A variable-condition gearbox fault diagnosis method based on fusion feature extraction network
By using WPT and deep fusion feature extraction network in deep learning fault diagnosis method, combined with ResNet and Transformer encoder, the migration fault diagnosis model is optimized, and the inadequate utilization of time and frequency information and poor feature adaptability in cross-domain diagnosis of deep learning methods is solved. The problem is to achieve high accuracy and strong generalization capabilities under varying operating conditions.
Patent Information
- Application Number
- CN202311220251.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-21
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2043-09-21
Smart Images

Figure CN117272172B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of mechanical equipment fault diagnosis, and in particular relates to a variable operating condition gearbox fault diagnosis method based on a fusion feature extraction network. Background Art
[0002] In industrial scenarios, rotating machinery and equipment are in a high-load working environment for a long time, which can easily cause wear of its parts, leading to equipment failure and even safety accidents. Establishing a fault diagnosis model for rotating machinery and equipment can ensure the safe and stable operation of the equipment and avoid the occurrence of safety accidents, which is of great practical significance.
[0003] Rotating machinery fault diagnosis methods can be divided into two categories: model-based and data-driven methods. The former requires a lot of prior knowledge to establish a mathematical model, and the diagnostic performance is easily affected by the environment. The latter extracts fault features through data processing, feature extraction and other steps, and analyzes the equipment status. In contrast, data-driven methods have more advantages because a large amount of available data can be collected from sensors. In recent years, data-driven fault diagnosis methods combined with deep learning have attracted the attention of researchers. These methods automatically extract deep features of vibration signals through deep neural networks, effectively solving the problem of traditional fault diagnosis models' dependence on empirical feature selection.
[0004] Fault diagnosis methods based on deep learning need to meet the requirement that the training set and the test set obey the same feature distribution, otherwise the trained model will fail. However, in actual industrial scenarios, due to changes in equipment operating conditions, there are often large differences between data. Transfer learning can extract domain-invariant features between two related fields, realizing the migration and reuse of diagnostic models between the two fields.
[0005] Research on deep transfer learning has made some progress in the field of fault diagnosis and improved the generalization ability of the model in cross-domain fault diagnosis scenarios. However, there are two issues that deserve further study. 1) The time-frequency information of the vibration signal contains the fault mechanism of the gearbox equipment. How to more effectively use the time-frequency information to improve the recognition accuracy of the diagnosis model. 2) The convolution-based deep learning network model pays more attention to the local features of the input information. How to further optimize the deep learning network to better extract the global time-frequency distribution features of the vibration signal, thereby improving the domain adaptability of the deep features. Summary of the invention
[0006] The purpose of the present invention is to provide a variable operating condition gearbox fault diagnosis method based on a fusion feature extraction network to solve the problems existing in the above-mentioned prior art.
[0007] To achieve the above object, the present invention provides a variable operating condition gearbox fault diagnosis method based on a fusion feature extraction network, comprising:
[0008] Acquire original vibration signals under different working conditions, and perform time-frequency analysis on the original vibration signals to obtain a time-frequency information grayscale image;
[0009] Build a lightweight ResNet network and a Transformer encoder, and combine the Transformer encoder with the lightweight ResNet network to form a deep fusion feature extraction network;
[0010] Obtaining deep fusion features based on the time-frequency information grayscale image and the deep fusion feature extraction network;
[0011] The deep fusion feature extraction network is combined with a classifier to construct a migration fault diagnosis model;
[0012] Constructing a total loss function of the migration fault diagnosis model, wherein the loss function includes a fault classification loss function of source domain data and a probability distribution difference loss function of deep fusion features of source domain and target domain;
[0013] By minimizing the total loss function, the migration fault diagnosis model is optimized using the back propagation and gradient descent methods; the variable operating condition gearbox fault diagnosis is realized through the optimized model.
[0014] Optionally, the process of obtaining the time-frequency information grayscale image includes:
[0015] The original vibration signal is subjected to four-layer wavelet packet decomposition to obtain a plurality of wavelet packet leaf nodes; a time-frequency information grayscale image can be obtained based on the wavelet packet coefficient of a leaf node, and a plurality of time-frequency information grayscale images can be obtained based on a plurality of wavelet packet leaf nodes.
