based on feature enhancement and a two-layer random architecture as described in claim 1, characterized in that, The specific process of step 1 is as follows: The bearing vibration signal is sliced ​​by a sliding window, and multiple product function components are obtained by extreme point envelope calculation. The 10-dimensional time-domain statistical features of the first k PF components are calculated respectively. The k×10-dimensional time-domain statistical features extracted from each sample are normalized to generate a shallow feature matrix. The time-domain statistical features include: mean, standard deviation, kurtosis, peak value, root mean square, impulse factor, waveform factor, margin factor, skewness, and peak value factor.

3. The composite fault diagnosis method based on feature enhancement and a two-layer random architecture as described in claim 1, characterized in that, Step 2 includes the following steps: Step 21: Map the shallow feature matrix to a high-dimensional feature space through a fully connected embedding layer. The specific formula is as follows: ； In the formula, For embedding feature sequences; and Assign its weights and biases; Step 22: Input the embedded feature sequence into the multi-head self-attention layer to extract coupling features, and then normalize the distribution through the first normalization layer to obtain the initial normalized features. The specific formula is as follows: ； In the formula, These are the initial normalized features; This is a coupling characteristic; and These are the mean and variance, respectively. and These are the learnable parameters; Step 23: The initial normalized features are transformed into higher-order nonlinear features through a two-layer fully connected feedforward network with ReLU activation, and then the normalized deep feature matrix is ​​output through the second normalization layer. Step 24: Use the multi-label one-hot matrix of composite faults as the supervision signal to perform network regression pre-training. During the training process, after the second normalization layer, a fully connected output layer and a Sigmoid activation layer are cascaded to decouple and map the predicted probabilities of each independent fault category. After training, the network parameters are fixed and the output fully connected layer and Sigmoid activation layer are stripped. The output of the second normalized layer is directly forward propagated as the deep feature matrix.

4. The composite fault diagnosis method based on feature enhancement and a two-layer random architecture as described in claim 1, characterized in that, In step 3, a cloud model feature template is generated based on deep features, wherein the expected value in the cloud model feature template is calculated using the following formula: ； The formula for calculating entropy in the cloud model feature template is: ； The formula for calculating hyperentropy in the cloud model feature template is: ； In the formula, For the first The number of class samples; For the first Class of samples Deep feature vectors of each sample; To calculate the variance by column; This is a preset, extremely small positive number; The dimensional membership degree calculated using the cloud model feature template is: ； In the formula, x represents the sample; Calculate the similarity matrix between the sample to be tested and the physical feature templates of each cloud model, specifically as follows: First, calculate the similarity vector based on the membership dimension. The formula is: ， in, ; Then, obtain the similarity matrix based on the similarity vector.

5. The composite fault diagnosis method based on feature enhancement and a two-layer random architecture as described in claim 4, characterized in that, The stochastic incremental learning construction method with global constraints in step 4 is specifically as follows: We begin by building a small network. Several candidate nodes are randomly generated within a range of random parameters. The generated nodes are then globally constrained using the Greville iteration method. Candidate nodes that satisfy the constraints are sorted, and the node with the largest constraint value is selected as the new node. The output weights and residuals after adding the new node are as follows: ； ； ； ； ； ， in, ; ； In the formula, These are the initial weights; The initial residual; For regularization parameters; For having The output of the hidden layer of each node; For having The output of the hidden layer of each node; These are the extracted deep features; For output data; Furthermore, the global constraint expression is: ， in, ; In the formula, , , These are the learning parameters.

6. The composite fault diagnosis method based on feature enhancement and a two-layer random architecture as described in claim 5, characterized in that, In step 4, when there are no candidate nodes in the candidate pool that meet the constraints, the learning parameters are scaled, and the generation range of the random parameters is expanded. The scaling formula for the learning parameters is: ； In the formula, for Random numbers within.

7. The composite fault diagnosis method based on feature enhancement and a two-layer random architecture as described in claim 1, characterized in that, In step 5, the weighted fusion confidence matrix is: ； In the formula, It is a probability matrix; This is a similarity matrix; These are weight hyperparameters; The fusion decision is: ； In the formula, The predicted value is from the counting model; X represents the label dimension; X represents the deep feature vector.

Images