Offshore wind turbine generator gearbox cross-domain fault diagnosis method based on wavelet transform and multi-scale residual network

Abstract

An offshore wind turbine gearbox cross-domain fault diagnosis method based on wavelet transform and a multi-scale residual network comprises the steps that collected data of different fault types serve as input, then wavelet transform is used for processing, and signals obtained after noise reduction are obtained; designing a multi-scale residual network, and inputting the denoised signal into the multi-scale residual network for feature extraction; in a full connection layer of the network, using a maximum mean difference (MMD) to reduce the difference between a source domain and a target domain; inputting the training sets of the source domain and the target domain into a multi-scale residual network to obtain a gearbox fault diagnosis model; and S5, inputting the target domain test set into the model in the S5 for fault classification, thereby identifying different fault types of the gearbox. The method comprises the following steps: carrying out data processing on an original vibration signal by using wavelet transform to obtain a denoised signal, constructing a multi-scale residual network to carry out feature extraction on the denoised signal, and classifying the extracted features through a Softmax function so as to realize identification of a fault type.

Description

Technical Field

[0001] The present invention belongs to the technical field of offshore wind power, and relates to a cross-domain fault diagnosis method for an offshore wind turbine gearbox based on wavelet transform and multi-scale residual network. Background Art

[0002] Offshore wind power development is trending from nearshore to offshore, and its momentum is expected to increase further. However, due to the complex and changing offshore environment and the constant influence of wind speed on wind turbine gearboxes, the vibration signals collected from them exhibit non-stationary characteristics. Therefore, fault diagnosis research on offshore wind turbine gearboxes is of great significance to ensure the safe operation of wind turbines.

[0003] Currently, wind turbine gearbox fault diagnosis methods can be divided into two categories: physical model-based and data-driven. Physical model-based methods require a familiarity with the machine's internal workings to establish a fault diagnosis model. These methods, however, require expert knowledge and manual feature extraction, making them inadequate for the demands of the "big data era."

[0004] With the development of artificial intelligence technology, data-driven fault diagnosis methods have become a new research hotspot, especially fault diagnosis methods based on deep learning. Deep learning can automatically extract features from input data without the need for manual processing. For example: Wu Shengli et al. proposed a gearbox fault diagnosis method based on symmetrical point patterns and multi-scale convolutional neural networks. This method improves the expressiveness of fault features through symmetrical point processing, and then extracts fault features through multi-scale convolutional networks, and finally realizes the classification of fault features. Zhou Dan et al. proposed a wind turbine gearbox fault diagnosis method based on continuous wavelet transform and Swin-Transformer. This method converts the vibration signal into a time-frequency graph through wavelet transform, and then uses Swin-Transformer for feature extraction. Finally, the effectiveness of this method is verified through experiments.

[0005] While the aforementioned methods achieve good fault diagnosis accuracy in their respective fault diagnosis tasks, they perform fault diagnosis under the same operating conditions. In real-world scenarios, gearboxes often operate under varying operating conditions. Therefore, improving wind farm gearbox fault diagnosis accuracy under these varying operating conditions remains a challenge. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a cross-domain fault diagnosis method for offshore wind turbine gearboxes based on wavelet transform and multi-scale residual network. The wavelet transform is used to process the original vibration signal to obtain the noise-reduced signal, and a multi-scale residual network is constructed to extract features of the noise-reduced signal. The extracted features are classified by the Softmax function, thereby realizing the identification of the fault type.

[0007] To solve the above technical problems, the technical solution adopted by the present invention is: a cross-domain fault diagnosis method for offshore wind turbine gearbox based on wavelet transform and multi-scale residual network, comprising the following steps: S1: Install the acceleration sensor in the horizontal direction of the gearbox to collect horizontal acceleration data of the gearbox, including health status, half missing of the big end of the gear, half missing of the small end of the gear, complete missing of the gear end, and gear wear; S2: Take the collected data of different fault types as input, and then process them using wavelet transform to obtain the noise-reduced signal; S3: Design a multi-scale residual network and input the denoised signal into the multi-scale residual network for feature extraction; S4: In the fully connected layer of the network, the maximum mean difference (MMD) is used to reduce the difference between the source domain and the target domain; S5: Input the training sets of the source domain and the target domain into the multi-scale residual network to obtain the gearbox fault diagnosis model; S6: Input the target domain test set into the model in S5 for fault classification, thereby identifying different fault types of the gearbox.