[0016] Optionally, the wavelet packet decomposition uses Dobesi wavelet as wavelet basis function.
[0017] Optionally, the lightweight ResNet network is a lightweight convolutional network structure including three layers, using a 3*3 convolution kernel for feature extraction, the convolution step size is set to 1, and the last layer is set to an average pooling layer.
[0018] Optionally, the process of obtaining deep fusion features includes:
[0019] The time-frequency information grayscale image is passed through the lightweight ResNet network to obtain local time-frequency features at multiple scales; the local time-frequency features are preprocessed, and the associated feature information of time-frequency features of different scales is extracted through the Transformer encoder to obtain global time-frequency features, and then the local time-frequency features are connected with the global time-frequency features to obtain deep fusion features.
[0020] Optionally, several time-frequency information grayscale images are respectively input into the lightweight ResNet network to extract local feature information of the time-frequency grayscale images at different scales.
[0021] Optional pre-processing steps include:
[0022] Linear mapping is performed on several local time-frequency features to obtain several mapping features, a random initialization flag is added to the head of each mapping feature to obtain an initial matrix, and position information is added to the initial matrix to obtain an input matrix fused with position information.
[0023] Optionally, the local time-frequency distribution feature is concatenated with the global time-frequency distribution feature to obtain local and global deep fusion features of the vibration signal, thereby enhancing the feature expression capability of the model.
[0024] Optional, state classification loss function L in the source domain C for:
[0025]
[0026] In the formula, m is the number of samples, n is the number of categories, S ic is a sign function with value 0 or 1, p ic is the probability that sample i belongs to category c.
[0027] Optionally, the distribution difference loss of deep fusion features between the source domain and the target domain is:
[0028]
[0029] In the formula, E is the expectation, the source domain deep fusion feature and the target domain deep fusion feature are denoted by F s and F t ,The mapping method adopts weighted kernel mapping of multiple kernel functions.
[0030] The technical effects of the present invention are:
[0031] The present invention proposes a novel gearbox migration fault diagnosis model under variable working conditions, uses WPT to construct a time-frequency information grayscale map as the input of a deep fusion feature extraction network, and more reasonably organizes the time-frequency information of the vibration signal; designs a deep fusion feature extraction network combining ResNet and Transformer, designs a lightweight ResNet, extracts local features from each sub-band information of the wavelet packet leaf node, and designs a Transformer encoder module to extract related features (global features) between the sub-band local features; introduces a multi-core maximum mean difference to optimize the deep feature extraction network, reduces the probability distribution difference of the deep features of the source domain and the target domain data, improves the variable working condition migration diagnosis capability of DAFDMRT, and improves the domain adaptation capability of the deep features under variable working conditions; the gearbox migration fault diagnosis model under variable working conditions of the present invention can achieve high diagnostic accuracy in various migration modes, and has strong adaptability to variable working condition scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:
[0033] Figure 1 It is the DAFDMRT fault diagnosis model architecture in the embodiment of the present invention;
[0034] Figure 2 Schematic diagram of the time-frequency analysis process in an embodiment of the present invention;
[0035] Figure 3 is a flow chart of local feature extraction in an embodiment of the present invention;
[0036] Figure 4 is a flow chart of global time-frequency distribution feature extraction in an embodiment of the present invention;
[0037] Figure 5 Figure 2 shows the confusion matrices and t-SNE graphs of the five ablation models under the G1->G2 migration task in this embodiment. (a) is the confusion matrix and corresponding t-SNE graph of the state classification results of the NW1 model, (b) is the confusion matrix and corresponding t-SNE graph of the state classification results of the NW3 model, (c) is the confusion matrix and corresponding t-SNE graph of the state classification results of the NW5 model, (d) is the confusion matrix and corresponding t-SNE graph of the state classification results of the NW6 model, and (e) is the confusion matrix and corresponding t-SNE graph of the state classification results of the NW7 model. DETAILED DESCRIPTION
[0038] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.
