Gearbox fault diagnosis method based on domain adaptation residual network and multi-channel fusion

Abstract

The present invention discloses a gearbox fault diagnosis method based on a domain adaptive residual network and multi-channel fusion, which is performed in the following steps: Step S1: Acceleration signals of gears and bearings under various working conditions and fault states are collected by acceleration sensors installed at different positions of the gearbox, and the input voltage signal and output voltage signal of the gearbox are used as sample data; Step S2: VMD decomposition is performed on the acceleration signal, input voltage signal and output voltage signal in the sample data respectively through three channels, and parameter optimization is performed to obtain optimized sample data; the present invention obtains higher diagnostic accuracy by simultaneously processing multi-channel data of multi-source sensors, and relies on multi-channel fusion data to obtain common features to perform gearbox fault classification prediction under multiple working conditions, which has strong generalization ability and reduces the demand for the number of samples.

Description

Technical Field

[0001] The present invention relates to the field of gearbox fault detection, and in particular to a gearbox fault diagnosis method based on domain adaptation residual network and multi-channel fusion. Background Art

[0002] Gearboxes are widely used in automobiles, machinery, wind power generation, robots and other fields. They are mainly used to realize functions such as power transmission and torque conversion. They are key components of rotating machinery. Due to long-term working conditions, gears and bearings in gearboxes are prone to failures, which may cause downtime and production stoppage, and may cause casualties, bringing serious disasters to production and life. Due to the large number of internal parts of gearboxes, the types of faults are complicated, including nonlinear original vibration signals and a large amount of noise. Relying on traditional signal processing and manual feature extraction is not only time-consuming and labor-intensive, but also has low diagnostic accuracy. With the rapid development of artificial intelligence, deep learning technology has gradually attracted attention, but fault diagnosis methods based on deep learning usually assume that the training set and test set data come from the same distribution and contain a large number of fault data samples. However, in actual industry, the operating environment of gearboxes is often variable speed and variable load, which may cause differences in the feature space of the data collected by the sensor, thereby reducing the fault diagnosis accuracy of the model. Summary of the invention

[0003] The purpose of the present invention is to provide a gearbox fault diagnosis method based on domain adaptation residual network and multi-channel fusion. The present invention obtains higher diagnostic accuracy by simultaneously processing multi-channel data of multi-source sensors, and relies on multi-channel fusion data to obtain common features to perform gearbox fault classification prediction under multiple working conditions, with strong generalization ability and reduced demand for sample quantity.

[0004] The technical solution of the present invention is a gearbox fault diagnosis method based on domain adaptive residual network and multi-channel fusion, which is carried out according to the following steps:

[0005] Step S1: Acceleration sensors installed at different positions of the gearbox are used to collect acceleration signals of gears and bearings under various working conditions and fault conditions, and the input voltage signal and output voltage signal of the gearbox are used as sample data;

[0006] Step S2: performing VMD decomposition and parameter optimization on the acceleration signal, input voltage signal and output voltage signal in the sample data using three channels respectively to obtain optimized sample data;

[0007] Step S3: Divide the optimized sample data into a training set, a test set and a validation set according to the fault status category, and perform normalization processing;

[0008] Step S4: Construct the residual network model training. The model performs multi-channel extraction, weight adjustment and feature fusion on the training set. The maximum mean difference and cross entropy classification loss are used as optimization targets. The softmax function is used to classify and output the predicted probability of each fault category. The model parameters are determined through the validation set and the test set.

[0009] Step S5: fine-tuning model parameters to perform feature migration to achieve gearbox fault classification prediction under multiple working conditions;

[0010] Step S6: Use the trained model to perform gearbox fault diagnosis.

[0011] In the above-mentioned gearbox fault diagnosis method based on domain adaptive residual network and multi-channel fusion, the acceleration sensor of step S1 is arranged at the fixed plate of the output shaft of the gearbox and the fixed plate position of the output shaft.

[0012] In the aforementioned gearbox fault diagnosis method based on domain adaptive residual network and multi-channel fusion, the fault state of step S1 includes any combination of one or more of gear longitudinal peeling fault, electrical erosion fault, tooth breakage fault, bearing rolling element fault, inner ring fault, outer ring fault, input shaft imbalance fault, and output shaft key shearing fault.

[0013] In the aforementioned gearbox fault diagnosis method based on domain adaptation residual network and multi-channel fusion, the VMD decomposition mathematical expression of step S2 is as follows:

[0014]

[0015] In the formula, a k (t) is the instantaneous amplitude of the intrinsic mode function obtained by VMD decomposition; is the instantaneous phase of the eigenmode function.