[0008] First, define the discrete wavelet function: ;(1) Among them, is the mother wavelet function, For scale size, is the location of the mother wavelet function. Then, the discrete wavelet transform can be expressed as: ;(2) in, is the original input signal, is the signal length; finally, perform inverse wavelet transform to obtain the denoised signal: ;(3) in, is the signal after noise reduction.

[0009] In S2, the denoised signal is obtained by inverse discrete wavelet transform to remove the noise interference.

[0010] In S3, the multi-scale residual network consists of a multi-scale residual module and a non-local feature extraction module; the multi-scale residual module first uses convolution kernels of sizes 3, 5, and 7 for feature extraction, then uses the channel attention mechanism to weight the extracted multi-scale features, and finally uses residual connections to connect the input and output together; the non-local feature extraction module can capture the long-range dependencies of features, thereby highlighting the characteristics of periodic faults.

[0011] In S3, the process of feature extraction using convolution kernels of size 3, 5, and 7 on the input can be described as: ;(4) ;(5) ;(6) in, For the Layer The feature map eigenvalues, is the convolution weight, is the bias, is the current layer, is the total number of eigenvalues, is the total number of feature maps, is a convolution kernel of scale 3, is a convolution kernel with a scale of 5, The convolution kernel size is 7.

[0012] In S3, the process of weighting the extracted multi-scale features by the channel attention mechanism can be described as: ;(7) ;(8) ;(9) Then, the multi-scale features are fused, and the process can be described as: ;(10) The process of connecting the input and output together using residual connections can be described as: ;(11) The non-local feature extraction module is able to capture the long-range dependencies of features, thereby highlighting periodic fault characteristics.

[0013] In S3, for the non-local feature extraction module, the most important thing is the selection of the similarity scale function; the Softmax function is selected as the scale function; for the input feature map , respectively, through three 1×1 convolutions for feature extraction; then, the extracted features are reconstructed in the spatial dimension and weighted by Softmax; finally, the input and output are connected using residual connections; the entire non-local feature extraction uses a small convolution kernel, which helps to reduce network parameters; the formula for the non-local feature extraction module is described as: ;(12) in, is the output feature map of the non-local operation, represents the input feature map, Represents the result of non-local feature operation; Obtained by the following formula: ; (13) ;(14) in, is the normalized exponential function, is the global information of the input feature map, and is the result of the convolution operation.

[0014] In S4, the maximum mean difference (MMD) is used in the fully connected layer of the network to reduce the difference between the source domain and the target domain: ;(15) in, is the mathematical expectation, To regenerate the Hilber space mapping, is the feature kernel, is the source domain, The target domain.

[0015] In S5, the experimental data comes from a roller gear fault simulation test platform, which consists of a three-phase AC asynchronous motor, a loading device, a transmission mechanism, a gearbox, and a frequency converter. The measured object is a bevel gear, and the collected fault states of the bevel gear are healthy state, half missing from the big end of the gear, half missing from the small end of the gear, complete missing from the gear end, and gear wear.

[0016] In S6, six Gaussian white noises with different signal-to-noise ratios are first added to the original signal to simulate environmental noise; the noise signal is only added to the source domain, and the diagnostic performance under different working conditions is evaluated by using the target domain.

[0017] The main beneficial effects of the present invention are: The vibration data was processed using wavelet transform, which can effectively remove noise interference in the original vibration signal.

[0018] A multi-scale residual network was constructed, which effectively mined fault features of different scales in the vibration signal of the wind turbine gearbox through multi-scale technology, thereby maximizing the extraction of domain-invariant features and improving the accuracy of fault diagnosis.

[0019] A non-local feature extraction module is constructed, which can capture the long-range dependencies between different features and thus highlight the periodic fault characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The present invention will be further described below with reference to the accompanying drawings and examples.

[0021] Figure 1 This is the multi-scale residual module diagram of the present invention.

[0022] Figure 2 This is a diagram of the non-local feature extraction module of the present invention.

[0023] Figure 3 This is the overall diagnostic flow chart of the present invention.

[0024] Figure 4 This is the diagnosis confusion matrix diagram of the present invention. DETAILED DESCRIPTION

[0025] like Figures 1 to 4 In this paper, a cross-domain fault diagnosis method for offshore wind turbine gearbox based on wavelet transform and multi-scale residual network is proposed. Example 1, First, the original vibration signal is processed using a wavelet transform to obtain a de-noised signal. A multi-scale residual network is then constructed to extract features from the de-noised signal. Finally, the extracted features are classified using a Softmax function to identify the fault type.

