Industrial intelligent predictive maintenance method based on interpretable depth feature fusion network
By combining the deep feature fusion network of discrete cosine enhanced attention mechanism and self-attention mechanism, the problem of poor noise robustness and interpretability in the prior art is solved, and efficient fault identification and transparent decision-making in strong noise environments are achieved. The process improves the accuracy and reliability of industrial intelligent predictive maintenance.
Patent Information
- Application Number
- CN202410995990.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-24
- Publication Date
- 2024-11-15
AI Technical Summary
The existing industrial intelligent predictive maintenance model based on deep learning has shortcomings in noise robustness and interpretability, and it is difficult to effectively extract fault characteristics and conduct in-depth analysis in a highly noisy environment.
The interpretable wavelet multi-scale feature extraction network based on discrete cosine-enhanced attention mechanism and the dual-branch feature fusion network based on self-attention mechanism are adopted. Combined with the multi-head self-attention mechanism, local and global features are extracted, and the output channel weights are visualized and Wavelet energy spectrum comparison analysis was performed for interpretability analysis.
It improves the noise immunity and interpretability of the model, can accurately identify faults in high noise environments, and enhances the transparency and accuracy of diagnostic results.
Smart Images

Figure CN118965259A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of industrial intelligent predictive maintenance, and in particular relates to an industrial intelligent predictive maintenance method based on an interpretable deep feature fusion network. Background Art
[0002] As the structure and automation functions of mechanical equipment become increasingly complex, mechanical equipment faces higher risks of failure and more serious consequences of failure. Some key mechanical components (such as bearings and gears) will inevitably experience fatigue degradation, cracks, deformation, peeling and other failures during the operation of mechanical equipment. If the faulty components are not repaired or replaced in time, it may cause serious damage to the entire mechanical system. Therefore, in order to ensure the normal operation of the mechanical system, it is very necessary to monitor the health status of the key components of the mechanical system.
[0003] Traditional machine learning methods rely heavily on researchers' prior experience and advanced signal preprocessing techniques because they are manually designed features based on raw data. In contrast, deep learning-based techniques reduce reliance on prior knowledge and provide end-to-end solutions. They have been widely used in machine vision, natural language processing and other fields. Due to the powerful ability of deep learning to extract key information hidden in data, many deep learning models have been widely used in mechanical fault diagnosis, such as deep belief networks (DBNs), convolutional neural networks (CNNs) and recurrent neural networks (RNNs). Although the above studies have achieved good performance in many fault diagnosis tasks, they still have the following problems. The fault diagnosis process based on deep learning is completely data-driven. As the network deepens, the feature maps of vibration signals usually have higher dimensions and lower resolutions. They do not have interpretable physical meanings as input signals. Many periodic pulse features associated with faults are likely to disappear in deep feature maps. It is difficult to understand what the model has learned and to conduct in-depth interpretability analysis.
[0004] Many studies and efforts have been made to enhance the interpretability of deep learning in intelligent fault diagnosis. One aspect is to integrate traditional signal processing methods into deep learning-based methods to improve the ex ante interpretability and diagnostic performance of diagnostic models. At the same time, in order to integrate the features of global and local information, enhance the noise robustness of the model and extract more abundant fault generation fault features, some studies aim to achieve global feature mining by directly stacking convolutional layers or combining recursive neural networks, but these methods make the model deeper and increase the training time cost due to the increase in the number of parameters, and the recursive structure may cause noise accumulation. The Transformer model based on the self-attention mechanism has been increasingly used in the field of fault diagnosis in recent years due to its excellent global information mining capabilities. Since the diagnostic signal is continuous, local signals cannot be ignored while paying attention to its global features. Therefore, many scholars have proposed combining CNN and Transformer to achieve the joint extraction of local features and global information of the original signal.