[0039] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.
[0040] Embodiment 1
[0041] like Figure 1-5 As shown, this embodiment provides a variable operating condition gearbox fault diagnosis method based on a fusion feature extraction network, including:
[0042] First, this paper uses WPT to perform time-frequency signal analysis on vibration signals. Then, in order to extract local and global features of time-frequency information, we design a deep feature extraction method that combines ResNet and Transformer encoders. Finally, in order to solve the problem of feature distribution differences between the source domain and the target domain in actual variable operating conditions, MK-MMD is introduced to establish a domain adaptive fault diagnosis model based on ResNet and Transformer, that is, a novel gearbox migration fault diagnosis model under variable operating conditions (Domain adaptive fault diagnosis modelbased on ResNet and Transformer, DAFDMRT).
[0043] like Figure 1 As shown in Figure 1, DAFDMRT consists of a time-frequency processing (TFP) module, a deep fusion feature extraction network (DFFEN), and a domain-adaptive feature migration module. In this work, WPT is used to extract the time-frequency information of vibration signals of different scales. Then, TFIGMs are constructed using the wavelet packet coefficients of the leaf nodes. Next, ResNet is used to extract the local time-frequency feature (TFF) of TFIGMs. L , and use the Transformer encoder to extract the deep joint features X between leaf nodes G , we get the deep feature X of the vibration signal fused with local and global TFF FFinally, in order to minimize the feature distribution difference between the source domain and the target domain, this paper also designs a domain adaptation strategy based on MK-MMD, which minimizes the classification loss and the feature distribution difference loss L between the source domain and the target domain. MK-MMD To optimize the network parameters of DFFEN. Therefore, DAFDMRT can provide high diagnostic accuracy and strong generalization ability under variable working condition migration tasks.
[0044] Signal time-frequency preprocessing module
[0045] The TFF of vibration signals contains fault mechanisms. WPT is an effective tool for TFP of non-stationary nonlinear vibration signals. Therefore, this paper uses WPT to preprocess the original vibration signals. Considering the excellent performance of Daubechies wavelet in processing signals with non-stationary and nonlinear characteristics, we selected Daubechie wavelet (db3) as the wavelet basis function, and the decomposition level was set to 4. After decomposing the vibration signal using wavelet transform, multiple wavelet packet leaf nodes can be obtained, and the wavelet packet coefficients can present the MSTFI of the vibration signal. Figure 2 As shown in the figure, the vibration signal has 4096 samples. After 4 layers of wavelet packet decomposition, 16 wavelet packet leaf nodes can be obtained. A TFIGM with a dimension of 16*16 is constructed using the 256 wavelet packet coefficients of a leaf node, and then each TFIGM is sent to a separate ResNet to extract the deep features of MSTFI.
[0046] Deep Fusion Feature Extraction Network
[0047] DFFEN includes two parts: local feature extraction and global feature extraction. First, ResNet is used to extract the local features of each wavelet packet leaf node TFIGM, and then the Transformer encoder is used to extract the correlation features between nodes. Finally, the local features are fused with the global features for subsequent fault diagnosis.
[0048] Local feature extraction module
[0049] In view of the small size of TFIGM, a 3-layer lightweight ResNet structure is designed here. The ResNet network uses 3*3 convolution kernels for feature extraction, the convolution step size is set to 1, and the pooling operation of the middle layer is cancelled, only the last average pooling layer is retained. The network structure is as follows: Figure 3 As shown. Through 16 independent lightweight ResNet networks, Local depth features , enter X WPT-TFG There are 16 16*16 TFIGMs, output X Lis a feature vector of 16 dimensions of 1 × 128. The ResNet parameter description is shown in Table 1.
[0050] Table 1
[0051]
[0052] Global feature extraction module
[0053] ResNet always has shortcomings such as lack of capturing the relationship between objects, equal treatment of all pixels, lack of targeting, etc. In addition, the local receptive field of the convolution kernel requires stacking many convolution layers to obtain global information. Therefore, it is necessary to design a Transformer encoder module for extracting global features. This paper introduces the deep local features extracted by ResNet. As the input of the Transformer encoder, the global time-frequency distribution features of the vibration signal are extracted, such as Figure 4 shown.
