A wind power rotating disc bearing fault diagnosis method based on WDCNN-Transformer
By using the WDCNN-Transformer model, combined with time-frequency domain feature extraction and fusion, the problem of low fault identification accuracy of acoustic emission signals from wind turbine turntable bearings under low-speed and heavy-load conditions was solved, and efficient early fault diagnosis was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIVERSITY OF SCIENCE AND TECHNOLOGY
- Filing Date
- 2026-03-03
- Publication Date
- 2026-06-09
AI Technical Summary
Existing technologies face significant challenges in extracting and modeling acoustic emission signal features from wind turbine slewing bearings under low-speed, heavy-load conditions, leading to low fault identification accuracy, especially since early fault features are easily obscured by noise.
The WDCNN-Transformer model is used to achieve fault diagnosis of wind turbine turntable bearings by combining time and frequency domain information through preprocessing, fast Fourier transform, time and frequency domain feature extraction and weighted matrix fusion.
It improves the accuracy and robustness of wind turbine turntable bearing fault diagnosis, effectively identifies early faults, and enhances the distinguishability of fault characteristics.
Smart Images
Figure CN122171208A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind power equipment condition monitoring and intelligent fault diagnosis technology, specifically a wind turbine turntable bearing fault diagnosis method based on WDCNN-Transformer. Background Technology
[0002] Wind turbine slewing bearings are critical load-bearing components in the yaw and pitch systems of wind turbine generators, and their operating status directly affects the safety and reliability of the wind turbine generator. Affected by complex operating conditions such as low speed, heavy load, and alternating loads, slewing bearings are prone to early damage such as raceway wear, pitting, and crack propagation. If these damages are not detected in time, they may lead to bearing failure or even complete turbine shutdown, resulting in serious economic losses.
[0003] Currently, slewing bearing fault diagnosis primarily relies on vibration signal analysis. However, under low-speed, heavy-load conditions, the vibration signal amplitude is small, and the signal-to-noise ratio is low, making early fault characteristics easily masked by noise, thus limiting diagnostic effectiveness. In contrast, acoustic emission signals are highly sensitive to early damage such as the initiation and propagation of microcracks in materials, making them more suitable for bearing condition monitoring under low-speed, heavy-load conditions. However, acoustic emission signals exhibit strong non-stationarity and time-frequency coupling characteristics, making feature extraction and modeling challenging.
[0004] In recent years, with the development of deep learning technology, models such as convolutional neural networks and Transformers have been gradually introduced into the field of mechanical fault diagnosis to achieve end-to-end feature learning.
[0005] Existing deep learning methods often focus on modeling a single time domain or frequency domain, making it difficult to simultaneously consider both the local transient characteristics and global temporal correlations of acoustic emission signals.
[0006] Therefore, it is necessary to propose an intelligent fault diagnosis method for acoustic emission signals of wind turbine turntable bearings to improve the accuracy and robustness of early fault identification under low-speed and heavy-load conditions. Summary of the Invention
[0007] The purpose of this invention is to provide a fault diagnosis method for wind turbine turntable bearings based on WDCNN-Transformer, in order to solve the problem of low fault identification accuracy caused by high noise and blurred fault features in the current traditional acoustic emission signal method under low speed and heavy load conditions.
[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for fault diagnosis of wind turbine turntable bearings based on WDCNN-Transformer, comprising at least the following steps:
[0009] S100: Employs an acoustic emission sensor to collect the raw acoustic emission signal of the wind turbine slewing bearing, and analyzes the raw acoustic emission signal... Preprocessing is performed, which includes at least sequential denoising, zero-mean normalization, and segmented preprocessing to obtain preprocessed signal samples.
[0010] S200: Construct a WDCNN-Transformer fault diagnosis model, which includes a WDCNN block, a Transformer block, a fully connected layer, and an output layer. The output of the WDCNN block is linearly mapped and used as the input of the Transformer block.
[0011] S300: Perform a Fast Fourier Transform on the preprocessed signal samples obtained in step S100 to convert the signal from the time domain to the frequency domain. Input the time domain signal and the frequency domain signal into the WDCNN block to extract local time-series features and local frequency domain features, respectively. Then, input the extracted features into the Transformer block, which captures the long-range global dependency and context information of the features to obtain the time-frequency domain feature representations in the time domain and frequency domain, respectively.
