Fan blade anomaly diagnosis method, device and equipment and storage medium

Abstract

The embodiment of the invention provides a fan blade anomaly diagnosis method, device and equipment and a storage medium. The method comprises the steps that time sequence segmentation is conducted on one-dimensional time sequence data of a wind power blade vibration signal, correlation analysis is conducted on time sequence segmentation results, and a graph structure is constructed; reconstructing the graph structure by using a graph automatic encoder, and performing parameter training on the graph automatic encoder to obtain a trained graph automatic encoder; analyzing a wind power blade vibration signal, and determining an asymmetric loss function; and constructing an end-to-end training frame based on the trained graph automatic encoder and the asymmetric loss function, determining prediction output according to the end-to-end training frame and a preset comparison model, adjusting frame parameters of the end-to-end training frame, obtaining an anomaly diagnosis model, and performing anomaly diagnosis on the wind power blade vibration signal according to the anomaly diagnosis model. Abnormal state recognition is achieved, the defect that a traditional method is insufficient in complex time sequence feature extraction is overcome, and diagnosis deviation caused by class imbalance is effectively solved.

Description

Technical Field

[0001] This disclosure relates to the field of computer modeling technology, and in particular to a method, apparatus, equipment and storage medium for diagnosing abnormalities in wind turbine blades. Background Technology

[0002] With the widespread application of wind power in renewable energy, wind turbine blades, as key structural components of wind power systems, directly affect the safety and operational efficiency of wind turbine units. However, during long-term operation, wind turbine blades often suffer structural damage such as cracks, torsion, and fatigue due to factors such as complex climate and load fluctuations. To achieve reliable operation of wind power systems and reduce maintenance costs, establishing efficient and accurate methods for detecting wind turbine blade anomalies has become a critical issue that urgently needs to be addressed in the field of wind power operation and maintenance.

[0003] Currently, traditional anomaly detection methods mainly rely on manual feature extraction and shallow classifiers, which struggle to effectively capture the complex temporal structural information latent in blade vibration signals. In recent years, graph neural networks have demonstrated powerful capabilities in modeling non-Euclidean structural data, effectively uncovering correlations between samples, capturing discriminative features in latent graph structures, and becoming a novel tool for processing temporal vibration data. However, under conditions of uneven sample labeling and class imbalance, commonly used loss functions are easily dominated by the majority class, leading to a decline in the model's ability to distinguish minority class samples, especially in minority class tasks such as wind turbine blade anomaly detection. To address these issues, there is still a lack of an efficient diagnostic method that balances overall recognition capability and minority class robustness under uneven sample conditions. Summary of the Invention

[0004] In order to solve the above-mentioned technical problems, or at least partially solve the above-mentioned technical problems, this disclosure provides a method, apparatus, equipment and storage medium for diagnosing abnormal wind turbine blades.

[0005] This disclosure provides a method for diagnosing abnormalities in wind turbine blades, the method comprising: The one-dimensional time-series data of wind turbine blade vibration signal is segmented into time series, the time series segmentation results are subjected to correlation analysis, and a graph structure is constructed based on the correlation analysis results and the time series segmentation results. The graph structure is reconstructed using a graph autoencoder to generate low-dimensional node embedding vectors. The graph autoencoder is then trained based on the parameters of the low-dimensional node embedding vectors to obtain the trained graph autoencoder. Analyze the vibration signal of wind turbine blades to determine the type of imbalance problem, and determine the asymmetric loss function based on the type of imbalance problem. An end-to-end training framework is constructed based on the trained image autoencoder and the asymmetric loss function. The predicted output is determined based on the end-to-end training framework and the preset comparison model. The framework parameters of the end-to-end training framework are adjusted based on the predicted output and the asymmetric loss function to obtain an anomaly diagnosis model. Anomaly diagnosis is performed on the vibration signal of the wind turbine blade based on the anomaly diagnosis model.

[0006] The method provided in this disclosure performs time-series segmentation on one-dimensional time-series data of wind turbine blade vibration signals, conducts correlation analysis on the time-series segmentation results, and constructs a graph structure based on the correlation analysis results and the time-series segmentation results, including: One-dimensional time-series data of wind turbine blade vibration signal is acquired, the wind turbine blade operating cycle is determined based on the one-dimensional time-series data, a fixed length is obtained, and the one-dimensional time-series data is time-series segmented based on the fixed length. Each time series segmentation result is converted into a frequency domain feature vector. Correlation analysis is performed on any two frequency domain feature vectors, and a correlation matrix is ​​generated based on all the correlation analysis results. A graph structure is constructed using the time-series segmentation results as nodes and elements in the correlation matrix that are greater than a preset correlation threshold as edges.

[0007] The method provided in this disclosure reconstructs the graph structure using a graph autoencoder to generate low-dimensional node embedding vectors, and trains the graph autoencoder parameters based on the low-dimensional node embedding vectors to obtain a trained graph autoencoder, including: The temporal normalized value of the corresponding node is determined based on the initial features of each node in the graph structure. A node feature matrix is ​​generated based on the temporal normalized values ​​of all nodes. At the same time, each edge in the graph structure is represented by an adjacency matrix. Input the graph structure, node feature matrix, and adjacency matrix into the encoder part of the graph autoencoder, and extract the low-dimensional node embedding vector set of each node to generate the embedding matrix. The decoder part of the graph autoencoder uses the embedding matrix to minimize the adjacency matrix to reconstruct the reconstructed matrix. The reconstructed matrix is ​​then compared with the adjacency matrix to calculate the reconstruction loss. Based on the reconstruction loss, the parameters of the decoder and encoder in the graph autoencoder are trained to obtain the trained graph autoencoder.

[0008] The method provided in this disclosure analyzes the vibration signal of a wind turbine blade, determines the type of imbalance problem, and determines an asymmetric loss function based on the type of imbalance problem, including: The vibration signals of wind turbine blades are statistically labeled with categories. The category imbalance ratio is calculated based on the category labels to determine the category imbalance problem. The category labels include normal labels and abnormal labels. Different label attention levels are set for signals corresponding to different categories of labels, and an asymmetric loss function is determined based on the label attention and category imbalance problem.

