Fault diagnosis method and system for wind turbine based on spatio-temporal coupling feature extraction

Abstract

The application discloses a wind turbine fault diagnosis method and system based on space-time coupling feature extraction, and belongs to the technical field of computer information processing. The method comprises the following steps: acquiring multi-source operation monitoring data of a wind turbine, wherein the multi-source operation monitoring data comprises various sample data; performing graph structure conversion processing on the sample data to obtain graph structure data; performing space feature extraction on the graph structure data to obtain space features; performing time feature extraction on sample data input to obtain time sequence features; splicing and fusing the space features and the time sequence features to obtain a fusion processing result; and obtaining a wind turbine fault diagnosis result according to the fusion processing result. The scheme can improve the wind turbine fault diagnosis precision.

Description

Technical Field

[0001] This invention relates to the field of computer information processing technology, and in particular to a method and system for diagnosing wind turbine faults based on spatiotemporal coupling feature extraction. Background Technology

[0002] In recent years, due to the vigorous development of clean energy such as wind power, the cumulative installed capacity of wind turbines has been continuously increasing. However, in complex natural environments, wind turbines are exposed to harsh conditions such as high humidity, high salt spray, high temperature, low temperature, and strong ultraviolet radiation for extended periods. Coupled with high-intensity workloads, this leads to various malfunctions in wind turbines, severely affecting their power generation efficiency and operational stability. Some older units, due to long operating times, harsh working environments, and insufficient accumulation of technical and management experience, have experienced a gradual decline in their operating capacity year by year. This not only further reduces the available life of the turbines but also causes significant power generation losses.

[0003] Wind turbine fault diagnosis technology can promptly and effectively identify potential safety hazards in the turbines, enabling early fault management and thus improving the reliability and economic efficiency of wind farm operations. Accurate fault diagnosis and classification allow for the rational development of turbine maintenance strategies, reducing operation and maintenance costs. By monitoring the operating status of wind turbines in real time, potential faults can be detected and addressed promptly, preventing further equipment damage and reducing maintenance costs and downtime. With the continuous expansion of wind farm scale and the accumulation of operational data, deep learning-based fault diagnosis methods have received widespread attention. The current design process for deep learning-based methods involves constructing corresponding deep learning identification and classification models using previously accumulated wind turbine operating data, inputting operational data into the models for training, and then deploying the trained models locally.

[0004] However, most current algorithms do not take into account the spatial characteristics of wind turbine data or do not extract sufficient spatial features, which can easily lead to poor identification and classification accuracy. Summary of the Invention

[0005] In order to fully extract the spatiotemporal features of wind turbine operation data and improve the accuracy of fault diagnosis and classification, this invention provides a wind turbine fault diagnosis method and system based on spatiotemporal coupling feature extraction.

[0006] A wind turbine fault diagnosis method based on spatiotemporal coupling feature extraction includes the following steps: Acquire multi-source operation monitoring data of wind turbine units, wherein the multi-source operation monitoring data includes various sample data; The sample data is subjected to graph structure transformation to obtain graph structure data; Spatial features are extracted from the graph structure data to obtain spatial features; Temporal features are extracted from the input sample data to obtain time-series features; The spatial features and the temporal features are concatenated and fused to obtain the fusion processing result; Based on the fusion processing results, the fault diagnosis results of the wind turbine are obtained.

[0007] Optionally, the sample data is subjected to graph structure transformation processing to obtain graph structure data, including: Determine an improved distance metric between each sample data point and other sample data points within the sample data; Based on the improved distance metric between sample data, the K nearest sample data nodes to the target data sample nodes are selected as their neighbor nodes. Each sample data point is then connected to its K neighbor nodes to form edges, resulting in graph structure data.

[0008] Optionally, spatial features are extracted from the graph structure data to obtain spatial features, including: The graph structure data is input into the first fully connected layer of a preset network for processing to extract shallow information of the original features; the shallow information of the original features is then subjected to k convolutions to obtain the feature data after k convolutions. Based on the feature data after k convolutions, a weighted summation is used to extract features from neighboring nodes to obtain spatial features.

[0009] Optionally, time features are extracted from the input sample data to obtain time-series features, including: The input feature X of the sample data is transformed from the time domain to the frequency domain, and the amplitude of each component of the input feature X is calculated. Select the top-ranked number of significant frequency components by amplitude and calculate the corresponding period length; Based on the selected significant frequency components and their corresponding period lengths, the input feature X is reshaped into a two-dimensional matrix; Adaptive aggregation of the two-dimensional matrix yields temporal features.