[0012] S400: The time-domain and frequency-domain feature representations obtained in step S300 are fused by weighted matrix addition to obtain the fused time-frequency domain feature matrix. The feature matrix is then input into the fully connected layer and the output layer for model training, and the probability values of different fault types are output to realize the fault diagnosis of wind turbine turntable bearings.
[0013] Furthermore, the original acoustic emission signal in S100 Preprocessing includes at least the following steps:
[0014] S101: First, denoise the original acoustic emission signal, determine a suitable frequency range, and set the frequency range to [value missing]. arrive ;
[0015] The original signal is filtered using a bandpass filter to obtain the denoised signal:
[0016]
[0017] in, This is the lower cutoff frequency of the bandpass filter; This is the upper cutoff frequency of the bandpass filter; This is the denoised signal after bandpass filtering; Indicates a bandpass filtering operation;
[0018] S102: Denoising signal Perform zero-mean normalization and calculate the mean of the signal. and standard deviation :
[0019]
[0020]
[0021] Normalize the signal:
[0022]
[0023] in, The signal after normalization For denoised signals The total number of sampling points, For the first The signal values at each sampling point; Reflects the overall energy center level of the acoustic emission signal in this segment;
[0024] S103: Normalized signal Segmentation;
[0025] The signal is divided into multiple small segments, each segment having a length of [length missing]. And process it using a sliding window;
[0026] Assume the total length of the signal is Step size is Then the first A small segment of the signal can be represented as:
[0027]
[0028] in, It is the first The start time of a small segment; It is the end time.
[0029] Furthermore, the WDCNN block uses convolutional kernels of different widths, enabling it to extract local features of the signal at multiple scales.
[0030] The Transformer block uses a self-attention mechanism to establish connections between features at different locations in the WDCNN-Transformer fault diagnosis model, and its multilayer perceptron further mixes features. After multiple stages of processing, the features are subjected to global average pooling, which averages the feature values of each channel to obtain a global feature vector and feeds it into the fully connected layer.
[0031] The global feature vectors of the fully connected layer are mapped to the output space of the model.
[0032] The output layer uses the Softmax activation function to convert the output into a probability distribution, determine the fault type of the wind turbine turntable bearing, and realize intelligent fault diagnosis of the wind turbine turntable bearing.
[0033] Furthermore, S300 includes at least the following steps:
[0034] S301: Performs a Fast Fourier Transform (FFT) on the acoustic emission signal, converting the AE signal from the time domain to the frequency domain.
[0035]
[0036] in, Represents a frequency domain signal; It is the imaginary unit; It is the length of the signal; It is a time-domain signal; It is a frequency domain index; It is a complex exponential basis function;
[0037] To facilitate subsequent processing, the frequency domain characteristics are represented by calculating the amplitude spectrum or power spectrum;
[0038] S302: Extracting local temporal features using WDCNN blocks in the temporal branch:
[0039] The original time-domain sequence The input to the variable-width one-dimensional convolutional neural network (WDCNN) block in the temporal branch extracts local temporal features from the AE signal through convolutional kernels with different receptive fields.
[0040] For the Layered convolution, with kernel length denoted as . Step size is Then the first layer The convolution output of each output channel can be represented as:
[0041]
[0042] in, For the next level Input characteristics of each channel; These are the kernel weights; For bias terms; Total number of input channels in the previous layer; For the first Layer The first output channel The local feature pixels output by each WDCNN block form the basic unit of the local feature sequence. The index mapping of input features reflects the sliding sampling logic of the convolution stride on the input features, and determines the size reduction ratio of the feature map; For the first Layer The bias terms of each output channel are trainable parameters of WDCNN, used to adjust the baseline of the feature output and improve the model's fitting ability.
