Wind turbine generator fault early warning method based on multi-modal multi-scale adaptive graph neural network

Abstract

The invention discloses a wind turbine generator fault early warning method based on a multi-modal multi-scale adaptive graph neural network. The method comprises the following steps: 1) constructing a multi-scale time-varying network for capturing operation fluctuation of a wind turbine generator; 2) obtaining SCADA multi-dimensional time sequence data X (orig), and inputting the data into the multi-scale time-varying network to obtain a multi-time-scale time-varying graph and a time-varying graph node feature set; 3) processing the time-varying graph structure by using a sensing feature prediction model based on a graph convolutional network and a self-attention mechanism to obtain a wind turbine generator SCADA sensor node prediction value in future Tp time; and 4) based on the wind turbine generator SCADA sensor node prediction value and the wind turbine generator SCADA sensor node actual measurement value, determining whether the wind turbine generator has a fault. According to the method, a framework based on an adaptive time-varying graph neural network (AT-GNN) is constructed, and a multi-scale prediction model of sensor behaviors is combined, so that the interpretability, the real-time performance and the robustness of an existing method are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of wind turbine fault warning, and specifically to a wind turbine fault warning method based on a multi-modal multi-scale adaptive graph neural network. Background Art

[0002] The fault warning methods of wind turbines can generally be classified into three categories: model-based methods, signal processing-based methods, and deep learning methods based on big data mining.

[0003] (1) Model-based methods: By establishing dynamic behavior models of key components of wind turbines to predict possible faults of the components. Although these methods can make certain inferences about the equipment status based on theoretical models, with the increase in equipment complexity, it becomes increasingly difficult to construct fault models. In addition, traditional model methods often have low processing efficiency when facing large-scale data, cannot meet the requirements of real-time warning, and lack sufficient robustness to handle complex interrelationships between equipment.

[0004] (2) Signal processing-based methods: This method relies on high-frequency signals (such as vibration, temperature, pressure, etc.) collected by sensors to monitor equipment health. However, these signals often contain noise, and in complex equipment, a large amount of computing resources are required for real-time processing, making it difficult to meet the high real-time requirements of large wind turbines for fault warning. In addition, signal processing methods fail to effectively decouple the complex dynamic relationships between sensors and are difficult to accurately capture potential fault correlations between different equipment.

[0005] (3) Deep learning methods based on big data mining: With the development of SCADA systems, the large-scale data of wind turbines provides the possibility for deep learning. Through big data analysis, deep learning-based methods can extract potential fault patterns from a large amount of sensor data and have a certain prediction ability. However, these methods usually rely on "black box" models, lack the decoupling of complex relationships between sensors, resulting in poor interpretability of the models and limited generalization ability of prediction results, especially difficult to achieve timely warning before the occurrence of faults.

[0006] Traditional wind turbine fault warning methods mainly have the following problems:

[0007] (1) Model-based fault warning methods are difficult to model and handle large-scale data: The complex coupling relationships between wind turbine components make it difficult for traditional models to accurately predict faults. When facing large-scale data, traditional models perform poorly in processing efficiency and prediction accuracy, restricting their application in real-time fault warning.

[0008] (2) The signal - to - noise ratio of the signal processing method is low and the computational consumption is large: The signals generated by a large number of sensors are often interfered by noise, making it difficult to identify early faults. At the same time, the real - time monitoring of high - frequency vibration signals requires a large amount of computational resources, and it is difficult to meet the high - efficiency early warning requirements of large - scale wind turbines.

[0009] (3) The deep - learning method lacks decoupling of sensor relationships and lacks interpretability and generalization ability: Existing deep - learning models lack effective modeling of the complex time - varying relationships between sensors, resulting in limited prediction ability for wind turbine faults. In addition, the model is usually in the form of a "black box" and lacks sufficient interpretability, making the reliability and generalization ability of the prediction results insufficient, and it is difficult to achieve accurate prediction and early warning of faults. Summary of the Invention

[0010] The purpose of the present invention is to provide a wind turbine fault early - warning method based on a multi - modal multi - scale adaptive graph neural network, including the following steps:

[0011] 1) Construct a multi - scale time - varying network for capturing the operation fluctuations of wind turbines;

[0012] 2) Obtain the SCADA multi - dimensional time - series data X (orig) , and input it into the multi - scale time - varying network to obtain time - varying graphs at multiple time scales and a set of time - varying graph node features;

[0013] 3) Use a sensing feature prediction model based on a graph convolutional network and a self - attention mechanism to process the time - varying graph structure to obtain the predicted values of the SCADA sensor nodes of the wind turbine in the future T p time;

[0014] 4) Based on the predicted values of the SCADA sensor nodes of the wind turbine and the measured values of the SCADA sensor nodes of the wind turbine, determine whether the wind turbine has a fault.

[0015] Furthermore, the multi - scale time - varying network uses a double - layer sliding window to capture the time correlation of the SCADA multi - dimensional time - series data X (orig) ;

[0016] The double - layer sliding window includes a first - level sliding window and a second - level sliding window;

[0017] The sampling length of the first - level sliding window is w1, and the sliding step is 1;

[0018] The sampling length of the second - level sliding window is w2.