[0054] like Figure 4 As shown, for 16 local depth features X L Perform linear mapping to obtain 16 mapping features Then, in X P Add a random initialization mark X to the head 0 , we can get In addition, in order to effectively utilize X C The order information in X C Add location information E pos , get the input matrix X T , which can be expressed as:
[0055]
[0056] Among them, X T is the input of the Transformer encoder. The multi-head self-attention mechanism is the calculation core of each layer of the Transformer encoder block, and its calculation formula is as follows:
[0057]
[0058] A n (X T )=Contact(H1,H2,L,H n )W 0 (3)
[0059] X MHSA =LayerNorm(A n (X T )+X T ) (4)
[0060] Among them, Q i =XW i q , K i =XW i k andV i =XW i v , where W i q , W i k , W i v By embedding X T Get Q i , K i , V i The i-th linear mapping, W O is a multi-head linear mapping, X MHA is the output of the multi-head self-attention module.
[0061] The output of the multi-head self-attention module passes through a two-layer feed-forward network whose hidden layer is activated by ReLU:
[0062] FF(X MHSA )=Relu(0,W1X MHSA +b1)W2+b2 (5)
[0063] Where W1, W2, b1, and b2 are the weights and biases of the two layers respectively. The output of the first block of the Transformer encoder can be expressed as:
[0064] X Block =LayerNorm(FF(X MHSA )+X MHSA ) (6)
[0065] X Block As the input of block2, repeat the above steps in sequence to extract the global feature X of the vibration signal G , X G is the output of DFFEN. Finally, X L With X G Perform splicing operations to obtain the deep fusion features of the vibration signal time-frequency information
[0066] This paper uses a four-layer wavelet transform to process the vibration signal and obtains 16 wavelet packet leaf nodes. A 16*16 grayscale image is constructed using the coefficients of each leaf node, and 16 grayscale images are used to represent the information of 16 sub-bands of the original vibration signal from low to high frequency. First, ResNet is used to extract features from each grayscale image to obtain local feature vectors (local features) of 16 sub-bands. Secondly, the local feature vector is used as input, and the Transformer encoder is used to extract the correlation information (global features) between the 16 sub-band features. Finally, by combining local features with global features, the deep fusion features of the vibration signal can be obtained.
[0067] Optimization parameter description
[0068] In this section, we construct the state classification loss L of the source domain in the DAFDMRT model. C and the distribution difference loss L between the source domain and the target domain deep features MK-MMD ,By minimizing the loss function, the network parameters are updated to ensure that the migration fault diagnosis model has a high fault identification accuracy and the ability to adapt to changing working conditions.
[0069] The source domain deep fusion features and target domain deep fusion features obtained by DFFEN are denoted as F s and F t , L MK-MMD Indicated as F s With F t The data distribution difference between them is lost.
[0070]
[0071] Where E is the expectation, and the mapping method uses weighted kernel mapping of multiple kernel functions. s After passing through the state classifier composed of the fully connected layer and the Softmax layer, the predicted label of the source domain sample can be obtained. According to the true label of the source domain data, the classification loss L can be calculated C :
[0072]
[0073] Among them, m is the number of samples, n is the number of categories, S ic is a sign function with value 0 or 1, p ic is the probability that sample i belongs to category c. Combined with L C and L MK-MMD The total loss L of the migration fault diagnosis model can be calculated DAFDMRT , as shown in formula (9):
[0074] L DAFDMRT =L C +λLMK-MMD (9)
[0075] Where λ is the weight coefficient. By minimizing L DAFDMRT , back-propagation and gradient descent methods are used to optimize the DFFEN model and classifier parameters.