[0043] The result obtained using a nonlinear activation function is:
[0044]
[0045] in, For the activation function, this invention chooses ReLU; The convolution output of the m-th channel at the p-th position in the l-th layer;
[0046] To enhance robustness and reduce redundancy, this invention may introduce max pooling operations after several convolutional layers:
[0047]
[0048] in, For the first Layer The first channel in the The pooling output value of each pooling window; For the first Layer The feature vectors of each channel before pooling; The index for the pooled window; This is the pooling window size; This is a local index within the pooling window, with a value range of [value range missing]. ; This is a maximum value operation, which selects the maximum value from each pooling window as the output of that window;
[0049] After passing through multiple WDCNN blocks, the temporal local feature sequence representation is obtained. ;
[0050] S303: Extracting local spectral structure features in the frequency domain branch using WDCNN blocks:
[0051] The frequency domain signal obtained by Fast Fourier Transform The WDCNN block input to the frequency domain branch extracts features along the frequency axis using one-dimensional convolutions. The convolution, activation, and pooling methods are consistent with those in the time domain branch, ultimately yielding a local feature sequence representation in the frequency domain. The extracted features include at least local frequency band energy concentration, harmonic peak structure, and local spectral variations;
[0052] S304: Capturing long-range global dependencies and context information using the Transformer encoder:
[0053] To further model the global dependencies of the AE signal over long time scales (time domain) or cross-band scales (frequency domain), a Transformer encoder module is cascaded onto the output feature sequence of the WDCNN block. Let the output feature sequence of the WDCNN block be:
[0054]
[0055] in, The sequence length; For feature dimensions;
[0056] By adding positional encoding to the sequence, we obtain:
[0057]
[0058] in, For the feature sequence after adding positional encoding, the dimension is... same; For the position encoding matrix, the dimensions are... The same applies; it is used to assign a unique code to each position in the sequence in order to preserve the sequence's order information.
[0059] Subsequently, in the self-attention layer of the Transformer, the query is obtained through linear mapping. ,key Sum :
[0060]
[0061] in, The linear transformation weight matrix is for learning; ;
[0062] Self-attention is calculated as follows:
[0063]
[0064] in, The query is the dot product of the key and the dimension. , representing the correlation score between each pair of positions; This is a scaling factor used to prevent the softmax gradient from vanishing due to an excessively large dot product result. The normalization function converts the correlation scores into a probability distribution; For self-attention output, the dimension is ;
[0065] After concatenating the multi-head attention signals, a linear transformation is used to obtain the final output of the multi-head attention signal.
[0066]
[0067] in, For the first The output of each attention head; To output the linear transformation weight matrix; The concatenation operation concatenates the outputs of all attention heads along the feature dimension; The final output of multi-head attention, with dimensions of ;
[0068] Combined with residual connection With layer normalization :
[0069]
[0070] Feedforward networks (FFNs) further process the output features:
[0071]
[0072] in, The input feature matrix (from the output of the previous layer); This is the weight matrix for the first linear transformation; This is the bias vector for the first linear transformation; For activation functions; This is the weight matrix for the first linear transformation; This is the bias vector for the first linear transformation; This is the output of the feedforward network;
[0073] Final output:
[0074]
[0075] in, For residual connection, For layer normalization, This is the final output of the Transformer layer.
[0076] Furthermore, when using amplitude spectrum in S301:
[0077]
[0078] in, For the first Amplitude spectrum values of each frequency component; The result obtained after Fast Fourier Transform (FFT) is the 1st... Complex representation of each frequency component; For complex numbers The real part, for The imaginary part.
[0079] Furthermore, when power spectrum is used in S301:
[0080]
[0081] in, For the first Power spectral values of each frequency component The total length of the signal. This represents the energy of that frequency component.
[0082] Furthermore, the S400 includes at least the following steps:
[0083] By fully combining the advantages of time-domain and frequency-domain signals, a weighted matrix addition fusion method is used to form the final time-frequency domain feature representation by weighted fusion of time-domain and frequency-domain features.
[0084] like Figure 3 As shown: By fully combining the advantages of time-domain and frequency-domain signals, the weighted matrix addition fusion method is used to form the final time-frequency domain feature representation by weighted fusion of time-domain and frequency-domain features.
[0085] Let the features extracted by the temporal branch after passing through the WDCNN block be... The features extracted from the frequency domain branch by the WDCNN block are: , and These represent the local feature representations in the time domain and frequency domain, respectively.
[0086] During fusion, firstly for and Assign fusion weights and ,satisfy The formula for weighted matrix addition fusion is:
[0087]
[0088] in, This represents the fused features. and These are the weighted coefficients of the time-domain and frequency-domain features, satisfying... ;
[0089] The fused feature matrix As input, a classifier (such as a fully connected layer and Softmax) is used to determine the fault category of the wind turbine turntable bearing;
[0090] The classification process uses the Softmax function to calculate the final output probability.
[0091]
[0092] in, For the first The probability of a type of failure. It is the number of fault categories. It is the output of the classifier; To index the original fractions.