[0026] S1: Install the acceleration sensor in the horizontal direction of the gearbox to collect horizontal acceleration data of the gearbox, including health status, half missing of the big end of the gear, half missing of the small end of the gear, complete missing of the gear end, and gear wear; S2: Take the collected data of different fault types as input, and then use wavelet transform to process it to obtain the noise-reduced signal; first define the discrete wavelet function: ;(1) in, is the mother wavelet function, For scale size, is the location of the mother wavelet function. Then, the discrete wavelet transform can be expressed as: ;(2) in, is the original input signal, is the signal length. Finally, perform inverse wavelet transform to obtain the denoised signal: ;(3) in, is the signal after noise reduction.

[0027] S3: Design a multi-scale residual network, and input the denoised signal into the multi-scale residual network for feature extraction; specifically, the multi-scale residual network consists of a multi-scale residual module and a non-local feature extraction module. The multi-scale residual module first uses convolution kernels of sizes 3, 5, and 7 for feature extraction, then uses the channel attention mechanism to weight the extracted multi-scale features, and finally uses the residual connection to connect the input and output together. The non-local feature extraction module can capture the long-range dependency of features, thereby highlighting the periodic fault characteristics. For the non-local feature extraction module, the most important thing is the selection of the similarity scaling function. The present invention selects the Softmax function as the scaling function. Specifically, for the input feature map , undergoes three 1×1 convolutions for feature extraction; then, the extracted features are reconstructed in the spatial dimension and weighted using Softmax; finally, a residual connection is used to connect the input and output. The entire non-local feature extraction uses a small convolution kernel, which helps reduce network parameters. The formula for the non-local feature extraction module is described as: ;(4) in, is the output feature map of the non-local operation, represents the input feature map, Represents the result of non-local feature operation. Obtained by the following formula: ;(5) ;(6) in, is the normalized exponential function, is the global information of the input feature map, and is the result of the convolution operation.

[0028] S4: In the fully connected layer of the network, the maximum mean difference (MMD) is used to reduce the difference between the source domain and the target domain: S5: Input the training sets of the source domain and the target domain into the multi-scale residual network to obtain the gearbox fault diagnosis model; S6: Input the target domain test set into the model in S5 for fault classification, thereby identifying different fault types of the gearbox.

[0029] Example 2, A cross-domain fault diagnosis method for offshore wind turbine gearbox based on wavelet transform and multi-scale residual network includes the following steps: S1: Install the acceleration sensor horizontally on the gearbox to collect horizontal acceleration data of the gearbox, including health status, half missing of the big end of the gear, half missing of the small end of the gear, complete missing of the gear end, and gear wear; S2: The collected data of different fault types are used as input, and then processed using wavelet transform to obtain the noise-reduced signal. The specific implementation is as follows. First, the discrete wavelet function is defined: ;(7) in, is the mother wavelet function, For scale size, is the location of the mother wavelet function. Then, the discrete wavelet transform can be expressed as: ;(8) in, is the original input signal, is the signal length. Finally, perform inverse wavelet transform to obtain the denoised signal: ;(9) in, The denoised signal is obtained by inverse discrete wavelet transform. Compared with the original gearbox vibration signal, the noise interference of this signal is effectively removed.

[0030] S3: Design a multi-scale residual network. The multi-scale residual network consists of a multi-scale residual module and a non-local feature extraction module. The multi-scale feature module is as follows: Figure 1 As shown in Figure 2, the multi-scale residual module first uses convolution kernels of sizes 3, 5, and 7 for feature extraction, then uses the channel attention mechanism to weight the extracted multi-scale features, and finally uses residual connections to connect the input and output. The process of feature extraction using convolution kernels of sizes 3, 5, and 7 on the input can be described as follows: ;(10) ;(11) ;(12) in, For the Layer The feature map eigenvalues, is the convolution weight, is the bias, is the current layer, is the total number of eigenvalues, is the total number of feature maps, is a convolution kernel of scale 3, is a convolution kernel with a scale of 5, The convolution kernel has a scale of 7. The process of weighting the extracted multi-scale features by the channel attention mechanism can be described as: ; (13) ;(14) ;(15) Then, the multi-scale features are fused, and the process can be described as: ; (16) The process of connecting the input and output together using residual connections can be described as: ; (17) The non-local feature extraction module can capture the long-range dependency of features, thereby highlighting the periodic fault characteristics. Figure 2 As shown. For the non-local feature extraction module, the most important thing is the selection of the similarity scale function. The present invention selects the Softmax function as the scale function. Specifically, for the input feature map , undergoes three 1×1 convolutions for feature extraction; then, the extracted features are reconstructed in the spatial dimension and weighted using Softmax; finally, a residual connection is used to connect the input and output. The entire non-local feature extraction uses a small convolution kernel, which helps reduce network parameters. The formula for the non-local feature extraction module is described as: ; (18) in, is the output feature map of the non-local operation, represents the input feature map, Represents the result of non-local feature operation. Obtained by the following formula: ; (19) ; (20) in, is the normalized exponential function, is the global information of the input feature map, and is the result of the convolution operation.

