Bearing composite fault diagnosis method based on deep learning

Abstract

The invention provides a bearing composite fault diagnosis method based on deep learning. The method comprises the following steps: separating a composite fault vibration signal into a single fault sub-signal by using a CNN-RNN hybrid network; quantifying an influence relationship between faults through a fault interaction attention module, dynamically generating a feature weight and re-weighting a feature matrix; and evaluating a network output influence system by adopting the fault influence degree, and adaptively compensating the diagnosis deviation in combination with a correction formula. Based on a deep learning model, by introducing a fault interaction attention mechanism, the method can automatically learn and quantify the mutual influence relationship between different fault types, adaptively adjust the feature weight, effectively reduce the interference between composite faults, and improve the diagnosis accuracy; and meanwhile, a fault interaction knowledge base and a self-adaptive correction algorithm are established through composite fault influence degree quantification and a self-adaptive correction mechanism, so that diagnosis deviation caused by mutual influence between faults is effectively compensated, and more reliable technical support is provided for bearing fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of bearing composite fault diagnosis technology, and specifically to a bearing composite fault diagnosis method based on deep learning. Background Technology

[0002] During the operation of mechanical equipment, bearings are critical components, and their operating status directly affects the overall performance and reliability of the equipment. Bearing fault diagnosis is of great significance for preventing equipment failures, reducing downtime, and improving production efficiency. However, in practical applications, compound bearing faults are quite common, meaning that bearings simultaneously exhibit multiple different types of faults. Because these compound faults interact with each other, the fault characteristics become complex and difficult to extract accurately, resulting in a low accuracy rate for existing diagnostic methods in diagnosing compound bearing faults.

[0003] With the continuous development of deep learning technology, it has demonstrated powerful feature extraction and pattern recognition capabilities in the field of fault diagnosis. However, most existing deep learning-based bearing fault diagnosis methods are designed for single fault types. When dealing with complex faults, they cannot effectively address the problem of mutual influence between faults, making it difficult to meet the needs of accurate diagnosis of complex bearing faults in practical applications. Therefore, there is an urgent need for a method and system that can effectively eliminate or mitigate the mutual influence of complex bearing faults and improve diagnostic accuracy. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based method for diagnosing complex bearing faults, in order to overcome the shortcomings of the existing technologies.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A deep learning-based method for diagnosing composite bearing faults includes the following steps:

[0007] S1. Collect vibration signals during bearing operation and preprocess them to obtain preprocessed vibration signals;

[0008] S2. Construct a CNN-RNN hybrid separation network based on feature reshaping and train and optimize it. The preprocessed vibration signal is separated into a single fault sub-signal through the trained hybrid separation network.

[0009] S3. Add a fault interaction attention mechanism module to the CNN network model. Use the improved CNN network model to extract deep features from fault sub-signals and obtain the feature vector of a single fault.

[0010] S4. A deep learning-based fault impact assessment network is constructed using a multilayer perceptron structure and trained and optimized. The trained fault impact assessment network is used to diagnose the feature vector of a single fault and adaptively correct the impact of a composite fault.

[0011] Furthermore, the vibration signal during the operation of the bearing is acquired by an accelerometer, and the preprocessing of the vibration signal includes, but is not limited to, bandpass filtering, wavelet denoising, and Z-score normalization.

[0012] Furthermore, the CNN-RNN hybrid separation network includes a one-dimensional convolutional layer, a dimension transpose layer, a bidirectional LSTM layer, and a fully connected layer connected in sequence; the pre-processed vibration signal is separated into a single fault sub-signal by the trained hybrid separation network, specifically including the following steps:

[0013] S21. Extract the spatial features of the vibration signal through a one-dimensional convolutional layer and output a first feature tensor with a size of (Batch, Channels, Time).

[0014] S22. The first feature tensor is converted into a second feature tensor of size (Batch, Time, Channels) through a dimension transpose layer.