[0009] The method provided in this disclosure, which constructs an end-to-end training framework based on the trained graph autoencoder and the asymmetric loss function, includes: By determining the asymmetric loss function as the training objective, and combining the graph structure reconstruction error with the classification loss, an end-to-end data structure is established from the graph structure input to the state classification output. The output of the trained graph autoencoder is fed into a global pooling layer to generate graph-level feature vectors, which are then input into the downstream classifier. The graph autoencoder, global pooling layer, and classifier are connected sequentially to establish a complete forward propagation path from graph structure input to state classification output. An end-to-end training framework is then constructed based on the complete forward propagation path and end-to-end data.

[0010] The method provided in this disclosure, which determines a predicted output based on an end-to-end training framework and a preset comparison model, adjusts the framework parameters of the end-to-end training framework based on the predicted output and an asymmetric loss function to obtain an anomaly diagnosis model, and performs anomaly diagnosis on the vibration signal of a wind turbine blade based on the anomaly diagnosis model, includes: Based on category labels, the vibration signal of the wind turbine blade is input into the end-to-end training framework and the preset comparison model. Feature extraction is performed to generate a deep feature representation of each sample of each model, and the deep feature representation is forward propagated to determine the prediction output of each sample of each model. The asymmetric loss function value is calculated based on the predicted output, and the frame parameters of the end-to-end training frame are adjusted through backpropagation to obtain the adjusted frame. The anomaly detection performance of the adjusted framework is evaluated based on an independent test set system. The adjusted framework and the comparison model are quantitatively compared on the evaluation index. The anomaly diagnosis model is determined based on the quantitative comparison results. The anomaly diagnosis model is then used to diagnose anomalies in the vibration signal of the wind turbine blade.

[0011] This disclosure also provides a wind turbine blade abnormality diagnosis device, the device comprising: The segmentation module is used to perform time-series segmentation on the one-dimensional time-series data of wind turbine blade vibration signals, perform correlation analysis on the time-series segmentation results, and construct a graph structure based on the correlation analysis results and the time-series segmentation results. The training module is used to reconstruct the graph structure using a graph autoencoder, generate low-dimensional node embedding vectors, and train the graph autoencoder parameters based on the low-dimensional node embedding vectors to obtain the trained graph autoencoder. The determination module is used to analyze the vibration signal of wind turbine blades, determine the type of imbalance problem, and determine the asymmetric loss function based on the type of imbalance problem. The diagnostic module is used to construct an end-to-end training framework based on the trained image autoencoder and the asymmetric loss function, determine the predicted output based on the end-to-end training framework and the preset comparison model, adjust the framework parameters of the end-to-end training framework based on the predicted output and the asymmetric loss function to obtain an anomaly diagnosis model, and perform anomaly diagnosis on the wind turbine blade vibration signal based on the anomaly diagnosis model.

[0012] The apparatus provided in this disclosure includes a segmentation module comprising: The acquisition unit is used to acquire one-dimensional time-series data of the vibration signal of the wind turbine blade, determine the operating cycle of the wind turbine blade based on the one-dimensional time-series data, obtain a fixed length, and perform time-series segmentation on the one-dimensional time-series data based on the fixed length. The generation unit is used to convert each time-series segmentation result into a frequency domain feature vector, perform correlation analysis on any two frequency domain feature vectors, and generate a correlation matrix based on all correlation analysis results. The building unit is used to construct a graph structure with the time-series segmentation results as nodes and elements in the correlation matrix that are greater than a preset correlation threshold as edges.

[0013] In some implementations, the training module includes: The temporal normalized value of the corresponding node is determined based on the initial features of each node in the graph structure. A node feature matrix is ​​generated based on the temporal normalized values ​​of all nodes. At the same time, each edge in the graph structure is represented by an adjacency matrix. Input the graph structure, node feature matrix, and adjacency matrix into the encoder part of the graph autoencoder, and extract the low-dimensional node embedding vector set of each node to generate the embedding matrix. The decoder part of the graph autoencoder uses the embedding matrix to minimize the adjacency matrix to reconstruct the reconstructed matrix. The reconstructed matrix is ​​then compared with the adjacency matrix to calculate the reconstruction loss. Based on the reconstruction loss, the parameters of the decoder and encoder in the graph autoencoder are trained to obtain the trained graph autoencoder.

[0014] In some implementations, the determining module includes: The vibration signals of wind turbine blades are statistically labeled with categories. The category imbalance ratio is calculated based on the category labels to determine the category imbalance problem. The category labels include normal labels and abnormal labels. Different label attention levels are set for signals corresponding to different categories of labels, and an asymmetric loss function is determined based on the label attention and category imbalance problem.

[0015] In some implementations, the diagnostic module includes: By determining the asymmetric loss function as the training objective, and combining the graph structure reconstruction error with the classification loss, an end-to-end data structure is established from the graph structure input to the state classification output. The output of the trained graph autoencoder is fed into a global pooling layer to generate graph-level feature vectors, which are then input into the downstream classifier. The graph autoencoder, global pooling layer, and classifier are connected sequentially to establish a complete forward propagation path from graph structure input to state classification output. An end-to-end training framework is then constructed based on the complete forward propagation path and end-to-end data.

[0016] In some implementations, the determining module includes: Based on category labels, the vibration signal of the wind turbine blade is input into the end-to-end training framework and the preset comparison model. Feature extraction is performed to generate a deep feature representation of each sample of each model, and the deep feature representation is forward propagated to determine the prediction output of each sample of each model. The asymmetric loss function value is calculated based on the predicted output, and the frame parameters of the end-to-end training frame are adjusted through backpropagation to obtain the adjusted frame. The anomaly detection performance of the adjusted framework is evaluated based on an independent test set system. The adjusted framework and the comparison model are quantitatively compared on the evaluation index. The anomaly diagnosis model is determined based on the quantitative comparison results. The anomaly diagnosis model is then used to diagnose anomalies in the vibration signal of the wind turbine blade.

[0017] This disclosure also provides an electronic device, the electronic device comprising: a processor; a memory for storing executable instructions of the processor; the processor being configured to read the executable instructions from the memory and execute the instructions to implement the wind turbine blade abnormality diagnosis method provided in this disclosure.

[0018] This disclosure also provides a computer-readable storage medium storing a computer program for executing the wind turbine blade anomaly diagnosis method provided in this disclosure.