[0010] Optionally, the spatial features and the temporal features are concatenated and fused to obtain a fusion processing result, including: according to The spatial feature output Y1 and the temporal feature output Y2 are concatenated at a feature layer to obtain the fused feature. .

[0011] Optionally, based on the fusion processing results, wind turbine fault diagnosis results are obtained, including: according to Determine the probability distribution of fault categories, where, This is the weight matrix. The bias is used, and softmax(·) is the activation function; The fault category with the highest probability distribution is selected as the final fault diagnosis result for the wind turbine.

[0012] Optionally, after obtaining the wind turbine fault diagnosis result based on the fusion processing result, the method further includes: The fault diagnosis results of the wind turbine are evaluated based on preset fault evaluation indicators to obtain the evaluation results. The preset fault evaluation indicators include: F1 score and recall rate R1. ; ; Wherein, TP represents the number of samples that are actually faults but are correctly identified as faults; FP represents the number of samples that are not actually faults but are incorrectly identified as faults; and FN represents the number of samples that are actually faults but are incorrectly identified as non-faults.

[0013] Embodiments of the present invention also provide a wind turbine fault diagnosis system based on spatiotemporal coupling feature extraction, comprising: The acquisition module is used to acquire multi-source operation monitoring data of wind turbine units, which includes various sample data. The processing module is used to perform graph structure transformation processing on the sample data to obtain graph structure data; extract spatial features from the graph structure data to obtain spatial features; extract temporal features from the sample data input to obtain temporal features; concatenate and fuse the spatial features and the temporal features to obtain a fusion processing result; and obtain a wind turbine fault diagnosis result based on the fusion processing result.

[0014] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above.

[0015] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above.

[0016] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: The wind turbine fault diagnosis method and system based on spatiotemporal coupling feature extraction described in this invention simultaneously captures the spatial coupling features and temporal evolution features of multi-source operation monitoring data of wind turbines. Compared with traditional diagnostic models that only use a single temporal network or a single graph convolutional network, it significantly improves the ability to identify and distinguish coupled faults of wind turbines. Through parallel extraction and fusion of spatiotemporal dual branches, it effectively overcomes the limitation of traditional methods that only focus on single-dimensional features, and significantly improves the accuracy of fault diagnosis. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the wind turbine fault diagnosis method based on spatiotemporal coupling feature extraction according to an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of the residual connection structure in an embodiment of the present invention.

[0019] Figure 3 This is a structural diagram of Timesblock in an embodiment of the present invention.

[0020] Figure 4 This is a block diagram of the modules of the wind turbine fault diagnosis system based on spatiotemporal coupling feature extraction in this embodiment of the invention. Detailed Implementation

[0021] The present invention will be further described in detail below with reference to all the accompanying drawings.

[0022] like Figure 1 As shown in the figure, this invention discloses a wind turbine fault diagnosis method based on spatiotemporal coupling feature extraction, including: S101, acquire multi-source operation monitoring data of the wind turbine, the multi-source operation monitoring data includes various sample data, the sample data includes one-dimensional sequence data such as generator rotor and gearbox speed, bearing and other key component temperature, blade pitch angle, output power, etc.

[0023] S102, Perform graph structure transformation processing on the sample data to obtain graph structure data; S103, Spatial features are extracted from the graph structure data to obtain spatial features; S104, Based on the changing trends of various sample data within a preset time period, time features are extracted to obtain time-series features; S105, the spatial features and the temporal features are spliced ​​and fused to obtain the fusion processing result; S106, Based on the fusion processing results, diagnose the wind turbine faults and determine the wind turbine fault diagnosis results.

[0024] In step S101 of this embodiment, the sample data is one-dimensional sequence data. In specific implementation, it is first necessary to obtain multi-source operation monitoring data from the wind turbine system. Taking a doubly-fed induction generator (DFIG) in a wind farm as an example, monitoring data is collected within a preset time period, with a data collection interval t, for example, 10 minutes. The acquired sample data may include generator rotor and gearbox speeds, temperatures of key components such as bearings, blade pitch angles, output power, etc., and these sample data serve as the raw input for subsequent processing.

