Wind turbine generator fault monitoring method and system

Abstract

The invention relates to the technical field of wind turbine generator fault monitoring. The invention provides a wind turbine generator fault monitoring method and system. The method comprises the following steps: synchronously acquiring gearbox and environment temperature and humidity data, and generating a time-frequency energy fusion matrix through adaptive wavelet packet transformation; adopting mutual information entropy weighted improved variational mode decomposition to screen out an intrinsic mode component set related to a fault mode; constructing a space-time double-flow residual network based on the intrinsic mode component set, and fusing two branch outputs of the space-time double-flow residual network through a dynamic feature gating mechanism to obtain a multi-dimensional feature vector; and inputting a multi-dimensional feature vector obtained by fusion into a lightweight fault classifier, and outputting a real-time fault probability and a component health degree evaluation index based on a sliding window mechanism. The problems of low efficiency, high false alarm rate, missing detection of early faults, reduction of prediction precision, incapability of mining multivariable coupling relations, need of massive annotation data, and high delay caused by insufficient edge side computing power existing in an existing wind turbine generator fault monitoring mode are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind turbine generator set fault monitoring, and in particular to a wind turbine generator set fault monitoring method and system. Background Art

[0002] As the global energy mix shifts toward low-carbon development, wind power has become a key pillar of renewable energy. However, wind turbines operate in harsh environments such as high humidity, salt spray, and extreme temperatures for extended periods, leading to high failure rates for core components like gearboxes, bearings, and blades.

[0003] Traditional troubleshooting methods based on manual inspections are inefficient and unable to meet the needs of large-scale operations and maintenance. Monitoring solutions that rely on single sensors suffer from high false alarm rates and missed early faults. Single-variable models based on SCADA systems experience reduced prediction accuracy when data volumes surge, and are unable to exploit multivariate coupling relationships. While deep learning models (such as LSTM and CNN) can improve diagnostic accuracy, they require massive amounts of labeled data, and insufficient edge computing power leads to high latency. Summary of the Invention

[0004] The purpose of the present invention is to provide a wind turbine fault monitoring method and system, aiming to solve the problems of low efficiency, high false alarm rate, missed early faults, reduced prediction accuracy, inability to mine multivariable coupling relationships, requirement for massive labeled data, and high latency caused by insufficient edge computing power in existing wind turbine fault monitoring methods.

[0005] The present invention is achieved through the following technical solutions:

[0006] A wind turbine fault monitoring method comprises the following steps:

[0007] Synchronously collect gearbox vibration signals, bearing temperature series, blade strain waveforms, and ambient temperature and humidity data, and generate a time-frequency energy fusion matrix through adaptive wavelet packet transform;

[0008] The improved variational mode decomposition weighted by mutual information entropy is used to perform multi-scale decomposition on the time-frequency energy fusion matrix and screen out the intrinsic mode component set related to the fault mode.

[0009] Based on the intrinsic mode component set, a spatiotemporal two-stream residual network is constructed. The two-branch output of the spatiotemporal two-stream residual network is fused through a dynamic feature gating mechanism to obtain a multi-dimensional feature vector. The spatial branch of the spatiotemporal two-stream residual network uses a deformable convolution kernel to extract non-uniformly distributed fault features, while the temporal branch of the spatiotemporal two-stream residual network uses a sparse gated recurrent unit to capture long-term degradation trends.

[0010] The fused multidimensional feature vector is input into a lightweight fault classifier, which outputs real-time fault probability and component health assessment indicators based on a sliding window mechanism. When the indicators exceed the dynamic threshold, a graded warning signal is triggered.

[0011] Optionally, the specific process of synchronously collecting the gearbox vibration signal, bearing temperature sequence, blade strain waveform and ambient temperature and humidity data, and generating the time-frequency energy fusion matrix through adaptive wavelet packet transform is as follows:

[0012] Through the synchronous triggering mechanism of multi-source heterogeneous sensors, the vibration signal, temperature series, strain waveform and environmental parameters are time-scale aligned, and the empirical wavelet denoising and sliding mean filtering are combined to preprocess the data of each channel;

[0013] According to the energy concentration of the vibration signal frequency band and the kurtosis of the strain waveform spectrum, the number of wavelet packet decomposition layers and the type of basis functions are dynamically adjusted, and the pre-processed multi-source signals are adaptively segmented in frequency bands.

[0014] The relative energy proportion of each sub-band is extracted and normalized, and then sorted by the energy entropy value of the frequency band to form a three-dimensional energy distribution matrix, where the first dimension corresponds to the type of physical quantity, the second dimension represents the frequency band division level, and the third dimension records the energy evolution characteristics within the time window;

[0015] The environmental temperature and humidity parameters are integrated as energy attenuation factors, and the three-dimensional energy distribution matrix is coupled with the attenuation factors through tensor contraction operation to generate a time-frequency energy fusion matrix with environmental adaptability.