[0019] Furthermore, the first - level sliding window samples the original dataset of the SCADA multi - dimensional time - series of the wind turbine , and the sampled dataset is denoted as n is the number of sensors, and T is the sampling duration;

[0020] Among them, the dataset is as follows:

[0021]

[0022] The secondary sliding window samples the dataset, and the sampled dataset is denoted as X2;

[0023] The dataset is shown as follows:

[0024]

[0025] Furthermore, for time point t, the time-varying graph structure G t =(V, E t ); V is the set of sensor nodes; E t ={E t (1, 2), E t (2, 3),..., E t (V - 1, V)} is the set of node relationships at time t; each node of the time-varying graph structure stores SCADA feature data;

[0026] Among them, the node relationship E t (i, j) between the i-th and j-th sensors is as follows:

[0027]

[0028] Among them, and are the time series data of the i-th and j-th sensors under the secondary sliding window respectively, and f is the dynamic correlation function.

[0029] Furthermore, the dynamic correlation function is obtained by using the Pearson correlation coefficient, mutual information, or dynamic weight learning based on a graph neural network.

[0030] Furthermore, the time-varying graph node feature set H is as follows:

[0031]

[0032] Among them, h i represents the feature vector of sensor i, and d is the dimension of the node features.

[0033] Furthermore, the predicted values of the SCADA sensor nodes of the wind turbine in the future T p time are as follows:

[0034]

[0035] Among them, L is the number of network layers; is the weight calculated by the self-attention mechanism; are the weight and bias term; f is the activation function; τ = 1, 2,..., T p ; T p is the time step of multi-step prediction; is the predicted value of sensor i at the future time t + τ; represents the feature vector of sensor i in the current time window t;

[0036] Among them, the weight is as follows:

[0037]

[0038] In the formula, W (l) , W (0) are the feature transformation matrices between the layer and the initial feature; a is the learnable attention vector; the superscript T represents the transpose; is the initial node feature; W (k ) is the feature transformation matrix of the k-th layer feature; is the k-th layer node feature;

[0039] Node feature is as follows:

[0040]

[0041] In the formula, W (l) is the weight matrix, represents vector concatenation, b (l) is the bias constant term; σ(·) is the non-linear activation function; is the (l - 1)-th layer node feature;

[0042] Aggregated feature is as follows:

[0043]

[0044] In the formula, and represent the feature information of nodes v and u in the (l - 1)-th layer. AGG (l) is the feature aggregation operation.

[0045] Furthermore, the sensing feature prediction model is trained with historical data;

[0046] During the training process, the loss function adopted by the sensing feature prediction model is as follows:

[0047]

[0048] In the formula, L MSE is the loss function. is the predicted value of sensor i at the future time t + τ; is the measured value of sensor i at the future time t + τ. T p is the time step of multi-step prediction; n is the number of sensors.

[0049] Furthermore, in step 4), the index for judging whether the wind turbine has a fault is the anomaly score f i ;

[0050] The anomaly score f i is as follows:

[0051]

[0052] where f i ∈[0,1] is the fault anomaly score of the system at the current moment. Among them is the prediction error of the i-th sensor at time (t + τ). is the predicted value of sensor i at the future time t + τ; is the measured value of sensor i at the future time t + τ.

[0053] Furthermore, when the fault anomaly score f i > θ t the wind turbine has a fault risk and triggers a fault warning signal;

[0054] where the threshold θ t is as follows:

[0055] θ t = μ + λσ (12)

[0056] where μ and σ are respectively the mean and standard deviation of historical anomaly scores; λ is a dynamic parameter for controlling the warning sensitivity.

[0057] The technical effect of the present invention is beyond doubt. The present invention proposes an innovative fault warning method for wind turbines. By constructing a framework based on an adaptive time-varying graph neural network (AT-GNN) and combining a multi-scale prediction model of sensor behavior, the interpretability, real-time performance, and robustness of the existing method are significantly improved.

[0058] The beneficial effects of the present invention are specifically as follows:

[0059] (1) Effectively simplifies the modeling complexity and improves the accuracy of fault warning

[0060] By introducing a multi-layer time-varying network and a graph neural network (GNN), the present invention decouples the time-varying dynamic relationships among the sensors of a wind turbine, avoiding the cumbersome mathematical modeling problems in traditional physics-based methods. This method can dynamically capture the behavioral characteristics of sensors and model the fault trends evolving over time. By modeling and fusing the node relationships on multiple time scales, the accuracy of fault warning is effectively improved. Especially when facing complex sensor data and large-scale systems, it has higher processing efficiency and detection accuracy.

[0061] (2) Reduces the impact of signal noise on the accuracy of fault warning and optimizes the consumption of computing resources

[0062] When traditional methods are used to process the SCADA system of a wind turbine, the signal noise is large and the consumption of computing resources is high, which easily leads to misjudgment and delay in early fault warning. The present invention uses a graph neural network to perform adaptive modeling and noise smoothing on the SCADA time-series data, which can effectively filter signal noise and enhance the effectiveness and stability of the data. In addition, by combining the self-attention mechanism and the embedded network optimization method, the present invention significantly optimizes the resource consumption in the process of fault warning calculation, reduces the system's dependence on computing resources, and meets the real-time online monitoring and fault warning requirements of large wind turbines.

[0063] (3) Improves the interpretability and generalization ability of the fault warning results

[0064] By introducing the graph self-attention mechanism, the present invention smoothly captures local and global features while modeling the dynamic relationships of sensors, avoiding the "black box" problem in traditional deep learning methods. By predicting the future state of sensors and dynamically comparing it with the actual operation data, this method can identify potential faults in real time and provide quantifiable and interpretable fault warning results. This design not only enhances the interpretability of the model but also improves the generalization ability of the model in different operating environments of wind turbines, ensuring the reliability and applicability of the warning results.