[0016] In the aforementioned gearbox fault diagnosis method based on domain adaptation residual network and multi-channel fusion, the maximum mean difference expression of step S4 is:

[0017]

[0018] In the formula, n is the number of samples in the source domain, m is the number of samples in the target domain, is the i-th sample in the source domain, is the jth sample in the target domain.

[0019] In the aforementioned gearbox fault diagnosis method based on domain adaptation residual network and multi-channel fusion, the model of step S4 includes a multi-channel extraction module, a feature fusion module, an attention mechanism module, a feature extraction module and a fault identification module;

[0020] The multi-channel extraction module extracts features from the input data to obtain feature maps, and then performs discarding and pooling operations to retain important features;

[0021] The feature fusion module fuses the features extracted by the multi-channel extraction module, and adjusts the weight of each channel in the feature map through the attention mechanism module;

[0022] The feature extraction module avoids gradient vanishing through an improved residual network, allows information to flow quickly through the model through cross-layer connections, and improves the learning rate of subtle changes in data and feature information;

[0023] The fault identification module transforms the high-level feature map into a one-dimensional shape through a flattening layer, and then inputs it into the fully connected layer, realizes classification through the softmax function, and combines the MMD loss and the cross entropy loss as the total loss function. During the back propagation of the residual network, the gradients of the MMD loss and the source task classification loss are calculated respectively, and the gradient descent algorithm is used to minimize the MMD loss to reduce the data distribution difference and realize parameter update.

[0024] In the aforementioned gearbox fault diagnosis method based on domain adaptation residual network and multi-channel fusion, the feature extraction process of the feature fusion module is performed through a convolution layer, and the convolution kernel size of the convolution layer is changed through an adaptive function, which is defined as follows:

[0025]

[0026] Where K is the coverage of cross-channel interaction, c is the channel dimension, γ=2, and b=1.

[0027] In the aforementioned gearbox fault diagnosis method based on domain adaptation residual network and multi-channel fusion, the mathematical expression of the cross-layer connection of the feature extraction module is as follows:

[0028] H(x)=F(x)+x;

[0029] Where H(x) is the unknown mapping; F(x) is the residual mapping; x is the identity mapping.

[0030] In the aforementioned gearbox fault diagnosis method based on domain adaptation residual network and multi-channel fusion, the gradient descent algorithm expression of the fault identification module is as follows:

[0031] L 总 =L 分类 +λL mmd ;

[0032] Where, L 总 Represents the total loss, and λ is a hyperparameter that adjusts the proportion of MMD loss.

[0033] In the aforementioned gearbox fault diagnosis method based on domain adaptation residual network and multi-channel fusion, the model training strategy in step S4 adopts Adam optimizer to accelerate the convergence speed, and uses particle swarm optimization algorithm to optimize the hyperparameters of the model. The classification effect of the model is evaluated by accuracy index, and the model training is carried out with the total loss function as the optimization criterion.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] 1. The acceleration signal, input voltage signal and output voltage signal are used as sample data, and then the data is classified and fused through multiple channels to improve the diversity and comprehensiveness of information, reduce the dependence on data distribution, and thus improve the classification performance of the model in the target field.

[0036] 2. The present invention decomposes and fuses multi-channel sample data through VMD, and then adjusts the channel weights, which can enhance weak gearbox fault signals and thus improve the model prediction accuracy.

[0037] 3. The present invention obtains general features by performing deeper feature extraction on multi-channel fused data through a model, and fine-tunes the general features for use in simulating fault classification and prediction of gearboxes under multiple working conditions. It has strong generalization ability and reduces the demand for sample quantity. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic diagram of the process of the present invention;

[0039] Figure 2 It is a schematic diagram of the flow of VMD decomposition of the present invention;

[0040] Figure 3 It is the result diagram of signal decomposition of the present invention;

[0041] Figure 4 It is a schematic diagram of the model framework of the present invention;

[0042] Figure 5 It is a visualized classification diagram after the source task of the present invention is migrated to the target task. DETAILED DESCRIPTION

[0043] The present invention is further described below in conjunction with the accompanying drawings and embodiments, but they are not intended to limit the present invention.