[0019] The technical solution provided in this disclosure has the following advantages compared with the prior art: The wind turbine blade anomaly diagnosis method provided in this disclosure converts the wind turbine blade vibration signal into a graph structure, uses the graph structure to train the parameters of a graph autoencoder, solves the class imbalance problem through an asymmetric loss function, and constructs an anomaly diagnosis model based on the trained graph structure autoencoder and the asymmetric loss function to achieve anomaly state recognition. This overcomes the shortcomings of traditional methods in extracting complex temporal features, effectively solves the diagnostic bias caused by class imbalance, and significantly improves the accuracy and robustness of blade anomaly detection. Attached Figure Description

[0020] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0021] Figure 1 A schematic flowchart of the wind turbine blade anomaly diagnosis method provided in this embodiment of the disclosure; Figure 2 A detailed flowchart of the wind turbine blade anomaly diagnosis method provided in the embodiments of this disclosure; Figure 3 The diagram below shows the structure of the automatic encoder model in an embodiment of this disclosure; Figure 4 Experimental diagram of abnormal fault types of wind turbine blades; Figure 5 Waveform diagram of abnormal fault types of wind turbine blades; Figure 6 The output result diagram of the anomaly diagnosis model provided in the embodiments of this disclosure; Figure 7 This is a schematic diagram of the structure of the wind turbine blade abnormality diagnosis device provided in the embodiments of this disclosure; Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0022] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0023] It should be understood that the steps described in the method embodiments of this disclosure may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of this disclosure is not limited in this respect.

[0024] The term "comprising" and its variations as used herein are open-ended inclusions, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the description below.

[0025] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0026] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0027] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0028] To address the aforementioned issues, this disclosure provides a method for diagnosing abnormalities in wind turbine blades. The method will be described below with reference to specific embodiments.

[0029] Figure 1 This is a flowchart illustrating a method for diagnosing abnormal wind turbine blades according to an embodiment of the present disclosure. The method can be executed by a wind turbine blade abnormality diagnosis device, which can be implemented using software and / or hardware and is generally integrated into an electronic device.

[0030] Example 1: This embodiment of the present disclosure provides a method for diagnosing abnormalities in wind turbine blades, such as... Figure 1 and Figure 2 As shown, the method includes: The one-dimensional time-series data of wind turbine blade vibration signal is segmented into time series, the time series segmentation results are subjected to correlation analysis, and a graph structure is constructed based on the correlation analysis results and the time series segmentation results. The graph structure is reconstructed using a graph autoencoder to generate low-dimensional node embedding vectors. The graph autoencoder is then trained based on the parameters of the low-dimensional node embedding vectors to obtain the trained graph autoencoder. Analyze the vibration signal of wind turbine blades to determine the type of imbalance problem, and determine the asymmetric loss function based on the type of imbalance problem. An end-to-end training framework is constructed based on the trained image autoencoder and the asymmetric loss function. The predicted output is determined based on the end-to-end training framework and the preset comparison model. The framework parameters of the end-to-end training framework are adjusted based on the predicted output and the asymmetric loss function to obtain an anomaly diagnosis model. Anomaly diagnosis is performed on the vibration signal of the wind turbine blade based on the anomaly diagnosis model.

[0031] In this embodiment, such as Figure 3 As shown, the automatic encoder in the figure includes a decoder section and an encoder section.

[0032] In this embodiment, such as Figure 4 and Figure 5 As shown, according to Figure 4 and Figure 5 One-dimensional time-series data of wind turbine blade vibration signals are acquired. This data refers to the sequence of amplitude changes recorded in chronological order by a vibration acceleration sensor installed at the blade root. For example, at a sampling frequency of 2560Hz, continuous acquisition for 10 minutes yields a sequence containing 1,536,000 data points: [0.15, -0.12, 0.23, ..., 0.08].

[0033] In this embodiment, the correlation analysis result is a quantified value of the degree of correlation between the frequency domain feature vectors of any two time-series segmentation results. The correlation matrix is ​​a matrix representation of the correlation coefficients among all time-series segmentation results.

[0034] In this embodiment, time-series segmentation is the process of dividing continuous time-series data into several segments of a fixed length. The result of time-series segmentation is a set of independent time-series segmentation results formed after segmentation.

[0035] In this embodiment, the graph structure is a non-Euclidean data structure composed of nodes and edges. It is represented as... Where G represents the graph structure; V represents the set of nodes; and E represents the set of adjacent edges. For example, a graph contains 35 nodes, corresponding to 35 time-series segmentation results, and 72 edges, connecting node pairs whose correlation analysis results are greater than a preset correlation threshold.

[0036] In this embodiment, the one-dimensional time-series signal is time-series divided into segments. N There are 3 equal-length subsequences, each with a fixed length of 1. L The formula corresponding to the graph structure is determined as follows: ,in, Indicates the first element in the original signal i One node; This represents the minimum value in the entire one-dimensional time-series signal; This represents the maximum value in the entire one-dimensional time-series signal; Represents the normalized i-th i The value of each node; Represents a node i With nodes j The edge weights between them; Represents the normalized i-th j The value of each node; Represents a node i With nodes j Similarity measure; This represents the average value of the nodes; j and i Indicates the index of the node.

[0037] In this embodiment, the low-dimensional node embedding vector is a condensed, low-dimensional feature vector of the encoder for each sensor node. This vector simultaneously contains the vibration characteristics of the node and its position information within the entire blade structure. For example, the encoder compresses 3D temporal vibration characteristics into a 2D embedding space. The embedding vector at the blade tip is [0.02, 0.85], which may capture high-frequency, large-amplitude modes; the embedding vector in the middle of the blade is [0.45, 0.45], which may capture transitional modes; and the embedding vector at the blade root is [0.88, 0.03], which may capture low-frequency, small-amplitude modes.

[0038] In this embodiment, the trained graph autoencoder is a graph neural network model that completes parameter optimization through unsupervised learning. Its encoder can compress graph structure data into low-dimensional node embedding vectors, and the decoder has the ability to reconstruct the graph structure from the embeddings, which can be used for feature extraction or anomaly detection.

[0039] In this embodiment, the temporal normalized value of the node is used as the node feature matrix, and the adjacency matrix represents the edge relationship. The low-dimensional node embedding matrix is ​​extracted by the encoder part of the graph autoencoder, and the adjacency matrix is ​​reconstructed and the reconstruction loss is calculated by the decoder part of the graph autoencoder. The graph autoencoder is then used to train the graph autoencoder parameters and to extract and reconstruct the features of the constructed graph.