[0065] (4) Enhances the robustness to non-stationary time-series data and improves the fault mode recognition ability

[0066] Aiming at the characteristics of non-stationary time-series data in the SCADA system of a wind turbine, the present invention combines multi-layer time-varying network modeling and graph neural network. By dynamically modeling and time-series smoothing of the sensor relationship network, it can effectively cope with the time-invariance within a short time window and the dynamic change characteristics on a long time scale of sensor data. In addition, by adopting an adaptive fault threshold mechanism, accurate identification and dynamic response of fault signals are realized, greatly enhancing the robustness and fault mode recognition ability of the model when processing non-stationary time-series data. Brief Description of the Drawings

[0067] Figure 1 It is a schematic diagram of the framework of the wind turbine operating status fault warning method based on the multi-scale adaptive graph neural network of the present invention;

[0068] Figure 2 It is a schematic diagram of the double-layer sliding window strategy of the present invention. Specific embodiments

[0069] The present invention will be further described below in conjunction with embodiments, but it should not be understood that the above-mentioned subject matter scope of the present invention is limited to the following embodiments. Without departing from the above technical idea of the present invention, various substitutions and changes made according to the common general knowledge and conventional means in the art shall be included within the protection scope of the present invention.

[0070] Embodiment 1:

[0071] Refer to Figures 1 to 2 , a wind turbine fault warning method based on a multi-modal multi-scale adaptive graph neural network, comprising the following steps:

[0072] 1) Construct a multi-scale time-varying network for capturing the operation fluctuations of wind turbines;

[0073] 2) Obtain the SCADA multi-dimensional time series data X (orig) , and input it into the multi-scale time-varying network to obtain a time-varying graph with multiple time scales and a set of time-varying graph node features;

[0074] 3) Use a sensing feature prediction model based on a graph convolutional network and a self-attention mechanism to process the time-varying graph structure to obtain the predicted values of the SCADA sensor nodes of the wind turbine in the future T p time;

[0075] 4) Based on the predicted values of the SCADA sensor nodes of the wind turbine and the measured values of the SCADA sensor nodes of the wind turbine, determine whether the wind turbine has a fault.

[0076] The multi-scale time-varying network uses a double-layer sliding window to capture the time correlation of the SCADA multi-dimensional time series data X (orig) ;

[0077] The double-layer sliding window includes a first-level sliding window and a second-level sliding window;

[0078] The sampling length of the first-level sliding window is w1, and the sliding step is 1;

[0079] The sampling length of the second-level sliding window is w2.

[0080] The first-level sliding window samples the original data set of the SCADA multi-dimensional time series of the wind turbine , and the sampled data set is denoted as n is the number of sensors, and T is the sampling duration;

[0081] Among them, the dataset is as follows:

[0082]

[0083] The secondary sliding window samples the dataset, and the sampled dataset is denoted as X2;

[0084] The dataset is shown in the following formula:

[0085]

[0086] For the time point t, the time-varying graph structure G t =(V, E t ); V is the set of sensor nodes; E t ={E t (1, 2), E t (2, 3),..., E t (V - 1, V)} is the set of node relationships at time t; each node of the time-varying graph structure stores SCADA feature data;

[0087] Among them, the node relationship E t (i, j) between the i-th and j-th sensors is as follows:

[0088]

[0089] Among them, and are the time series data of the i-th and j-th sensors under the secondary sliding window respectively, and f is the dynamic correlation function.

[0090] The dynamic correlation function is obtained by using the Pearson correlation coefficient, mutual information, or dynamic weight learning based on a graph neural network.

[0091] The time-varying graph node feature set H is as follows:

[0092]

[0093] Among them, h i represents the feature vector of sensor i in the current time window t, and d is the dimension of the node features.

[0094] The predicted values of the SCADA sensor nodes of the wind turbine in the future T p time are as follows:

[0095]

[0096] Among them, L is the number of network layers; is the weight calculated by the self-attention mechanism; are the weight and the bias term; f is the activation function; τ = 1, 2,..., T p ; T p is the time step of multi-step prediction; is the predicted value of sensor i at the future time t + τ; represents the feature vector of sensor i in the current time window t;

[0097] Among them, the weight is as follows:

[0098]

[0099] In the formula, W (l) , W (0) are the feature transformation matrices between the layer and the initial feature; a is the learnable attention vector; the superscript T represents the transpose; is the initial node feature; W (k ) is the feature transformation matrix of the k-th layer feature; is the node feature of the k-th layer;

[0100] The node feature is as follows:

[0101]

[0102] In the formula, W (l) is the weight matrix, represents vector concatenation, b (l) is the bias constant term; σ(·) is the non-linear activation function; is the node feature of the (l - 1)-th layer;

[0103] The aggregated feature is as follows:

[0104]

[0105] In the formula, and represent the feature information of nodes v and u in the (l - 1)-th layer. AGG (l) is the feature aggregation operation.

[0106] The sensing feature prediction model is obtained by training with historical data;

[0107] During the training process, the loss function adopted by the sensing feature prediction model is as follows:

[0108]

[0109] Wherein, L MSE is the loss function. is the predicted value of sensor i at the future time t + τ; is the measured value of sensor i at the future time t + τ. T p is the time step of multi-step prediction; n is the number of sensors.

[0110] In step 4), the index for judging whether the wind turbine has a fault is the anomaly score f i ;

[0111] The anomaly score f i is as follows:

[0112]

[0113] where f i ∈[0, 1] is the fault anomaly score of the system at the current moment. Where is the prediction error of the i-th sensor at time (t + τ). is the predicted value of sensor i at the future time t + τ; is the measured value of sensor i at the future time t + τ.