[0093] Compared with the prior art, the beneficial effects of the present invention are:
[0094] This invention proposes a novel fault diagnosis method by employing dual-branch time-frequency domain parallel feature extraction, combined with a WDCNN-Transformer serial structure and time-frequency domain feature fusion. This method effectively improves the accuracy and robustness of wind turbine turntable bearing fault diagnosis. It utilizes both time and frequency domain branches, automatically extracting features using deep networks to ensure comprehensive mining of signal information from multiple perspectives. The concatenation of WDCNN and Transformer enhances the utilization efficiency of temporal and spatial features. Feature fusion fully leverages the complementary advantages of time-frequency information, enhancing the discriminative power of fault features. Attached Figure Description
[0095] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0096] Figure 1 This is a schematic diagram of the overall process of the present invention;
[0097] Figure 2 This is a flowchart of the fault feature extraction process of the present invention;
[0098] Figure 3 This is a schematic diagram of the time-frequency fault feature fusion and fault classification and identification method of the present invention. Detailed Implementation
[0099] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0100] Please see Figure 1 A method for diagnosing wind turbine turntable bearing faults based on WDCNN-Transformer includes at least the following steps:
[0101] S100: Uses an acoustic emission sensor to collect the raw acoustic emission signal of the wind turbine turntable bearing, and then performs preprocessing such as noise reduction and normalization on the raw signal;
[0102] S200: Construct the WDCNN-Transformer model, using the WDCNN after linear mapping as the input to the Transformer;
[0103] S300: Performs Fast Fourier Transform (FFT) on the acoustic emission signal of the wind turbine turntable bearing to convert the signal from the time domain to the frequency domain. Inputs the time domain and frequency domain signals into WDCNN for local time-series and frequency domain feature extraction. The Transformer module captures long-distance global dependencies and contextual information to finally obtain time-frequency domain feature representation.
[0104] S400: The feature matrices obtained from the time domain branch and frequency domain branch in step S300 are fused by weighted matrix addition to achieve the fusion of time and frequency domain information. The fusion is then input into the model to train the model and obtain the probability values of different faults, thereby realizing the fault diagnosis of wind turbine turntable bearings based on WDCNN-Transformer.
[0105] Original acoustic emission signal in S100 Preprocessing includes at least the following steps:
[0106] S101: First, denoise the original acoustic emission signal, determine a suitable frequency range, and set the frequency range to [value missing]. arrive ;
[0107] The original signal is filtered using a bandpass filter to obtain the denoised signal:
[0108]
[0109] in, This is the lower cutoff frequency of the bandpass filter; This is the upper cutoff frequency of the bandpass filter; This is the denoised signal after bandpass filtering; Indicates a bandpass filtering operation;
[0110] S102: Denoising signal Perform zero-mean normalization and calculate the mean of the signal. and standard deviation :
[0111]
[0112]
[0113] Normalize the signal:
[0114]
[0115] in, The signal after normalization For denoised signals The total number of sampling points, For the first The signal values at each sampling point; Reflects the overall energy center level of the acoustic emission signal in this segment;
[0116] S103: Normalized signal Segmentation;
[0117] The signal is divided into multiple small segments, each segment having a length of [length missing]. And process it using a sliding window;
[0118] Assume the total length of the signal is Step size is Then the first A small segment of the signal can be represented as:
[0119]
[0120] in, It is the first The start time of a small segment; It is the end time.
[0121] WDCNN blocks use convolutional kernels of different widths, enabling them to extract local features of signals at multiple scales.
[0122] The Transformer block employs a self-attention mechanism to establish connections between features at different locations in the WDCNN-Transformer fault diagnosis model. Its multilayer perceptron further mixes the features. After multiple stages of processing, the features undergo global average pooling, averaging the feature values of each channel to obtain a global feature vector, which is then fed into the fully connected layer.
[0123] The global feature vectors of the fully connected layer are mapped to the output space of the model.
[0124] The output layer uses the Softmax activation function to convert the output into a probability distribution, which determines the fault type of the wind turbine turntable bearing and realizes intelligent fault diagnosis of the wind turbine turntable bearing.
[0125] The WDCNN-Transformer fault diagnosis model consists of a WDCNN block, a Transformer block, a fully connected layer, and an output layer.
