A current-guided vibration feature enhanced gearbox fault diagnosis method
By employing wavelet packet decomposition and current-guided fusion strategies, the challenge of fusing vibration and current signals in wind turbine gearbox fault diagnosis was solved, achieving efficient fault feature extraction and improved diagnostic accuracy, thereby enhancing the accuracy and reliability of wind turbine gearbox fault diagnosis.
Patent Information
- Application Number
- CN202310384885.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-04-12
- Publication Date
- 2025-12-19
- Estimated Expiration
- 2043-04-12
AI Technical Summary
Existing technologies have low diagnostic accuracy when fusing vibration and current signals, and traditional methods cannot effectively utilize the characteristic information of the two signals, resulting in insufficient accuracy and reliability in wind turbine gearbox fault diagnosis.
By generating a time-frequency feature matrix through wavelet packet decomposition, a current-guided fusion module and a frequency band pooling module are designed. The vibration signal features are reweighted using the attention of the current signal, and adaptive weighting is performed at the decision layer to optimize the model training process and achieve efficient fusion of vibration and current signals.
It improves the performance and robustness of wind turbine gearbox fault diagnosis, effectively extracts and fuses fault diagnosis information from different sensors, and enhances diagnostic accuracy and reliability.
Smart Images
Figure CN116383764B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of wind power gear box fault diagnosis, and in particular to a gear box fault diagnosis method with current-guided vibration feature enhancement. BACKGROUND
[0002] In recent years, with the deepening of the concept of low-carbon environmental protection, the new energy field has developed rapidly, and wind power is an important part of new energy. In the operation of wind power generation, the operation and maintenance cost accounts for about 20%-25% of the balanced energy cost of wind power generation assets. In order to reduce the operation and maintenance cost and make wind power more competitive in the electricity market, an effective way is to implement state-based maintenance. In order to achieve effective maintenance of the state, a condition monitoring system (CMS) is needed to monitor the health status of the wind turbine online. Among various sub-component failures and high maintenance costs, gear box failure causes the longest downtime of the wind turbine. Therefore, many modern large-scale wind turbines are equipped with a CMS for the gear box, and most of them use vibration signals. One disadvantage of vibration monitoring is that the cost of vibration sensors and related data acquisition and processing equipment is high. Compared with vibration monitoring, current monitoring has the advantages of easy acquisition, high stability, etc., but the fault information is weak. In addition, there is a coupling relationship between vibration signals and current signals in fault diagnosis.
[0003] In order to effectively utilize the fault-related features extracted from multiple sensors, various information fusion technology-based methods have been developed. These methods include data-level fusion, feature-level fusion, and decision-level fusion. Feature-level fusion is the most widely used and efficient method in recent years. However, current technical methods encounter the following difficulties when fusing vibration and current signals: fault information in vibration signals is easier to extract, and the diagnostic accuracy is much higher than that of current signals. During model training, the convergence of the two types of signals is difficult to align. The traditional fusion method inevitably adds more interference information in the current signal at any level of fusion, which improves the diagnostic accuracy but cannot fully utilize the feature information of the two types of signals.
[0004] Therefore, in order to more efficiently utilize the fault-related features extracted from multiple sensors, there is an urgent need to design an improved fusion technology method to improve the accuracy and reliability of fault diagnosis. SUMMARY
[0005] The technical problem to be solved by the present application is to provide a gear box fault diagnosis method with current-guided vibration feature enhancement, which efficiently fuses the vibration signals and current signals extracted from sensors to effectively improve the performance and robustness of wind power gear box fault diagnosis.
[0006] To solve the above technical problems, the technical solution adopted by the present application is: a gear box fault diagnosis method with current-guided vibration feature enhancement, comprising the following steps:
[0007] Step 1: collect the original data of vibration and current signals in the wind turbine gearbox, and convert the original data into a plurality of time-frequency feature matrices through wavelet packet decomposition;
[0008] Step 2: design a current-guided fusion module and a frequency band pooling module for extracting important features in the frequency band of the current signal and applying attention to the frequency band features of the vibration signal;
[0009] Step 3: input the time-frequency feature matrices of the preprocessed vibration and current signals respectively, and obtain a feature space matrix through three times of feature extraction and feature fusion;
[0010] Step 4: pass the feature space matrix through a classification module and adaptively weight it at the decision layer to obtain a final diagnostic result, and finally complete the training process through the addition of the losses of two classification tasks to jointly optimize the model.