[0005] Most of the ex ante interpretable fault diagnosis models are based on one-dimensional convolutional networks, which lack attention to global information in a strong noise environment, and due to the sensitivity of the convolution kernel to noise, it is easy to cause the loss of key fault information in a strong noise environment. Secondly, the current advanced fault diagnosis model based on the fusion of CNN and Transformer has achieved initial results in intelligent fault diagnosis in a strong noise environment, but the ex ante interpretability research on various new self-attention-based variant models is still lacking. Summary of the invention
[0006] In view of the problems of poor noise robustness and interpretability of current industrial intelligent predictive maintenance, the present invention provides an industrial intelligent predictive maintenance method based on an interpretable deep feature fusion network, comprising:
[0007] S1: Establish an interpretable neural network model, which includes an interpretable wavelet multi-scale feature extraction network based on discrete cosine enhanced attention mechanism, a dual-branch feature fusion network based on self-attention mechanism, and a classifier MLP;
[0008] S2: Acquire the vibration data of the gearbox containing noise;
[0009] S3: The interpretable wavelet multi-scale feature extraction network based on discrete cosine enhanced attention mechanism is used to capture the interpretable features of gearbox vibration data containing noise while suppressing and filtering the noise signal;
[0010] S4: Deep fusion of features is achieved through a dual-branch feature fusion network based on the self-attention mechanism;
[0011] The dual-branch feature fusion network based on the self-attention mechanism includes: a global convolution self-attention branch for global modeling and a local self-attention branch for focusing on local features;
[0012] S5: Apply convolutional layers to stabilize the learned features, use classifier MLP to reduce the dimension of the features and output the final fault identification results;
[0013] S6: The final fault identification results are analyzed for interpretability using output channel weight visualization and wavelet energy spectrum comparative analysis.
[0014] Beneficial effects of the present invention:
[0015] The present invention integrates CNN and multi-head self-attention mechanism more deeply, combines their respective advantages in local and global feature extraction, and makes the decision-making process of the model more transparent and explainable, establishes an explainable deep feature fusion network for industrial intelligent predictive maintenance, and improves the noise resistance of the model by extracting local and sparse features at different positions of complex vibration signals and filtering out irrelevant information; and by realizing the transmission of global and local spatial feature information, a deeper fusion of high-dimensional abstract features is performed, so that features at different levels can adaptively adjust their contributions to the final diagnosis results, making the industrial intelligent predictive maintenance noise more robust and explainable. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 A flowchart of an industrial intelligent predictive maintenance method based on an interpretable deep feature fusion network of the present invention;
[0017] Figure 2 A model framework diagram of the interpretable neural network model of the present invention;
[0018] Figure 3 This is a fault identification result diagram of an embodiment of the present invention;
[0019] Figure 4 An interpretable analysis diagram of an embodiment of the present invention. DETAILED DESCRIPTION
[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0021] An industrial intelligent predictive maintenance method based on interpretable deep feature fusion network, such as Figure 1 As shown, including:
[0022] S1: Build an interpretable neural network model, such as Figure 2 As shown, the model includes an interpretable wavelet multi-scale feature extraction network based on discrete cosine enhanced attention mechanism, a dual-branch feature fusion network based on self-attention mechanism, and a classifier MLP;
[0023] S2: Acquire the vibration data of the gearbox containing noise;
[0024] S3: The interpretable wavelet multi-scale feature extraction network based on discrete cosine enhanced attention mechanism is used to capture the interpretable features of gearbox vibration data containing noise while suppressing and filtering the noise signal;
[0025] S4: Deep fusion of features is achieved through a dual-branch feature fusion network based on the self-attention mechanism;
[0026] The dual-branch feature fusion network based on the self-attention mechanism includes: a global convolution self-attention branch for global modeling and a local self-attention branch for focusing on local features;
[0027] S5: Apply convolutional layers to stabilize the learned features, use classifier MLP to reduce the dimension of the features and output the final fault identification results;
[0028] S6: The final fault identification results are analyzed for interpretability using output channel weight visualization and wavelet energy spectrum comparative analysis.
[0029] Construct an interpretable wavelet multi-scale feature extraction network based on discrete cosine enhanced attention mechanism, including:
[0030] The discrete wavelet transform is embedded into the neural network to expand the feature learning space from the time domain to the wavelet domain; cross-channel multi-scale convolution is used to increase the local receptive field of the model while reducing the number of model parameters and floating-point operations; discrete cosine transform is introduced to construct a cosine enhanced channel attention mechanism, and a dual-scale feature fusion network based on the self-attention mechanism is obtained.