[0126] WDCNN block: Employs a four-level convolutional block structure, using convolutional kernels of varying widths to achieve multi-scale local feature extraction. Specific parameters are as follows:
[0127] Convolutional block 1: Convolutional kernel width 64, number of channels 16, stride 2, ReLU activation followed by max pooling (kernel size 2).
[0128] Convolutional block 2: kernel width 32, number of channels 32, stride 2, ReLU activation followed by max pooling;
[0129] Convolutional block 3: kernel width 16, number of channels 64, stride 2, ReLU activation followed by max pooling;
[0130] Convolutional block 4: kernel width 8, number of channels 128, stride 1, ReLU activation;
[0131] The input to the WDCNN block is the preprocessed time-domain / frequency-domain signal, and the output is a multi-scale local feature sequence. The time-domain branch outputs features with dimensions of (batch_size, 128, L1), and the frequency-domain branch outputs features with dimensions of (batch_size, 128, L2).
[0132] Linear mapping layer: Set up a linear mapping layer to map the time-domain and frequency-domain local feature sequences output by the WDCNN block to a unified feature dimension d_model=256, ensuring that the features are adapted to the input of the Transformer block;
[0133] The Transformer block includes a position encoding module, a multi-head self-attention layer, and a feedforward network layer. It also features residual connections and layer normalization. Specific parameters are: number of multi-head self-attention heads (nhead) = 8, and the hidden layer dimension of the feedforward network is... =512; The positional encoding module adds positional encoding to the feature sequence, preserving the sequence order information; The multi-head self-attention layer captures the long-distance global dependencies of the features; The feedforward network layer further fuses and nonlinearly transforms the features; Residual connections and layer normalization avoid the vanishing gradient of the model and improve training stability;
[0134] Fully connected layer: maps the fused time-frequency domain feature matrix to the model's output space, achieving feature dimensionality reduction and classification mapping;
[0135] Softmax output layer: The Softmax activation function is used to convert the output of the fully connected layer into a probability distribution of different fault types, thereby realizing fault type discrimination.
[0136] See Figure 2 The S300 includes at least the following steps:
[0137] S301: Performs a Fast Fourier Transform (FFT) on the acoustic emission signal, converting the AE signal from the time domain to the frequency domain.
[0138]
[0139] in, Represents a frequency domain signal; It is the imaginary unit; It is the length of the signal; It is a time-domain signal; It is a frequency domain index; It is a complex exponential basis function;
[0140] To facilitate subsequent processing, the frequency domain characteristics are represented by calculating the amplitude spectrum or power spectrum;
[0141] S302: Extracting local temporal features using WDCNN blocks in the temporal branch:
[0142] The original time-domain sequence The input to the variable-width one-dimensional convolutional neural network (WDCNN) block in the temporal branch extracts local temporal features from the AE signal through convolutional kernels with different receptive fields.
[0143] For the Layered convolution, with kernel length denoted as . Step size is Then the first layer The convolution output of each output channel can be represented as:
[0144]
[0145] in, For the next level Input characteristics of each channel; These are the kernel weights; For bias terms; Total number of input channels in the previous layer; For the first Layer The first output channel The local feature pixels output by each WDCNN block form the basic unit of the local feature sequence. The index mapping of input features reflects the sliding sampling logic of the convolution stride on the input features, and determines the size reduction ratio of the feature map; For the first Layer The bias terms of each output channel are trainable parameters of WDCNN, used to adjust the baseline of the feature output and improve the model's fitting ability.
[0146] The result obtained using a nonlinear activation function is:
[0147]
[0148] in, For the activation function, this invention chooses ReLU; The convolution output of the m-th channel at the p-th position in the l-th layer;
[0149] To enhance robustness and reduce redundancy, this invention may introduce max pooling operations after several convolutional layers:
[0150]
[0151] in, For the first Layer The first channel in the The pooling output value of each pooling window; For the first Layer The feature vectors of each channel before pooling; The index for the pooled window; This is the pooling window size; This is a local index within the pooling window, with a value range of [value range missing]. ; This is a maximum value operation, which selects the maximum value from each pooling window as the output of that window;
[0152] After passing through multiple WDCNN blocks, the temporal local feature sequence representation is obtained. ;
[0153] S303: Extracting local spectral structure features in the frequency domain branch using WDCNN blocks:
[0154] The frequency domain signal obtained by Fast Fourier Transform The WDCNN block input to the frequency domain branch extracts features along the frequency axis using one-dimensional convolutions. The convolution, activation, and pooling methods are consistent with those in the time domain branch, ultimately yielding a local feature sequence representation in the frequency domain. The extracted features include at least local frequency band energy concentration, harmonic peak structure, and local spectral variations;
[0155] S304: Capturing long-range global dependencies and context information using the Transformer encoder:
[0156] To further model the global dependencies of the AE signal over long time scales (time domain) or cross-band scales (frequency domain), a Transformer encoder module is cascaded onto the output feature sequence of the WDCNN block. Let the output feature sequence of the WDCNN block be:
[0157]
[0158] in, The sequence length; For feature dimensions;
[0159] By adding positional encoding to the sequence, we obtain:
[0160]
[0161] in, For the feature sequence after adding positional encoding, the dimension is... same; For the position encoding matrix, the dimensions are... The same applies; it is used to assign a unique code to each position in the sequence in order to preserve the sequence's order information.