[0031] S4: Input the denoised signal into the multi-scale residual network for feature extraction. The overall diagnostic flow chart is as follows: Figure 3 As shown; S5: In the fully connected layer of the network, the maximum mean difference (MMD) is used to reduce the difference between the source domain and the target domain: ;(twenty one) in, is the mathematical expectation, To regenerate the Hilber space mapping, is the feature kernel, is the source domain, The target domain.

[0032] S6: Input the training data from the source and target domains into a multiscale residual network to obtain a gearbox fault diagnosis model. Specifically, the experimental data comes from a roller gear fault simulation test platform, which consists of a three-phase AC asynchronous motor, a loading device, a transmission mechanism, a gearbox, and a frequency converter. The test objects in this invention are bevel gears, and the collected fault states of the bevel gears include healthy state, half missing gear end, half missing gear end, complete gear end missing, and gear wear. The source domain data is 1HP data at 20Hz, and the target domain data is 1HP data at 30Hz. The detailed classification of the data set is shown in Table 1.

[0033] Table 1

[0034] S7: Input the target domain test set into the model in S5 for fault classification, thereby identifying different gearbox fault types. The experimental results are shown below: The present invention first adds six Gaussian white noises with different signal-to-noise ratios to the original signal to simulate environmental noise.

[0035] It should be noted that only noise signals are added to the source domain, and the diagnostic performance of the present invention under different working conditions is evaluated by using the target domain. The experimental results are shown in Table 2. As can be seen from Table 2, when the SNR value increases from -2dB to 8dB, the diagnostic accuracy of the proposed method gradually improves, and its diagnostic accuracy is higher than that of the other four comparison methods. In particular, when the signal-to-noise ratio is 8dB, the diagnostic accuracy of the proposed method is 90.17% and 91.00% respectively, while the diagnostic accuracy of the other comparison methods is less than 90%. This shows that the proposed method not only has good noise resistance, but also is capable of fault identification tasks under multiple loads.

[0036] In order to further show the classification of different types of faults, the confusion matrix is ​​used to display the cross-domain fault diagnosis results, such as Figure 4 As shown in the figure, the proposed method misclassifies the target domain gear small end half missing fault, but does not misclassify other fault types, which shows that the proposed method is suitable for cross-domain fault diagnosis tasks.

[0037] Table 2

[0038] In the above method, wavelet transform is used to process the vibration data. Wavelet transform can effectively remove noise interference in the original vibration signal.

[0039] A multi-scale residual network was constructed, which effectively mined fault features of different scales in the vibration signal of the wind turbine gearbox through multi-scale technology, thereby maximizing the extraction of domain-invariant features and improving the accuracy of fault diagnosis.

[0040] A non-local feature extraction module is constructed, which can capture the long-range dependencies between different features and thus highlight the periodic fault characteristics.

[0041] The above embodiments are merely preferred technical solutions of the present invention and should not be construed as limiting the present invention. The embodiments and features in the embodiments of this application may be arbitrarily combined with each other unless they conflict. The scope of protection of the present invention shall be the technical solutions described in the claims, including equivalent alternatives to the technical features of the technical solutions described in the claims. Equivalent alternatives and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A cross-domain fault diagnosis method for offshore wind turbine gearbox based on wavelet transform and multi-scale residual network, characterized by: The steps include: S1: Install the acceleration sensor in the horizontal direction of the gearbox to collect horizontal acceleration data of the gearbox, including health status, half missing of the big end of the gear, half missing of the small end of the gear, complete missing of the gear end, and gear wear; S2: Take the collected data of different fault types as input, and then process them using wavelet transform to obtain the noise-reduced signal; S3: Design a multi-scale residual network and input the denoised signal into the multi-scale residual network for feature extraction; S4: In the fully connected layer of the network, the maximum mean difference (MMD) is used to reduce the difference between the source domain and the target domain; S5: Input the training sets of the source domain and the target domain into the multi-scale residual network to obtain the gearbox fault diagnosis model; S6: Input the target domain test set into the model in S5 for fault classification, thereby identifying different fault types of the gearbox.