[0076] Experimental verification
[0077] In order to verify and analyze the fault diagnosis effect of the DAFDMRT model proposed in this paper under variable working conditions, we use the gearbox dataset in this section to experimentally verify the model. The dataset contains four fault types and one health state, namely chipped fault (CF), missing fault (MF), root fault (RF), surface fault (SUF) and health state. At the same time, two speed-load conditions are set in the dataset, namely 20-0 and 30-2 (for 20-0, 20 means the speed is 20r / s, and 0 means the load is 0HP), which are respectively represented as datasets G1 and G2. The settings of the dataset are shown in Table 2. Under the working conditions of G1 and G2, the data of the above five states are collected, 250 data samples are set under each working condition, and the sample length is 4096.
[0078] Table 2
[0079]
[0080] Experimental setup
[0081] In order to verify the effectiveness of the DAFDMRT model, we set up migration experiments between different speeds and loads of different equipment. The specific settings of the working condition migration task are G1->G2 and G2->G1 (G1->G2 means that the source domain data is G1 and the target domain data is G2). In addition, we also introduced a variety of ablation experiment models and existing migration diagnosis models for comparative experiments. The specific description of the comparison method is shown in Table 3. Group 1 is the ablation experiment model, and Group 2 is the existing migration fault diagnosis model.
[0082] Table 3
[0083]
[0084] Compared with DAFDMRT, NW1 cancels the work of WPT for time-frequency information extraction and directly sends the original vibration signal to the DFFEN network. Other designs are consistent with DAFDMRT. Compared with DAFDMRT, NW2 and NW3 use 3-layer CNNs to replace ResNet, and NW2 cancels the step of Transformer encoder extracting global features. NW4 and NW5 use 5-layer and 7-layer ResNet to replace the 3-layer ResNet network in DAFDMRT respectively. NW6 cancels the step of Transformer encoder extracting global features in the DAFDMRT model.
[0085] In Group 2, the DAFDMRT model proposed in this paper is compared with other transfer learning methods, including transfer component analysis (TCA), joint distribution adaptation (JDA), correlation alignment for convolutional neural network (CNN-CORAL), maximum mean discrepancy for convolutional neural network (CNN-MMD) and adversarial discriminative domain adaptation (ADDA).
[0086] Ablation Experiment Results and Analysis
[0087] The experimental results of the ablation model comparison under different working condition migration tasks are shown in Table 4. DAFDMRT (NW7) can obtain higher average accuracy than other methods, with an average accuracy of 96.06%. Compared with NW1, the average accuracy of DAFDMRT has increased by more than 12%, and the results under each migration task have been greatly improved, which shows that it is necessary to perform time-frequency processing on the original vibration signal. Compared with NW3, the average accuracy of DAFDMRT has increased by 8.63%, indicating that ResNet can more effectively extract the local depth features of the vibration signal than the traditional convolution method. The comparison of the results of NW4, NW5 and NW7 shows that simply increasing the depth of the network cannot improve the accuracy of the diagnostic model, but will cause overfitting due to the increase in network complexity. At the same time, this also verifies the effectiveness of designing a lightweight ResNet based on the small size of TFIGM. Compared with NW6 and NW7, the average accuracy of NW7 has increased by 4.4%, indicating that it is necessary to use the Transformer encoder to further extract the global features of the time-frequency information of the vibration signal. Similarly, the comparison results of NW2 and NW3 can also illustrate this point.
[0088] Table 4
[0089]
[0090] Figure 5 The confusion matrix and t-SNE deep feature visualization of the diagnosis results of each model in the migration task G1->G2. Including the confusion matrix and t-SNE deep feature visualization of NW1, NW3, NW5, NW6, and NW7. The left figure is the confusion matrix of the model state classification result under the migration task G1->G2; the right figure is the corresponding t-SNE figure.
[0091] From the results of the confusion matrix, we can see that DAFDMRT has achieved the highest recognition accuracy; through t-SNE, we can see that DAFDMRT can obtain the maximum inter-class distance and the minimum intra-class distance, making the deep features of the target domain have better state separability. The experimental comparison results further show that DAFDMRT can effectively extract the time-frequency features of the vibration signal by integrating WPT, Resnet, and Transformer encoder, and use MK-MMD to optimize the deep feature extraction network, which can better narrow the distribution difference between the deep features of the source domain and the target domain.