[0114] When the fault anomaly score f i > θ t , the wind turbine has a fault risk and triggers a fault warning signal;

[0115] wherein, the threshold θ t is as follows:

[0116] θ t = μ + λσ (12)

[0117] where μ and σ are respectively the mean and standard deviation of historical anomaly scores; λ is a dynamic parameter for controlling the warning sensitivity.

[0118] Embodiment 2:

[0119] A wind turbine fault warning method based on a multi-modal multi-scale adaptive graph neural network, comprising the following steps:

[0120] 1) Construct a multi-scale time-varying network for capturing the operation fluctuations of the wind turbine;

[0121] 2) Obtain the SCADA multi-dimensional time series data X (orig) , and input it into the multi-scale time-varying network to obtain a time-varying graph with multiple time scales and a set of time-varying graph node features;

[0122] 3) Process the time-varying graph structure using a sensing feature prediction model based on a graph convolutional network and a self-attention mechanism to obtain the predicted values of the SCADA sensor nodes of the wind turbine for the next T p time;

[0123] 4) Based on the predicted values of the SCADA sensor nodes of the wind turbine and the measured values of the SCADA sensor nodes of the wind turbine, determine whether the wind turbine has failed.

[0124] Example 3:

[0125] A wind turbine fault warning method based on a multi-modal multi-scale adaptive graph neural network, the technical content is the same as that of Example 2. Further, the multi-scale time-varying network uses a double-layer sliding window to capture the temporal correlation of the SCADA multi-dimensional time series data X (orig) ;

[0126] The double-layer sliding window includes a first-level sliding window and a second-level sliding window;

[0127] The sampling length of the first-level sliding window is w1, and the sliding step is 1;

[0128] The sampling length of the second-level sliding window is w2.

[0129] Example 4:

[0130] A wind turbine fault warning method based on a multi-modal multi-scale adaptive graph neural network, the technical content is the same as any one of Examples 2-3. Further, the first-level sliding window samples the original dataset of the SCADA multi-dimensional time series of the wind turbine and the sampled dataset is denoted as

[0131] where the dataset is as follows:

[0132]

[0133] The second-level sliding window samples the dataset, and the sampled dataset is denoted as X2;

[0134] The dataset is shown by the following formula:

[0135]

[0136] Example 5:

[0137] A wind turbine fault warning method based on a multi-modal multi-scale adaptive graph neural network, the technical content is the same as any one of Examples 2-4. Further, for time point t, the time-varying graph structure G t =(V, Et ); V is the set of sensor nodes; E t = {E t (1, 2), E t (2, 3),..., E t (V - 1, V)} is the set of node relationships at time t; each node in the time-varying graph structure stores SCADA feature data;

[0138] Among them, the node relationship E t (i, j) between the i-th and j-th sensors is as follows:

[0139]

[0140] Among them, and are the time series data of the i-th and j-th sensors under the secondary sliding window respectively, and f is the dynamic correlation function.

[0141] Example 6:

[0142] A wind turbine fault warning method based on a multi-modal multi-scale adaptive graph neural network, the technical content is the same as any one of Examples 2 - 5. Further, the dynamic correlation function is obtained by using the Pearson correlation coefficient, mutual information, or dynamic weight learning based on the graph neural network.

[0143] Example 7:

[0144] A wind turbine fault warning method based on a multi-modal multi-scale adaptive graph neural network, the technical content is the same as any one of Examples 2 - 6. Further, the time-varying graph node feature set is as follows:

[0145]

[0146] Among them, h i represents the feature vector of sensor i in the current time window t, and d is the dimension of the node feature.

[0147] Example 8:

[0148] A wind turbine fault warning method based on a multi-modal multi-scale adaptive graph neural network, the technical content is the same as any one of Examples 2 - 7. Further, the predicted values of the SCADA sensor nodes of the wind turbine in the future T p time are as follows:

[0149]

[0150] Among them, L is the number of network layers; α i (l) is the weight calculated through the self-attention mechanism; are weights and bias terms; f is the activation function; τ = 1, 2,..., T p ; T p is the time step of multi-step prediction; is the predicted value of sensor i at the future time t + τ;

[0151] Among them, the weights are as follows:

[0152]

[0153] In the formula, W (l) , W (0) are the feature transformation matrices between the layer and the initial features; a is the learnable attention vector;

[0154] The node features are as follows:

[0155]

[0156] In the formula, W (l) is the weight matrix, represents vector concatenation, b (l) is the bias constant term; σ(·) is the non-linear activation function;

[0157] The aggregated features are as follows:

[0158]

[0159] In the formula, and represent the feature information of nodes v and u in the (l - 1)-th layer.

[0160] Example 9:

[0161] A wind turbine fault warning method based on a multi-modal multi-scale adaptive graph neural network, the technical content is the same as any one of Examples 2 - 8. Further, the sensing feature prediction model is trained through historical data;

[0162] During the training process, the loss function adopted by the sensing feature prediction model is as follows:

[0163]

[0164] In the formula, L MSE is the loss function.

[0165] Example 10:

[0166] A wind turbine fault warning method based on a multi-modal multi-scale adaptive graph neural network, the technical content is the same as any one of Embodiments 2-9. Further, in step 4), the index for judging whether a wind turbine has a fault is the abnormal score f i ;

[0167] The abnormal score f i is as follows:

[0168]

[0169] where f i ∈[0,1] is the fault abnormality score of the system at the current moment. Among them is the prediction error of the i-th sensor at time (t+τ).