[0011] Further improvement of the technical scheme of the application is that step 1 includes the following steps:
[0012] Step 11: the obtained wind turbine gearbox original data is labeled according to the number of fault types;
[0013] Step 12: the original data is cut into equal segments with a length of M using a sliding window;
[0014] Step 13: use a "db1" wavelet to perform three-layer wavelet packet decomposition on the cut segments, each segment is cut into eight sub-bands with a length of M / 8, and a plurality of time-frequency feature matrices with a size of (8, M / 8) are obtained by concatenating in the frequency band dimension;
[0015] Step 14: the time-frequency feature matrix is standardized in the frequency band dimension using the maximum and minimum normalization method to obtain a data set for network training.
[0016] Further improvement of the technical scheme of the application is that step 2 includes the following steps:
[0017] Step 21: design a current-guided fusion module, taking the current feature matrix, the fusion feature matrix and the vibration feature matrix as input, and taking the enhanced fusion feature matrix as output;
[0018] Step 22: design a frequency band pooling module, use a mean pooling layer to compress the input data in the channel dimension and the time dimension respectively, and then pass it through a Sigmoid activation function to obtain a frequency band distribution feature representing the input data, the calculation formula is as follows:
[0019]
[0020] wherein is the output of the frequency band pooling module, Y b，t，c is the input information of the frequency band pooling module, T is the number of time dimensions of the input information, C is the number of channel dimensions of the input information, and sigma(.) is a Sigmoid activation function;
[0021] Step 23: input the current feature matrix into the frequency band pooling module to obtain a frequency band distribution feature representing the current, multiply the fusion feature matrix by the frequency band distribution feature, reweight the fusion feature matrix in the frequency band dimension using attention, and obtain a fusion feature matrix enhanced by the current;
[0022] Step 24: introduce a parameter gamma that can be trained as the number of iterations increases, multiply the fusion feature matrix by the parameter gamma, multiply the vibration feature matrix by 1-gamma, and add the two to obtain a fusion feature matrix enhanced by the vibration;
[0023] Step 25: add the corresponding values of the fusion feature matrix enhanced by the current and the fusion feature matrix enhanced by the vibration to obtain a final enhanced fusion feature matrix as the output of the fusion module.
[0024] Further improvement of the technical scheme of the application is that step 3 includes the following steps:
[0025] Step 31: use a two-dimensional convolution layer, a Relu activation function and a maximum pooling layer to extract features from the time-frequency feature matrices of the vibration and the current, respectively, to obtain vibration feature matrices and current feature matrices;
[0026] Step 32: use the vibration feature matrices and the current feature matrices obtained from step 1 as the input of the current-guided fusion module, and use the vibration feature matrices obtained by the first feature extraction as the fusion feature matrices of the current-guided fusion module;
[0027] Step 33: repeat steps 31 and 32 three times, i.e., perform three times of feature extraction and feature fusion, to obtain vibration, current and fusion feature space matrices.
[0028] Further improvement of the technical scheme of the application is that step 4 includes the following steps:
[0029] Step 41: input the vibration feature space matrix and the fusion feature space matrix into a classification layer containing two fully connected layers, a Relu activation function and a Softmax activation function, and adaptively weight the output in the decision layer to obtain the main task output of the classification module, which is also the output of the entire model, and the calculation formula is as follows:
[0030] y pred = ξ v y v + ξ f yf
[0031] [ξ v ξ f ]=σ soft (FC([F v F f ])
[0032] Where y pred It is the output of the main task, y v and y f These are the outputs of the vibration feature space matrix and the fused feature space matrix at the classification layer, respectively, F. v and F f These are the vibration feature space matrix and the fused feature space matrix outputs in the second fully connected layer, respectively. FC represents a fully connected layer with 2 output channels, and σ... soft It is the Softmax activation function;
[0033] Step 42: Input the current feature space matrix into a classification layer containing two fully connected layers, a ReLU activation function, and a Softmax activation function to obtain the auxiliary task output of the classification module. The training is also optimized using the cross-entropy loss function, the formula for which is as follows:
[0034]
[0035] Where M is the number of fault types, y i p is a symbolic function i It is the model's prediction score for the i-th type of fault;
[0036] The total loss function is expressed as follows:
[0037] Loss total =Loss main +Loss aid
[0038] Loss total It is the total loss, Loss main It's the main task loss, Los. aid The auxiliary task loss, the main task loss, and the auxiliary task loss are added together to optimize the entire training process.