[0025] Multi-source operational monitoring data is processed to obtain input features X. Input features X can be a data tensor of the multi-source operational monitoring data, which refers to a three-dimensional monitoring data array arranged by batch-time-physical parameter. First, a spatiotemporal input module is constructed, combining pre-processed multi-source operational monitoring data from T consecutive sampling times into a two-dimensional data matrix in chronological order (rows correspond to sampling times, and columns correspond to different physical parameters such as generator rotor and gearbox speeds, bearing and other key component temperatures, blade pitch angles, and output power). Then, to meet the requirements of parallel network processing, multiple two-dimensional matrices are stacked along the batch dimension to form a three-dimensional structured data block. Finally, a Fast Fourier Transform is performed along the time dimension (row direction) of this data block to convert the time-domain waveform values ​​of each physical parameter into frequency-domain spectrum values, thereby extracting the main fluctuation periods implicit in the historical monitoring data.

[0026] In S102, after acquiring multi-source operation monitoring data, the sample data is constructed into graph structure data, so that the one-dimensional sequence data that originally did not have topological relationships can obtain a graph structure representation of nodes and edges, which is used to extract spatial features.

[0027] This invention proposes the above-mentioned technical solution. The operating data of wind turbine generators is essentially multiple one-dimensional, multi-source time-series data, containing both spatial correlations between different monitoring parameters (such as the inherent coupling between speed and temperature) and the variation patterns of the same parameter over time. In embodiments of this invention, one-dimensional sequence data is transformed into graph-structured data, allowing explicit modeling of the relationships between different monitoring points. Based on this, graph convolution operations are used to transfer and aggregate neighborhood information along the graph structure, extracting spatial coupling features between data. Fast Fourier Transform is used to identify multi-period patterns in the data, and the one-dimensional time-series data is folded into a two-dimensional tensor before convolution operations are used to extract the variation patterns over time. The two branches are extracted in parallel and then spliced ​​and fused, achieving comprehensive capture of the spatiotemporal coupling features of wind turbine generator operating data.

[0028] In an optional embodiment of the present invention, in S102, performing graph structure transformation processing on the sample data to obtain graph structure data may include: S1021, Determine the improved distance metric between each sample data point and other sample data points in the sample data; S1022, Based on the improved distance metric between sample data, select the K nearest sample data nodes to the target data sample nodes as their neighbor nodes, and connect each sample data point to its K neighbor nodes to form edges, thus obtaining graph structure data.

[0029] In S1021 of this embodiment, for each sample data, an improved distance metric is calculated between each sample data and other sample data. The formula for calculating the improved distance metric is as follows: ; In the formula, x iK It is sample data x i The specific value of x under the Kth sample data jK It is sample x j The specific value for the Kth sample data, s iK Let be the standard deviation of the specific value of the Kth sample data; by calculating the weighted distance between the sample data, the adjacency relationship of each sample data is determined.

[0030] In S1022, after calculating the improved distance metric between all sample data pairs, for each sample data point, the K nearest samples are selected as its neighbor nodes. Then, each sample data point is connected to its K neighbor nodes to form edges, ultimately constructing a complete graph structure. In this graph structure, each node corresponds to a sample data point, and edges represent the connection relationships between nodes.

[0031] In this embodiment, the dimensions and numerical ranges of the various samples in the wind turbine monitoring data differ significantly. For example, the magnitudes of temperature data and speed data may differ by tens of times. This solution introduces the standard deviation of the specific values ​​of each sample data as a normalization factor to standardize the differences in each sample data, making different sample data comparable in distance measurement. This improved distance measurement can more reasonably reflect the true similarity between sample data, thereby constructing a data structure graph topology that better reflects the inherent structure of the data.

[0032] In an optional embodiment of the present invention, in S103, spatial feature extraction of the graph structure data to obtain spatial features may include: S1031, The graph structure data is input into the first fully connected layer of the preset network for processing to extract shallow information of the original features; S1032, perform k convolutions on the shallow information of the original features to obtain the feature data after k convolutions; S1033: Based on the feature data after k convolutions, use weighted summation to extract features from neighboring nodes to obtain spatial features.