[0031] Captures interpretable features of noisy gearbox vibration data, including:
[0032] The noisy gearbox vibration data is expanded from the time domain to the wavelet domain through discrete wavelet transform: the input signal is decomposed into high and low frequency bands through discrete wavelet transform, and the high and low frequency bands are further processed;
[0033] Use multi-scale convolution to capture features of different scales in the wavelet domain;
[0034] The discrete cosine enhanced channel attention mechanism is used to adaptively weight the features of different frequency bands to effectively filter out noise and irrelevant information.
[0035] The formula of the discrete wavelet transform includes:
[0036]
[0037] Among them, t is the time series number, f(t) is the original signal, j is the number of decomposition layers, H and G are the low-frequency and high-frequency wavelet decomposition filters in the time domain, respectively, and A j is the wavelet coefficient of the low-frequency part of the signal f(t) at the jth layer; D j is the wavelet coefficient of the high-frequency part of the signal f(t) in the jth layer.
[0038] Further processing of the high and low frequency bands includes:
[0039]
[0040] Among them, x L and x H are the low-frequency features and high-frequency features after discrete wavelet decomposition; x cat For x L and x H The output features after stacking along the channel dimension; Concat() represents x L and x H Perform splicing operations along the channel dimension; For x cat The final output size is 2C output channels, and the length of each channel is L / 2.
[0041] Multi-scale convolution is used to capture features of different scales in the wavelet domain, including:
[0042]
[0043] in, is the output of the jth channel-by-channel convolution, λ is the Gaussian error activation function GELU, μ is the batch normalization BN, z out It is the output after discrete wavelet transform DWT and multi-scale feature extraction, and Concat() is the concatenation operation of each multi-scale convolution branch along the channel dimension.
[0044] The discrete cosine enhanced channel attention mechanism is used to adaptively weight the features of different frequency bands and effectively filter out noise and irrelevant information, including:
[0045] After cosine transform of features in different frequency bands, the channel attention mechanism is weighted. The frequency channel features output after 1D-DCT of each channel component are stacked, and convolution operation is performed to output the attention weight. A larger weight is applied to the frequency band related to the fault feature, and a smaller weight is applied to the noise band to obtain the final output.
[0046] The process of achieving deep fusion of features includes:
[0047] The global multi-head convolution self-attention branch is used to further extract global features and spatial information:
[0048]
[0049] Among them, ξ is the position embedding, i∈(1,2,3,...,H) represents the attention head number, F() represents the convolution operation with a kernel size of 1, and Q i is the query vector of the i-th attention head, K i is the key value vector of the i-th attention head, V i is the value vector of the i-th attention head, Q, K, and V are respectively the query vector, key vector, and value vector obtained after the signal is processed by the feature extraction module and input to the global multi-head convolution self-attention branch, and the unfold() table performs an expansion operation on the input; F att represents the global feature output obtained by the global multi-head self-attention branch, ε is the connection operation of H attention heads, T represents the transposition operation of the input key value vector, and d is used to scale the operation result, and its value is the product of the number of attention heads and the feature dimension of each attention head;
[0050] In the local convolutional attention branch, global maximum pooling and discrete cosine transform are performed in parallel to highlight high-frequency features while aggregating full-band features to achieve the purpose of further refining features:
[0051] F conv =F(x)*SoftMax(F(x a *x b ))
[0052] Among them, x is the input of the local convolutional attention branch, x a ,x b are the output features after GMP and DCT processing, respectively. F() represents the convolution operation with a kernel size of 1. conv is the local feature output after local convolution attention processing.
[0053] The convolution layer is used to stabilize the learned features, and the MLP is used to reduce the dimension of the features and output the final fault recognition results, including:
[0054] Residual connection and Dropout methods are introduced to reduce the risk of overfitting. At the same time, two learnable parameters are introduced in the residual connection to adaptively adjust the importance of features. Then, a feedforward block consisting of two convolutional layers is used to integrate stable output features. Finally, a multi-layer perceptron (MLP) is used to reduce the dimension of the features and output the identified fault type.