[0092] Comparative Analysis of Migration Fault Diagnosis Models
[0093] The experimental results of the comparison between DAFDMRT and other migration models in Group 2 on the gearbox test platform are shown in Table 5. DAFDMRT achieved the highest average accuracy, with the highest average accuracy improvement of 16%, and DAFDMRT achieved the best performance in various working condition migration tasks. In general, under variable working condition scenarios, compared with other migration fault diagnosis models, DAFDMRT can better extract fault features that reflect the local and global time-frequency characteristics of vibration signals, and can more effectively reduce the distribution difference between the source domain and the target domain, and has a better diagnostic effect.
[0094] Table 5
[0095]
[0096]
[0097] The above is only a preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.
Claims
1. A variable-condition gearbox fault diagnosis method based on a fusion feature extraction network, characterized in that: The following steps are involved: Acquire original vibration signals under different working conditions, and perform time-frequency analysis on the original vibration signals to obtain a time-frequency information grayscale image; The process of obtaining the grayscale image of time-frequency information includes: The original vibration signal is subjected to four-layer wavelet packet decomposition to obtain a plurality of wavelet packet leaf nodes; a time-frequency information grayscale image can be obtained based on the wavelet packet coefficient of a leaf node, and a plurality of time-frequency information grayscale images can be obtained based on a plurality of wavelet packet leaf nodes; Wavelet packet decomposition uses Dobesi wavelet as wavelet basis function; Several time-frequency information grayscale images are input into the lightweight ResNet network to extract the local feature information of the time-frequency grayscale images at different scales; Build a lightweight ResNet network and a Transformer encoder, and combine the Transformer encoder with the lightweight ResNet network to form a deep fusion feature extraction network; Obtaining deep fusion features based on the time-frequency information grayscale image and the deep fusion feature extraction network; Optionally, the lightweight ResNet network is a lightweight convolutional network structure including three layers, a 3*3 convolution kernel is used for feature extraction, the convolution step is set to 1, and the last layer is set to an average pooling layer; Optionally, the process of obtaining deep fusion features includes: The time-frequency information grayscale image is processed by the lightweight ResNet network to obtain local time-frequency features at multiple scales; the local time-frequency features are preprocessed, and the associated feature information of time-frequency features at different scales is extracted by the Transformer encoder to obtain global time-frequency features, and then the local time-frequency features are connected with the global time-frequency features to obtain deep fusion features; The local time-frequency features are concatenated with the global time-frequency features to obtain local and global deep fusion features of the vibration signal, thereby enhancing the feature expression capability of the model; Optional pre-processing steps include: Linear mapping is performed on several local time-frequency features to obtain several mapping features, a random initialization flag is added to the head of each mapping feature to obtain an initial matrix, and position information is added to the initial matrix to obtain an input matrix fused with the position information; The deep fusion feature extraction network is combined with a classifier to construct a migration fault diagnosis model; Constructing a total loss function of the migration fault diagnosis model, wherein the loss function includes a fault classification loss function of source domain data and a probability distribution difference loss function of deep fusion features of source domain and target domain; By minimizing the total loss function, the migration fault diagnosis model is optimized using the back propagation and gradient descent methods; the variable operating condition gearbox fault diagnosis is realized through the optimized model.
2. The variable operating condition gearbox fault diagnosis method based on fusion feature extraction network according to claim 1 is characterized in that: The state classification loss function L in the source domain C for: In the formula, m is the number of samples, n is the number of categories, S ic is a sign function with value 0 or 1, p ic is the probability that sample i belongs to category c.
3. The variable operating condition gearbox fault diagnosis method based on fusion feature extraction network according to claim 1 is characterized in that: The distribution difference loss of deep fusion features between source domain and target domain is: In the formula, E is the expectation, the source domain deep fusion feature and the target domain deep fusion feature are denoted by F s and F t ,The mapping method adopts weighted kernel mapping of multiple kernel functions.