[0033] like Figure 2As shown, G represents a graph convolutional layer, A represents an activation function, T represents batch normalization, In represents the input feature, and 1 indicates that the input feature first enters the first processing unit (graph convolutional layer). The residual graph convolutional network extracts spatial features from the graph structure data to obtain spatial feature output. The residual structure is embedded into each layer of the pre-defined network KNN-ResGCN to construct a deep and scalable graph neural network architecture, ensuring the stability of deep network training while extracting spatial features from multi-source operational monitoring data. Figure 2 The CCP encompasses two paths: a nonlinear transformation path and an identity mapping path. The nonlinear transformation path serves as the main path. The input feature X first undergoes feature extraction and spatial transformation through a graph convolutional layer or a regular convolutional layer, followed by a batch normalization layer and a ReLU activation function to obtain the nonlinearly transformed output. This path is responsible for learning the residual mapping between the input and output, i.e., fitting the difference between the objective function and the input.

[0034] The linear identity mapping path is a jumper wire, which transmits the input feature X to the output through a linear identity mapping (i.e., direct jump connection, without any parameter learning), ensuring lossless transmission of the original information. The outputs of the two paths are added element by element to obtain the final output of the residual block.

[0035] In step S1031 of this embodiment, the graph structure data is processed by a first fully connected layer to extract shallow information of the original features. The calculation formula for the first fully connected layer is: In the formula, W (0) b is the weight matrix of the first fully connected layer; (0) This is the bias matrix of the first fully connected layer; For the ReLU (·) activation function, x i (0) Let x be the initial feature vector of the multidimensional node corresponding to the i-th sample data. This initial feature vector is composed of all monitored parameters, including the generator rotor and gearbox speed, the temperature of key components such as bearings, the blade pitch angle, and the output power. i (1) This is the output of the fully connected layer, which contains shallow information about the original features.

[0036] In S1032, the output of the first fully connected layer is input to the residual graph convolutional network to extract spatial features from the graph structure data. The general mathematical prototype calculation formula is as follows: In the formula: w i,j x represents the weight of the edge; i (k) The depth space feature vector is the output after k convolutions of the i-th sample data point. To find the maximum value of a differentiable function, , It is a multilayer perceptron.

[0037] Based on the improved distance metric, each edge (i,j) in the graph structure is assigned an edge weight w. i,j The calculation formula is: Nodes that are closer in distance have larger edge weights, reflecting a high degree of similarity in their operational states. This physically weighted graph structure data serves as subsequent input. During graph convolution, the node features between sample data are weighted according to their respective weights w. i,j By aggregating data, sample data with similar fault characteristics in the original space maintain a tight coupling relationship in the hidden space of the graph neural network, thereby achieving accurate spatial feature extraction.

[0038] After the multi-source operation monitoring data of wind turbine units is constructed into a graph structure, the feature vector of each node contains multi-dimensional information of that monitoring point, while the edges encode the similarity relationships between different monitoring points. Graph convolutional networks, through convolution operations on the graph, enable the features of each node to transmit information to its neighboring nodes along the edges, thereby extracting spatial correlation features between nodes. This scheme first sets a fully connected layer before entering graph convolution to map the original input features to a higher-dimensional feature space, extracting shallow information and providing richer feature representations for subsequent graph convolution. During graph convolution, the intensity of information transmission is controlled by the edge weights, and a multilayer perceptron is used to perform nonlinear transformations on the node features, enabling the network to learn complex spatial coupling patterns.

[0039] In S1033, spatial correlation features of the graph structure data are extracted layer by layer. Weighted summation is used to extract features from neighboring nodes. The specific calculation formula for weighted summation and normalization is as follows: In the formula, is the weight matrix; deg(i) and deg(j) are the actual number of neighbors (i.e., node degree) of node i and node j in the constructed graph structure data, respectively. A residual mechanism is added to each graph convolutional layer. The residual connection directly maps the input to the output layer, and finally, spatial features are extracted through the residual graph convolutional network. The enhanced calculation formula with the addition of residual connections is as follows: In the formula: ɑ l The residual decay factor is the raw data characteristic of multi-source operation monitoring data. I is the decay factor of the weight matrix; n It is an identity matrix.

[0040] Deep graph convolutional networks (GCNNs) are prone to gradient vanishing or network degradation after stacking multiple layers, meaning that training accuracy decreases as the number of layers increases. Residual connections, by directly mapping the input of the previous layer to the output layer and adding it to the transformation result of the current layer, allow gradients to be directly backpropagated via a shortcut path, effectively mitigating the gradient vanishing problem in deep network training. In weighted summation graph convolutions, introducing a node degree normalization factor prevents nodes with high degrees from occupying excessive weights in information aggregation, resulting in more balanced feature extraction. Residual decay factors and weight matrix decay factors provide the ability to adjust the contribution of residuals and the magnitude of weight updates, making network training more stable.