[0016] Optionally, the specific process of using the improved variational mode decomposition weighted by mutual information entropy to perform multi-scale decomposition on the time-frequency energy fusion matrix and screen out the intrinsic mode component set related to the fault mode is as follows:

[0017] Based on the energy entropy distribution of each sub-band of the time-frequency energy fusion matrix, the number of modes and penalty factors of variational mode decomposition are dynamically initialized;

[0018] Expand the matrix into two-dimensional time series slices along the frequency band dimension, calculate the mutual information entropy value between each slice and the historical fault mode library, and construct the frequency band sensitivity weight vector;

[0019] A frequency band weight constraint term is introduced into the objective function of variational mode decomposition, and the optimal eigenmode component is iteratively solved using the alternating direction multiplier method. The convergence condition of the modal center frequency is related to the energy attenuation factor of the corresponding frequency band.

[0020] Perform a joint kurtosis-envelope spectrum analysis on the decomposed intrinsic mode component set to screen out components that simultaneously meet the kurtosis threshold greater than 3.5 and the envelope spectrum peak frequency matches the fault characteristic frequency library;

[0021] The filtered components are converted into time-frequency image feature blocks through the Gram angle field and input into the pre-trained twin network for fault mode similarity matching. The intrinsic mode components with similarity higher than the set threshold are retained to form the intrinsic mode component set.

[0022] Optionally, the specific construction process of the pre-trained twin network is:

[0023] A time-frequency image training set containing historical fault samples and their corresponding fault type labels is constructed. Random scaling and salt-and-pepper noise are applied to each sample to enhance data diversity.

[0024] Build a feature extraction subnetwork consisting of parallel depthwise separable convolutional layers and channel attention modules. Each subnetwork contains three cascaded convolutional blocks, each of which consists of deformable convolution, batch normalization, and LeakyReLU activation functions.

[0025] A dual-path symmetrical structure is used to build the twin network backbone. The outputs of the two feature extraction sub-networks are globally averaged and pooled before being input into the similarity measurement layer. The Mahalanobis distance between the two is calculated as the fault mode difference indicator.

[0026] The network parameters are optimized based on the triplet loss function. Anchor samples, positive samples and negative samples are collected from the training set to form training triplets. The distance between the anchor point and the positive sample is constrained to be less than the distance between the anchor point and the negative sample by at least one marginal coefficient.

[0027] By introducing prior knowledge of fault feature frequencies and adding a frequency-sensitive constraint term to the last convolutional layer of the feature extraction subnetwork, the network weight update direction is positively correlated with the energy distribution of the fault feature frequency band.

[0028] The training cycle is controlled by the sliding validation set early stopping strategy. When the F1 score on the validation set does not improve for three consecutive training cycles, the training is terminated and the optimal weight parameters are saved for online similarity matching.

[0029] Optionally, the specific process of constructing a spatiotemporal dual-stream residual network based on the intrinsic mode component set and fusing the two-branch outputs of the spatiotemporal dual-stream residual network through a dynamic feature gating mechanism to obtain a multidimensional feature vector is as follows:

[0030] The eigenmode component set is divided into overlapping segments according to the time dimension and input into the spatial branch and the time branch for feature extraction respectively;

[0031] The spatial branch consists of multiple cascaded deformable residual blocks, each of which contains a deformable convolution layer, a batch normalization layer, and an adaptive threshold ReLU activation function. The offset of the deformable convolution kernel is dynamically generated by the feature map of the previous layer through a lightweight fully connected layer to capture non-uniformly distributed local fault modes.

[0032] The time branch uses stacked sparse gated recurrent units, compresses the unit's internal state matrix through weight pruning and quantization operations, and embeds a temporal attention mechanism in the hidden state update equation to enhance the memory retention of key time nodes in long-term degradation trends.

[0033] During the feature fusion phase of the spatiotemporal dual-stream network, a dynamic feature gating mechanism is established: the high-dimensional feature map output by the spatial branch is channel-compressed and then tensor-concatenated with the hidden state vector of the temporal branch. The spatial-temporal feature coupling coefficient is calculated using a learnable gating weight matrix, and the sigmoid function is used to generate adaptive fusion weights.

[0034] The weighted fusion features are subjected to multi-scale pooling operations to extract statistical features of different granularities. Finally, after dimensionality reduction through the fully connected layer, a multi-dimensional feature vector containing spatiotemporal correlation information is output.

[0035] Optionally, the specific process of inputting the fused multidimensional feature vector into a lightweight fault classifier and outputting the real-time fault probability and component health assessment index based on a sliding window mechanism is as follows:

[0036] Construct a hybrid classifier that is an integration of depthwise separable convolutional layers and random forests. The convolutional layers are used to extract local correlations of feature vectors, while the random forest classifier achieves dimensionality compression through feature importance sampling.

[0037] A sliding window mechanism is used to perform time-series sliding interception of multi-dimensional feature vectors, and the window length matches the typical fault evolution cycle of wind turbines.

[0038] A dual-channel evaluation module is set at the output end of the classifier; the first channel of the dual-channel evaluation module calculates the real-time probability distribution of each fault mode through the Softmax function; the second channel of the dual-channel evaluation module calculates the deviation between the current state and the health benchmark library based on the Mahalanobis distance of the multidimensional feature vector as the health indicator.