[0040] In this embodiment, the corresponding formula for the asymmetric loss function is: ,in, This represents the asymmetric loss value between the final input features and the normal label; h Represents input features; y This indicates a normal label, signifying whether the current sample is a positive class. p This represents the probability of being predicted as positive, that is, the probability of being predicted as a normal label; This represents the moderating factor for the positive class, i.e., the moderating factor for the normal label; This represents the regulation factor for negative classes, i.e., the regulation factor for abnormal labels.

[0041] In this embodiment, the ratio of normal to abnormal category labels in the wind turbine blade vibration signal is statistically analyzed to identify category imbalance. Based on the category imbalance, different label attention levels are assigned to signals corresponding to different category labels, thus constructing an asymmetric loss function.

[0042] In this embodiment, the class imbalance problem is derived from the class imbalance ratio, which refers to the ratio of the number of samples of different classes in the dataset. For example, if there are 950 normal label samples and 50 abnormal label samples, the imbalance ratio is 950:50 = 19:1, indicating that the number of abnormal label samples is only 1 / 19 of the number of normal label samples, and there is a serious class imbalance problem.

[0043] In this embodiment, the end-to-end training framework includes a unified training architecture encompassing the entire data processing flow. For example, the PyTorch framework, which integrates graph encoders, pooling layers, and classifiers, supports end-to-end training from raw graph data to classification results.

[0044] In this embodiment, the asymmetric loss function is used as the training target, and the graph structure reconstruction error is combined with the classification and discrimination loss. An end-to-end training framework is established through the forward path of encoder-pooling layer-classifier.

[0045] In this embodiment, adjusting the framework parameters refers to the process of iteratively improving the model parameters through an optimization algorithm. For example, in 100 epochs of training, with 32 samples per batch, millions of parameters of the encoder and classifier are updated using gradient descent.

[0046] In this embodiment, such as Figure 6 The output result diagram of the abnormal diagnosis model is shown. Anomalies in the wind turbine blades are judged based on the output result diagram.

[0047] The working principle and beneficial effects of this embodiment are as follows: By converting the vibration signal of the wind turbine blade into a graph structure, the graph structure is used to train the parameters of the graph autoencoder. The class imbalance problem is solved by using an asymmetric loss function. An anomaly diagnosis model is constructed based on the trained graph structure autoencoder and the asymmetric loss function to realize the identification of abnormal states. This overcomes the shortcomings of traditional methods in extracting complex temporal features, effectively solves the diagnostic bias caused by class imbalance, and significantly improves the accuracy and robustness of blade anomaly detection.

[0048] Example 2: The method provided in this embodiment of the present disclosure performs time-series segmentation on one-dimensional time-series data of wind turbine blade vibration signals, performs correlation analysis on the time-series segmentation results, and constructs a graph structure based on the correlation analysis results and the time-series segmentation results, including: One-dimensional time-series data of wind turbine blade vibration signal is acquired, the wind turbine blade operating cycle is determined based on the one-dimensional time-series data, a fixed length is obtained, and the one-dimensional time-series data is time-series segmented based on the fixed length. Each time series segmentation result is converted into a frequency domain feature vector. Correlation analysis is performed on any two frequency domain feature vectors, and a correlation matrix is ​​generated based on all the correlation analysis results. A graph structure is constructed using the time-series segmentation results as nodes and elements in the correlation matrix that are greater than a preset correlation threshold as edges.

[0049] In this embodiment, the fixed length is the signal segment length determined based on the blade's operating characteristics. For example, when the blade rotates at 15 RPM, a complete rotation cycle is 4 seconds. The fixed length is 32 seconds, corresponding to 8 cycles, which corresponds to 81,920 data points at a sampling rate of 2560Hz.

[0050] In this embodiment, the preset correlation threshold is a threshold value for determining whether two nodes should establish a connection edge. For example, if the preset correlation threshold is 0.65, only strong correlations where the correlation analysis result is greater than the preset correlation threshold are retained for edge construction.

[0051] In this embodiment, the frequency domain feature vector is the frequency domain amplitude feature obtained after each time-series segmentation result undergoes a Fast Fourier Transform. For example, an 8192-point FFT is performed on segment 1, and the amplitudes of the first 4096 frequency components are taken to form a 4096-dimensional vector: [1.2, 0.8, 0.5, ..., 0.02].

[0052] The working principle and beneficial effects of this embodiment are as follows: A fixed length is determined by the operating cycle of the wind turbine blades, and the vibration time-series data is segmented and converted into frequency domain feature vectors. Correlation between vectors is calculated to form a correlation matrix. A graph structure is constructed using the corresponding time-series segmentation results as nodes and strong correlations as edges, effectively capturing the spatiotemporal pattern characteristics of the blade's operating state. This provides high-quality input for subsequent deep analysis using graph neural networks, improving the accuracy of anomaly detection. It has significant advantages in capturing complex nonlinear and non-stationary dynamic patterns, providing a solid structured representation foundation for subsequent graph embedding and relational reasoning.

[0053] Example 3: The method provided in this embodiment of the present disclosure reconstructs the graph structure using a graph autoencoder to generate low-dimensional node embedding vectors, and trains the graph autoencoder parameters based on the low-dimensional node embedding vectors to obtain a trained graph autoencoder, including: The temporal normalized value of the corresponding node is determined based on the initial features of each node in the graph structure. A node feature matrix is ​​generated based on the temporal normalized values ​​of all nodes. At the same time, each edge in the graph structure is represented by an adjacency matrix. Input the graph structure, node feature matrix, and adjacency matrix into the encoder part of the graph autoencoder, and extract the low-dimensional node embedding vector set of each node to generate the embedding matrix. The decoder part of the graph autoencoder uses the embedding matrix to minimize the adjacency matrix to reconstruct the reconstructed matrix. The reconstructed matrix is ​​then compared with the adjacency matrix to calculate the reconstruction loss. Based on the reconstruction loss, the parameters of the decoder and encoder in the graph autoencoder are trained to obtain the trained graph autoencoder.