[0170] Embodiment 11:

[0171] A wind turbine fault warning method based on a multi-modal multi-scale adaptive graph neural network, the technical content is the same as any one of Embodiments 2-10. Further, when the fault abnormality score f i >θ t the wind turbine has a fault risk and triggers a fault warning signal;

[0172] where the threshold θ t is as follows:

[0173] θ t =μ + λσ (12)

[0174] where μ and σ are the mean and standard deviation of the historical abnormal scores respectively; λ is a dynamic parameter for controlling the warning sensitivity.

[0175] Embodiment 12:

[0176] A wind turbine fault warning method based on a multi-modal multi-scale adaptive graph neural network. This method captures the dynamic relationships of SCADA sensors in the normal operating state through an adaptive graph neural network and multi-modal data fusion technology, and predicts and warns of faults based on these relationships. This warning framework is mainly divided into four core modules, as Figure 1 shown, specifically including the following four steps:

[0177] (1) Multi-scale time-varying network modeling;

[0178] (2) Graph convolutional self-attention mechanism feature fusion;

[0179] (3) SCADA sensor behavior prediction and evolution;

[0180] (4) Adaptive fault warning and response.

[0181] First, a multi-scale time-varying network modeling module is adopted. By combining a dynamic sliding window and an adaptive time scale, the system can simultaneously capture short-term fluctuations and long-term trends during the operation of wind turbines. On this basis, by combining the graph convolutional network (GCN) with the self-attention mechanism, the evolution process of sensors at different time scales is accurately modeled, and the sensor relationship network is dynamically adjusted to avoid information loss caused by a fixed time scale. Then, the system extracts and fuses multi-modal sensor data through the graph convolutional self-attention mechanism (GC-SAM). It can not only effectively capture the dynamic relationships between sensors but also automatically assign adaptive weights to sensors according to different scenarios, thereby enhancing the sensitivity and accuracy of fault warning. In addition, the system also introduces external meteorological information and equipment maintenance historical data to enhance the prediction ability for potential faults and the adaptability to complex environments. Subsequently, based on the GCN and the self-attention mechanism, the system makes multi-step predictions on the behavior of sensors, and based on the evolution law of the time-varying graph network, identifies potential fault patterns and issues early warning signals in advance. By predicting the future device state, the system can identify possible fault risks in advance and issue early warnings in a timely manner, thus providing decision-making support for maintenance personnel. Finally, the present invention designs an adaptive threshold mechanism to dynamically adjust the sensitivity of fault warning according to the real-time operation state and historical data of wind turbines. Especially when the state of the wind turbine changes violently, it can enhance the warning sensitivity and avoid false alarms and missed alarms. By comparing the predicted values with the actual operation data, the system can identify faults in a timely manner and correct the data of faulty sensors through a self-calibration mechanism to ensure the stability and accuracy of the system.

[0182] 2. Specific Invention Content

[0183] 2.1 Multi-scale Time-varying Network Modeling

[0184] 2.1.1 Double-layer Sliding Window Processing Strategy

[0185] To overcome the limitations of traditional single sliding window methods in capturing the temporal correlation of sensor data, especially the problem that the empirical covariance matrix cannot be reliably estimated when the data volume is limited, the present invention innovatively designs a double-layer sliding window strategy to accurately capture the dynamic evolution trend of the relationships between sensors in the SCADA system of wind turbines in the time dimension and provide data support for fault warning. The double-layer sliding window reconstructs the multi-dimensional time series data of the SCADA system by introducing two different time scale processing methods: the intra-layer sliding window and the extra-layer sliding window, so as to realize multi-level and multi-time scale data analysis. The specific process is as Figure 2 shown.

[0186] (1) First-level Sliding Window:

[0187] The original dataset of the multi-dimensional time series of the SCADA system of wind turbines is Let \(n\) be the number of sensors and \(T\) be the sampling duration. In the first-level sliding window stage, a dataset is constructed with a sampling length of \(w1\) and a sliding step of 1 where \(X1\) (t) The data at time point \(t\) is represented by \(X\) (org) It is expressed as:

[0188]

[0189] This stage starts from the breadth of the SCAD data structure, providing a sufficient amount of data for the calculation of the empirical covariance and avoiding the singularity problem that may occur in the calculation of the covariance matrix.

[0190] (2) Two-level sliding window:

[0191] Based on the data obtained from the first-level window, the second-level sliding window has a length of \(w2\) and a step of 1. From the perspective of the depth of the data structure, it will further analyze the time-level correlation between sensors. This process will provide richer association information for mining the relationship between sensors during the short-term operation of wind turbines. This process further performs vertical connection on the window data sequence of length \(w1\) by applying a sliding window of length \(w2\) to obtain the dataset \(X2\). The dataset \(X1\) represents the time point \(t\) As shown in the following formula:

[0192]

[0193] The data matrix \(X2\) processed by the double-layer sliding window strategy not only provides richer time series information for wind turbines, but also enhances the data structure through the method of vertical superposition. The data within each window is no longer independent, but is correlated with the data in other time windows. The double-layer sliding window strategy can observe the dynamic trend of sensor data changing over time and provides a deeper data foundation for the subsequent multi-layer time-varying network.

[0194] 2.1.2 Construction of multi-layer wind turbine time-varying network

[0195] The present invention proposes a multi-layer time-varying network modeling method based on graph neural network (GNN) for the complex time series data of the wind turbine SCADA system. By combining the double-layer sliding window strategy with multi-scale time series data, the system can accurately capture the evolution of the dynamic relationship between sensors and provide multi-level and time-varying network modeling support for fault warning.