[0055] The final fault identification results are analyzed for interpretability by using output channel weight visualization and wavelet energy spectrum comparative analysis. The interpretable analysis diagram is shown in the figure below. Figure 4 As shown, including:
[0056] The eigenvalues of the final output channel of the model feature extraction network are obtained, and the eigenvalues of each channel are summed and normalized, and then sorted from low to high frequency to obtain the wavelet energy spectrum. At the same time, the original input signal is fast Fourier transformed to obtain its spectrum diagram;
[0057] The wavelet energy spectrum and the spectrum diagram of the original signal are compared and analyzed to determine whether the main frequency components in the obtained wavelet energy spectrum match the main frequency components in the original signal spectrum, which is used to show that the learned features conform to the accurate physical meaning in the experiment; in addition, there are differences in the distribution of the main frequency components of the features of different health states, indicating that the network has indeed learned interpretable features that can be used for classification.
[0058] Embodiment 1:
[0059] In this embodiment, the gearbox vibration signal is sampled by the sensor, and a total of five different health states are selected, each of which is assigned a different label; the vibration signal is divided by a sliding window, each sample contains 1024 sampling points, and the ratio of the training data set to the test set is 7:3; in order to simulate the actual working scenario of the gearbox, 0dB, -4dB and -6db Gaussian noise are added to the signal respectively; to simulate three noise levels: weak, medium and strong.
[0060] Construct an interpretable deep feature fusion model and train the model using a training dataset. The neural network model can interpret the feature extraction module, which consists of a layer of discrete wavelet convolution (DWT-CNN), a layer of lightweight multi-scale convolution, and a layer of discrete cosine enhanced channel attention module. The global branch of the dual-branch feature fusion module consists of global multi-head convolution self-attention, and the local branch uses global maximum pooling (GMP) and discrete cosine transform (DCT) to process local features in parallel. In addition, a dropout layer is introduced in the deep feature fusion network, and its parameters are set to 0.1; as well as residual connections to prevent model overfitting and gradient explosion problems. The specific model structure is as follows:
[0061]
[0062] d represents the output dimension, h represents the number of self-attention heads, k represents the CNN kernel size, s represents the step size of CNN, and N represents the number of health status types.
[0063] During the training of the interpretable deep feature fusion model, the training data is divided into a training data set and a validation data set. The mean absolute error loss function is selected as the training error in the training, and Adam is used as the optimization method. The initial learning rate is set to 0.001. If the learning rate does not decrease every 50 epochs, the learning rate is adjusted to the original 0.1.
[0064] In order to verify the feasibility and accuracy of this method, a test experiment was conducted, and this model was compared with three advanced new end-to-end fault recognition models based on CNN-Transformer architecture: Convfomer-NSE, LiConvFormer, and CLFormer. The fault recognition results are shown in the figure Figure 3 By carrying out identification experiments on gearbox faults under three different noise conditions, the comparison results are shown in Table 1:
[0065] Table 1 Comparison of fault recognition results of different neural networks
[0066]
[0067] In the constant speed gearbox dataset, the proposed method achieved the best diagnostic results under all noise conditions, and the advantages of the proposed method became more apparent as the noise increased. Under 0dB conditions, LiConvformer and our method both achieved 100% diagnostic accuracy under constant speed conditions, but our model complexity is smaller than that of the other two methods, so it has a smaller computational overhead and a higher cost-effectiveness. Under -4dB and -6dB conditions, the highest diagnostic accuracy of the comparison method is 0.19% and 0.39% higher, respectively, and compared with the best results of the three comparison methods based on the CNN-Transformer architecture, the proposed method is 1.09% and 1.10% higher, respectively. Overall, the proposed method shows satisfactory performance on the constant speed dataset. In the variable speed data, as the noise increases, the proposed advantages gradually become apparent, and under -4dB and -6dB conditions, it is 6.66% and 4.65% higher than the best method, respectively. This is because as the noise increases, the damage to non-uniform speed signals is more serious than that to constant speed signals, and the correlation of local characteristics of the signals is weakened to varying degrees. Therefore, it is crucial to establish a global dependence on the signal.
