Interpretable deep feature fusion network-based industrial intelligent predictive maintenance method

Abstract

The present invention relates to the field of industrial intelligent predictive maintenance, and in particular to an interpretable deep feature fusion network-based industrial intelligent predictive maintenance method, comprising: acquiring gearbox vibration data comprising noise; performing preliminary extraction and noise suppression on features of the acquired data by establishing an interpretable feature extraction module having a physical information constraint; integrating multi-scale features comprising long-distance and local dependencies by means of a dual-branch feature fusion module having global and local feature fusion capabilities; performing dimensionality reduction on a high-dimensional feature and generating an output by means of a classifier to obtain a final fault identification result; and performing interpretability analysis on a diagnosis process of a model. In the present invention, by embedding the signal processing technology having a well-defined physical theory support into a deep neural network, the interpretability and reliability of model inference results are effectively improved while the fault identification accuracy of the model is improved.

Description

Industrial intelligent predictive maintenance method based on interpretable deep feature fusion network TECHNICAL FIELD

[0001] The present application belongs to the field of industrial intelligent predictive maintenance, and particularly relates to an industrial intelligent predictive maintenance method based on an interpretable deep feature fusion network. BACKGROUND

[0002] With the increasing complexity of mechanical equipment structure and automation function, mechanical equipment faces higher failure risk and more serious failure consequences. Some key mechanical components (such as bearings and gears) inevitably appear fatigue degradation, cracks, deformation, spalling and other failures during the operation of mechanical equipment. If the failure 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 necessary to monitor the health status of the key components of the mechanical system.

[0003] Traditional machine learning methods are heavily dependent on researchers' previous experience and advanced signal preprocessing techniques, because they are manually designed features based on raw data. In contrast, deep learning-based techniques reduce the dependence on prior knowledge and provide end-to-end solutions. They have been widely applied in the fields of machine vision, natural language processing, etc. Due to the powerful ability of deep learning to extract hidden key information in data, many deep learning models have been widely applied in mechanical fault diagnosis, such as deep belief network (DBN), convolutional neural network (CNN) and recurrent neural network (RNN). Although the above researches 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, and with the deepening of the network, the feature map of the vibration signal usually has higher dimension and lower resolution. They do not have interpretable physical meaning as input signals. Many periodic pulse features related to faults are likely to disappear in deep feature maps. It is difficult to understand what the model has learned and it is difficult to conduct in-depth explainability analysis.

[0004] Many researches and efforts have been made to enhance the explainability of deep learning in intelligent fault diagnosis. One aspect is to integrate traditional signal processing methods into deep learning-based methods to improve the prior explainability and diagnostic performance of the diagnostic model. Meanwhile, in order to fuse global and local information features, enhance the noise robustness of the model and extract more rich fault features, some researches aim to realize global feature mining by directly stacking convolutional layers or combining recurrent neural networks, but these methods will make the model deeper and increase the training time cost due to the increase of the number of parameters, and the recursive structure may cause noise accumulation. The Transformer model based on the self-attention mechanism has been applied to the field of fault diagnosis in recent years due to its excellent global information mining capability. Since the diagnostic signal has continuity, the local signal should not be ignored while focusing on its global features. Therefore, many scholars have proposed to combine CNN and Transformer to realize the joint extraction of local features and global information of the original signal.

[0005] Most prior explainable fault diagnosis models are based on one-dimensional convolutional networks, which lack attention to global information in a strong noise environment and are sensitive to noise due to the convolution kernel. In a strong noise environment, it is easy to cause the loss of key fault information. Secondly, the current advanced fault diagnosis model based on the fusion of CNN and Transformer has achieved preliminary results in intelligent fault diagnosis in a strong noise environment, but the prior explainability research of various new self-attention-based variant models is still lacking. SUMMARY

[0006] In view of the problems of poor noise robustness and explainability of current industrial intelligent predictive maintenance, the present application provides an industrial intelligent predictive maintenance method based on an explainable deep feature fusion network, comprising:

[0007] S1: establishing an explainable neural network model, which includes an explainable wavelet multi-scale feature extraction network based on a discrete cosine enhanced attention mechanism, a double-branch feature fusion network based on a self-attention mechanism, and a classifier MLP;

[0008] S2: acquiring gear box vibration data containing noise;

[0009] S3: capturing the explainable features of the gear box vibration data containing noise through the explainable wavelet multi-scale feature extraction network based on the discrete cosine enhanced attention mechanism, while suppressing and filtering the noise signal;

[0010] S4: realizing deep fusion of features through the double-branch feature fusion network based on the self-attention mechanism;

[0011] The double-branch feature fusion network based on the self-attention mechanism comprises a global convolutional self-attention branch for global modeling and a local self-attention branch for focusing on local features.