[0041] In an optional embodiment of the present invention, in S104, extracting time features from the input sample data to obtain time-series features may include: S1041. Perform a Fast Fourier Transform on the input feature X of the sample data to convert it from the time domain to the frequency domain, and calculate the amplitude of each component of the input feature X. S1042. Select the top-preset number of significant frequency components in terms of amplitude and calculate the corresponding period length. S1043. Based on the selected significant frequency components and their corresponding period lengths, the input feature X is reshaped into a two-dimensional matrix; S1044. Perform adaptive aggregation on the two-dimensional matrix to obtain the temporal features.

[0042] In this embodiment, the aforementioned input features of the sample data are subjected to temporal feature extraction to obtain a temporal feature output. Time blocks are stacked in a residual manner. The time blocks capture various temporal change trends from tensors of different shapes and fuse them together based on normalized amplitudes.

[0043] In step S1041 of this embodiment, the input feature X is subjected to a Fast Fourier Transform to convert the data from the time domain to the frequency domain, and the amplitude of each component of the sample data is calculated using the following formula: In the formula, Amp(·) is the amplitude calculation, Avg(·) is the average value taken on the channel, and FFT(·) is the fast Fourier transform.

[0044] First, the continuous Multi-source operation monitoring data at each sampling time are processed using the formula Build as Two-dimensional time domain matrix : ,in Indicates the first Class of monitoring parameters in continuous The time-series numerical vectors within each time step. Then, for the matrix... Each column Perform a Fast Fourier Transform to extract the frequency domain fluctuation characteristics of each physical parameter, that is... Convert to frequency domain This addresses the problem that traditional time series models struggle to effectively capture complex "multi-period" patterns.

[0045] In S1042, after the Fourier transform is completed, all frequencies are sorted in descending order according to their amplitude, retaining only the n most significant frequency components with the largest amplitudes. For these n selected frequencies, on the one hand, their corresponding true amplitudes are extracted (without normalization) to characterize the energy intensity of each major periodic fluctuation; on the other hand, according to the principles of digital signal processing, each selected frequency f is... i Convert to the corresponding actual period length This period length will be directly used in the subsequent operation of folding one-dimensional time series data into a two-dimensional spatial matrix. The calculation formula is as follows: In the formula, p i Let f be the period length corresponding to the i-th significant frequency component, T be the sequence length, and f be the period length. i These are significant frequency components.

[0046] In S1043, a multi-path parallel folding strategy is adopted for the selected n significant frequency components. For the i-th significant frequency component and its corresponding period length pi, the original input sequence is folded... ( For sequence length, (As a feature dimension) is independently folded along the time axis according to the period pi, and reshaped into a two-dimensional spatial matrix. This operation is performed in parallel on all n significant frequency components, thereby generating n two-dimensional feature maps that retain different periodic evolution patterns. This provides diverse input views for subsequent two-dimensional convolutional neural networks to extract multi-scale temporal texture features. The calculation formula is as follows: In the formula, This indicates that the rows are rearranged according to the period length. This indicates that trailing zeros are padded to make the data neat. Specifically, for each significant period pi, a sequence folding transformation is performed: from the original time series data... Starting from the first sampling point, data segments are extracted with a fixed step size of pi. Each data segment is arranged as a row from top to bottom, forming a two-dimensional spatial matrix. If Total length If the value is not divisible by pi, then zero values ​​are padded to the end of the last incomplete data segment, with the number of zero values ​​padded being [number missing]. This padding operation ensures that the last row has the same length as the other rows. This padding is solely to meet the matrix's regularity requirements and does not alter the original, valid monitoring data values.

[0047] In S1044, the two-dimensional matrix is ​​restored to a one-dimensional sequence after undergoing the Inception dilation and convolution operation. The calculation formula is as follows: In the formula, This represents the i-th periodic pattern, and Inception(·) represents the dilated convolution operation. Taking the first significant frequency component as an example, the Inception convolution and recovery operation are illustrated. In actual execution, the i-th periodic pattern is... {1,2...K top Each periodic pattern undergoes the same processing to obtain its feature vector. , as input for adaptive aggregation.