[0039] Optionally, the specific process of triggering a graded warning signal when the indicator exceeds a dynamic threshold is:

[0040] A dynamic threshold update algorithm is established to determine the initial threshold based on the 3σ criterion of historical health data of the same wind farm, and an exponential decay factor of ambient temperature and humidity parameters is introduced to perform online correction of the threshold.

[0041] When the health index exceeds the dynamic threshold for three consecutive samples, the dominant fault type is determined based on the maximum value of the fault probability distribution, and a graded signal including warning level, fault location and maintenance suggestions is generated;

[0042] The graded warning signal is transmitted to the edge gateway after compressed sensing encoding, synchronously triggering the on-site sound and light alarm and the push of work orders on the remote operation and maintenance platform.

[0043] Based on the same inventive concept, the present invention further provides a wind turbine fault monitoring system for implementing the wind turbine fault monitoring method, comprising:

[0044] The multi-source synchronous acquisition module is used to collect gearbox vibration signals, bearing temperature series, blade strain waveforms, and ambient temperature and humidity data through a synchronous triggering mechanism of multi-source heterogeneous sensors, and perform time-scale alignment and combined noise reduction preprocessing on the data of each channel;

[0045] Adaptive time-frequency processing module, used to dynamically adjust the wavelet packet decomposition parameters according to the signal frequency band characteristics, generate a three-dimensional energy distribution matrix containing the physical quantity type, frequency band level and time window energy evolution, and couple it with the environmental parameters to form a time-frequency energy fusion matrix;

[0046] The multi-scale decomposition module is used to perform an improved variational modal decomposition of the time-frequency energy fusion matrix based on a mutual information entropy weighting strategy, and to screen out the fault-related intrinsic modal component set through joint kurtosis-envelope spectrum analysis and twin network matching;

[0047] The spatiotemporal feature extraction module is used to construct a spatial branch containing a deformable residual block and a temporal branch with a sparse gated recurrent unit to extract non-uniform fault features and long-term degradation trends, respectively.

[0048] A dynamic fusion decision module is used to calculate the spatiotemporal feature coupling coefficient through a learnable gating weight matrix, realize the adaptive weighted fusion of the dual-stream network output, and generate a multi-dimensional feature vector containing multi-scale statistical features;

[0049] The edge warning execution module is used to input the fusion features into the lightweight fault classifier for real-time probability prediction, dynamically adjust the health assessment threshold according to the sliding window mechanism, and trigger the transmission of graded warning signals to the edge execution terminal.

[0050] Based on the same inventive concept, the present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the above-mentioned wind turbine fault monitoring method.

[0051] Based on the same inventive concept, the present invention further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the above-mentioned wind turbine fault monitoring method is implemented.

[0052] The technical solution of the present invention has at least the following advantages and beneficial effects:

[0053] Synchronously collect gearbox vibration signals, bearing temperature series, blade strain waveforms and ambient temperature and humidity data to comprehensively obtain wind turbine operating status information. Compared with a single sensor monitoring solution, it can more accurately reflect the overall condition of the unit, effectively reduce the false alarm rate and the problem of missed early faults, and provide a richer and more reliable data basis for subsequent fault monitoring. By generating a time-frequency energy fusion matrix through adaptive wavelet packet transform, it can fully explore the characteristics of the data at different time and frequency scales, provide a more effective feature representation for subsequent analysis, and help to more accurately identify fault modes.

[0054] The improved variational modal decomposition weighted by mutual information entropy is used to perform multi-scale decomposition of the time-frequency energy fusion matrix and screen out the set of intrinsic modal components related to the fault mode. Compared with traditional methods, it can more effectively decompose the signal, extract characteristic components closely related to the fault, and improve the accuracy and pertinence of fault feature extraction.

[0055] A spatiotemporal dual-stream residual network is constructed based on the intrinsic mode component set. The deformable convolution kernel of the spatial branch is used to extract non-uniformly distributed fault features. This network can adapt to the complex and changeable operating environment of wind turbines and capture fault information at different locations. The temporal branch uses a sparse gated recurrent unit to capture long-term degradation trends, effectively tracking the slow changes in unit performance over time, and comprehensively exploring the characteristics of faults in the temporal and spatial dimensions, providing richer feature information for subsequent fault classification.

[0056] The two-branch outputs of the spatiotemporal dual-stream residual network are fused through a dynamic feature gating mechanism to obtain a multi-dimensional feature vector, which can fully utilize the feature information of both time and space, improve the rationality and effectiveness of feature fusion, and further enhance the accuracy of fault diagnosis.