[0044] Embodiment 1: A gearbox fault diagnosis method based on domain adaptation residual network and multi-channel fusion, as shown in the attached Figure 1 As shown, follow the steps below:

[0045] Step S1: Acceleration sensors installed at different positions of the gearbox are used to collect acceleration signals of gears and bearings under various working conditions and fault conditions, and the input voltage signal and output voltage signal of the gearbox are used as sample data; the faulty bearing and gear are installed in the gearbox, and the acceleration sensor is set at the fixed plate of the output shaft of the gearbox and the fixed plate position of the output shaft. The connected tachometer generates 10 pulses per revolution to provide accurate zero-crossing information. The sampling frequency is 66.7KHz. Three sets of data are collected at different speeds and loads for 8 seconds. The collected data include 5 different shaft speeds, namely 30 Hz, 35Hz, 40Hz, 45Hz and 50Hz, and including high load and low load; the fault state includes any combination of one or more of gear longitudinal peeling fault, electrical erosion fault, broken tooth fault, bearing rolling element fault, inner ring fault, outer ring fault, input shaft imbalance fault, and output shaft key shearing fault. In order to simulate the real operating environment of the gearbox and the state when a fault occurs, in this embodiment, eight composite fault types are set, and each fault type is tested under five different working conditions. The sample configuration of each working condition is shown in Table 1, and the sample configuration of the fault label is shown in Table 2.

[0046]

[0047] Table 1

[0048]

[0049]

[0050] Table 2

[0051] Step S2: Put the sample data into three channels respectively. Since the acceleration sensor collects non-stationary vibration signals, which contain a lot of noise that is irrelevant to the gear and bearing fault types, it will interfere with the final fault identification. Therefore, the signal is first decomposed by VMD to filter the noise, and then the decomposed signal is extracted through the convolution layer. In order to avoid overfitting, the feature map is discarded and pooled to retain important features. The mathematical expression of VMD decomposition is:

[0052]

[0053] In the formula, a k (t) is the instantaneous amplitude of the intrinsic mode function (IMF) obtained by VMD decomposition; is the instantaneous phase of the eigenmode function;

[0054] As attached Figure 2As shown in the figure, the decomposition process first inputs a signal, which can be a one-dimensional or multi-dimensional signal, then determines the parameters required for decomposition and uses random parameters for initialization. Different frequency modes in the signal are gradually extracted through an iterative method. At the same time, the center frequency and bandwidth of each modal component are continuously updated until convergence. Figure 3 As shown in the figure, the signal is decomposed into 6 IMFs, and the center frequency of each IMF does not show aliasing, which helps to better capture the detailed information in the signal and remove high-frequency noise during the signal noise reduction process.

[0055] Step S3: Divide the optimized sample data into data sets according to the fault categories, and divide them into training set, validation set and test set in a ratio of 8:1:1. The training set is used for iterative training of the model, the test set is used to evaluate the change of model accuracy during the training process, and the validation set is used to evaluate the generalization effect of the model. In order to reduce the amount of model training, the data is normalized using the maximum and minimum value methods. The mathematical expression is:

[0056]

[0057] In the formula: x' represents the value of a single data; min is the minimum value of the column where the data is located; max is the maximum value of the column where the data is located.

[0058] Step S4: Construct the residual network model training. The model performs multi-channel extraction, weight adjustment and feature fusion on the training set. The maximum mean difference and cross entropy classification loss are used as optimization targets. The softmax function is used to classify and output the predicted probability of each fault category. The model parameters are determined through the validation set and the test set. Figure 4 As shown, the model includes a multi-channel extraction module, a feature fusion module, an attention mechanism module, a feature extraction module and a fault identification module;

[0059] The expression for the maximum mean difference is:

[0060]

[0061] In the formula, n is the number of samples in the source domain, m is the number of samples in the target domain, is the i-th sample in the source domain, is the jth sample in the target domain;

[0062] The input of the multi-channel extraction module is the non-stationary vibration signal collected by the acceleration sensor, which contains a lot of noise that is irrelevant to the gear and bearing fault types, which will interfere with the final fault identification. Therefore, the signal is first decomposed to filter the noise, and then the decomposed signal is subjected to feature extraction through the convolution layer. In order to avoid overfitting, the feature map is discarded and pooled to retain important features;

[0063] The feature fusion module fuses the extracted low-level features and dynamically adjusts the weight of each channel in the feature map through the ECA attention mechanism. Compared with weighted multi-channel fusion, this method reduces the computational cost and improves the efficiency and performance of the model. The feature extraction process is performed through the convolution layer, and the convolution kernel size of the convolution layer is changed through an adaptive function, which is defined as follows:

[0064]

[0065] Where K is the coverage of cross-channel interaction, c is the channel dimension, γ = 2, b = 1

[0066] The feature extraction module uses an improved residual network to avoid the problem of gradient vanishing. Through cross-layer connections, information can flow through the model more easily, making it easier to learn subtle changes and feature information in the data. The mathematical expression of cross-layer connections is as follows:

[0067] H(x)=F(x)+x;

[0068] Where: H(x) is the unknown mapping; F(x) is the residual mapping; x is the identity mapping.