[0015] S23, a cyclic bidirectional LSTM layer, receives the second feature tensor and models the temporal dependency, and outputs the separated fault sub-signals through a fully connected layer.

[0016] Furthermore, the training and optimization of the CNN-RNN hybrid separation network includes: using vibration signal data with known fault types as training samples, taking the minimization of the error between the original vibration signal and the reconstructed signal after separation as the objective function, and using the stochastic gradient descent algorithm to adjust the network parameters.

[0017] Furthermore, the step of extracting deep features from fault sub-signals and obtaining the feature vector of a single fault using an improved CNN network model specifically includes the following steps:

[0018] S31. Perform a short-time Fourier transform on each fault sub-signal to generate a time-frequency graph, which serves as the input to the improved CNN network model.

[0019] S32. By using an improved CNN network model to extract features from the input time-frequency graph, the feature vectors of each individual fault are obtained and mapped to a unified dimension. At the same time, the feature vectors of each individual fault are concatenated into a composite feature matrix in the channel dimension.

[0020] The composite feature matrix can be represented by the following formula (1):

[0021] F∈R d×n (1);

[0022] In formula (1), d represents the feature dimension; n represents the number of fault types; R represents the set of real numbers; and F represents the composite feature matrix.

[0023] S33. Based on the composite feature matrix, calculate the fault feature correlation matrix through the fault interaction attention mechanism module;

[0024] The fault feature correlation matrix can be represented by the following formula (2):

[0025]

[0026] In formula (2): Q and K represent the linear projection of the composite feature matrix F, which are the query matrix Query and the key matrix Key, respectively; Let K be the transpose of K, representing the matching dimension; d represents the feature dimension. W represents the scaling factor; softmax represents the softmax function, which is a normalized exponential function; a Represents the fault characteristic correlation matrix;

[0027] S34. Generate adaptive weights using the fault feature correlation matrix and reweight the original features.

[0028] This step can be represented by the following formulas (3)-(4):

[0029] W = Sigmoid(W a •Attention) (3);

[0030]

[0031] In formula (3)-(4): W a represents the fault feature correlation matrix; Sigmoid represents the Sigmoid activation function; Attention represents the attention feature matrix, which is the denoised feature representation of the composite feature matrix F after dimensionality reduction; W represents the adaptive weights. represents the transpose operation of adaptive weights to match the dimension; ⊙ represents element-wise multiplication; F′ represents the reweighted feature matrix.

[0032] Furthermore, the training and optimization of the fault diagnosis network includes: using labeled feature vector data as training samples, taking the minimization of the error between the diagnosis result and the actual fault label as the objective function, and using the backpropagation algorithm to train the network parameters.

[0033] Furthermore, the fault impact assessment network includes:

[0034] The input layer is used to receive one-hot codes for fault combinations;

[0035] Two fully connected hidden layers, with 64 and 32 neurons respectively;

[0036] The output layer outputs the impact coefficient for each fault type, with a value range of [0,1].

[0037] The influence coefficient is calculated using the following formula (5):

[0038] α i =FIAN(f1,f2,…,f j (5);

[0039] In formula (5): α i f represents the impact coefficient of the i-th type of fault; FIAN represents the fault impact assessment network; f j This indicates the existence state of the j-th type of fault.

[0040] Furthermore, the step of using a trained fault impact assessment network to diagnose a single fault's feature vector and adaptively correcting the impact of a composite fault specifically includes the following steps:

[0041] S41. Input the feature vector of a single fault into a multilayer perceptron to obtain the preliminary fault probability.

[0042] S42. Generate impact coefficients through the fault impact assessment network;

[0043] S43. Use the influence coefficient to adaptively correct the preliminary diagnosis results and compensate for the diagnostic bias caused by the mutual influence between faults.