[0054] In this embodiment, the initial feature is the sequence of raw vibration signals collected by each sensor node. For example, the vibration amplitude collected by the leaf tip sensor at three time points within a certain period of time is [0.5, 1.2, 2.1]; the data collected by the leaf mid-node sensor is [0.3, 0.8, 1.5]; and the data collected by the leaf root sensor is [0.1, 0.3, 0.9].

[0055] In this embodiment, the time-series normalization value is used to eliminate the influence of differences in the measurement ranges of each sensor and extreme values, and to standardize the time-series data of each sensor.

[0056] In this embodiment, the node feature matrix is ​​a matrix composed of the normalized features of all sensor nodes, which is used as the input to the model.

[0057] In this embodiment, the adjacency matrix is ​​a matrix that describes the physical connections between different parts of the leaf. For example, based on the physical structure of the leaf, the leaf root is connected to the leaf midsection, and the leaf midsection is connected to the leaf tip.

[0058] In this embodiment, the embedding matrix is ​​the set of embedding vectors of all low-dimensional sensor nodes, and is the output matrix of the node. The input matrix of the node includes the node feature matrix and the adjacency matrix.

[0059] In this embodiment, minimizing reconstruction is the goal of the decoder to restore the original connection relationships between nodes as accurately as possible using node embedding vectors.

[0060] In this embodiment, the reconstruction matrix is ​​the node connection matrix predicted by the decoder. For example, the decoder outputs a matrix representing the connection probabilities.

[0061] In this embodiment, the reconstruction loss measures the difference between the reconstructed matrix and the true adjacency matrix. For example, it compares the predicted connectivity probabilities with the actual connectivity probabilities and calculates the difference. The loss function penalizes incorrect predictions if the model predicts a high connectivity probability between the leaf root and leaf tip, even though they are not actually directly connected.

[0062] In this embodiment, the corresponding formula for the embedding matrix is: ,in, Represents the embedding matrix; Represents a nonlinear activation function; A This indicates an adjacency matrix with added self-connections; D Indicates correspondence A The degree matrix is ​​defined as ; X Represents the node feature matrix; W This represents the weight matrix that needs to be learned.

[0063] In this embodiment, the formula for the corresponding elements in the reconstruction matrix is: ,in, Represents a node i With nodes j The probability of reconnection; Represents a node i The transpose of the low-dimensional node embedding vector; Represents a node j The low-dimensional node embedding vector.

[0064] The working principle and beneficial effects of this embodiment are as follows: the temporal normalized value of the node is used as the node feature matrix, the adjacency matrix represents the edge relationship, the low-dimensional node embedding matrix is ​​extracted by the encoder part of the graph autoencoder, the adjacency matrix is ​​reconstructed by the decoder part of the graph autoencoder and the reconstruction loss is calculated, and the parameters of the graph autoencoder are trained in this way. The graph autoencoder is introduced to extract and reconstruct the features of the constructed graph, and the temporal semantic information of the node and its positional relationship in the graph space are integrated, thereby enhancing the model's ability to characterize the location of abnormal signals and their propagation path within the system, and effectively improving the accuracy and overall robustness of the fault detection results.

[0065] Example 4: The method provided in this embodiment of the present disclosure analyzes the vibration signal of a wind turbine blade, determines the type of imbalance problem, and determines an asymmetric loss function based on the type of imbalance problem, including: The vibration signals of wind turbine blades are statistically labeled with categories. The category imbalance ratio is calculated based on the category labels to determine the category imbalance problem. The category labels include normal labels and abnormal labels. Different label attention levels are set for signals corresponding to different categories of labels, and an asymmetric loss function is determined based on the label attention and category imbalance problem.

[0066] In this embodiment, the category label refers to the state category marked for each wind turbine blade vibration signal sample. For example, the normal label is 0, which represents that the blade is in a healthy operating state; the abnormal label is 1, which represents that the blade has fault conditions such as cracks or imbalance.

[0067] In this embodiment, the class imbalance ratio refers to the ratio of the number of samples of different classes in the dataset. For example, if there are 950 normal label samples and 50 abnormal label samples, the imbalance ratio is 950:50 = 19:1, indicating that the number of abnormal label samples is only 1 / 19 of the number of normal label samples, indicating a serious class imbalance problem.

[0068] In this embodiment, label attention refers to the weight coefficients set for different categories in the loss function, used to increase the model's attention to the minority class. For example, a normal label, positive class 0, corresponds to a weight of 1.0; an abnormal label sample, negative class 1, corresponds to a weight of 5.0. This means the model will pay more attention to correctly classifying abnormal label samples, and the cost of misclassifying them is 5 times that of normal label samples.

[0069] The working principle and beneficial effects of this embodiment are as follows: The proportion of normal and abnormal category labels in the vibration signals of wind turbine blades is statistically analyzed to identify class imbalance. Based on the class imbalance, different label attention levels are assigned to signals corresponding to different category labels. An asymmetric loss function is constructed to specifically optimize the class distribution imbalance, focusing on enhancing the ability to distinguish between difficult-to-classify samples and easily confused categories. This significantly improves the model's sensitivity and recall in minority class anomaly and weak signal anomaly detection tasks.

[0070] Example 5: The method provided in this embodiment of the present disclosure constructs an end-to-end training framework based on the trained graph autoencoder and the asymmetric loss function, including: By determining the asymmetric loss function as the training objective, and combining the graph structure reconstruction error with the classification loss, an end-to-end data structure is established from the graph structure input to the state classification output. The output of the trained graph autoencoder is fed into a global pooling layer to generate graph-level feature vectors, which are then input into the downstream classifier. The graph autoencoder, global pooling layer, and classifier are connected sequentially to establish a complete forward propagation path from graph structure input to state classification output. An end-to-end training framework is then constructed based on the complete forward propagation path and end-to-end data.

[0071] In this embodiment, the training objective refers to the objective function that needs to be minimized during the model optimization process.

[0072] In this embodiment, the graph structure reconstruction error refers to the difference between the adjacency matrix reconstructed by the graph autoencoder and the original adjacency matrix. For example, the difference between the reconstructed matrix and the adjacency matrix is ​​calculated using binary cross-entropy to obtain the corresponding loss value.

[0073] In this embodiment, the classification loss refers to the measure of the difference between the model's predicted class and the true class. For example, a weighted cross-entropy loss is used, with the normal class weight = 1.0, the abnormal class weight = 5.0, and the current sample loss = 2.3.