[0069] The fault recognition model transforms the high-level feature map into a one-dimensional shape through a flattening layer, and then inputs it into the fully connected layer, and realizes classification through the softmax function. At the same time, the MMD loss and the cross entropy loss are combined as the total loss function. During the back propagation of the residual network, the gradients of the MMD loss and the source task classification loss are calculated respectively, and the gradient descent algorithm is used to minimize the MMD loss, thereby reducing the data distribution difference and realizing parameter update. The mathematical expression is as follows:

[0070] L 总 =L 分类 +λL mmd ;

[0071] Among them, L 总 Represents the total loss, λ is a hyperparameter that adjusts the proportion of MMD loss;

[0072] Step S5: Fine-tune the model parameters to perform gearbox fault classification prediction under multiple working conditions; randomly initialize the parameters and input sample data into the model, extract the gearbox fault features layer by layer through forward propagation, obtain the fault prediction result, calculate the error between the predicted value and the true value and the sample distribution difference between the source domain and the target domain, update the model parameters according to the error gradient during back propagation, and output the diagnostic accuracy of the source domain pre-trained model if the convergence condition is met. When performing source domain migration for cross-working condition experiments, it is only necessary to freeze the feature extraction part of the pre-trained model in the target task and fine-tune the last few layers and hyperparameters of the model.

[0073] In order to visually observe the classification effect after the source task is migrated to the target task, this embodiment uses the T-SNE algorithm to project the data in the original high-dimensional space into a two-dimensional plane. Figure 5 (a) shows the visualization result of the test set data of target task B after dimensionality reduction. It can be seen that the eight compound faults in the original gearbox overlap on the two-dimensional plane and are difficult to classify. Figure 5 (b) shows the classification of the acceleration sensor after the convolutional layer feature extraction and fusion. It is also difficult to clearly distinguish the various fault types. Figure 5 (c) shows the classification after the improved model processing. It can be clearly seen that most of the composite fault types have been separated, and only a few points representing fault types 5 and 7 overlap. Table 2 shows that these two composite fault types are very similar and are difficult to classify correctly when the speed is similar; Appendix Figure 5 (d) shows the final classification result. Fault types 5 and 7 are projected to similar positions in the two-dimensional space, while other compound fault types have been well clustered, indicating that the pre-trained model successfully completes the fault classification task after migration.

[0074] Step S6: Use the trained model to perform gearbox fault diagnosis.

[0075] Furthermore, the method proposed in the present invention was tested for comparison with several other methods, all of which used 3 channel data and the same number of samples to perform small sample migration under different working conditions. The results are shown in Table 3.

[0076] 30hz → 35hz 30hz → 40hz 30hz→45hz 30hz→50hz Balanced distribution adaptation 33.12％ 26.71％ 24.50％ 22.85％ Migration component analysis 23.15％ 35.00％ 22.00％ 27.25％ Deep Convolutional Transfer Learning Network 57.20％ 66.52％ 61.00％ 62.30％ Weighted multi-channel data fusion 93.50％ 95.83％ 96.25％ 97.50％ The model proposed in this paper 96.88％ 100％ 99.38％ 98.75％

[0077] Table 3

[0078] The diagnostic accuracy of algorithms such as balanced distribution adaptation, migration component analysis, and deep convolutional transfer learning network are all below 70%, indicating that the model failed to reduce the data distribution difference between the source domain and the target domain during transfer learning, and therefore could not extract common deep fault features, resulting in a very low final classification accuracy. In contrast, the weighted multi-channel data fusion algorithm calculates the fusion weights of each channel by information entropy, effectively reducing the influence of irrelevant data, thereby greatly improving the diagnostic accuracy. The classification accuracy of the diagnostic method proposed in this paper under four different migration tasks is improved by 3.38%, 4.17%, 3.13% and 1.25% respectively compared with the weighted multi-channel data fusion algorithm, indicating that the proposed model has stronger information mining capabilities.

[0079] In summary, the present invention obtains higher diagnostic accuracy by simultaneously processing multi-channel data from multi-source sensors, and relies on multi-channel fusion data to obtain common features to perform gearbox fault classification prediction under multiple working conditions. It has strong generalization ability and reduces the demand for the number of samples.