[0048] After applying the softmax activation function and adaptive aggregation, the temporal feature extraction result is obtained, which is the temporal feature output Y2. The calculation formula is as follows: In the formula, Let be the unnormalized true amplitude energy corresponding to the i-th significant frequency component fi. This amplitude is directly calculated by Fast Fourier Transform, and its magnitude objectively reflects the energy intensity of this periodic fluctuation in the original monitoring signal. It is used as an adaptive fusion weight, along with the two-dimensional spatial features extracted from the corresponding period. Multiplying and then summing results in stronger and more significant periodic features appearing in the final time series characteristics. The larger the proportion of data in the process, the more data-driven multi-period feature adaptive aggregation can be achieved.

[0049] like Figure 3 As shown, the Inception module structure used in the TimesNet branch of this invention is used to extract multi-scale spatial features from the two-dimensional feature matrix generated after periodic folding. First, the two-dimensional spatial matrix obtained after folding... This serves as the input to the module. The rows of this two-dimensional matrix represent the macroscopic evolution between different periods, while the columns represent the microscopic fluctuations within a single period.

[0050] The Inception module contains multiple parallel convolutional branches, each employing different kernel sizes (e.g., 1×1, 3×3, 5×5) to perform convolution operations on the input feature maps: the 1×1 branch integrates cross-channel features, reducing channel dimensionality and computational cost; the 3×3 branch extracts mid-scale spatial texture features within the local neighborhood; and the 5×5 branch extracts long-range spatial dependency features over a larger receptive field. Each branch's convolutional operation is followed by a batch normalization layer and a ReLU activation function to accelerate convergence and enhance non-linear expressiveness.

[0051] Spatial features extracted from each branch at different scales are concatenated along the channel dimension to achieve multi-scale information fusion. This fusion operation enables the network to simultaneously capture fluctuation patterns at different cycle lengths in wind turbine monitoring data, taking into account both short-term abrupt changes and long-term trends. The fused multi-scale feature map is then subjected to global average pooling for dimensionality reduction and restored to a temporal feature vector through a reshape operation for subsequent adaptive aggregation.

[0052] In an optional embodiment of the present invention, step S105 involves concatenating and fusing the spatial features and the temporal features to obtain a fusion processing result, including: S1051, according to The spatial feature output Y1 and the temporal feature output Y2 are concatenated at a feature layer to obtain the fused feature. .

[0053] In an optional embodiment of the present invention, in S106, obtaining the wind turbine fault diagnosis result based on the fusion processing result may include: S1061, according to Determine the probability distribution of fault categories, where, This is the weight matrix. The bias is used, and softmax(·) is the activation function; S1062, take the fault category with the highest probability distribution as the final fault diagnosis result for the wind turbine.

[0054] In S1061 of this embodiment, the spliced ​​sequence result is processed by a linear layer and a softmax activation function to output the probability distribution of each fault category. The calculation formula is as follows: In the formula, softmax(·) is the activation function. Feature fusion Through a fully connected linear layer (weight matrix) Bias Mapping to the fault category space, outputting the unnormalized score for each category.

[0055] In S1062, based on the probability distribution of the determined fault categories, the fault category with the highest probability distribution is selected as the final fault diagnosis result for the wind turbine. For example, there are 4 fault categories: [normal, gearbox fault, generator bearing fault, transformer fault]; if classification=[0.02,0.91,0.05,0.02], then the highest probability of 0.91 corresponds to a gearbox fault, so the current fault category is determined to be a gearbox fault.

[0056] In an optional embodiment of the present invention, after obtaining the wind turbine fault diagnosis result based on the fusion processing result, the method may further include: S107. Evaluate the fault diagnosis results of the wind turbine unit according to the preset fault evaluation indicators, and obtain the evaluation results.

[0057] In this embodiment, the operating states such as normal operation, gearbox failure, generator bearing failure and transformer failure are selected as the research objects. The actual operating state of the wind turbine is obtained through manual on-site inspection and other methods, and the actual operating state is matched one by one with the corresponding wind turbine fault diagnosis results.

[0058] When the actual operating status is the same as the corresponding wind turbine fault diagnosis result, it is identified as a true positive example (TP); when the actual operating status is not a fault but the wind turbine fault diagnosis result identifies it as a fault, it is identified as a false positive example (FP); when the actual operating status is a fault but the wind turbine fault diagnosis result identifies it as a non-fault, it is identified as a false negative example (FP).