[0057] The fused multi-dimensional feature vector is input into a lightweight fault classifier, and real-time fault probability and component health assessment indicators are output based on a sliding window mechanism. The lightweight design enables the algorithm to run quickly even with limited computing power on the edge side, reducing latency. The real-time output evaluation indicators can promptly reflect the operating status of the unit. When the indicators exceed the dynamic threshold, a graded warning signal is triggered, achieving real-time warning of faults. This helps operation and maintenance personnel take timely measures to avoid further expansion of the fault, improve the operating reliability and safety of wind turbines, and meet the needs of large-scale operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 Schematic diagram of a flow chart of a wind turbine fault monitoring method according to an embodiment of the present invention;

[0059] Figure 2 Schematic diagram of the structure of a wind turbine fault monitoring system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The following is a specific implementation method with reference to the accompanying drawings.

[0061] Reference Figure 1 , a wind turbine fault monitoring method, comprising the following steps:

[0062] Step 1: Synchronously collect the gearbox vibration signal, bearing temperature series, blade strain waveform and ambient temperature and humidity data, and generate a time-frequency energy fusion matrix through adaptive wavelet packet transform.

[0063] In some embodiments, the specific process of synchronously collecting the gearbox vibration signal, bearing temperature series, blade strain waveform and ambient temperature and humidity data and generating the time-frequency energy fusion matrix through adaptive wavelet packet transform is as follows:

[0064] Through the synchronous triggering mechanism of multi-source heterogeneous sensors, the vibration signal, temperature series, strain waveform and environmental parameters are time-scale aligned, and the empirical wavelet denoising and sliding mean filtering are combined to preprocess the data of each channel;

[0065] According to the energy concentration of the vibration signal frequency band and the kurtosis of the strain waveform spectrum, the number of wavelet packet decomposition layers and the type of basis functions are dynamically adjusted, and the pre-processed multi-source signals are adaptively segmented in frequency bands.

[0066] The relative energy proportion of each sub-band is extracted and normalized, and then sorted by the energy entropy value of the frequency band to form a three-dimensional energy distribution matrix, where the first dimension corresponds to the type of physical quantity, the second dimension represents the frequency band division level, and the third dimension records the energy evolution characteristics within the time window;

[0067] The environmental temperature and humidity parameters are integrated as energy attenuation factors, and the three-dimensional energy distribution matrix is coupled with the attenuation factors through tensor contraction operation to generate a time-frequency energy fusion matrix with environmental adaptability.

[0068] Step 2: Use the improved variational mode decomposition weighted by mutual information entropy to perform multi-scale decomposition on the time-frequency energy fusion matrix and screen out the set of intrinsic mode components related to the fault mode.

[0069] In some embodiments, the improved variational mode decomposition weighted by mutual information entropy is used to perform multi-scale decomposition on the time-frequency energy fusion matrix, and the specific process of screening out the intrinsic mode component set related to the fault mode is as follows:

[0070] Based on the energy entropy distribution of each sub-band of the time-frequency energy fusion matrix, the number of modes and penalty factors of variational mode decomposition are dynamically initialized;

[0071] Expand the matrix into two-dimensional time series slices along the frequency band dimension, calculate the mutual information entropy value between each slice and the historical fault mode library, and construct the frequency band sensitivity weight vector;

[0072] A frequency band weight constraint term is introduced into the objective function of variational mode decomposition, and the optimal eigenmode component is iteratively solved using the alternating direction multiplier method. The convergence condition of the modal center frequency is related to the energy attenuation factor of the corresponding frequency band.

[0073] Perform a joint kurtosis-envelope spectrum analysis on the decomposed intrinsic mode component set to screen out components that simultaneously meet the kurtosis threshold greater than 3.5 and the envelope spectrum peak frequency matches the fault characteristic frequency library;

[0074] The filtered components are converted into time-frequency image feature blocks through the Gram angle field and input into the pre-trained twin network for fault mode similarity matching. The intrinsic mode components with similarity higher than the set threshold are retained to form the intrinsic mode component set.

[0075] In some embodiments, the specific construction process of the pre-trained Siamese network is:

[0076] Construct a time-frequency image training set containing historical fault samples and their corresponding fault type labels, and apply random scale scaling and salt and pepper noise to each sample to enhance data diversity; given the original sample The enhanced sample x ′ As shown in the following formula (1):

[0077]

[0078] Among them, T aug (x) represents the data enhancement function; α represents the randomly generated scaling factor matrix; ⊙ represents the element-by-element multiplication, which is used to adjust the local brightness of the original sample x; Represents the salt and pepper noise function, which randomly sets some pixels in the original sample x to the minimum or maximum value (such as 0 or 1);

[0079] Establish a feature extraction subnetwork consisting of parallel depth-wise separable convolutional layers and channel attention modules; each subnetwork contains three cascaded convolutional blocks, each of which consists of deformable convolution, batch normalization, and LeakyReLU activation functions; the enhanced sample x ′ As input, the output of the lth deformable convolution block is shown in the following formula (2):

[0080]

[0081] In fact, F l-1 Represents the output feature map of the l-1th layer, which is used as the input of the current layer;

[0082] Represents a lightweight fully connected layer, input F l-1 , output the offset Δpl of the deformable convolution; Represents the weight parameter of the l-th layer offset generation layer;