[0068] In summary, through the diagnosis experiments on gear faults with different noise levels, it can be seen that the industrial intelligent predictive maintenance methods based on interpretable deep feature fusion networks can better realize the correct classification of faults, greatly improving the accuracy and reliability of fault diagnosis.
[0069] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. An industrial intelligent predictive maintenance method based on an interpretable deep feature fusion network, characterized in that: include: S1: Establish an interpretable neural network model, which includes an interpretable wavelet multi-scale feature extraction network based on discrete cosine enhanced attention mechanism, a dual-branch feature fusion network based on self-attention mechanism, and a classifier MLP; S2: Acquire the vibration data of the gearbox containing noise; S3: The interpretable wavelet multi-scale feature extraction network based on discrete cosine enhanced attention mechanism is used to capture the interpretable features of gearbox vibration data containing noise while suppressing and filtering the noise signal; S4: Deep fusion of features is achieved through a dual-branch feature fusion network based on the self-attention mechanism; The dual-branch feature fusion network based on the self-attention mechanism includes: a global convolution self-attention branch for global modeling and a local self-attention branch for focusing on local features; S5: Apply convolutional layers to stabilize the learned features, use classifier MLP to reduce the dimension of the features and output the final fault identification results; S6: The final fault identification results are analyzed for interpretability using output channel weight visualization and wavelet energy spectrum comparative analysis.
2. The industrial intelligent predictive maintenance method based on an interpretable deep feature fusion network according to claim 1 is characterized in that: Construct an interpretable wavelet multi-scale feature extraction network based on discrete cosine enhanced attention mechanism, including: The discrete wavelet transform is embedded into the neural network to expand the feature learning space from the time domain to the wavelet domain; cross-channel multi-scale convolution is used to increase the local receptive field of the model while reducing the number of model parameters and floating-point operations; discrete cosine transform is introduced to construct a cosine enhanced channel attention mechanism, and a dual-scale feature fusion network based on the self-attention mechanism is obtained.
3. The industrial intelligent predictive maintenance method based on an interpretable deep feature fusion network according to claim 1 is characterized in that: Captures interpretable features of noisy gearbox vibration data, including: The noisy gearbox vibration data is expanded from the time domain to the wavelet domain through discrete wavelet transform: the input signal is decomposed into high and low frequency bands through discrete wavelet transform, and the high and low frequency bands are further processed; Use multi-scale convolution to capture features of different scales in the wavelet domain; The discrete cosine enhanced channel attention mechanism is used to adaptively weight the features of different frequency bands to effectively filter out noise and irrelevant information.
4. The industrial intelligent predictive maintenance method based on an interpretable deep feature fusion network according to claim 3 is characterized in that: The formula of the discrete wavelet transform includes: Among them, t is the time series number, f(t) is the original signal, j is the number of decomposition layers, H and G are the low-frequency and high-frequency wavelet decomposition filters in the time domain, respectively, and A j is the wavelet coefficient of the low-frequency part of the signal f(t) at the jth layer; D j is the wavelet coefficient of the high-frequency part of the signal f(t) in the jth layer.
5. The industrial intelligent predictive maintenance method based on an interpretable deep feature fusion network according to claim 3 is characterized in that: Further processing of the high and low frequency bands includes: Among them, x L and x H are the low-frequency features and high-frequency features after discrete wavelet decomposition; x cat For x L and x H Output features after stacking along the channel dimension; Concat() is a concatenation operation; For x cat The final output size is 2C output channels, and the length of each channel is L / 2.
6. The industrial intelligent predictive maintenance method based on an interpretable deep feature fusion network according to claim 3 is characterized in that: Multi-scale convolution is used to capture features of different scales in the wavelet domain, including: in, is the output of the jth channel-by-channel convolution, λ is the Gaussian error activation function GELU, μ is the batch normalization BN, z out It is the output after discrete wavelet transform DWT and multi-scale feature extraction, and Concat() is the concatenation operation.