[0196] (1) Modeling principle of multi-layer time-varying network

[0197] During the operation of a wind turbine, the sensor data in the SCADA system exhibits obvious time-evolution characteristics. In the present invention, the SCADA data is mapped into a time-series graph structure Gt = (V, Et) to model the relationships between sensors at different time points t, where:

[0198] V is the set of sensor nodes, representing all sensors in the system;

[0199] Et is the set of node relationships at time t, reflecting the time-varying correlation strength between sensors.

[0200] The relationship weights between sensor nodes are calculated through a dynamic correlation function f, which is specifically defined as:

[0201]

[0202] where and are the time-series data of the i-th and j-th sensors under the secondary sliding window respectively, and f can adopt the Pearson correlation coefficient, mutual information, or dynamic weight learning based on a graph neural network.

[0203] (2) Construction process of the multi-scale time-varying network

[0204] Input data preparation: The SCADA multi-dimensional time-series data X (orig) is reconstructed through a double-layer sliding window to obtain multi-scale time-window data and which provides a basis for the construction of the time-varying network.

[0205] Single-moment network modeling: For each time point t, a corresponding graph structure Gt is constructed. Through a graph neural network (GNN), the system performs feature modeling on the nodes and edges of Gt to capture the dynamic relationships between sensors.

[0206] Evolution modeling of the time-varying network: By introducing the time dimension, the system forms a series of time-varying networks G1, G2, …, GT to describe the dynamic evolution process of the relationships between sensors over time. Based on the evolution law of the time-varying graph, the system realizes multi-step prediction of sensor states.

[0207] 2.2 Graph convolutional self-attention mechanism feature fusion;

[0208] The present invention proposes a feature fusion method based on the combination of a graph convolutional network (GCN) and a self-attention mechanism (SAM) to capture the dynamic relationships of sensor data in the SCADA system of a wind turbine, and perform feature extraction and adaptive weight assignment according to different time scales and environmental characteristics.

[0209] 2.2.1 Principle and operation steps of the graph convolutional network

[0210] In the present invention, the initial aggregation of node features is achieved through graph convolutional operations (GCN), and the self-attention mechanism is combined to dynamically model the association strength between nodes, enabling the system to efficiently aggregate and update node features in both the time and space dimensions. The specific steps are as follows:

[0211] (1) Feature Aggregation: In each layer of the network, nodes collect the feature information of neighboring nodes through graph convolutional operations (GCN) and aggregate it with their own features. In an undirected graph, the feature aggregation of node v is represented as follows:

[0212]

[0213] Equation (4) represents the update process of each node v at the l-th layer, and represent the feature representations of nodes v and u at the (l - 1)-th layer. This step realizes the initial aggregation of node features through the information interaction between nodes and their neighbors, and gradually extracts the structural information of the entire graph.

[0214] (2) Information Update: After node v aggregates the information of its neighbors, it combines the information with its own features to update its state. In an undirected graph, not only the information of adjacent nodes is considered, but also the historical state of the node itself may be included. This process is jointly represented by the previous state information and the aggregated feature as follows:

[0215]

[0216] (3) Non-linear Transformation: To improve the non-linear expression ability of the model, the updated node feature will be transformed by applying a non-linear activation function σ(·), such as ReLU or tanh, so as to effectively capture and model the complex relationships between node features, as shown in the following equation:

[0217]

[0218] In Equation (6), W (l) is the weight matrix, denotes vector concatenation, and b (l) is the bias constant term. This process concatenates the previous state of the node itself and the information of the aggregated neighbors and performs a non-linear operation, enabling the model to capture sudden or unconventional operating state changes of the wind turbine.

[0219] (4) Message Propagation: Ensures the propagation of node features between different network layers, enabling each node to obtain information from its neighbors, neighbors' neighbors, etc., and then through a fully connected layer, performing related tasks such as classification, regression, prediction, etc.

[0220] 2.2.2 Graph Node-Level Feature Extraction

[0221] During the node-level feature extraction process, in order to further capture the dynamic correlation relationships between sensor nodes and considering external environmental factors, the present invention integrates multi-modal data such as meteorological data (such as wind speed, temperature) and equipment maintenance history into the graph structure.

[0222] Before training the graph convolutional network, randomly initialize the nw2 sensor nodes in the time window w2 to generate an embedding vector as which will then be trained and updated along with the message propagation process of the GNN model.

[0223]

[0224] The above formula includes the influence degree of SCADA sensors on the current relationship network layer and other sensors in the cross-network layer, where is a trainable weight matrix, d is the dimension transformed by the sensor variable, and represents the feature representation for each sensor. represents the self-influence degree of sensor node i in the s-th layer, is the influence degree of sensor node i in the s-th layer on sensor node j in other r-th layers. This cross-layer attention mechanism helps the GNN model capture the dependency relationships of sensors at different time levels, and both are scalars. According to the matrix multiplication rule, so transforms the nw2 sensor nodes into d-dimensional feature representations. Therefore, according to Equation (5), the node-level embedding representation can be obtained. In Equation (8), the calculation process of the coefficient γ of the intra-layer and inter-layer self-attention learning term is:

[0225]

[0226] In Equation (8), through the cross-time self-attention mechanism, weights are dynamically assigned to fuse the current features and historical features of the nodes, capturing the evolutionary relationships of sensor nodes in the time dimension. Among them is the input sample data of the SCADA system in the t time window; represents the vector concatenation symbol, used to connect the sensor initial embedding vector and the transformed features in the input sample data and Are respectively represented as the linear combination of the sensor node i in the s layer, the remaining layers, and the input embedding vector.