[0039] The technical progress achieved by this invention is as follows: by using wavelet packet decomposition to preprocess the vibration and current signals respectively, and constructing a time-frequency feature matrix, the vibration and current signals can be aligned in the frequency band dimension, highlighting the complementarity between the two signals.
[0040] By proposing a fusion strategy of current guiding vibration, the attention map obtained by using current signal itself is used to emphasize the necessary area in the vibration signal, that is, attention reweighting is used in the frequency band dimension of the vibration signal, which solves the problem of strong heterogeneity of different signals to a certain extent while effectively extracting and fusing fault features;
[0041] By using the adaptive weighting strategy in the classification stage of the model, the outputs of different branches are fused again in the decision layer to obtain the output of the main task of the model, and the robustness of the model is increased.
[0042] Therefore, the present application can extract more efficient fusion of fault diagnosis information in different sensors, can enhance the performance and precision of fault diagnosis, and provides a new way for multi-sensor fusion of wind turbine gearboxes. BRIEF DESCRIPTION OF DRAWINGS
[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings;
[0044] Figure 1 is a flow chart of the method of the present application;
[0045] Figure 2 is a flow chart of the vibration signal and current signal fusion network;
[0046] Figure 3 is a flow chart of the current-guided fusion module;
[0047] Figure 4 is a flow chart of the frequency band pooling module; DETAILED DESCRIPTION
[0048] The present application will be further described in detail below in combination with the embodiments:
[0049] Embodiment one
[0050] A gearbox fault simulation test bench is adopted, which is composed of a 3.7kW motor, a two-stage parallel gearbox with a speed increasing ratio of 20, a 3kW permanent magnet synchronous motor and a 3kW load box. The vibration data and current data are collected under the conditions that the motor speed is 400r / min and 600r / min and no load. Each working condition contains eight state types including the healthy state.
[0051] A current-guided vibration feature enhancement gearbox fault diagnosis method, the flow is as shown in Figure 1 ,
[0052] Step 1: Collect the original data of vibration and current signals in the gearbox, and convert the original data into multiple time-frequency feature matrices through wavelet packet decomposition, including the following specific steps:
[0053] Step 11: The obtained gearbox original data is labeled according to the number of fault types;
[0054] Step 12: The original data is divided into equal segments with a length of 2048 using a sliding window;
[0055] Step 13: The segmented segments are decomposed into eight sub-bands with a length of 256 using a "db1" wavelet, and multiple time-frequency feature matrices with a size of (8, 256) are obtained by concatenating in the frequency band dimension;
[0056] Step 14: The time-frequency feature matrix is standardized in the frequency band dimension using the maximum and minimum normalization method to obtain a data set for network training;
[0057] Step 2: As shown in Figure 3 , a current-guided fusion module and a frequency band pooling module are designed to extract important features in the frequency band of the current signal and apply attention to the frequency band features of the vibration signal. The designed current-guided fusion module includes:
[0058] Step 21: The designed current-guided fusion module takes the current feature matrix, the fusion feature matrix, and the vibration feature matrix as input, and outputs an enhanced fusion feature matrix.
[0059] Step 22: As shown in Figure 4 , the designed frequency band pooling module uses a mean pooling layer to compress the input data in the channel dimension and the time dimension, respectively, and then obtains the frequency band distribution features representing the input data through a Sigmoid activation function. The calculation formula is as follows:
[0060]
[0061] Where is the output of the frequency band pooling module, Y b，t，c is the input information of the frequency band pooling module, T is the value of the time dimension of the input information, C is the value of the channel dimension of the input information, and σ(.) represents the Sigmoid activation function.