7. The industrial intelligent predictive maintenance method based on an interpretable deep feature fusion network according to claim 3 is characterized in that: The discrete cosine enhanced channel attention mechanism is used to adaptively weight the features of different frequency bands and effectively filter out noise and irrelevant information, including: After cosine transform of features in different frequency bands, the channel attention mechanism is weighted. The frequency channel features are output after 1D-DCT of each channel component, and each frequency channel feature is stacked and convolution operation is performed to output the attention weight. A larger weight is applied to the frequency band related to the fault feature, and a smaller weight is applied to the noise band to obtain the final output.
8. The industrial intelligent predictive maintenance method based on an interpretable deep feature fusion network according to claim 1 is characterized in that: The process of achieving deep fusion of features includes: The global multi-head convolution self-attention branch is used to further extract global features and spatial information: Among them, ξ is the position embedding, i∈(1,2,3,...,H) represents the attention head number, F() represents the convolution operation with a kernel size of 1, and Q i is the query vector of the i-th attention head, K i is the key value vector of the i-th attention head, V i is the value vector of the i-th attention head, Q, K, and V are respectively the query vector, key value vector, and value vector obtained after the signal is processed by the feature extraction module and input to the global multi-head convolution self-attention branch, and the unfold() table performs an expansion operation on the input; F att represents the global feature output obtained by the global multi-head self-attention branch, ε is the connection operation of H attention heads, T represents the transposition operation, and d is used to scale the operation result. Its value is the product of the number of attention heads and the feature dimension of each attention head. ; In the local convolutional attention branch, global maximum pooling and discrete cosine transform are performed in parallel to highlight high-frequency features while aggregating full-band features to further refine features: F conv =fF(x)*SoftMax(F(x a *x b )) Among them, x is the input of the local convolutional attention branch, x a ,x b are the output features after GMP and DCT processing, F() represents the convolution operation with a kernel size of 1, and F conv It is the local feature output after local convolution attention processing.
9. The industrial intelligent predictive maintenance method based on an interpretable deep feature fusion network according to claim 1 is characterized in that: The convolution layer is used to stabilize the learned features, and the MLP is used to reduce the dimension of the features and output the final fault recognition results, including: Residual connection and Dropout methods are introduced to reduce the risk of overfitting. At the same time, two learnable parameters are introduced in the residual connection to adaptively adjust the importance of features. Then, a feedforward block consisting of two convolutional layers is used to integrate stable output features. Finally, a multi-layer perceptron (MLP) is used to reduce the dimension of the features and output the identified fault type.
10. The industrial intelligent predictive maintenance method based on an interpretable deep feature fusion network according to claim 1, characterized in that: The final fault identification results are analyzed for interpretability using output channel weight visualization and wavelet energy spectrum comparative analysis, including: The eigenvalues of the final output channel of the model feature extraction network are obtained, and the eigenvalues of each channel are summed and normalized, and then sorted from low to high frequency to obtain the wavelet energy spectrum. At the same time, the original input signal is fast Fourier transformed to obtain its spectrum diagram; The wavelet energy spectrum and the spectrum diagram of the original signal are compared and analyzed to determine whether the main frequency components in the obtained wavelet energy spectrum match the main frequency components in the original signal spectrum, which is used to show that the learned features conform to the accurate physical meaning in the experiment; in addition, there are differences in the distribution of the main frequency components of the features of different health states, indicating that the network has indeed learned interpretable features that can be used for classification.
Citation Information
Patent Citations
Gearbox composite fault diagnosis method and system
CN116628591A
Attention twinning intelligent migration interpretability diagnosis method suitable for high-end equipment
CN117574259A
Remaining service life prediction of bi-directional Transform model based on frequency spectrum
CN118278263A
Cross-domain mechanical fault diagnosis method based on multi-channel feature fusion of CBAM and use thereof
US20240142342A1
Cited By
Complex equipment fault prediction method and system based on multi-dimensional time sequence characteristic modulation
CN120123850A
A Complex Equipment Fault Prediction Method and System Based on Multi-Dimensional Time-Series Feature Modulation
CN120123850B
Interpretable deep feature fusion network-based industrial intelligent predictive maintenance method
WO2026021130A1