[0012] S5: applying a convolutional layer to stabilize the learned features, using a classifier MLP to reduce the dimension of the features and output, to obtain a final fault identification result;

[0013] S6: performing an interpretable analysis on the final fault identification result by using output channel weight visualization and wavelet energy spectrum comparative analysis.

[0014] The beneficial effects of the present application are:

[0015] The present application integrates CNN and multi-head self-attention mechanism more deeply, fuses the advantages of local and global feature extraction of each other, makes the decision-making process of the model more transparent and more interpretable, establishes an interpretable deep feature fusion network for industrial intelligent predictive maintenance, extracts local and sparse features at different positions of complex vibration signals, filters out irrelevant information, thereby improving the noise resistance of the model, and realizes the transmission of global and local spatial feature information, and deeper fusion of high-dimensional abstract features, so that features at different levels can adaptively adjust their contribution in the final diagnostic result, so that the industrial intelligent predictive maintenance noise has high robustness and interpretability. BRIEF DESCRIPTION OF DRAWINGS

[0016] Fig. 1 is a flowchart of an industrial intelligent predictive maintenance method based on an interpretable deep feature fusion network according to the present application;

[0017] Fig. 2 is a model framework diagram of an interpretable neural network model according to the present application;

[0018] Fig. 3 is a fault identification result diagram of an embodiment of the present application;

[0019] Fig. 4 is an interpretable analysis diagram of an embodiment of the present application. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0021] An industrial intelligent predictive maintenance method based on an interpretable deep feature fusion network, as shown in Fig. 1, comprises:

[0022] S1: An interpretable neural network model is established, as shown in FIG. 2, which includes an interpretable wavelet multi-scale feature extraction network based on a discrete cosine enhanced attention mechanism, a double-branch feature fusion network based on a self-attention mechanism, and a classifier MLP;

[0023] S2: Obtain gear box vibration data containing noise;

[0024] S3: Capture the interpretable features of the gear box vibration data containing noise while suppressing and filtering the noise signal through the interpretable wavelet multi-scale feature extraction network based on the discrete cosine enhanced attention mechanism;

[0025] S4: Deeply fuse the features through the double-branch feature fusion network based on the self-attention mechanism;

[0026] The double-branch feature fusion network based on the self-attention mechanism includes a global convolutional self-attention branch for global modeling and a local self-attention branch for focusing on local features;

[0027] S5: Apply a convolutional layer to stabilize the learned features, use a classifier MLP to reduce the dimension of the features and output, and obtain the final fault recognition result;

[0028] S6: Perform interpretable analysis on the final fault recognition result using output channel weight visualization and wavelet energy spectrum comparative analysis.

[0029] The interpretable wavelet multi-scale feature extraction network based on the discrete cosine enhanced attention mechanism is constructed, including:

[0030] Embed the discrete wavelet transform into the neural network to expand the feature learning space from the time domain to the wavelet domain; use cross-channel multi-scale convolution to increase the local receptive field of the model while reducing the parameter quantity and floating point operation quantity of the model; introduce the discrete cosine transform to construct the cosine enhanced channel attention mechanism, and obtain the double-scale feature fusion network based on the self-attention mechanism.

[0031] Capture the interpretable features of the gear box vibration data containing noise, including:

[0032] Expand the gear box vibration data containing noise from the time domain to the wavelet domain through the discrete wavelet transform: decompose the input signal into high and low frequency bands through the discrete wavelet transform, and further process the high and low frequency bands;

[0033] Use multi-scale convolution to capture features of different scales in the wavelet domain;

[0034] Use the discrete cosine enhanced channel attention mechanism to adaptively weight the features of different frequency bands, effectively filtering out noise and irrelevant information.

[0035] The formula of the discrete wavelet transform includes:

[0036] Wherein, t is a time series number, f(t) is an original signal, j is a decomposition layer number, H and G are low-frequency and high-frequency wavelet decomposition filters in time domain, A j is a wavelet coefficient of a low-frequency part of the signal f(t) at the jth layer; D j is a wavelet coefficient of a high-frequency part of the signal f(t) at the jth layer.