[0074] In this embodiment, the state classification output refers to the model's final classification result of the input graph structure. For example, the output of the Softmax layer [0.85, 0.15] represents a normal probability of 85% and an abnormal probability of 15%.

[0075] In this embodiment, end-to-end data refers to the complete data stream from the original input to the final output. For example, node feature matrix X (35×4096) + adjacency matrix A (35×35) → graph encoder → classifier → probability output (2D).

[0076] In this embodiment, the global pooling layer is a network layer that aggregates node-level features into graph-level features. For example, the global average pooling layer embeds the 512 dimensions of 35 nodes into a single 512-dimensional graph feature vector.

[0077] In this embodiment, the graph-level feature vector is an aggregated vector of the features of the entire graph structure. For example, a 512-dimensional vector [0.12, -0.05, 0.33, ..., 0.08] comprehensively represents the overall state of all 35 time-series segmentation results.

[0078] In this embodiment, the downstream classifier is the network module that classifies the final state. For example, a two-layer fully connected network (512→64→2) using ReLU activation and Dropout=0.3.

[0079] In this embodiment, the complete forward propagation path is the complete transmission path of data from input to output. For example, graph structure → encoder → node embedding (35×512) → pooling layer → graph features (512-dimensional) → classifier → classification output (2-dimensional).

[0080] The working principle and beneficial effects of this embodiment are as follows: Using an asymmetric loss function as the training objective, the graph structure reconstruction error is combined with the classification discriminant loss. An end-to-end training framework is established through the encoder-pooling layer-classifier forward path, realizing the mapping from graph structure to state classification. Asymmetric loss is utilized to enhance sensitivity to anomalous labeled samples, improving the model's diagnostic accuracy and generalization ability under class imbalance conditions.

[0081] Example 6: The method provided in this embodiment of the present disclosure determines the predicted output based on an end-to-end training framework and a preset comparison model, adjusts the framework parameters of the end-to-end training framework based on the predicted output and an asymmetric loss function to obtain an anomaly diagnosis model, and performs anomaly diagnosis on the vibration signal of wind turbine blades based on the anomaly diagnosis model, including: Based on category labels, the vibration signal of the wind turbine blade is input into the end-to-end training framework and the preset comparison model. Feature extraction is performed to generate a deep feature representation of each sample of each model, and the deep feature representation is forward propagated to determine the prediction output of each sample of each model. The asymmetric loss function value is calculated based on the predicted output, and the frame parameters of the end-to-end training frame are adjusted through backpropagation to obtain the adjusted frame. The anomaly detection performance of the adjusted framework is evaluated based on an independent test set system. The adjusted framework and the comparison model are quantitatively compared on the evaluation index. The anomaly diagnosis model is determined based on the quantitative comparison results. The anomaly diagnosis model is then used to diagnose anomalies in the vibration signal of the wind turbine blade.

[0082] In this embodiment, the preset comparison model refers to the benchmark model used for performance comparison. For example, classic models such as Graph Convolutional Network (GCN), Graph Attention Network (GAT), and Variational Autoencoder (VAE).

[0083] In this embodiment, deep feature representation refers to high-level abstract features extracted through a deep neural network. For example, the 512-dimensional node embedding vector extracted by the GAAD model, and the 256-dimensional graph representation vector output by the GCN model.

[0084] In this embodiment, backpropagation refers to the algorithmic process of updating model parameters based on the gradient of the loss function. For example, using the Adam optimizer, the gradient is backpropagated along the computation graph with a learning rate of 0.001 to update the weight matrix.

[0085] In this embodiment, the adjusted framework refers to the complete model system after parameter optimization. For example, the trained GAAD framework includes fixed graph encoder weights and optimized classifier parameters.

[0086] In this embodiment, anomaly detection performance refers to the model's ability to identify abnormal states. For example, it achieves an accuracy of 95.2%, a recall of 89.7%, and a precision of 91.3% on the test set.

[0087] In this embodiment, the evaluation metric refers to a standard that quantifies the model's performance. Examples include accuracy, precision, and recall.

[0088] In this embodiment, quantitative comparison refers to a comparative analysis of model performance based on numerical indicators. As shown in Table 1,

[0089] Table 1.

[0090] In this embodiment, such as Figure 2As shown, the anomaly diagnosis process refers to the complete workflow of using a trained model to determine the state. Starting with signal acquisition, the one-dimensional time-series data is partitioned and processed using FFT to become frequency domain data, which is then converted into graph node data and input into the GNN model. Through loss function calculation and anomaly identification, the state is determined to be normal. If an anomaly is detected, an alarm is triggered and the system is shut down; otherwise, continuous monitoring resumes.

[0091] The working principle and beneficial effects of this embodiment are as follows: Vibration signals are input into an end-to-end frame and a comparison model, depth features are extracted, and forward propagation is used to obtain the predicted output. The loss value is calculated using an asymmetric loss function, the frame parameters are adjusted using backpropagation, and the final anomaly diagnosis model is determined after evaluation on an independent test set. Through the collaborative optimization of the end-to-end training frame and the comparison model, combined with the asymmetric loss function, the sensitivity of the model to anomaly features is effectively improved, significantly enhancing diagnostic accuracy and robustness.

[0092] To achieve the above embodiments, this disclosure also proposes a wind turbine blade abnormality diagnosis device.

[0093] Figure 7 This is a schematic diagram of the structure of a wind turbine blade abnormality diagnosis device provided in an embodiment of this disclosure. This device can be implemented by software and / or hardware, and is generally integrated into an electronic device. Figure 7 As shown, the device includes: a segmentation module, a training module, a determination module, and a diagnosis module, wherein, The segmentation module is used to perform time-series segmentation on the one-dimensional time-series data of wind turbine blade vibration signals, perform correlation analysis on the time-series segmentation results, and construct a graph structure based on the correlation analysis results and the time-series segmentation results. The training module is used to reconstruct the graph structure using a graph autoencoder, generate low-dimensional node embedding vectors, and train the graph autoencoder parameters based on the low-dimensional node embedding vectors to obtain the trained graph autoencoder. The determination module is used to analyze the vibration signal of wind turbine blades, determine the type of imbalance problem, and determine the asymmetric loss function based on the type of imbalance problem. The diagnostic module is used to construct an end-to-end training framework based on the trained image autoencoder and the asymmetric loss function, determine the predicted output based on the end-to-end training framework and the preset comparison model, adjust the framework parameters of the end-to-end training framework based on the predicted output and the asymmetric loss function to obtain an anomaly diagnosis model, and perform anomaly diagnosis on the wind turbine blade vibration signal based on the anomaly diagnosis model.