[0083] Represents a deformable convolution operation, using a convolution kernel and the offset Δpl to F l-1 Perform convolution; F l bn Represents the feature map after batch normalization; represents the weight parameter of the lth layer of deformable convolution; F l deform Represents the feature map after deformable convolution; μ l and σ l Represent the mean and standard deviation of the current batch data, respectively, for batch normalization; γ l and β l denote the learnable scaling and translation parameters of batch normalization, respectively; F l act Represents the activated feature map; LeakyReLU(F l bn ,α=0.1) represents the linear unit activation function with leakage correction, α=0.1 represents the leakage coefficient; A l represents the channel attention weight vector of layer l; σ represents the Sigmoid activation function; GAP(F l act ) represents global average pooling, which is used to compress the spatial dimension of the feature map to 1×1; and Represent the weight matrix and bias term of the channel attention module respectively; F l out Represents the final output feature map of the lth convolutional block; Represents element-wise multiplication;

[0084] A dual-path symmetrical structure is used to build the twin network backbone, and the input pair (x i ,x j ), and extract features through a dual-path network, as shown in the following formula (3):

[0085]

[0086] Among them, (x i ′ ,x j ′ ) represents the input sample pair after data enhancement; G represents the feature extraction subnetwork with parameter θ; (h i ,h j ) represents the input pair (x i ,x j ) Feature vector extracted by the two-way network;

[0087] The outputs of the two feature extraction sub-networks are globally averaged and pooled before being input into the similarity measurement layer. The Mahalanobis distance between the two is calculated as the fault mode difference index, as shown in the following formula (4):

[0088]

[0089] Among them, Σ -1 Represents the inverse matrix of the covariance matrix of the training set feature vector, which is used to standardize the correlation between different dimensions; d M represents the Mahalanobis distance;

[0090] The network parameters are optimized based on the triplet loss function. Anchor samples, positive samples, and negative samples are collected from the training set to form training triplets. The distance between the anchor point and the positive sample is constrained to be less than the distance between the anchor point and the negative sample by at least one marginal coefficient. Given the anchor point a, positive sample p, and negative sample n, the loss function is shown in the following formula (5):

[0091]

[0092] in, represents the triplet loss function;

[0093] The prior knowledge of fault feature frequency is introduced, and a frequency-sensitive constraint term is added to the last convolutional layer of the feature extraction subnetwork, so that the network weight update direction is positively correlated with the energy distribution of the fault feature frequency band. The frequency-sensitive constraint term is shown in the following formula (6):

[0094]

[0095] in, represents the frequency-sensitive constraint term; λ represents a hyperparameter used to balance the weight of the frequency-sensitive constraint term and the triplet loss; E represents the energy distribution matrix of the fault feature frequency band, reflecting the energy distribution prior knowledge of the fault-related frequency band

[0096] The overall loss function is shown in the following formula (7):

[0097]

[0098] in, represents the overall loss function; is the expectation operation on the training triple (a, p, n), which means averaging the losses of all training samples;

[0099] The network parameters are optimized by back propagation as shown in the following equation (8):

[0100]

[0101] Among them, θ represents the parameters to be optimized of the twin network; θ* Represents the overall loss function The optimal parameters for minimization;

[0102] The training cycle is controlled by the sliding validation set early stopping strategy. When the F1 score on the validation set does not improve for three consecutive training cycles, the training is terminated and the optimal weight parameters are saved for online similarity matching.

[0103] Step 3: Based on the intrinsic mode component set, a spatiotemporal dual-stream residual network is constructed. The two-branch outputs of the spatiotemporal dual-stream residual network are fused through a dynamic feature gating mechanism to obtain a multi-dimensional feature vector. The spatial branch of the spatiotemporal dual-stream residual network extracts non-uniformly distributed fault features through a deformable convolution kernel; the temporal branch of the spatiotemporal dual-stream residual network uses a sparse gated recurrent unit to capture long-term degradation trends.

[0104] In some embodiments, based on the intrinsic mode component set, a spatiotemporal dual-stream residual network is constructed, and the two-branch outputs of the spatiotemporal dual-stream residual network are fused through a dynamic feature gating mechanism to obtain a multidimensional feature vector. The specific process is:

[0105] The eigenmode component set is divided into overlapping segments according to the time dimension and input into the spatial branch and the time branch for feature extraction respectively;

[0106] The spatial branch consists of multiple cascaded deformable residual blocks, each of which contains a deformable convolution layer, a batch normalization layer, and an adaptive threshold ReLU activation function. The offset of the deformable convolution kernel is dynamically generated by the feature map of the previous layer through a lightweight fully connected layer to capture non-uniformly distributed local fault modes.

[0107] The time branch uses stacked sparse gated recurrent units, compresses the unit's internal state matrix through weight pruning and quantization operations, and embeds a temporal attention mechanism in the hidden state update equation to enhance the memory retention of key time nodes in long-term degradation trends.

[0108] During the feature fusion phase of the spatiotemporal dual-stream network, a dynamic feature gating mechanism is established: the high-dimensional feature map output by the spatial branch is channel-compressed and then tensor-concatenated with the hidden state vector of the temporal branch. The spatial-temporal feature coupling coefficient is calculated using a learnable gating weight matrix, and the sigmoid function is used to generate adaptive fusion weights.