[0059] The F1 score and recall rate R1 are used as evaluation indicators for fault diagnosis accuracy, and the calculation formulas are as follows: ; ; In the formula, TP represents the number of samples that are actually faults and are correctly identified as faults; FP represents the number of samples that are not actually faults but are incorrectly identified as faults; and FN represents the number of samples that are actually faults but are incorrectly identified as non-faults. The F1 score, as the harmonic mean of precision and recall R1, can comprehensively evaluate the classification performance of the model across all categories and is suitable for class imbalance scenarios in fault diagnosis.

[0060] The implementation principle of a wind turbine fault diagnosis method based on spatiotemporal coupling feature extraction in this application is as follows: The sample data of a wind turbine is essentially one-dimensional sequence data such as the rotational speed of the generator rotor and gearbox, the temperature of key components like bearings, the blade pitch angle, and the output power. This one-dimensional sequence data is transformed into graph-structured data. Then, spatial features are extracted from the graph-structured data, and temporal features are extracted from the sample data. Finally, the spatial and temporal features are concatenated through a splicing and fusion method, thereby more comprehensively depicting the early signs of potential faults. This achieves accurate identification of fault modes and generates diagnostic results. The entire process, through the complementarity of spatiotemporal features, significantly improves the reliability and generalization ability of the diagnosis.

[0061] like Figure 4 As shown in the figure, this embodiment of the invention also discloses a wind turbine fault diagnosis system 400 based on spatiotemporal coupling feature extraction, comprising: The acquisition module 401 is used to acquire multi-source operation monitoring data of the wind turbine, wherein the multi-source operation monitoring data includes various sample data; The processing module 402 is used to perform graph structure transformation processing on the sample data to obtain graph structure data; extract spatial features from the graph structure data to obtain spatial features; extract temporal features from the sample data input to obtain temporal features; concatenate and fuse the spatial features and the temporal features to obtain a fusion processing result; and obtain a wind turbine fault diagnosis result based on the fusion processing result.

[0062] Optionally, the sample data is subjected to graph structure transformation processing to obtain graph structure data, including: Determine an improved distance metric between each sample data point and other sample data points within the sample data; Based on the improved distance metric between sample data, the K nearest sample data nodes to the target data sample nodes are selected as their neighbor nodes. Each sample data point is then connected to its K neighbor nodes to form edges, resulting in graph structure data.

[0063] Optionally, spatial features are extracted from the graph structure data to obtain spatial features, including: The graph structure data is input into the first fully connected layer of a preset network for processing to extract shallow information of the original features; the shallow information of the original features is then subjected to k convolutions to obtain the feature data after k convolutions. Based on the feature data after k convolutions, a weighted summation is used to extract features from neighboring nodes to obtain spatial features.

[0064] Optionally, time features are extracted from the input sample data to obtain time-series features, including: The input feature X of the sample data is transformed from the time domain to the frequency domain, and the amplitude of each component of the input feature X is calculated. Select the top-ranked number of significant frequency components by amplitude and calculate the corresponding period length; Based on the selected significant frequency components and their corresponding period lengths, the input feature X is reshaped into a two-dimensional matrix; Adaptive aggregation of the two-dimensional matrix yields temporal features.

[0065] Optionally, the spatial features and the temporal features are concatenated and fused to obtain a fusion processing result, including: according to The spatial feature output Y1 and the temporal feature output Y2 are concatenated at a feature layer to obtain the fused feature. .

[0066] Optionally, based on the fusion processing results, wind turbine fault diagnosis results are obtained, including: according to Determine the probability distribution of fault categories, where, This is the weight matrix. The bias is used, and softmax(·) is the activation function; The fault category with the highest probability distribution is selected as the final fault diagnosis result for the wind turbine.

[0067] Optionally, the processing module 402 is further configured to evaluate the fault diagnosis results of the wind turbine unit according to preset fault evaluation indicators, and obtain evaluation results. The preset fault evaluation indicators include F1 score and recall rate R1. ; ; Wherein, TP represents the number of samples that are actually faults but are correctly identified as faults; FP represents the number of samples that are not actually faults but are incorrectly identified as faults; and FN represents the number of samples that are actually faults but are incorrectly identified as non-faults.