[0094] The apparatus provided in this disclosure includes a segmentation module comprising: The acquisition unit is used to acquire one-dimensional time-series data of the vibration signal of the wind turbine blade, determine the operating cycle of the wind turbine blade based on the one-dimensional time-series data, obtain a fixed length, and perform time-series segmentation on the one-dimensional time-series data based on the fixed length. The generation unit is used to convert each time-series segmentation result into a frequency domain feature vector, perform correlation analysis on any two frequency domain feature vectors, and generate a correlation matrix based on all correlation analysis results. The building unit is used to construct a graph structure with the time-series segmentation results as nodes and elements in the correlation matrix that are greater than a preset correlation threshold as edges.

[0095] In some implementations, the training module includes: The temporal normalized value of the corresponding node is determined based on the initial features of each node in the graph structure. A node feature matrix is ​​generated based on the temporal normalized values ​​of all nodes. At the same time, each edge in the graph structure is represented by an adjacency matrix. Input the graph structure, node feature matrix, and adjacency matrix into the encoder part of the graph autoencoder, and extract the low-dimensional node embedding vector set of each node to generate the embedding matrix. The decoder part of the graph autoencoder uses the embedding matrix to minimize the adjacency matrix to reconstruct the reconstructed matrix. The reconstructed matrix is ​​then compared with the adjacency matrix to calculate the reconstruction loss. Based on the reconstruction loss, the parameters of the decoder and encoder in the graph autoencoder are trained to obtain the trained graph autoencoder.

[0096] In some implementations, the determining module includes: The vibration signals of wind turbine blades are statistically labeled with categories. The category imbalance ratio is calculated based on the category labels to determine the category imbalance problem. The category labels include normal labels and abnormal labels. Different label attention levels are set for signals corresponding to different categories of labels, and an asymmetric loss function is determined based on the label attention and category imbalance problem.

[0097] In some implementations, the diagnostic module includes: By determining the asymmetric loss function as the training objective, and combining the graph structure reconstruction error with the classification loss, an end-to-end data structure is established from the graph structure input to the state classification output. The output of the trained graph autoencoder is fed into a global pooling layer to generate graph-level feature vectors, which are then input into the downstream classifier. The graph autoencoder, global pooling layer, and classifier are connected sequentially to establish a complete forward propagation path from graph structure input to state classification output. An end-to-end training framework is then constructed based on the complete forward propagation path and end-to-end data.

[0098] In some implementations, the determining module includes: Based on category labels, the vibration signal of the wind turbine blade is input into the end-to-end training framework and the preset comparison model. Feature extraction is performed to generate a deep feature representation of each sample of each model, and the deep feature representation is forward propagated to determine the prediction output of each sample of each model. The asymmetric loss function value is calculated based on the predicted output, and the frame parameters of the end-to-end training frame are adjusted through backpropagation to obtain the adjusted frame. The anomaly detection performance of the adjusted framework is evaluated based on an independent test set system. The adjusted framework and the comparison model are quantitatively compared on the evaluation index. The anomaly diagnosis model is determined based on the quantitative comparison results. The anomaly diagnosis model is then used to diagnose anomalies in the vibration signal of the wind turbine blade.

[0099] The wind turbine blade abnormality diagnosis device provided in this disclosure can execute the wind turbine blade abnormality diagnosis method provided in any embodiment of this disclosure, and has the corresponding functional modules and beneficial effects of executing the method.

[0100] To implement the above embodiments, this disclosure also proposes a computer program product, including a computer program / instruction, which, when executed by a processor, implements the wind turbine blade abnormality diagnosis method in the above embodiments.

[0101] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure.

[0102] The following is a detailed reference. Figure 8 The diagram illustrates a structural schematic suitable for implementing the electronic device 300 in the embodiments of this disclosure. The electronic device 300 in the embodiments of this disclosure may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 8 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0103] like Figure 8As shown, the electronic device 300 may include a processor (e.g., a central processing unit, a graphics processing unit, etc.) 301, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 302 or a program loaded from a memory 308 into a random access memory (RAM) 303. The RAM 303 also stores various programs and data required for the operation of the electronic device 300. The processor 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0104] Typically, the following devices can be connected to I / O interface 305: input devices 306 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 307 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 308 including, for example, magnetic tapes, hard disks, etc.; and communication devices 309. Communication device 309 allows electronic device 300 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 8 An electronic device 300 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0105] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 309, or installed from a memory 308, or installed from a ROM 302. When the computer program is executed by the processor 301, it performs the functions defined in the wind turbine blade anomaly diagnosis method of embodiments of this disclosure.

[0106] It should be noted that the computer-readable medium described in this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0107] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0108] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0109] The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the aforementioned wind turbine blade abnormality diagnosis method.

[0110] Electronic devices can be programmed with computer program code in one or more programming languages ​​or a combination thereof to perform the operations of this disclosure. These programming languages ​​include, but are not limited to, object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as "C" or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0111] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0112] The units described in the embodiments of this disclosure can be implemented in software or hardware. The names of the units are not, in some cases, intended to limit the specific unit.

[0113] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0114] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0115] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

[0116] Furthermore, while the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. In certain environments, multitasking and parallel processing may be advantageous. Similarly, while several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0117] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A method for diagnosing abnormalities in wind turbine blades, characterized in that, include: The one-dimensional time-series data of wind turbine blade vibration signal is segmented into time series, the time series segmentation results are subjected to correlation analysis, and a graph structure is constructed based on the correlation analysis results and the time series segmentation results. The graph structure is reconstructed using a graph autoencoder to generate low-dimensional node embedding vectors. The graph autoencoder is then trained based on the parameters of the low-dimensional node embedding vectors to obtain the trained graph autoencoder. Analyze the vibration signal of wind turbine blades to determine the type of imbalance problem, and determine the asymmetric loss function based on the type of imbalance problem. An end-to-end training framework is constructed based on the trained image autoencoder and the asymmetric loss function. The predicted output is determined based on the end-to-end training framework and the preset comparison model. The framework parameters of the end-to-end training framework are adjusted based on the predicted output and the asymmetric loss function to obtain an anomaly diagnosis model. Anomaly diagnosis is performed on the vibration signal of the wind turbine blade based on the anomaly diagnosis model.