[0044] This step can be represented by the following formula (6):

[0045] p′ i =p i (1-α i )+α i max(p j (6);

[0046] In formula (6): α i p represents the influence coefficient of the i-th type of fault; i Let p′ represent the original predicted probability of the i-th type of fault. i p represents the corrected probability. j This represents the original predicted probability of the j-th type of fault.

[0047] Furthermore, the step of diagnosing a single fault's feature vector using a trained fault impact assessment network and adaptively correcting the impact of a composite fault also includes: establishing a fault interaction impact knowledge base based on historical diagnostic data, wherein the fault interaction impact knowledge base includes fault types and impact coefficients; and when diagnosing a new sample, obtaining the corresponding impact coefficients from the knowledge base according to the detected fault combinations.

[0048] As can be seen from the above technical solutions, the present invention has the following advantages compared with the prior art:

[0049] 1. This invention is based on a deep learning model. By introducing a fault interaction attention mechanism, it can automatically learn and quantify the mutual influence relationship between different fault types, adaptively adjust feature weights, effectively reduce interference between complex faults, and improve diagnostic accuracy.

[0050] 2. This invention establishes a fault interaction knowledge base and adaptive correction algorithm through the quantification of the impact of complex faults and the adaptive correction mechanism. This effectively compensates for the diagnostic bias caused by the mutual influence between faults, significantly improves the model's diagnostic robustness for complex faults and the interpretability of the results, and provides more reliable technical support for bearing fault diagnosis. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the steps of the fault diagnosis method of the present invention. Detailed Implementation

[0052] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.

[0053] like Figure 1 The method for diagnosing composite bearing faults based on deep learning, as shown, includes the following steps:

[0054] S1. Collect vibration signals during bearing operation and preprocess them to obtain preprocessed vibration signals.

[0055] Specifically, the vibration signal during the operation of the bearing is collected by an accelerometer, and the preprocessing of the vibration signal includes, but is not limited to, bandpass filtering (100-2000Hz), wavelet denoising (db4 wavelet basis), and Z-score normalization.

[0056] S2. Construct a CNN-RNN hybrid separation network based on feature reshaping and train and optimize it. The preprocessed vibration signal is separated into a single fault sub-signal through the trained hybrid separation network.

[0057] The CNN-RNN hybrid separation network described in this preferred embodiment includes a one-dimensional convolutional layer, a dimension transpose layer, a bidirectional LSTM layer, and a fully connected layer connected in sequence. Separating the preprocessed vibration signal into a single fault sub-signal using the trained hybrid separation network specifically includes the following steps:

[0058] S21. Extract the spatial features of the vibration signal through a one-dimensional convolutional layer and output a first feature tensor with a size of (Batch, Channels, Time).

[0059] S22. The first feature tensor is converted into a second feature tensor of size (Batch, Time, Channels) through a dimension transpose layer.

[0060] S23, a cyclic bidirectional LSTM layer, receives the second feature tensor and models the temporal dependency, and outputs the separated fault sub-signals through a fully connected layer.

[0061] For example, if the original vibration signal has dimensions [64, 1, 1024], following the steps above, the output dimension of the one-dimensional convolutional layer is [64, 64, 128], the output dimension of the dimension transpose layer is [64, 128, 64], the output dimension of the bidirectional LSTM is [64, 128, 256], and the dimension of the separated sub-signals is [64, 3, 1024].

[0062] Furthermore, the training and optimization of the CNN-RNN hybrid separation network includes: using vibration signal data with known fault types as training samples, taking the minimization of the error between the original vibration signal and the reconstructed signal after separation as the objective function, and using the stochastic gradient descent algorithm to adjust the network parameters.

[0063] S3. Add a fault interaction attention mechanism module to the CNN network model (such as the ResNet-18 model), and use the improved CNN network model to perform deep feature extraction on the fault sub-signals to obtain the feature vector of a single fault.