[0037] Further processing of the low-frequency band and the high-frequency band includes:

[0038] Wherein, x L and x H are low-frequency features and high-frequency features after discrete wavelet decomposition; x cat is an output feature after stacking x L and x H along the channel dimension; Concat() represents a splicing operation along the channel dimension of x L and x H . x cat is the final output size, the output channel number is 2C, and the length of each channel is L / 2.

[0039] Capturing features of different scales in the wavelet domain by using multi-scale convolution includes:

[0040] Wherein, is the output of the jth channel-by-channel convolution, λ is a Gaussian error activation function GELU, μ is batch normalization BN, z out is an output after discrete wavelet transform DWT and multi-scale feature extraction, and Concat() is a splicing operation along the channel dimension of each multi-scale convolution branch.

[0041] Enhancing the channel attention mechanism by using discrete cosine to adaptively weight the features of different frequency bands, effectively filtering out noise and irrelevant information, includes:

[0042] After performing cosine transformation on the features of different frequency bands, the channel attention mechanism is weighted, the frequency channel features are output after performing 1D-DCT on each channel component, each frequency channel feature is stacked and convolution operation is performed, the attention weight is output, a larger weight is applied to the frequency band related to the fault feature, and a smaller weight is applied to the noise frequency band, and finally the output is obtained.

[0043] The process of realizing deep fusion of features includes:

[0044] Further extract global features and spatial information through a global multi-head convolution self-attention branch:

[0045] Wherein, 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, 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, V are respectively the query vector, key-value vector and value vector obtained by inputting the signal processed by the feature extraction module into the global multi-head convolution self-attention branch, unfold() represents the unfolding operation on the input; F att represents the global feature output obtained through the global multi-head self-attention branch, ε is the connection operation on the H attention heads, T represents the transposition operation on the input key-value vector, 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;

[0046] In the local convolution attention branch, the global maximum pooling and discrete cosine transform are executed in parallel to highlight the high-frequency features while aggregating the full-band features, so as to further refine the features: F conv =F(x)*SoftMax(F(x a *x b ))

[0047] Wherein, x is the input of the local convolution attention branch, x a ,x b are respectively the output features processed by GMP and DCT, F() represents the convolution operation with a kernel size of 1, F conv is the local feature output processed by the local convolution attention.

[0048] The learned features are stabilized by applying a convolution layer, and the features are reduced in dimension by using an MLP, to obtain the final fault recognition result, including:

[0049] The residual connection and Dropout method are introduced to reduce the risk of overfitting, and two learnable parameters are introduced in the residual connection to adaptively adjust the importance of the features. Then, a feedforward block composed of two convolution layers is used to integrate and stabilize the output features, and finally a multilayer perceptron (MLP) is used to reduce the dimension of the features to obtain the recognized fault type.

[0050] The output channel weight visualization and wavelet energy spectrum comparative analysis are used to analyze the interpretability of the final fault recognition result, and the interpretability analysis diagram is shown in FIG. 4, including:

[0051] The feature values of the final output channel of the model feature extraction network are obtained, the feature values of each channel are summed and normalized, and then sorted by frequency from low to high to obtain a wavelet energy spectrum, and a fast Fourier transform is performed on the original input signal to obtain a frequency spectrum of the original input signal;

[0052] The wavelet energy spectrum and the frequency spectrum 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 meet the accurate physical meaning in experiments; in addition, the distribution of the main frequency components of the features of different health states is different, which shows that the network has indeed learned interpretable features that can be used for classification.

[0053] Embodiment 1:

[0054] In this embodiment, the gearbox vibration signals are collected by sensors, and a total of five different health states are selected, each of which is assigned a different label; the vibration signals are divided by a sliding window, each sample contains 1024 sampling points, and the training data set and the test set are in a ratio of 7:3; in order to simulate the actual working scene of the gearbox, 0dB, -4dB and -6db Gaussian noise are added to the signal respectively; to simulate weak, medium and strong three noise levels.

[0055] An interpretable deep feature fusion model is constructed, and the model is trained using the training data set. The neural network model interpretable feature extraction module is composed of a layer of discrete wavelet convolution (DWT-CNN), a layer of lightweight multi-scale convolution, and a layer of discrete cosine enhancement channel attention module. The global branch of the double-branch feature fusion module is composed of a 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. And a dropout layer is introduced in the deep feature fusion network, and its parameters are set to 0.1 respectively; and a residual connection is introduced to prevent model overfitting and gradient explosion problem. The specific model structure is as follows:

[0056] d represents the output dimension, h is the number of self-attention heads, k is the CNN kernel size, s is the CNN step size, and N represents the number of health state types.

