Sensor drift self-compensation wind turbine generator state prediction method and system

Abstract

The invention discloses a wind turbine generator state prediction method and system for sensor drift self-compensation, and belongs to the technical field of wind turbine generator operation and maintenance. According to the wind turbine generator set state prediction method with sensor drift self-compensation, a sensing data processing model, a sensing drift simulation model, a sensing drift compensation model, a sensing fusion model and a wind power state prediction model are constructed, and spatial correlation and drift non-stationary characteristics of multiple types of sensors are fully considered; and a wind turbine generator state prediction result is obtained by capturing a spatial dependency relationship in a multi-sensor fusion signal sequence and dynamic evolution characteristics of a time sequence, so that the drift distance can be accurately identified and high-reliability prediction input can be generated in a multi-class sensor scene, and the precision of a prediction model can be effectively improved. Therefore, the method can effectively improve the utilization rate of the wind turbine generator data, has high compatibility and real-time performance, and provides technical support for long-term stable operation of the wind turbine generator.

Description

Technical Field

[0001] This invention relates to a method and system for predicting the state of wind turbines with sensor drift self-compensation, belonging to the field of wind turbine operation and maintenance technology. Background Technology

[0002] Chinese Patent (Publication No.: CN119062524A) discloses a method for load inversion and health status monitoring of a hybrid wind turbine based on a dynamic tilt sensor at the top of the turbine. This method first installs a dynamic tilt sensor at the top of the wind turbine to monitor the tilt angle changes in real time. The sensor data is transmitted to the wind turbine's main control system via a wired connection. An automatic calibration algorithm is then used to correct the tilt sensor data, eliminating accumulated errors caused by sensor drift. The real-time absolute tilt angle of the wind turbine is calculated using the calibrated and compensated tilt angle data. Then, based on the real-time absolute tilt angle, an inversion algorithm is used to calculate the actual bending moment load at different interface heights of the wind turbine. By monitoring the actual load on the wind turbine, the ultimate limit state and fatigue state of the wind turbine during operation are analyzed. Finally, the health status of the wind turbine is assessed in real time, and potential structural problems are predicted.

[0003] The above solution can calibrate the data from the tilt sensor, but it does not disclose how to calibrate other types of sensors. Modern wind turbine operation status monitoring systems typically deploy multiple types of sensors, including oil temperature sensors, oil level sensors, nacelle vibration sensors, gas-insulated switchgear (GIS) partial discharge sensors, and transformer oil chromatography sensors. The data from these sensors are used to assess the health status of the equipment in real time, predict potential faults, and provide a basis for operation and maintenance strategies.

[0004] However, during long-term operation, these sensors are often affected by factors such as changes in ambient temperature, electromagnetic interference, aging, and wear, resulting in problems such as zero-point drift, range drift, sudden drift, and slow drift. This causes sensor data to deviate from the true state, thereby reducing the accuracy of state prediction, increasing the false alarm rate, and in severe cases, potentially affecting the safe operation of wind turbines. Therefore, when predicting the state of wind turbines, it is necessary to fully consider the drift problems of various types of sensors.

[0005] However, existing wind turbine state prediction methods, such as moving average, Kalman filtering, and Bayesian estimation, usually ignore the spatial correlation and drift non-stationary characteristics between sensors. Therefore, in multi-sensor scenarios, they cannot accurately identify the drift amount, and thus cannot generate highly reliable prediction inputs, which significantly affects the accuracy of the prediction model.

[0006] The information disclosed in this background section is only for understanding the background of the inventive concept, and therefore may include information that does not constitute prior art. Summary of the Invention

[0007] To address the aforementioned problems, or one of them, the present invention aims to provide a wind turbine state prediction method and system with sensor drift self-compensation. This method constructs a sensor data processing model, a sensor drift simulation model, a sensor drift compensation model, a sensor fusion model, and a wind turbine state prediction model. It fully considers the spatial correlation and non-stationary drift characteristics of multiple types of sensors. Based on the relationship data between sensor nodes, a low-frequency drift index is calculated to quantify the global drift degree of the original sensor observation signals. Adaptive compensation is then performed on the original sensor observation signals according to the low-frequency drift index, resulting in multiple sensor compensation signals. Simultaneously, dynamic weighting coefficients are introduced to adaptively fuse the multiple sensor compensation signals, generating a multi-sensor fused signal sequence. The spatial dependencies and dynamic evolution characteristics of the time series in the multi-sensor fused signal sequence are then captured to obtain the wind turbine state prediction result. This allows for accurate identification of drift amounts and generation of highly reliable prediction inputs in multi-sensor scenarios, thus effectively improving the accuracy of the prediction model.