[0227]

[0228] In the above formula, is the combined feature of the sensor node calculated according to formula (8) and Use the LeakyReLU function to increase the non-linear expression ability of the model, and perform normalization to obtain the final node attention coefficients of each layer's network diagram and

[0229] 2.3 SCADA Sensor Behavior Prediction and Evolution

[0230] In the node-level feature extraction of Figure 2.2.2, we extracted the multi-layer time-varying network features of SCADA sensor nodes through the intra-layer and inter-layer self-attention mechanisms. Based on these node features, this section designs a multi-step prediction and evolution mechanism for sensor behavior to achieve accurate prediction of the future state of SCADA system sensors and provide important data support for fault warning.

[0231] (1) Prediction Target and Node Feature Representation

[0232] The node features extracted from the time-varying network of wind turbine SCADA sensors are represented as:

[0233]

[0234] Where:

[0235] h i Represents the feature vector of sensor i in the current time window t, and d is the dimension of the node feature;

[0236] H is the set of features of all sensor nodes.

[0237] (2) Sensor Behavior Modeling Based on Multi-step Prediction

[0238] Based on the node feature H, the present invention introduces a multi-step time series prediction mechanism to predict the future behavior of wind turbine SCADA sensor nodes.

[0239] ① Time Series Prediction Model

[0240] The future behavior of node i Can be modeled through the current node feature h i And the network embedding vector, and is represented as:

[0241]

[0242] Wherein:

[0243] The predicted value of sensor i at the future time t+τ;

[0244] W out and b out : The weight matrix and bias term of the fully connected network;

[0245] f: Activation function;

[0246] T p : The time step of multi-step prediction.

[0247] ② Loss function design

[0248] To improve the prediction accuracy, the present invention adopts the mean square error (MSE) as the loss function, which is defined as follows:

[0249]

[0250] By minimizing the loss function L MSE , the sensor behavior prediction model can accurately fit the time evolution trend of the SCADA system sensors.

[0251] (3) Node feature evolution and multi-layer fusion

[0252] To further improve the accuracy of the prediction results, the model combines the evolution results of the multi-layer time-varying network node features and adopts the self-attention mechanism to dynamically fuse the prediction outputs of different levels. The final fused prediction value is expressed as follows:

[0253]

[0254] Wherein:

[0255] L: The number of layers of the time-varying network;

[0256] The weight calculated by the self-attention mechanism, indicating the contribution degree of the node features of the l-th layer to the prediction:

[0257]

[0258] W (l) ,W (0) : The feature transformation matrix between the layer and the initial feature;

[0259] a: Learnable attention vector;

[0260] By fusing the node features of the multi-layer time-varying network, the model can further improve the accuracy and robustness of the prediction results.

[0261] 2.4 Adaptive fault warning and response

[0262] Based on the SCADA sensor behavior prediction model, this invention designs an adaptive fault warning mechanism by comparing the predicted behavior and the actual behavior of the sensors, so as to achieve dynamic detection and response to the fault status of wind turbines.

[0263] (1) Fault determination and abnormal score calculation

[0264] By comparing the predicted value of the sensor with the actual measured value calculate the error Error, which is defined as:

[0265]

[0266] where is the prediction error of the i-th sensor at time (t + τ).

[0267] To eliminate the difference in error scales between different sensors, this invention adopts a normalization processing method to normalize the prediction errors of all sensors into an abnormal score f i :

[0268]

[0269] where f t ∈[0,1] is the fault abnormal score of the system at the current moment.

[0270] (2) Adaptive fault warning mechanism

[0271] According to the fault abnormal score f i , design an adaptive threshold θ t , and dynamically adjust the fault warning sensitivity:

[0272] When f t > θ t , the system determines that the sensor has a fault risk and triggers a fault warning signal;

[0273] When f t ≤θ t , the system is in a normal state.

[0274] The threshold θ t is dynamically adjusted according to the historical behavior and current state of the sensor, and is specifically defined as:

[0275] θ t = μ + λσ (18)

[0276] where:

[0277] μ and σ are the mean and standard deviation of the historical abnormal scores respectively;

[0278] λ is a dynamic parameter for controlling the warning sensitivity.

[0279] In summary, the specific method framework of the present invention is as follows:

[0280] (1) Combination of multiscale time-varying network and adaptive graph neural network: To address the complex time-varying coupling relationships among sensors in wind turbines, the present invention proposes a framework that combines a multiscale time-varying network (Multiscale Temporal Network) and an adaptive graph neural network (Adaptive GNN). By constructing multiple time-varying networks at different time scales, this framework can capture both short-term fluctuations and long-term trends during equipment operation, thereby improving the accuracy of fault warning. The system can dynamically select an appropriate time-scale window based on real-time data, accurately model the behavior of each sensor, and adjust the network structure to adapt to the time-varying characteristics of the equipment state, ensuring accurate fault prediction and warning under different operating conditions.

[0281] (2) Multimodal data fusion based on graph convolutional self-attention mechanism: To improve the accuracy of fault warning, the present invention introduces a graph convolutional self-attention mechanism (GC-SAM) for multimodal data fusion. By combining a graph convolutional network (GCN) with a self-attention mechanism, it can effectively capture the complex relationships among sensors and adaptively adjust the weights of each sensor according to different situations. This fusion method enhances the sensitivity and accuracy of fault warning. At the same time, by combining auxiliary data such as meteorological information and equipment maintenance history, the system enhances the prediction ability of the wind turbine operation state, making the warning more accurate in a changing environment.