[0064] The Fault Interaction Attention Module (FIAM) described in this preferred embodiment can automatically learn the mutual influence relationships between different fault types and adaptively adjust feature weights, effectively reducing interference between complex faults and improving diagnostic accuracy. This step, which involves extracting deep features from fault sub-signals using an improved CNN network model to obtain the feature vector of a single fault, specifically includes the following steps:

[0065] S31. Perform a short-time Fourier transform on each fault sub-signal to generate a time-frequency graph, which serves as the input to the improved CNN network model.

[0066] S32. By using an improved CNN network model to extract features from the input time-frequency map, the feature vectors of each individual fault are obtained and mapped to a unified dimension. At the same time, the feature vectors of each individual fault are concatenated into a composite feature matrix in the channel dimension. This not only preserves the independence of each fault feature, but also provides a foundation for learning the interaction relationship between faults and achieving accurate diagnosis of composite faults through the fault interaction attention module.

[0067] The composite feature matrix can be represented by the following formula (1):

[0068] F∈R d×n (1);

[0069] In formula (1), d represents the feature dimension; n represents the number of fault types; R represents the set of real numbers; and F represents the composite feature matrix.

[0070] S33. Based on the composite feature matrix, calculate the fault feature correlation matrix through the fault interaction attention mechanism module;

[0071] The fault feature correlation matrix can be represented by the following formula (2):

[0072]

[0073] In formula (2): Q and K represent the linear projection of the composite feature matrix F, which are the query matrix Query and the key matrix Key, respectively; Let K be the transpose of K, representing the matching dimension; d represents the feature dimension. W represents the scaling factor to prevent gradient vanishing; softmax represents the softmax function, which is a normalized exponential function; a Represents the fault characteristic correlation matrix, such as W a [2,3] = 0.9 indicates that fault 3 strongly interferes with fault 2;

[0074] S34. Generate adaptive weights using the fault feature correlation matrix and reweight the original features.

[0075] This step can be represented by the following formulas (3)-(4):

[0076] W = Sigmoid(W a •Attention) (3);

[0077]

[0078] In formula (3)-(4): W arepresents the fault feature association matrix; Sigmoid represents the Sigmoid activation function, which can map the weighted result to the (0,1) interval and generate interpretable weights; Attention represents the attention feature matrix, which is the denoised feature expression after dimensionality reduction of the composite feature matrix F; W represents the adaptive weight, where high weights are key features and low weights are noise pollution features. represents the transpose operation of adaptive weights to match the dimension; ⊙ represents element-wise multiplication; F′ represents the reweighted feature matrix.

[0079] The expression for the attention feature matrix is ​​Attention∈R n×d It is through the original characteristic moments F∈R d×n Transpose Then obtain through linear projection (W v (For trainable projection matrix), noise interference is filtered out through compressed key feature representation.

[0080] The reweighted enhanced feature matrix retains the original dimensions but highlights key features, which can effectively enhance robustness to the impact of fault interactions.

[0081] S4. A deep learning-based fault impact assessment network is constructed using a multilayer perceptron (MLP) structure. The trained fault impact assessment network is used to diagnose the feature vector of a single fault and adaptively correct the impact of a composite fault.

[0082] The training and optimization of the fault diagnosis network includes: using labeled feature vector data as training samples, minimizing the error between the diagnosis result and the actual fault label as the objective function, and training the network parameters using the backpropagation algorithm. The fault impact assessment network includes:

[0083] The input layer is used to receive one-hot codes for fault combinations;

[0084] Two fully connected hidden layers, with 64 and 32 neurons respectively;

[0085] The output layer outputs the impact coefficient for each fault type, with a value range of [0,1].

[0086] The influence coefficient is calculated using the following formula (5):

[0087] α i =FIAN(f1,f2,…,f j (5);

[0088] In formula (5): α if represents the impact coefficient of the i-th type of fault; FIAN represents the fault impact assessment network; f j This indicates the existence state of the j-th type of fault.