[0162] Subsequently, in the self-attention layer of the Transformer, the query is obtained through linear mapping. ,key Sum :
[0163]
[0164] in, The linear transformation weight matrix is for learning; ;
[0165] Self-attention is calculated as follows:
[0166]
[0167] in, The query is the dot product of the key and the dimension. , representing the correlation score between each pair of positions; This is a scaling factor used to prevent the softmax gradient from vanishing due to an excessively large dot product result. The normalization function converts the correlation scores into a probability distribution; For self-attention output, the dimension is ;
[0168] After concatenating the multi-head attention signals, a linear transformation is used to obtain the final output of the multi-head attention signal.
[0169]
[0170] in, For the first The output of each attention head; To output the linear transformation weight matrix; The concatenation operation concatenates the outputs of all attention heads along the feature dimension; The final output of multi-head attention, with dimensions of ;
[0171] Combined with residual connection With layer normalization :
[0172]
[0173] Feedforward networks (FFNs) further process the output features:
[0174]
[0175] in, The input feature matrix (from the output of the previous layer); This is the weight matrix for the first linear transformation; This is the bias vector for the first linear transformation; For activation functions; This is the weight matrix for the first linear transformation; This is the bias vector for the first linear transformation; This is the output of the feedforward network;
[0176] Final output:
[0177]
[0178] in, For residual connection, For layer normalization, This is the final output of the Transformer layer.
[0179] When using amplitude spectrum in S301:
[0180]
[0181] in, For the first Amplitude spectrum values of each frequency component; The result obtained after Fast Fourier Transform (FFT) is the 1st... Complex representation of each frequency component; For complex numbers The real part, for The imaginary part.
[0182] When using power spectrum in S301:
[0183]
[0184] in, For the first Power spectral values of each frequency component The total length of the signal. This represents the energy of that frequency component.
[0185] S400 includes at least the following steps:
[0186] By fully combining the advantages of time-domain and frequency-domain signals, a weighted matrix addition fusion method is used to form the final time-frequency domain feature representation by weighted fusion of time-domain and frequency-domain features.
[0187] like Figure 3 As shown: By fully combining the advantages of time-domain and frequency-domain signals, the weighted matrix addition fusion method is used to form the final time-frequency domain feature representation by weighted fusion of time-domain and frequency-domain features.
[0188] Let the features extracted by the temporal branch after passing through the WDCNN block be... The features extracted from the frequency domain branch by the WDCNN block are: , and These represent the local feature representations in the time domain and frequency domain, respectively.
[0189] During fusion, firstly for and Assign fusion weights and ,satisfy The formula for weighted matrix addition fusion is:
[0190]
[0191] in, This represents the fused features. and These are the weighted coefficients of the time-domain and frequency-domain features, satisfying... ;
[0192] The fused feature matrix As input, a classifier (such as a fully connected layer and Softmax) is used to determine the fault category of the wind turbine turntable bearing;
[0193] The classification process uses the Softmax function to calculate the final output probability.
[0194]
[0195] in, For the first The probability of a type of failure. It is the number of fault categories. It is the output of the classifier; To index the original fractions.