[0068] It should be noted that this system is the same as the method described above. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0069] This invention also discloses a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0070] This invention also discloses a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and achieve the same technical effect. Computer-readable storage media include, for example, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0071] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art can still combine, add, delete, or otherwise adjust the features of the various embodiments of the present invention according to the circumstances without conflict or creative effort, thereby obtaining different technical solutions that do not fundamentally depart from the concept of the present invention. These technical solutions also fall within the scope of protection of the present invention.

Claims

1. A method for fault diagnosis of wind turbine units based on spatiotemporal coupling feature extraction, characterized in that, Includes the following steps: Acquire multi-source operation monitoring data of wind turbine units, wherein the multi-source operation monitoring data includes various sample data; The sample data is subjected to graph structure transformation processing to obtain graph structure data; Spatial features are extracted from the graph structure data to obtain spatial features; Temporal features are extracted from the input sample data to obtain time-series features; The spatial features and the temporal features are concatenated and fused to obtain the fusion processing result; Based on the fusion processing results, the fault diagnosis results of the wind turbine are obtained.

2. The wind turbine fault diagnosis method based on spatiotemporal coupling feature extraction according to claim 1, characterized in that, The sample data is subjected to graph structure transformation to obtain graph structure data, including: Determine an improved distance metric between each sample data point and other sample data points within the sample data; Based on the improved distance metric between sample data, the K nearest sample data nodes to the target data sample nodes are selected as their neighbor nodes. Each sample data point is then connected to its K neighbor nodes to form edges, resulting in graph structure data.

3. The wind turbine fault diagnosis method based on spatiotemporal coupling feature extraction according to claim 2, characterized in that, Spatial features are extracted from the graph structure data to obtain spatial features, including: The graph structure data is input into the first fully connected layer of a preset network for processing to extract shallow information of the original features; Perform k convolutions on the shallow information of the original features to obtain the feature data after k convolutions; Based on the feature data after k convolutions, a weighted summation is used to extract features from neighboring nodes to obtain spatial features.

4. The wind turbine fault diagnosis method based on spatiotemporal coupling feature extraction according to claim 1, characterized in that, Temporal features are extracted from the input sample data to obtain time-series features, including: The input feature X of the sample data is transformed from the time domain to the frequency domain, and the amplitude of each component of the input feature X is calculated. Select the top-ranked number of significant frequency components by amplitude and calculate the corresponding period length; Based on the selected significant frequency components and their corresponding period lengths, the input feature X is reshaped into a two-dimensional matrix; Adaptive aggregation of the two-dimensional matrix yields temporal features.

5. The wind turbine fault diagnosis method based on spatiotemporal coupling feature extraction according to claim 1, characterized in that, The spatial features and the temporal features are concatenated and fused to obtain the fusion result, including: according to The spatial feature output Y1 and the temporal feature output Y2 are concatenated at a feature layer to obtain the fused feature. .

6. The wind turbine fault diagnosis method based on spatiotemporal coupling feature extraction according to claim 1, characterized in that, Based on the fusion processing results, the wind turbine fault diagnosis results are obtained, including: according to Determine the probability distribution of fault categories, where, This is the weight matrix. The bias is used, and softmax(·) is the activation function; The fault category with the highest probability distribution is selected as the final fault diagnosis result for the wind turbine.

7. The wind turbine fault diagnosis method based on spatiotemporal coupling feature extraction according to claim 1, characterized in that, After obtaining the wind turbine fault diagnosis results based on the fusion processing results, the process also includes: The fault diagnosis results of the wind turbine are evaluated based on preset fault evaluation indicators to obtain the evaluation results. The preset fault evaluation indicators include F1 score and recall rate R1. ； ； Wherein, TP represents the number of samples that are actually faults but are correctly identified as faults; FP represents the number of samples that are not actually faults but are incorrectly identified as faults; and FN represents the number of samples that are actually faults but are incorrectly identified as non-faults.

8. A wind turbine fault diagnosis system based on spatiotemporal coupling feature extraction, characterized in that, include: The acquisition module is used to acquire multi-source operation monitoring data of wind turbine units, which includes various sample data. The processing module is used to perform graph structure transformation processing on the sample data to obtain graph structure data; and to extract spatial features from the graph structure data to obtain spatial features. Temporal features are extracted from the input sample data to obtain temporal features; the spatial features are concatenated and fused with the temporal features to obtain a fusion processing result; based on the fusion processing result, the wind turbine fault diagnosis result is obtained.

9. A computing device, characterized in that, include: A processor, a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The system stores instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1 to 7.

Images