[0089] Specifically, the step of diagnosing a single fault's feature vector using a trained fault impact assessment network and adaptively correcting the impact of a composite fault includes the following steps:

[0090] S41. Input the feature vector of a single fault into a multilayer perceptron to obtain the preliminary fault probability; through network calculation and judgment, determine the type of fault existing in the bearing and the severity of each fault, and realize the diagnosis of bearing compound faults.

[0091] S42. Generate impact coefficients through the fault impact assessment network;

[0092] S43. Use the influence coefficient to adaptively correct the preliminary diagnosis results and compensate for the diagnostic bias caused by the mutual influence between faults.

[0093] This step can be represented by the following formula (6):

[0094] p′ i =p i (1-α i )+α i max(p j (6);

[0095] In formula (6): α i p represents the influence coefficient of the i-th type of fault; i Let p′ represent the original predicted probability of the i-th type of fault. i p represents the corrected probability. j This represents the original predicted probability of the j-th type of fault.

[0096] Furthermore, the step of diagnosing a single fault's feature vector using a trained fault impact assessment network and adaptively correcting for composite fault impact also includes: establishing a fault interaction impact knowledge base based on historical diagnostic data, which includes fault types and impact coefficients; and when diagnosing new samples, obtaining the corresponding impact coefficients from the knowledge base based on the detected fault combinations. This composite fault impact quantification and adaptive correction mechanism, through the establishment of a fault interaction knowledge base and an adaptive correction algorithm, effectively compensates for diagnostic biases caused by the mutual influence between faults, further improving the accuracy and reliability of the diagnosis.

[0097] To further verify the effectiveness of the deep learning-based bearing composite fault diagnosis method of this invention, the following supplementary quantitative experimental data are provided from the dimensions of diagnostic accuracy, module effectiveness, and noise resistance.

[0098] The experiment used the bearing dataset from Case Western Reserve University (CWRU) and a self-made composite fault dataset from the laboratory. This dataset included composite fault types such as inner ring + outer ring, inner ring + rolling element, and outer ring + rolling element, with speeds covering 1797 r / min, 1730 r / min, and 1750 r / min. Comparison methods included traditional methods (wavelet decomposition + SVM), existing deep learning methods (CNN-LSTM, ResNet-18), and the method of this invention (CNN-RNN separation network + FIAM + influence correction). Evaluation metrics included diagnostic accuracy, signal-to-noise ratio (SNR), and F1 score (combining precision and recall). Experimental data are shown in Tables 1-4 below.

[0099] Table 1. Accuracy (%) of bearing composite fault diagnosis by various methods at different speeds.

[0100] Rotational speed (r / min) Wavelet + SVM CNN-LSTM ResNet18 Method of the present invention 1730 72.3 85.6 88.9 96.2 1750 68.5 82.1 86.4 94.8 1797 65.7 79.3 83.7 93.5 Average 68.8 82.3 86.3 94.8

[0101] Analysis: The diagnostic accuracy of the method in this invention is significantly higher than that of the comparative method at different speeds, with an average improvement of 8.5%-26%. In particular, it maintains an accuracy of 93.5% even at high speeds (1797 r / min), verifying its robustness to speed changes. This is attributed to the effective separation of complex fault signals by the CNN-RNN hybrid separation network and the suppression of fault interaction interference by the FIAM module.

[0102] To verify the effectiveness of the Fault Interactive Attention Mechanism (FIAM) module, an ablation experiment was designed to compare the performance of the "method of the present invention" and the "method of removing FIAM" in composite fault diagnosis.

[0103] Table 2 Ablation Experiment Results of FIAM Module

[0104]

[0105]

[0106] Analysis: After removing FIAM, the accuracy of composite fault diagnosis decreased significantly (average decrease of 9.3%), especially the diagnostic accuracy of outer ring + rolling element faults decreased by 10.6%, indicating that the FIAM module can effectively learn the interaction relationship between faults, reduce interference, and improve diagnostic accuracy.