[0282] (3) Adaptive fault warning mechanism: The present invention designs an adaptive threshold mechanism that can dynamically adjust the warning sensitivity of fault warning according to the real-time operation state and historical data of the wind turbine. Specifically, the system adjusts the threshold in combination with real-time feedback to ensure that the warning sensitivity is increased when the wind turbine state changes significantly, thereby reducing false alarms and missed alarms. At the same time, through the multi-step prediction function of the graph neural network, it can identify potential risks before a fault occurs and issue a warning signal in advance, providing preventive measures for equipment maintenance and operators. If some sensors fail, the system will automatically calibrate the data to ensure the stability and accuracy of the warning system.

Claims

1. A fault warning method for wind turbines based on a multi-modal multi-scale adaptive graph neural network, characterized in that, It includes the following steps: 1) Construct a multi-scale time-varying network for capturing the operation fluctuations of wind turbines; 2) Obtain the SCADA multi-dimensional time series data X (orig) , and input it into the multi-scale time-varying network to obtain the time-varying graphs at multiple time scales and the set of node features of the time-varying graphs; 3) Process the time-varying graph structure using a sensing feature prediction model based on a graph convolutional network and a self-attention mechanism to obtain the predicted values of the SCADA sensor nodes of the wind turbine for the next T p time steps. 4) Based on the predicted values and measured values of the SCADA sensor nodes of the wind turbine, determine whether the wind turbine has a fault.

2. The wind turbine fault warning method based on a multi-modal multi-scale adaptive graph neural network according to claim 1, wherein The multi-scale time-varying network uses a double-layer sliding window to capture the temporal correlation of SCADA multi-dimensional time series data X (orig) ; The double-layer sliding window includes a first-level sliding window and a second-level sliding window; The sampling length of the first-level sliding window is w1, and the sliding step size is 1; The sampling length of the second-level sliding window is w2.

3. The fault warning method for a wind turbine based on a multi-modal multi-scale adaptive graph neural network according to claim 2, characterized in that, The first-level sliding window is applied to the original dataset of the wind turbine SCADA multi-dimensional time series for sampling, and the sampled dataset is denoted as where n is the number of sensors and T is the sampling duration; Among them, the dataset is as follows: The second-level sliding window samples the data set, and the sampled data set is denoted as X2; Dataset As shown in the following formula:

4. A wind turbine fault warning method based on a multi-modal multi-scale adaptive graph neural network according to claim 1, characterized in that For time point t, the time-varying graph structure G t =(V, E t ); V is the set of sensor nodes; E t ={E t (1, 2), E t (2, 3),..., E t (V - 1, V)} is the set of node relationships at time t; each node of the time-varying graph structure stores SCADA feature data; Among them, the node relationship E t t between the i-th and j-th sensors is as follows: wherein, and are the time series data of the i-th and j-th sensors under the secondary sliding window respectively, and f is the dynamic correlation function.

5. The wind turbine fault warning method based on a multi-modal multi-scale adaptive graph neural network according to claim 4, wherein, The dynamic correlation function is obtained by using the Pearson correlation coefficient, mutual information or dynamic weight learning based on a graph neural network.

6. The wind turbine fault warning method based on a multi-modal multi-scale adaptive graph neural network according to claim 1, characterized in that The time-varying graph node feature set H is as follows: where h i represents the feature vector of sensor i, and d is the dimension of node features.

7. A wind turbine fault warning method based on a multi-modal multi-scale adaptive graph neural network according to claim 1, characterized in that, Future T p The predicted values of the SCADA sensor nodes of the wind turbine at time are as follows: Among them, L is the number of network layers; is the weight calculated by the self-attention mechanism; are the weight and the bias term; f is the activation function; τ = 1, 2,..., T p ; T p is the time step of multi-step prediction; is the predicted value of sensor i at the future time t + τ; represents the feature vector of sensor i in the current time window t; Among them, the weight is as follows: Where, W (l) , W (0) is the feature transformation matrix of the initial feature within the layer; a is the learnable attention vector; the superscript T represents the transpose; is the initial node feature; W (k ) is the feature transformation matrix of the k-th layer feature; is the k-th layer node feature; Node features are as follows: where, W (l) is the weight matrix, denotes vector concatenation, b (l) is the bias constant term; σ(·) is the non-linear activation function; is the node feature of the (l-1)-th layer; Aggregate feature As shown below: wherein, and represent the feature information of nodes v and u at the (l-1)-th layer; AGG (l) is a feature aggregation operation.

8. A wind turbine fault warning method based on a multi-modal multi-scale adaptive graph neural network according to claim 1, characterized in that The sensing feature prediction model is trained by historical data; During the training process, the loss function adopted by the sensing feature prediction model is as follows: where L MSE is the loss function; is the predicted value of sensor i at the future time t + τ; is the measured value of sensor i at the future time t + τ; T p is the time step of multi-step prediction; n is the number of sensors.

9. The wind turbine fault warning method based on a multi-modal multi-scale adaptive graph neural network according to claim 1, characterized in that In step 4), the index for judging whether a wind turbine has a fault is the abnormal score f i ; Abnormal score f i As follows: where f i ∈ [0, 1] is the fault anomaly score at the current moment of the system; Error i (t+τ) is the prediction error of the i-th sensor at time (t + τ); is the predicted value of sensor i at the future time t + τ; is the measured value of sensor i at the future time t + τ.

10. A wind turbine fault warning method based on a multi-modal multi-scale adaptive graph neural network according to claim 9, characterized in that, When the fault anomaly score f i > θ t , the wind turbine has a fault risk and triggers a fault warning signal; Among them, the threshold θ t is as follows: θ t = μ + λσ (12) where μ and σ are the mean and standard deviation of the historical anomaly scores respectively; λ is a dynamic parameter for controlling the warning sensitivity.

Images