[0196] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
Claims
1. A wind power rotating disc bearing fault diagnosis method based on WDCNN-Transformer, characterized by: At least the following steps are included: S100: Collecting the original acoustic emission signal of the wind power rotary bearing by using the acoustic emission sensor, and preprocessing the original acoustic emission signal The preprocessing at least includes denoising, zero-mean normalization and segmentation preprocessing in sequence, to obtain the preprocessed signal sample; S200: Construct a WDCNN-Transformer fault diagnosis model, which includes a WDCNN block, a Transformer block, a fully connected layer, and an output layer. The output of the WDCNN block is linearly mapped and used as the input of the Transformer block. S300: Perform a Fast Fourier Transform on the preprocessed signal samples obtained in step S100 to convert the signal from the time domain to the frequency domain. Input the time domain signal and the frequency domain signal into the WDCNN block to extract local time-series features and local frequency domain features, respectively. Then, input the extracted features into the Transformer block, which captures the long-range global dependency and context information of the features to obtain the time-frequency domain feature representations in the time domain and frequency domain, respectively. S400: The time-domain and frequency-domain feature representations obtained in step S300 are fused by weighted matrix addition to obtain the fused time-frequency domain feature matrix. The feature matrix is then input into the fully connected layer and the output layer for model training, and the probability values of different fault types are output to realize the fault diagnosis of wind turbine turntable bearings.
2. The method for fault diagnosis of wind turbine turntable bearings based on WDCNN-Transformer according to claim 1, characterized in that: The raw acoustic emission signals in the S100 The preprocessing of the raw acoustic emission signals in the S100 comprises at least the following steps: S101: First, the original acoustic emission signal is denoised, the appropriate frequency range is determined, and the frequency range is set to to ; The original signal is filtered using a bandpass filter to obtain the denoised signal: wherein, is a lower cut-off frequency of the band-pass filter; is an upper cut-off frequency of the band-pass filter; is a de-noised signal after band-pass filtering; denotes a band-pass filtering operation; S102: Denoising signal Perform zero-mean normalization and calculate the mean of the signal. and standard deviation : Normalize the signal: in, The signal after normalization For denoised signals The total number of sampling points, For the first The signal values at each sampling point; Reflects the overall energy center level of the acoustic emission signal in this segment; S103: Normalized signal Segmentation; The signal is divided into multiple small segments, each segment having a length of [length missing]. And process it using a sliding window; Assume the total length of the signal is Step size is Then the first A small segment of the signal can be represented as: in, It is the first The start time of a small segment; It is the end time.
3. The method for fault diagnosis of wind turbine turntable bearings based on WDCNN-Transformer according to claim 2, characterized in that: The WDCNN block uses convolutional kernels of different widths, which enables the extraction of local features of the signal at multiple scales; The Transformer block employs a self-attention mechanism to establish connections between features at different locations in the WDCNN-Transformer fault diagnosis model, and its multilayer perceptron further mixes features. After multiple stages of processing, the features are subjected to global average pooling, which averages the feature values of each channel to obtain a global feature vector, which is then fed into the fully connected layer. The global feature vectors of the fully connected layer are mapped to the output space of the model. The output layer uses the Softmax activation function to convert the output into a probability distribution, determine the fault type of the wind turbine turntable bearing, and realize intelligent fault diagnosis of the wind turbine turntable bearing.
4. The method for fault diagnosis of wind turbine turntable bearings based on WDCNN-Transformer according to claim 3, characterized in that: The S300 includes at least the following steps: S301: Performs a Fast Fourier Transform on the acoustic emission signal, converting the AE signal from the time domain to the frequency domain. in, Represents a frequency domain signal; It is the imaginary unit; It is the length of the signal; It is a time-domain signal; It is a frequency domain index; It is a complex exponential basis function; To facilitate subsequent processing, the frequency domain characteristics are represented by calculating the amplitude spectrum or power spectrum; S302: Extracting local temporal features using WDCNN blocks in the temporal branch: The original time-domain sequence The input to the variable-width one-dimensional convolutional neural network (WDCNN) block in the temporal branch extracts local temporal features from the AE signal through convolutional kernels with different receptive fields. For the Layered convolution, with kernel length denoted as . Step size is Then the first layer The convolution output of each output channel can be represented as: in, For the next level Input characteristics of each channel; These are the kernel weights; For bias terms; Total number of input channels in the previous layer; For the first Layer The first output channel The local feature pixels output by each WDCNN block form the basic unit of the local feature sequence. The index mapping of input features reflects the sliding sampling logic of the convolution stride on the input features, and determines the size reduction ratio of the feature map; For the first Layer The bias terms of each output channel