Claims

1. A gearbox fault diagnosis method based on domain adaptation residual network and multi-channel fusion, characterized by: Follow these steps: Step S1: Acceleration sensors installed at different positions of the gearbox are used to collect acceleration signals of gears and bearings under various working conditions and fault conditions, and the input voltage signal and output voltage signal of the gearbox are used as sample data; Step S2: performing VMD decomposition and parameter optimization on the acceleration signal, input voltage signal and output voltage signal in the sample data using three channels respectively to obtain optimized sample data; Step S3: Divide the optimized sample data into a training set, a test set and a validation set according to the fault status category, and perform normalization processing; Step S4: Construct the residual network model training. The model performs multi-channel extraction, weight adjustment and feature fusion on the training set. The maximum mean difference and cross entropy classification loss are used as optimization targets. The softmax function is used to classify and output the predicted probability of each fault category. The model parameters are determined through the validation set and the test set. Step S5: fine-tuning model parameters to perform feature migration to achieve gearbox fault classification prediction under multiple working conditions; Step S6: Use the trained model to perform gearbox fault diagnosis.

2. The gearbox fault diagnosis method based on domain adaptation residual network and multi-channel fusion according to claim 1 is characterized in that: The acceleration sensor in step S1 is disposed at the fixed plate of the output shaft of the gear box and the fixed plate position of the output shaft.

3. The gearbox fault diagnosis method based on domain adaptive residual network and multi-channel fusion according to claim 1 is characterized in that: The fault state of step S1 includes any combination of one or more of gear longitudinal peeling fault, electrical erosion fault, tooth breakage fault, bearing rolling element fault, inner ring fault, outer ring fault, input shaft imbalance fault, and output shaft key shearing fault.

4. The gearbox fault diagnosis method based on domain adaptive residual network and multi-channel fusion according to claim 1 is characterized in that: The VMD decomposition mathematical expression of step S2 is as follows: In the formula, a k (t) is the instantaneous amplitude of the intrinsic mode function obtained by VMD decomposition; is the instantaneous phase of the eigenmode function.

5. The gearbox fault diagnosis method based on domain adaptive residual network and multi-channel fusion according to claim 1 is characterized in that: The maximum mean difference expression of step S4 is: In the formula, n is the number of samples in the source domain, m is the number of samples in the target domain, is the i-th sample in the source domain, is the jth sample in the target domain.

6. The gearbox fault diagnosis method based on domain adaptive residual network and multi-channel fusion according to claim 1 is characterized in that: The model of step S4 includes a multi-channel extraction module, a feature fusion module, an attention mechanism module, a feature extraction module and a fault identification module; The multi-channel extraction module extracts features from the input data to obtain feature maps, and then performs discarding and pooling operations to retain important features; The feature fusion module fuses the features extracted by the multi-channel extraction module, and adjusts the weight of each channel in the feature map through the attention mechanism module; The feature extraction module avoids gradient vanishing through an improved residual network, allows information to flow quickly through the model through cross-layer connections, and improves the learning rate of subtle changes in data and feature information; The fault identification module transforms the high-level feature map into a one-dimensional shape through a flattening layer, and then inputs it into the fully connected layer, realizes classification through the softmax function, and combines the MMD loss and the cross entropy loss as the total loss function. During the back propagation of the residual network, the gradients of the MMD loss and the source task classification loss are calculated respectively, and the gradient descent algorithm is used to minimize the MMD loss to reduce the data distribution difference and realize parameter update.

7. The gearbox fault diagnosis method based on domain adaptive residual network and multi-channel fusion according to claim 6 is characterized in that: The feature extraction process of the feature fusion module is performed through a convolution layer, and the convolution kernel size of the convolution layer is changed through an adaptive function, which is defined as follows: Where K is the coverage of cross-channel interaction, c is the channel dimension, γ=2, and b=1.

8. The gearbox fault diagnosis method based on domain adaptive residual network and multi-channel fusion according to claim 6 is characterized in that: The mathematical expression of the cross-layer connection of the feature extraction module is as follows: H(x)=F(x)+x; Where H(x) is the unknown mapping; F(x) is the residual mapping; x is the identity mapping.

9. The gearbox fault diagnosis method based on domain adaptive residual network and multi-channel fusion according to claim 6 is characterized in that: The gradient descent algorithm expression of the fault identification module is as follows: THE 总 =L 分类 +λL mmd ; Where, L 总 Represents the total loss, and λ is a hyperparameter that adjusts the proportion of MMD loss.

10. The gearbox fault diagnosis method based on domain adaptive residual network and multi-channel fusion according to claim 1, characterized in that: The model training strategy in step S4 adopts the Adam optimizer to accelerate the convergence speed, and uses the particle swarm optimization algorithm to optimize the hyperparameters of the model. The classification effect of the model is evaluated using the accuracy index, and the model training is performed with the total loss function as the optimization criterion.

Images