[0062] Step 23: The current feature matrix is input into the frequency band pooling module to obtain the frequency band distribution features representing the current, and the fusion feature matrix is multiplied by the current-enhanced fusion feature matrix to reweight the fusion feature matrix in the frequency band dimension;
[0063] Step 24: Introducing the parameter γ which can be trained with the increase of iteration number multiplied with the fusion feature matrix, meanwhile multiplying the vibration feature matrix with 1-γ, adding the above two to get the fusion feature matrix enhanced by vibration;
[0064] Step 25: Adding the corresponding values of the fusion feature matrix enhanced by current and the fusion feature matrix enhanced by vibration to get the final enhanced fusion feature matrix as the output of the fusion module;
[0065] Step 3: Taking the time-frequency feature matrix of the pre-processed vibration and current signal as input respectively, and obtaining the feature space matrix through three times of feature extraction and feature fusion, including the following specific steps:
[0066] Step 31: Using a two-dimensional convolution layer, a Relu activation function and a maximum pooling layer to extract the features of the time-frequency feature matrix of vibration and current respectively to obtain the vibration feature matrix and the current feature matrix;
[0067] Step 32: Taking the vibration feature matrix and the current feature matrix obtained in S1 as the input of the current-guided fusion module, and taking the vibration feature matrix extracted for the first time as the fusion feature matrix of the first current-guided fusion module.
[0068] Step 33: Repeating S31 and S32 three times, i.e. three times of feature extraction and feature fusion, to obtain the vibration, current and fusion feature space matrix;
[0069] Step 4: Passing the feature space matrix through the classification module and obtaining the final diagnostic result through adaptive weighting in the decision layer, and finally completing the training process through the addition of the loss of two classification tasks to jointly optimize the model, including the following specific steps:
[0070] Step 41: Inputting the vibration feature space matrix and the fusion feature space matrix into the classification layer containing two fully connected layers, a Relu activation function and a Softmax activation function respectively, and adaptively weighting the output in the decision layer to obtain the output of the main task of the classification module, which is also the output of the whole model, and the calculation formula is as follows:
[0071] y pred =ξ v y v +ξ f y f
[0072] [ξ v , ξ f ]=σ soft (FC([F v , F f ])
[0073] where ypred is the output of the main task, y v and y f are the outputs of the vibration feature space matrix and the fusion feature space matrix at the classification layer, respectively, F v and F f are the outputs of the vibration feature space matrix and the fusion feature space matrix at the second fully connected layer, respectively, FC represents a fully connected layer with an output channel of 2, and σ soft is a Softmax activation function.
[0074] Step 42: input the current feature space matrix into a classification layer containing two fully connected layers, a Relu activation function, and a Softmax activation function, to obtain the auxiliary task output of the classification module. The cross-entropy loss function is also used for optimization training, and the calculation formula of the cross-entropy loss function is as follows:
[0075]
[0076] where M is the number of fault types, y i is a sign function, p i is the predicted score of the model for the i-th fault type;
[0077] The total loss function expression is as follows:
[0078] Loss total = Loss main + Loss aid
[0079] where Loss total is the total loss, Loss main is the main task loss, and Loss aid is the auxiliary task loss. The latter two are added together to optimize the entire training process.
[0080] The following comparison methods are set in this embodiment:
[0081] Vibration network: the original vibration signal is decomposed by wavelet packet, and then extracted three times with the same parameters as the proposed method, and the results are output by the classification module.
[0082] Current network: the original current signal is decomposed by wavelet packet, and then extracted three times with the same parameters as the proposed method, and the results are output by the classification module.
[0083] Fusion network: the original vibration and current signals are decomposed by wavelet packet, and then extracted three times with the same parameters as the proposed method, and the results are output by the classification module.
[0084] The accuracy of the results of the method of the present application and the control method is shown in Table 1, and all the results are the average of 10 random training times. The method of the present application has good performance under two working conditions, which shows that the current-guided vibration feature enhanced gearbox fault diagnosis method can effectively utilize the fault-related features extracted from the vibration and current signals, and can better diagnose and predict the gearbox fault.
[0085] Table 1 Comparison of test results of the method of the present application and the test results of the control method
[0086]
[0087] The above-described embodiments are merely preferred embodiments of the present application and are not intended to limit the scope of the present application. Various modifications and improvements to the technical solutions of the present application made by those of ordinary skill in the art without departing from the design spirit of the present application shall fall within the scope of protection of the present application as defined by the claims.