[0109] The weighted fusion features are subjected to multi-scale pooling operations to extract statistical features of different granularities. Finally, after dimensionality reduction through the fully connected layer, a multi-dimensional feature vector containing spatiotemporal correlation information is output.

[0110] Step 4: Input the fused multidimensional feature vector into the lightweight fault classifier, and output the real-time fault probability and component health assessment indicators based on the sliding window mechanism. When the indicators exceed the dynamic threshold, a graded warning signal is triggered.

[0111] In some embodiments, the fused multi-dimensional feature vector is input into a lightweight fault classifier, and the specific process of outputting real-time fault probability and component health assessment indicators based on a sliding window mechanism is as follows:

[0112] Construct a hybrid classifier that is an integration of depthwise separable convolutional layers and random forests. The convolutional layers are used to extract local correlations of feature vectors, while the random forest classifier achieves dimensionality compression through feature importance sampling.

[0113] A sliding window mechanism is used to perform time-series sliding interception of multi-dimensional feature vectors, and the window length matches the typical fault evolution cycle of wind turbines.

[0114] A dual-channel evaluation module is set at the output end of the classifier; the first channel of the dual-channel evaluation module calculates the real-time probability distribution of each fault mode through the Softmax function; the second channel of the dual-channel evaluation module calculates the deviation between the current state and the health benchmark library based on the Mahalanobis distance of the multidimensional feature vector as the health indicator.

[0115] In some embodiments, the specific process of triggering a graded warning signal when an indicator exceeds a dynamic threshold is as follows:

[0116] A dynamic threshold update algorithm is established to determine the initial threshold based on the 3σ criterion of historical health data of the same wind farm, and an exponential decay factor of ambient temperature and humidity parameters is introduced to perform online correction of the threshold.

[0117] When the health index exceeds the dynamic threshold for three consecutive samples, the dominant fault type is determined based on the maximum value of the fault probability distribution, and a graded signal including warning level, fault location and maintenance suggestions is generated;

[0118] The graded warning signal is transmitted to the edge gateway after compressed sensing encoding, synchronously triggering the on-site sound and light alarm and the push of work orders on the remote operation and maintenance platform.

[0119] Based on the same inventive concept, corresponding to any of the above embodiments, refer to Figure 2 The present invention provides a wind turbine fault monitoring system for implementing the aforementioned wind turbine fault monitoring method, comprising:

[0120] The multi-source synchronous acquisition module is used to collect gearbox vibration signals, bearing temperature series, blade strain waveforms, and ambient temperature and humidity data through a synchronous triggering mechanism of multi-source heterogeneous sensors, and perform time-scale alignment and combined noise reduction preprocessing on the data of each channel;

[0121] Adaptive time-frequency processing module, used to dynamically adjust the wavelet packet decomposition parameters according to the signal frequency band characteristics, generate a three-dimensional energy distribution matrix containing the physical quantity type, frequency band level and time window energy evolution, and couple it with the environmental parameters to form a time-frequency energy fusion matrix;

[0122] The multi-scale decomposition module is used to perform an improved variational modal decomposition of the time-frequency energy fusion matrix based on a mutual information entropy weighting strategy, and to screen out the fault-related intrinsic modal component set through joint kurtosis-envelope spectrum analysis and twin network matching;

[0123] The spatiotemporal feature extraction module is used to construct a spatial branch containing a deformable residual block and a temporal branch with a sparse gated recurrent unit to extract non-uniform fault features and long-term degradation trends, respectively.

[0124] A dynamic fusion decision module is used to calculate the spatiotemporal feature coupling coefficient through a learnable gating weight matrix, realize the adaptive weighted fusion of the dual-stream network output, and generate a multi-dimensional feature vector containing multi-scale statistical features;

[0125] The edge warning execution module is used to input the fusion features into the lightweight fault classifier for real-time probability prediction, dynamically adjust the health assessment threshold according to the sliding window mechanism, and trigger the transmission of graded warning signals to the edge execution terminal.

[0126] Based on the same inventive concept, corresponding to any of the above embodiments, the present invention provides an electronic device, including a memory and a processor, the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the wind turbine fault monitoring method of the embodiment.

[0127] Optionally, the above-mentioned electronic device may be a server.

[0128] In addition, this embodiment further provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the wind turbine fault monitoring method of the embodiment is implemented.

[0129] It is understood that the processor in the embodiments of the present invention may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field programmable gate arrays (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. The general-purpose processor may be a microprocessor or any conventional processor.

[0130] The method steps in the embodiments of the present invention can be implemented by hardware or by a processor executing software instructions. The software instructions can be composed of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, mobile hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be an integral part of the processor. The processor and the storage medium can be located in an ASIC.