are trainable parameters of WDCNN, used to adjust the baseline of the feature output and improve the model's fitting ability. The result obtained using a nonlinear activation function is: in, For the activation function, this invention chooses ReLU; The convolution output of the m-th channel at the p-th position in the l-th layer; To enhance robustness and reduce redundancy, this invention may introduce max pooling operations after several convolutional layers: in, For the first Layer The first channel in the The pooling output value of each pooling window; For the first Layer The feature vectors of each channel before pooling; The index for the pooled window; This is the pooling window size; This is a local index within the pooling window, with a value range of [value range missing]. ; This is a maximum value operation, which selects the maximum value from each pooling window as the output of that window; After passing through multiple WDCNN blocks, the temporal local feature sequence representation is obtained. ; S303: Extracting local spectral structure features in the frequency domain branch using WDCNN blocks: The frequency domain signal obtained by Fast Fourier Transform The WDCNN block input to the frequency domain branch extracts features along the frequency axis using one-dimensional convolutions. The convolution, activation, and pooling methods are consistent with those in the time domain branch, ultimately yielding a local feature sequence representation in the frequency domain. The extracted features include at least local frequency band energy concentration, harmonic peak structure, and local spectral variations; S304: Capturing long-range global dependencies and context information using the Transformer encoder: To further model the global dependencies of the AE signal over long time scales or cross-band scales, a Transformer encoder module is cascaded onto the output feature sequence of the WDCNN block. Let the output feature sequence of the WDCNN block be: in, The sequence length; For feature dimensions; By adding positional encoding to the sequence, we obtain: in, For the feature sequence after adding positional encoding, the dimension is... same; For the position encoding matrix, the dimensions are... The same applies; it is used to assign a unique code to each position in the sequence in order to preserve the sequence's order information. Subsequently, in the self-attention layer of the Transformer, the query is obtained through linear mapping. ,key Sum : in, The linear transformation weight matrix is for learning; ; Self-attention is calculated as follows: in, The query is the dot product of the key and the dimension. , representing the correlation score between each pair of positions; This is a scaling factor used to prevent the softmax gradient from vanishing due to an excessively large dot product result. The normalization function converts the correlation scores into a probability distribution; For self-attention output, the dimension is ; After concatenating the multi-head attention signals, a linear transformation is used to obtain the final output of the multi-head attention signal. in, For the first The output of each attention head; To output the linear transformation weight matrix; The concatenation operation concatenates the outputs of all attention heads along the feature dimension; The final output of multi-head attention, with dimensions of ; Combined with residual connection With layer normalization : The feedforward network further processes the output features: in, The input feature matrix; This is the weight matrix for the first linear transformation; This is the bias vector for the first linear transformation; For activation functions; This is the weight matrix for the first linear transformation; This is the bias vector for the first linear transformation; This is the output of the feedforward network; Final output: in, For residual connection, For layer normalization, This is the final output of the Transformer layer.
5. The method for fault diagnosis of wind turbine turntable bearings based on WDCNN-Transformer according to claim 4, characterized in that: When using amplitude spectrum in S301: in, For the first Amplitude spectrum values of each frequency component; The first one obtained after Fast Fourier Transform Complex representation of each frequency component; For complex numbers The real part, for The imaginary part.
6. The method for fault diagnosis of wind turbine turntable bearings based on WDCNN-Transformer according to claim 4, characterized in that: When power spectrum is used in S301: in, For the first Power spectral values of each frequency component The total length of the signal. This represents the energy of that frequency component.
7. The method for fault diagnosis of wind turbine turntable bearings based on WDCNN-Transformer according to claim 4, characterized in that: The S400 includes at least the following steps: By fully combining the advantages of time-domain and frequency-domain signals, a weighted matrix addition fusion method is used to form the final time-frequency domain feature representation by weighted fusion of time-domain and frequency-domain features. Let the features extracted by the temporal branch after passing through the WDCNN block be... The features extracted from the frequency domain branch by the WDCNN block are: , and These represent the local feature representations in the time domain and frequency domain, respectively. During fusion, firstly for and Assign fusion weights and ,satisfy The formula for weighted matrix addition fusion is: in, This represents the fused features. and These are the weighted coefficients of the time-domain and frequency-domain features, satisfying... ; The fused feature matrix As input, a classifier is used to determine the fault type of the wind turbine turntable bearing; The classification process uses the Softmax function to calculate the final output probability. in, For the first The probability of a type of failure. It is the number of fault categories. It is the output of the classifier; To index the original fractions.