2. The cross-domain fault diagnosis method for offshore wind turbine gearboxes based on wavelet transform and multi-scale residual network according to claim 1 is characterized by: In S2, first define the discrete wavelet function: ；(1) in, is the mother wavelet function, For scale size, is the location of the mother wavelet function; then, the discrete wavelet transform can be expressed as: ；(2) in, is the original input signal, is the signal length; finally, perform inverse wavelet transform to obtain the denoised signal: ；(3) in, is the signal after noise reduction.

3. The cross-domain fault diagnosis method for offshore wind turbine gearboxes based on wavelet transform and multi-scale residual network according to claim 1 is characterized by: In S2, the denoised signal is obtained by inverse discrete wavelet transform to remove the noise interference.

4. The cross-domain fault diagnosis method for offshore wind turbine gearboxes based on wavelet transform and multi-scale residual network according to claim 1 is characterized by: In S3, the multi-scale residual network consists of a multi-scale residual module and a non-local feature extraction module; The multi-scale residual module first uses convolution kernels of sizes 3, 5, and 7 for feature extraction, then uses the channel attention mechanism to weight the extracted multi-scale features, and finally uses residual connections to connect the input and output together; the non-local feature extraction module can capture the long-range dependencies of features, thereby highlighting the characteristics of periodic faults.

5. The cross-domain fault diagnosis method for offshore wind turbine gearboxes based on wavelet transform and multi-scale residual network according to claim 4 is characterized by: In S3, the process of feature extraction using convolution kernels of size 3, 5, and 7 on the input can be described as: ；(4) ；(5) ；(6) in, For the Layer The feature map eigenvalues, is the convolution weight, is the bias, is the current layer, is the total number of eigenvalues, is the total number of feature maps, is a convolution kernel of scale 3, is a convolution kernel with a scale of 5, The convolution kernel size is 7.

6. The cross-domain fault diagnosis method for offshore wind turbine gearboxes based on wavelet transform and multi-scale residual network according to claim 4 is characterized by: In S3, the process of weighting the extracted multi-scale features by the channel attention mechanism can be described as: ；(7) ；(8) ； (9) Then, the multi-scale features are fused, and the process can be described as: ；(10) The process of connecting the input and output together using residual connections can be described as: ；(11) The non-local feature extraction module is able to capture the long-range dependencies of features, thereby highlighting periodic fault characteristics.

7. The cross-domain fault diagnosis method for offshore wind turbine gearboxes based on wavelet transform and multi-scale residual network according to claim 6 is characterized by: In S3, for the non-local feature extraction module, the most important thing is the selection of the similarity scaling function; Select the Softmax function as the scale function; for the input feature map , respectively, through three 1×1 convolutions for feature extraction; then, the extracted features are reconstructed in the spatial dimension and weighted by Softmax; finally, the input and output are connected using residual connections; the entire non-local feature extraction uses a small convolution kernel, which helps to reduce network parameters; the formula for the non-local feature extraction module is described as: ；(12) in, is the output feature map of the non-local operation, represents the input feature map, Represents the result of non-local feature operation; Obtained by the following formula: ；(13) ；(14) in, is the normalized exponential function, is the global information of the input feature map, and is the result of the convolution operation.

8. The cross-domain fault diagnosis method for offshore wind turbine gearboxes based on wavelet transform and multi-scale residual network according to claim 1 is characterized by: In S4, the maximum mean difference (MMD) is used in the fully connected layer of the network to reduce the difference between the source domain and the target domain: ； (15) in, is the mathematical expectation, To regenerate the Hilber space mapping, is the feature kernel, is the source domain, The target domain.

9. The cross-domain fault diagnosis method for offshore wind turbine gearboxes based on wavelet transform and multi-scale residual network according to claim 1 is characterized by: In S5, the experimental data comes from a roller gear fault simulation test platform, which consists of a three-phase AC asynchronous motor, a loading device, a transmission mechanism, a gearbox, and a frequency converter. The measured object is a bevel gear, and the collected fault states of the bevel gear are healthy state, half missing from the big end of the gear, half missing from the small end of the gear, complete missing from the gear end, and gear wear.

10. The cross-domain fault diagnosis method for offshore wind turbine gearbox based on wavelet transform and multi-scale residual network according to claim 1 is characterized in that: In S6, six Gaussian white noises with different signal-to-noise ratios are first added to the original signal to simulate environmental noise; the noise signal is only added to the source domain, and the diagnostic performance under different working conditions is evaluated by using the target domain.

Images