[0057] 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 during training, the Adam optimization method is used, the initial learning rate is set to 0.001, and when the learning rate does not decrease every 50 epochs, the learning rate is adjusted to 0.1 of the original.

[0058] To verify the feasibility and accuracy of the method, test experiments were conducted, and the model was compared with Convformer-NSE and LiConvFormer, CLFormer three advanced new end-to-end fault identification models based on CNN-Transformer architecture, and the fault identification result diagram is shown in Figure 3. Through the identification experiment of the gearbox fault under three different noise conditions, the comparison results are shown in Table 1:

[0059] Table 1 Comparison table of fault identification results of different neural networks

[0060] In the constant speed gearbox data set, the method in this paper achieved the best diagnostic results under all noise conditions, and the advantage of the proposed method was further manifested as the noise increased. Under the condition of 0dB, LiConvformer and our method reached 100% diagnostic accuracy under constant speed conditions, but the complexity of our model was smaller than the other two methods, so it had smaller calculation overhead and higher cost performance. Under the conditions of-4dB and-6dB, the highest diagnostic accuracy was 0.19% and 0.39% higher than the comparison method, and compared with the best results of the three comparison methods based on CNN-Transformer architecture, the proposed method was 1.09% and 1.10% higher, respectively. Overall, the proposed method showed satisfactory performance on the constant speed data set. In the variable speed data, the advantage of the proposed method gradually appeared as the noise increased, which was 6.66% and 4.65% higher than the best method under the conditions of-4dB and-6dB, respectively. This is because the damage to non-uniform speed signals is more serious than constant speed signals as the noise increases, and the local feature correlation of the signal is weakened to varying degrees, so it is crucial to establish the global dependence of the signal.

[0061] In summary, through the diagnosis experiment of gear faults under different noise levels, it can be seen that the industrial intelligent predictive maintenance method based on interpretable deep feature fusion network can better realize the correct classification of faults and greatly improve the accuracy and reliability of fault diagnosis.