[0107] Furthermore, the effectiveness of the adaptive correction mechanism was verified by comparing the diagnostic results of "uncorrected" and "corrected by the impact assessment network".

[0108] Table 3. Fault Impact Correction Results (Accuracy %)

[0109] Meanwhile, by adding Gaussian noise of different intensities (SNR = 5dB, 10dB, ...) to the vibration signal...

[0110]

[0111] (15dB), comparing the noise immunity performance of the FIAM module of this invention with that of the traditional channel attention mechanism (CAM).

[0112] Table 4. Comparison of diagnostic accuracy (%) under different SNRs

[0113] SNR (dB) Traditional CAM This invention FIAM Advantages (%) 5 76.3 89.5 13.2 10 82.7 93.2 10.5 15 88.1 95.6 7.5

[0114] Analysis: In a low signal-to-noise ratio (5dB) environment, the diagnostic accuracy of the FIAM module is 13.2% higher than that of the traditional CAM, indicating that its adaptive weights generated by learning fault interaction relationships can more effectively filter out noise interference and enhance the model's noise resistance.

[0115] In practical applications, the bearing composite fault diagnosis method described in this invention can be implemented through the following bearing composite fault diagnosis system:

[0116] The signal acquisition module includes an accelerometer and a signal conditioning circuit.

[0117] Edge computing units are used to deploy embedded hardware to implement the aforementioned deep learning-based bearing composite fault diagnosis method.

[0118] The diagnostic results visualization terminal is used to output the fault type, severity, and confidence curve.

[0119] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A deep learning-based method for diagnosing composite bearing faults, characterized in that, Includes the following steps: S1. Collect vibration signals during bearing operation and preprocess them to obtain preprocessed vibration signals; S2. Construct a CNN-RNN hybrid separation network based on feature reshaping and train and optimize it. The preprocessed vibration signal is separated into a single fault sub-signal through the trained hybrid separation network. S3. Add a fault interaction attention mechanism module to the CNN network model. Use the improved CNN network model to extract deep features from fault sub-signals and obtain the feature vector of a single fault. S4. A deep learning-based fault impact assessment network is constructed using a multilayer perceptron structure and trained and optimized. The trained fault impact assessment network is used to diagnose the feature vector of a single fault and adaptively correct the impact of a composite fault.

2. The bearing composite fault diagnosis method based on deep learning according to claim 1, characterized in that, The vibration signal during the operation of the bearing is collected by an accelerometer. The preprocessing of the vibration signal includes, but is not limited to, bandpass filtering, wavelet denoising, and Z-score normalization.

3. The bearing composite fault diagnosis method based on deep learning according to claim 1, characterized in that, The CNN-RNN hybrid separation network comprises a one-dimensional convolutional layer, a dimension transpose layer, a bidirectional LSTM layer, and a fully connected layer connected in sequence. Separating the preprocessed vibration signal into a single fault sub-signal using the trained hybrid separation network specifically includes the following steps: S21. Extract the spatial features of the vibration signal through a one-dimensional convolutional layer and output a first feature tensor with a size of (Batch, Channels, Time). S22. The first feature tensor is converted into a second feature tensor of size (Batch, Time, Channels) through a dimension transpose layer. S23, a cyclic bidirectional LSTM layer, receives the second feature tensor and models the temporal dependency, and outputs the separated fault sub-signals through a fully connected layer.

4. The bearing composite fault diagnosis method based on deep learning according to claim 3, characterized in that, The training and optimization of the CNN-RNN hybrid separation network includes: using vibration signal data with known fault types as training samples, taking the minimization of the error between the original vibration signal and the reconstructed signal after separation as the objective function, and using the stochastic gradient descent algorithm to adjust the network parameters.