[0131] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions according to the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a storage medium or transmitted via a storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

Claims

1. A wind turbine fault monitoring method, characterized in that: The following steps are involved: Synchronously collect gearbox vibration signals, bearing temperature series, blade strain waveforms, and ambient temperature and humidity data, and generate a time-frequency energy fusion matrix through adaptive wavelet packet transform; The improved variational mode decomposition weighted by mutual information entropy is used to perform multi-scale decomposition on the time-frequency energy fusion matrix and screen out the intrinsic mode component set related to the fault mode. Based on the intrinsic mode component set, a spatiotemporal two-stream residual network is constructed. The two-branch output of the spatiotemporal two-stream residual network is fused through a dynamic feature gating mechanism to obtain a multi-dimensional feature vector. The spatial branch of the spatiotemporal two-stream residual network uses a deformable convolution kernel to extract non-uniformly distributed fault features, while the temporal branch of the spatiotemporal two-stream residual network uses a sparse gated recurrent unit to capture long-term degradation trends. The fused multidimensional feature vector is input into a lightweight fault classifier, which outputs real-time fault probability and component health assessment indicators based on a sliding window mechanism. When the indicators exceed the dynamic threshold, a graded warning signal is triggered.

2. The wind turbine fault monitoring method according to claim 1, wherein: The specific process of synchronously collecting the gearbox vibration signal, bearing temperature sequence, blade strain waveform and ambient temperature and humidity data and generating the time-frequency energy fusion matrix through adaptive wavelet packet transform is as follows: Through the synchronous triggering mechanism of multi-source heterogeneous sensors, the vibration signal, temperature series, strain waveform and environmental parameters are time-scale aligned, and the empirical wavelet denoising and sliding mean filtering are combined to preprocess the data of each channel; According to the energy concentration of the vibration signal frequency band and the kurtosis of the strain waveform spectrum, the number of wavelet packet decomposition layers and the type of basis functions are dynamically adjusted, and the pre-processed multi-source signals are adaptively segmented in frequency bands. The relative energy proportion of each sub-band is extracted and normalized, and then sorted by the energy entropy value of the frequency band to form a three-dimensional energy distribution matrix, where the first dimension corresponds to the type of physical quantity, the second dimension represents the frequency band division level, and the third dimension records the energy evolution characteristics within the time window; The environmental temperature and humidity parameters are integrated as energy attenuation factors, and the three-dimensional energy distribution matrix is coupled with the attenuation factors through tensor contraction operation to generate a time-frequency energy fusion matrix with environmental adaptability.

3. The wind turbine fault monitoring method according to claim 1, wherein: The specific process of using the improved variational mode decomposition weighted by mutual information entropy to perform multi-scale decomposition on the time-frequency energy fusion matrix and screen out the intrinsic mode component set related to the fault mode is as follows: Based on the energy entropy distribution of each sub-band of the time-frequency energy fusion matrix, the number of modes and penalty factors of variational mode decomposition are dynamically initialized; Expand the matrix into two-dimensional time series slices along the frequency band dimension, calculate the mutual information entropy value between each slice and the historical fault mode library, and construct the frequency band sensitivity weight vector; A frequency band weight constraint term is introduced into the objective function of variational mode decomposition, and the optimal eigenmode component is iteratively solved using the alternating direction multiplier method. The convergence condition of the modal center frequency is related to the energy attenuation factor of the corresponding frequency band. Perform a joint kurtosis-envelope spectrum analysis on the decomposed intrinsic mode component set to screen out components that simultaneously meet the kurtosis threshold greater than 3.5 and the envelope spectrum peak frequency matches the fault characteristic frequency library; The filtered components are converted into time-frequency image feature blocks through the Gram angle field and input into the pre-trained twin network for fault mode similarity matching. The intrinsic mode components with similarity higher than the set threshold are retained to form the intrinsic mode component set.

4. The wind turbine fault monitoring method according to claim 3, wherein: The specific construction process of the pre-trained twin network is as follows: A time-frequency image training set containing historical fault samples and their corresponding fault type labels is constructed. Random scaling and salt-and-pepper noise are applied to each sample to enhance data diversity. Build a feature extraction subnetwork consisting of parallel depthwise separable convolutional layers and channel attention modules. Each subnetwork contains three cascaded convolutional blocks, each of which consists of deformable convolution, batch normalization, and LeakyReLU activation functions. A dual-path symmetrical structure is used to build the twin network backbone. The outputs of the two feature extraction sub-networks are globally averaged and pooled before being input into the similarity measurement layer. The Mahalanobis distance between the two is calculated as the fault mode difference indicator. The network parameters are optimized based on the triplet loss function. Anchor samples, positive samples and negative samples are collected from the training set to form training triplets. The distance between the anchor point and the positive sample is constrained to be less than the distance between the anchor point and the negative sample by at least one marginal coefficient. By introducing prior knowledge of fault feature frequencies and adding a frequency-sensitive constraint term to the last convolutional layer of the feature extraction subnetwork, the network weight update direction is positively correlated with the energy distribution of the fault feature frequency band. The training cycle is controlled by the sliding validation set early stopping strategy. When the F1 score on the validation set does not improve for three consecutive training cycles, the training is terminated and the optimal weight parameters are saved for online similarity matching.