[0062] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the application, the scope of which 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, The method comprises the following steps: S1: establishing an interpretable neural network model, which comprises an interpretable wavelet multi-scale feature extraction network based on a discrete cosine enhanced attention mechanism, a double-branch feature fusion network based on a self-attention mechanism, and a classifier MLP; S2: obtaining gear box vibration data containing noise; S3: capturing interpretable features of the gear box vibration data containing noise through the interpretable wavelet multi-scale feature extraction network based on the discrete cosine enhanced attention mechanism, while suppressing and filtering noise signals; S4: realizing deep fusion of the features through the double-branch feature fusion network based on the self-attention mechanism; The double-branch feature fusion network based on the self-attention mechanism comprises a global convolution self-attention branch for global modeling and a local self-attention branch for focusing on local features; S5: applying a convolution layer to stabilize the learned features, using the classifier MLP to reduce the dimension of the features and outputting a final fault recognition result; S6: performing interpretable analysis on the final fault recognition result through output channel weight visualization and wavelet energy spectrum comparative analysis. 2.The industrial intelligent predictive maintenance method based on the interpretable deep feature fusion network according to claim 1, characterized in that, The interpretable wavelet multi-scale feature extraction network based on the discrete cosine enhanced attention mechanism is constructed, comprising: The discrete wavelet transform is embedded into the neural network to expand the feature learning space from the time domain to the wavelet domain; the cross-channel multi-scale convolution is used to increase the local receptive field of the model while reducing the parameter quantity and floating point operation quantity of the model; the discrete cosine transform is introduced to construct the cosine enhanced channel attention mechanism, and the double-scale feature fusion network based on the self-attention mechanism is obtained. 3.The industrial intelligent predictive maintenance method based on the explainable deep feature fusion network according to claim 1, characterized in that, The interpretable features of the gear box vibration data containing noise are captured, comprising: The gear box vibration data containing noise is expanded from the time domain to the wavelet domain through the discrete wavelet transform: the input signal is decomposed into high and low frequency bands through the discrete wavelet transform, and the high and low frequency bands are further processed; The multi-scale convolution is used 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, effectively filtering out noise and irrelevant information. 4.The industrial intelligent predictive maintenance method based on the interpretable deep feature fusion network according to claim 3, characterized in that, The formula of the discrete wavelet transform includes: where t is a time series number, f(t) is an original signal, j is a decomposition layer number, H, G are low-frequency and high-frequency wavelet decomposition filters in time domain, A j is a wavelet coefficient of a low-frequency part of the signal f(t) at the jth layer; D j is a wavelet coefficient of a high-frequency part of the signal f(t) at the jth layer. 5.The industrial intelligent predictive maintenance method based on the explainable deep feature fusion network according to claim 3, characterized in that, Further processing of the high and low frequency bands includes: wherein x L and x H are low-frequency features and high-frequency features after discrete wavelet decomposition, respectively; x cat is an output feature after stacking x L and x H along the channel dimension; and Concat() is a concatenation operation. For x cat The final output size is 2C output channels, each of length L / 2. 6.The industrial intelligent predictive maintenance method based on the explainable deep feature fusion network according to claim 3, characterized in that, Utilizing multi-scale convolution to capture features of different scales in wavelet domain, including: wherein, For the output of the jth channel-wise convolution, λ is the Gaussian Error Linear Unit (GELU) activation function, μ is the Batch Normalization (BN), and z out For the output after Discrete Wavelet Transform (DWT) and multi-scale feature extraction, Concat() is the concatenation operation. 7.The industrial intelligent predictive maintenance method based on the explainable deep feature fusion network according to claim 3, characterized in that, The discrete cosine enhanced channel attention mechanism is used to adaptively weight the features of different frequency bands, effectively filtering out noise and irrelevant information, comprising: After the cosine transform of the features of different frequency bands, the channel attention mechanism is weighted, the frequency channel features output after the 1D-DCT of each channel component are stacked, and the convolution operation is performed to output the attention weight, a larger weight is applied to the frequency band related to the fault feature, a smaller weight is applied to the noise frequency band, and the final output is obtained. 8.The industrial intelligent predictive maintenance method based on the explainable deep feature fusion network according to claim 1, characterized in that, The process of realizing deep fusion of the features, comprising: Further extract global features and spatial information through a global multi-head convolution self-attention branch: wherein, ξ is the position embedding, i∈(1, 2, 3,..., H) represents the attention head number, F() represents performing a convolution operation with a kernel size of 1, 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, V are respectively the query vector, the key-value vector and the value vector obtained by inputting the signal processed by the feature extraction module into the global multi-head convolution self-attention branch, unfold() represents performing an unfolding operation on the input; F att represents the global feature output obtained by the global multi-head self-attention branch, ε is a connection operation on H attention heads, T represents a transposition operation, 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. In the local convolution attention branch, the global maximum pooling and the discrete cosine transform are executed in parallel to highlight the high-frequency features and aggregate the full-band features, so as to further refine the features: F conv = F(x) * SoftMax(F(x a *x b )) wherein x is the input of the local convolution attention branch, x a b are the output features after GMP and DCT processing, respectively, F() represents a convolution operation with a kernel size of 1, and F conv is the local feature output after local convolution attention processing.​ 9.The industrial intelligent predictive maintenance method based on the explainable deep feature fusion network according to claim 1, characterized in that, The convolution layer is applied to stabilize the learned features, the MLP is used to reduce the dimension of the features, and the final fault recognition result is obtained, comprising: The residual connection and Dropout method are introduced to reduce the risk of overfitting, and two learnable parameters are introduced in the residual connection to adaptively adjust the importance of features. Then, a feedforward block composed of two convolutional layers is used to integrate stable output features. Finally, a multilayer perceptron (MLP) is used to reduce the dimension of the features and output the recognized fault type. 10.The industrial intelligent predictive maintenance method based on the interpretable deep feature fusion network according to claim 1, characterized in that, The output channel weight visualization and wavelet energy spectrum comparative analysis are used to analyze the interpretability of the final fault recognition results, including: The feature values of the final output channel of the model feature extraction network are obtained, and the feature values of each channel are summed and normalized. Then, the wavelet energy spectrum is obtained by sorting the frequency from low to high. Meanwhile, the frequency spectrum of the original input signal is obtained by fast Fourier transform. The wavelet energy spectrum and the frequency spectrum of the original signal are compared and analyzed to determine whether the main frequency components in the wavelet energy spectrum match those in the original signal frequency spectrum, which is used to show that the learned features meet the accurate physical meaning in the experiment. In addition, the distribution of the main frequency components of the features of different health states is different, which shows that the network has indeed learned interpretable features that can be used for classification.

Images