2. The method according to claim 1, characterized in that, The one-dimensional time-series data of wind turbine blade vibration signals are time-series segmented, correlation analysis is performed on the time-series segmentation results, and a graph structure is constructed based on the correlation analysis results and the time-series segmentation results, including: One-dimensional time-series data of wind turbine blade vibration signal is acquired, the wind turbine blade operating cycle is determined based on the one-dimensional time-series data, a fixed length is obtained, and the one-dimensional time-series data is time-series segmented based on the fixed length. Each time series segmentation result is converted into a frequency domain feature vector. Correlation analysis is performed on any two frequency domain feature vectors, and a correlation matrix is ​​generated based on all the correlation analysis results. A graph structure is constructed using the time-series segmentation results as nodes and elements in the correlation matrix that are greater than a preset correlation threshold as edges.

3. The method according to claim 1, characterized in that, The graph structure is reconstructed using a graph autoencoder to generate low-dimensional node embedding vectors. The graph autoencoder is then trained using these low-dimensional node embedding vectors to obtain the trained graph autoencoder, which includes: The temporal normalized value of the corresponding node is determined based on the initial features of each node in the graph structure. A node feature matrix is ​​generated based on the temporal normalized values ​​of all nodes. At the same time, each edge in the graph structure is represented by an adjacency matrix. Input the graph structure, node feature matrix, and adjacency matrix into the encoder part of the graph autoencoder, and extract the low-dimensional node embedding vector set of each node to generate the embedding matrix. The decoder part of the graph autoencoder uses the embedding matrix to minimize the adjacency matrix to reconstruct the reconstructed matrix. The reconstructed matrix is ​​then compared with the adjacency matrix to calculate the reconstruction loss. Based on the reconstruction loss, the parameters of the decoder and encoder in the graph autoencoder are trained to obtain the trained graph autoencoder.

4. The method according to claim 1, characterized in that, Analyze the vibration signal of wind turbine blades to determine the type of imbalance problem, and determine the asymmetric loss function based on the type of imbalance problem, including: The vibration signals of wind turbine blades are statistically labeled with categories. The category imbalance ratio is calculated based on the category labels to determine the category imbalance problem. The category labels include normal labels and abnormal labels. Different label attention levels are set for signals corresponding to different categories of labels, and an asymmetric loss function is determined based on the label attention and category imbalance problem.

5. The method according to claim 1, characterized in that, An end-to-end training framework is constructed based on the trained graph autoencoder and the asymmetric loss function, including: By determining the asymmetric loss function as the training objective, and combining the graph structure reconstruction error with the classification loss, an end-to-end data structure is established from the graph structure input to the state classification output. The output of the trained graph autoencoder is fed into a global pooling layer to generate graph-level feature vectors, which are then input into the downstream classifier. The graph autoencoder, global pooling layer, and classifier are connected sequentially to establish a complete forward propagation path from graph structure input to state classification output. An end-to-end training framework is then constructed based on the complete forward propagation path and end-to-end data.

6. The method according to claim 4, characterized in that, The predicted output is determined based on the end-to-end training framework and the preset comparison model. The framework parameters of the end-to-end training framework are adjusted based on the predicted output and the asymmetric loss function to obtain an anomaly diagnosis model. Anomaly diagnosis of the wind turbine blade vibration signal is then performed based on this model, including: Based on category labels, the vibration signal of the wind turbine blade is input into the end-to-end training framework and the preset comparison model. Feature extraction is performed to generate a deep feature representation of each sample of each model, and the deep feature representation is forward propagated to determine the prediction output of each sample of each model. The asymmetric loss function value is calculated based on the predicted output, and the frame parameters of the end-to-end training frame are adjusted through backpropagation to obtain the adjusted frame. The anomaly detection performance of the adjusted framework is evaluated based on an independent test set system. The adjusted framework and the comparison model are quantitatively compared on the evaluation index. The anomaly diagnosis model is determined based on the quantitative comparison results. The anomaly diagnosis model is then used to diagnose anomalies in the vibration signal of the wind turbine blade.

7. A device for diagnosing abnormalities in wind turbine blades, characterized in that, include: The segmentation module is used to perform time-series segmentation on the one-dimensional time-series data of wind turbine blade vibration signals, perform correlation analysis on the time-series segmentation results, and construct a graph structure based on the correlation analysis results and the time-series segmentation results. The training module is used to reconstruct the graph structure using a graph autoencoder, generate low-dimensional node embedding vectors, and train the graph autoencoder parameters based on the low-dimensional node embedding vectors to obtain the trained graph autoencoder. The determination module is used to analyze the vibration signal of wind turbine blades, determine the type of imbalance problem, and determine the asymmetric loss function based on the type of imbalance problem. The diagnostic module is used to construct an end-to-end training framework based on the trained image autoencoder and the asymmetric loss function, determine the predicted output based on the end-to-end training framework and the preset comparison model, adjust the framework parameters of the end-to-end training framework based on the predicted output and the asymmetric loss function to obtain an anomaly diagnosis model, and perform anomaly diagnosis on the wind turbine blade vibration signal based on the anomaly diagnosis model.

8. The apparatus according to claim 7, characterized in that, The segmentation module includes: The acquisition unit is used to acquire one-dimensional time-series data of the vibration signal of the wind turbine blade, determine the operating cycle of the wind turbine blade based on the one-dimensional time-series data, obtain a fixed length, and perform time-series segmentation on the one-dimensional time-series data based on the fixed length. The generation unit is used to convert each time-series segmentation result into a frequency domain feature vector, perform correlation analysis on any two frequency domain feature vectors, and generate a correlation matrix based on all correlation analysis results. The building unit is used to construct a graph structure with the time-series segmentation results as nodes and elements in the correlation matrix that are greater than a preset correlation threshold as edges.

9. An electronic device, characterized in that, include: Memory; processor; as well as Computer programs; The computer program is stored in the memory and configured to be executed by the processor to implement the wind turbine blade abnormality diagnosis method as described in any one of claims 1-6.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, causes the processor to implement the wind turbine blade abnormality diagnosis method provided in any one of claims 1-6.

Images