5. The bearing composite fault diagnosis method based on deep learning according to claim 1, characterized in that, The process of extracting deep features from fault sub-signals and obtaining feature vectors for single faults using an improved CNN network model specifically includes the following steps: S31. Perform a short-time Fourier transform on each fault sub-signal to generate a time-frequency graph, which serves as the input to the improved CNN network model. S32. By using an improved CNN network model to extract features from the input time-frequency graph, the feature vectors of each individual fault are obtained and mapped to a unified dimension. At the same time, the feature vectors of each individual fault are concatenated into a composite feature matrix in the channel dimension. The composite feature matrix can be represented by the following formula (1): F∈R d×n (1); In formula (1), d represents the feature dimension; n represents the number of fault types; R represents the set of real numbers; and F represents the composite feature matrix. S33. Based on the composite feature matrix, calculate the fault feature correlation matrix through the fault interaction attention mechanism module; The fault feature correlation matrix can be represented by the following formula (2): In formula (2): Q and K represent the linear projection of the composite feature matrix F, which are the query matrix Query and the key matrix Key, respectively; Let K be the transpose of K, representing the matching dimension; d represents the feature dimension. W represents the scaling factor; softmax represents the softmax function, which is a normalized exponential function; a Represents the fault characteristic correlation matrix; S34. Generate adaptive weights using the fault feature correlation matrix and reweight the original features. This step can be represented by the following formulas (3)-(4): W=Sigmoid(W a ·Attention) (3); In formula (3)-(4): W a represents the fault feature correlation matrix; Sigmoid represents the Sigmoid activation function; Attention represents the attention feature matrix, which is the denoised feature representation of the composite feature matrix F after dimensionality reduction; W represents the adaptive weights. represents the transpose operation of adaptive weights to match the dimension; ⊙ represents element-wise multiplication; F′ represents the reweighted feature matrix.

6. The bearing composite fault diagnosis method based on deep learning according to claim 1, characterized in that, The training and optimization of the fault diagnosis network includes: using labeled feature vector data as training samples, taking the minimization of the error between the diagnosis result and the actual fault label as the objective function, and using the backpropagation algorithm to train the network parameters.

7. The bearing composite fault diagnosis method based on deep learning according to claim 1, characterized in that, The fault impact assessment network includes: The input layer is used to receive one-hot codes for fault combinations; Two fully connected hidden layers, with 64 and 32 neurons respectively; The output layer outputs the impact coefficient for each fault type, with a value range of [0,1]. The influence coefficient is calculated using the following formula (5): α i FIAN(f1,f2,…,f j ) (5); In formula (5): α i f represents the impact coefficient of the i-th type of fault; FIAN represents the fault impact assessment network; f j This indicates the existence state of the j-th type of fault.

8. The bearing composite fault diagnosis method based on deep learning according to claim 1, characterized in that, The process of diagnosing a single fault's feature vector using a trained fault impact assessment network and adaptively correcting the impact of a composite fault specifically includes the following steps: S41. Input the feature vector of a single fault into a multilayer perceptron to obtain the preliminary fault probability. S42. Generate impact coefficients through the fault impact assessment network; S43. Use the influence coefficient to adaptively correct the preliminary diagnosis results and compensate for the diagnostic bias caused by the mutual influence between faults. This step can be represented by the following formula (6): p′ i =p i (1-a i )+a i max(p j ) (6); In formula (6): α i p represents the influence coefficient of the i-th type of fault; i Let p′ represent the original predicted probability of the i-th type of fault. i p represents the corrected probability. j This represents the original predicted probability of the j-th type of fault.

9. The bearing composite fault diagnosis method based on deep learning according to claim 8, characterized in that, The method of diagnosing a single fault's feature vector using a trained fault impact assessment network and adaptively correcting the impact of a composite fault further includes: establishing a fault interaction impact knowledge base based on historical diagnostic data, wherein the fault interaction impact knowledge base includes fault types and impact coefficients; and when diagnosing a new sample, obtaining the corresponding impact coefficients from the knowledge base based on the detected fault combinations.

Images