Claims
1. A method for diagnosing gearbox faults by enhancing vibration characteristics guided by current, characterized in that, The steps include the following: Step 1: Collect raw data of vibration and current signals in the wind turbine gearbox, and transform the raw data into multiple time-frequency feature matrices through wavelet packet decomposition; Step 2: Design a current-guided fusion module and a band pooling module to extract important features in the current signal band and apply attention to the vibration signal band features; Step 2 includes the following specific steps: Step 21: Design a current-guided fusion module, using the current feature matrix, fusion feature matrix, and vibration feature matrix as inputs, and the enhanced fusion feature matrix as output; Step 22: Design a frequency band pooling module to compress the input data using mean pooling layers in both the channel and time dimensions, and then obtain the frequency band distribution characteristics representing the input data through the Sigmoid activation function. The calculation formula is as follows: ； in It is the output of the band pooling module. This refers to the input information for the band pooling module, where T is the time dimension value of the input information and C is the channel dimension value of the input information. Use the Sigmoid activation function; Step 23: Input the current feature matrix into the frequency band pooling module to obtain the frequency band distribution features representing the current. Multiply the fusion feature matrix with the fusion feature matrix and reweight the fusion feature matrix using attention in the frequency band dimension to obtain the fusion feature matrix using current enhancement. Step 24: Introduce training parameters that can be increased with the number of iterations. Multiply with the fused feature matrix, and simultaneously combine the vibration feature matrix with... Multiply the two and add them together to obtain the fused feature matrix using vibration enhancement; Step 25: Add the corresponding values of the current-enhanced fusion feature matrix and the vibration-enhanced fusion feature matrix to obtain the final enhanced fusion feature matrix as the output of the fusion module; Step 3: Using the time-frequency feature matrices of the preprocessed vibration and current signals as inputs, the feature space matrix is obtained through three feature extraction and feature fusion processes; Step 4: The feature space matrix is processed by the classification module and adaptively weighted at the decision layer to obtain the final diagnostic result. Finally, the model is optimized by adding the losses of the two classification tasks to complete the training process.
2. The gearbox fault diagnosis method with enhanced vibration characteristics guided by current as described in claim 1, characterized in that, Step 1 includes the following specific steps: Step 11: Mark the obtained raw data of the wind turbine gearbox according to the number of fault types; Step 12: Use a sliding window to divide the original data into equal segments of length M; Step 13: Use the "db1" wavelet to perform three-level wavelet packet decomposition on the segmented segments. Each segment is divided into eight sub-bands with a length of M / 8. The sub-bands are then concatenated to obtain multiple time-frequency feature matrices of size (8, M / 8). Step 14: Standardize the time-frequency feature matrix in the frequency band dimension using the max-min normalization method to obtain a dataset that can be used for network training.
3. The gearbox fault diagnosis method with enhanced vibration characteristics guided by current according to claim 2, characterized in that, Step 3 includes the following specific steps: Step 31: Extract the vibration and current time-frequency feature matrices using a two-dimensional convolutional layer, a ReLU activation function, and a max pooling layer, respectively, to obtain the vibration feature matrix and the current feature matrix. Step 32: Use the vibration feature matrix and current feature matrix obtained from Step 1 as input to the current-guided fusion module, and use the vibration feature matrix extracted in the first feature extraction as the fusion feature matrix of the first input current-guided fusion module; Step 33: Repeat steps 31 and 32 three times, that is, perform feature extraction and feature fusion three times to obtain the vibration, current and fused feature space matrix.
4. The gearbox fault diagnosis method with enhanced vibration characteristics guided by current according to claim 3, characterized in that, Step 4 includes the following specific steps: Step 41: Input the vibration feature space matrix and the fused feature space matrix into a classification layer containing two fully connected layers, a ReLU activation function, and a Softmax activation function, respectively. Adaptively weight the outputs in the decision layer to obtain the main task output of the classification module, which also serves as the output of the entire model. The calculation formula is as follows: ； ； in It is the output of the main task. and These are the outputs of the vibration feature space matrix and the fused feature space matrix at the classification layer, respectively. and These are the vibration feature space matrix and the fused feature space matrix outputs of the second fully connected layer, respectively. This indicates a fully connected layer with 2 output channels. It is the Softmax activation function; Step 42: Input the current feature space matrix into a classification layer containing two fully connected layers, a ReLU activation function, and a Softmax activation function to obtain the auxiliary task output of the classification module. The training is also optimized using the cross-entropy loss function, the formula for which is as follows: ； Where M represents the number of fault types. For symbolic functions, The model is for the first Predicted scores for different fault types; The total loss function is expressed as follows: ； in It is the total loss. It's a loss in the main task. The auxiliary task loss, the main task loss, and the auxiliary task loss are added together to optimize the entire training process.