5. The wind turbine fault monitoring method according to claim 1, wherein: The specific process of constructing a spatiotemporal dual-stream residual network based on the intrinsic mode component set and fusing the two-branch outputs of the spatiotemporal dual-stream residual network through a dynamic feature gating mechanism to obtain a multidimensional feature vector is as follows: The eigenmode component set is divided into overlapping segments according to the time dimension and input into the spatial branch and the time branch for feature extraction respectively; The spatial branch consists of multiple cascaded deformable residual blocks, each of which contains a deformable convolution layer, a batch normalization layer, and an adaptive threshold ReLU activation function. The offset of the deformable convolution kernel is dynamically generated by the feature map of the previous layer through a lightweight fully connected layer to capture non-uniformly distributed local fault modes. The time branch uses stacked sparse gated recurrent units, compresses the unit's internal state matrix through weight pruning and quantization operations, and embeds a temporal attention mechanism in the hidden state update equation to enhance the memory retention of key time nodes in long-term degradation trends. During the feature fusion phase of the spatiotemporal dual-stream network, a dynamic feature gating mechanism is established: the high-dimensional feature map output by the spatial branch is channel-compressed and then tensor-concatenated with the hidden state vector of the temporal branch. The spatial-temporal feature coupling coefficient is calculated using a learnable gating weight matrix, and the sigmoid function is used to generate adaptive fusion weights. The weighted fusion features are subjected to multi-scale pooling operations to extract statistical features of different granularities. Finally, after dimensionality reduction through the fully connected layer, a multi-dimensional feature vector containing spatiotemporal correlation information is output.

6. The wind turbine fault monitoring method according to claim 1, wherein: The specific process of inputting the fused multi-dimensional feature vector into the lightweight fault classifier and outputting the real-time fault probability and component health evaluation index based on the sliding window mechanism is as follows: Construct a hybrid classifier that is an integration of depthwise separable convolutional layers and random forests. The convolutional layers are used to extract local correlations of feature vectors, while the random forest classifier achieves dimensionality compression through feature importance sampling. A sliding window mechanism is used to perform time-series sliding interception of multi-dimensional feature vectors, and the window length matches the typical fault evolution cycle of wind turbines. A dual-channel evaluation module is set at the output end of the classifier; the first channel of the dual-channel evaluation module calculates the real-time probability distribution of each fault mode through the Softmax function; the second channel of the dual-channel evaluation module calculates the deviation between the current state and the health benchmark library based on the Mahalanobis distance of the multidimensional feature vector as the health indicator.

7. The wind turbine fault monitoring method according to claim 6, wherein: The specific process of triggering the graded warning signal when the indicator exceeds the dynamic threshold is as follows: A dynamic threshold update algorithm is established to determine the initial threshold based on the 3σ criterion of historical health data of the same wind farm, and an exponential decay factor of ambient temperature and humidity parameters is introduced to perform online correction of the threshold. When the health index exceeds the dynamic threshold for three consecutive samples, the dominant fault type is determined based on the maximum value of the fault probability distribution, and a graded signal including warning level, fault location and maintenance suggestions is generated; The graded warning signal is transmitted to the edge gateway after compressed sensing encoding, synchronously triggering the on-site sound and light alarm and the push of work orders on the remote operation and maintenance platform.

8. A wind turbine fault monitoring system, used to implement the wind turbine fault monitoring method according to any one of claims 1 to 7, characterized in that: include: The multi-source synchronous acquisition module is used to collect gearbox vibration signals, bearing temperature series, blade strain waveforms, and ambient temperature and humidity data through a synchronous triggering mechanism of multi-source heterogeneous sensors, and perform time-scale alignment and combined noise reduction preprocessing on the data of each channel; Adaptive time-frequency processing module, used to dynamically adjust the wavelet packet decomposition parameters according to the signal frequency band characteristics, generate a three-dimensional energy distribution matrix containing the physical quantity type, frequency band level and time window energy evolution, and couple it with the environmental parameters to form a time-frequency energy fusion matrix; The multi-scale decomposition module is used to perform an improved variational modal decomposition of the time-frequency energy fusion matrix based on a mutual information entropy weighting strategy, and to screen out the fault-related intrinsic modal component set through joint kurtosis-envelope spectrum analysis and twin network matching; The spatiotemporal feature extraction module is used to construct a spatial branch containing a deformable residual block and a temporal branch with a sparse gated recurrent unit to extract non-uniform fault features and long-term degradation trends, respectively. A dynamic fusion decision module is used to calculate the spatiotemporal feature coupling coefficient through a learnable gating weight matrix, realize the adaptive weighted fusion of the dual-stream network output, and generate a multi-dimensional feature vector containing multi-scale statistical features; The edge warning execution module is used to input the fusion features into the lightweight fault classifier for real-time probability prediction, dynamically adjust the health assessment threshold according to the sliding window mechanism, and trigger the transmission of graded warning signals to the edge execution terminal.

9. An electronic device, characterized in that: The electronic device comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the wind turbine fault monitoring method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the wind turbine fault monitoring method according to any one of claims 1 to 7 is implemented.

Images