[0008] To address the aforementioned problems or one of them, the second objective of this invention is to provide a wind turbine state prediction method and system with sensor drift self-compensation. Through data purification, fusion enhancement, and accurate prediction, this method effectively addresses the industry pain points of low state prediction accuracy and high false alarm rate caused by multi-sensor drift in gas-insulated switchgear (GIS) wind turbines. It can improve the utilization rate of wind turbine data, while also possessing strong compatibility and real-time performance, providing technical support for the long-term stable operation of wind turbines.

[0009] To achieve one of the above objectives, the first technical solution of the present invention is as follows: A method for predicting the state of a wind turbine with sensor drift self-compensation includes the following steps: Step 1: Collect raw observation signals from multiple sensors using a pre-created sensor data processing model, and process the raw observation signals to obtain data on the relationships between sensor nodes. Step 2: Using a pre-created sensor drift simulation model, a low-frequency drift index is calculated based on the relationship data between sensor nodes. This index is used to quantify the global drift degree of the original sensor observation signal. Step 3: Based on the pre-created sensor drift compensation model, adaptive compensation is performed on the original sensor observation signal according to the low-frequency drift index to obtain multiple sensor compensation signals. Step 4: Using a pre-created sensor fusion model, dynamic weight coefficients are introduced to adaptively fuse multiple sensor compensation signals to generate a multi-sensor fusion signal sequence. Step 5: Using a pre-created wind power state prediction model, capture the spatial dependence and dynamic evolution characteristics of the time series in the multi-sensor fusion signal sequence to obtain the wind turbine state prediction results.

[0010] This invention constructs a sensor data processing model, a sensor drift simulation model, a sensor drift compensation model, a sensor fusion model, and a wind power state prediction model. It fully considers the spatial correlation and non-stationary drift characteristics of multiple types of sensors. Based on the relationship data between sensor nodes, a low-frequency drift index is calculated to quantify the global drift degree of the original sensor observation signals. Adaptive compensation is then performed on the original sensor observation signals according to the low-frequency drift index, resulting in multiple sensor compensation signals. Simultaneously, dynamic weighting coefficients are introduced to adaptively fuse the multiple sensor compensation signals, generating a multi-sensor fused signal sequence. Finally, the spatial dependencies and dynamic evolution characteristics of the time series in the multi-sensor fused signal sequence are captured to obtain the wind turbine state prediction results. This allows for accurate identification of drift amounts and generation of highly reliable prediction inputs in multi-sensor scenarios, thus effectively improving the accuracy of the prediction model.

[0011] Furthermore, by constructing a wind power condition prediction model, this invention can simultaneously mine the "spatial dependence" (such as the physical correlation between oil temperature and oil pressure sensors) and "temporal dynamic evolution characteristics" (such as the periodic changes of vibration signals) of multi-sensor fusion signals. Compared with traditional prediction models that only focus on the time dimension, this invention can effectively reduce the prediction error of critical states and facilitate troubleshooting by operation and maintenance personnel.

[0012] Furthermore, this invention addresses the industry pain points of low state prediction accuracy and high false alarm rate caused by multi-sensor drift in gas-insulated switchgear and GIS wind turbine units through data purification, fusion enhancement, and accurate prediction. It can effectively improve the utilization rate of wind turbine unit data, while also possessing strong compatibility and real-time performance, providing technical support for the long-term stable operation of wind turbine units.

[0013] As a preferred technical measure: Step one involves acquiring raw observation signals from multiple sensors using a pre-created sensor data processing model, and then processing these raw signals to obtain the relationship data between sensor nodes. The method is as follows: Collect raw observation signals from one or more sensors, and preprocess the raw observation signals to obtain a standardized time series matrix and node feature matrix; All monitoring sensors of the wind turbine are mapped one by one to graph nodes, ensuring that each sensor corresponds to a unique graph node; Create a sensor graph structure based on the sensor's physical topology and statistical relationships; The time series matrix and node feature matrix are mapped onto the sensor graph structure to generate relationship data between sensor nodes.

[0014] As a preferred technical measure: The method for acquiring raw observation signals from one or more sensors and preprocessing these signals to obtain a standardized time series matrix and node feature matrix is ​​as follows: The system collects raw observation signals from multiple sensors in real time through the interface of the wind turbine monitoring and data acquisition system. These signals include at least oil temperature signal, oil quantity signal, nacelle vibration signal, partial discharge signal of gas-insulated switchgear, and transformer oil chromatography signal. The sampling frequency is set to once every Y minutes to generate a multidimensional time series vector. Outliers are removed from the raw observation signals of multiple types of sensors to obtain several preliminary raw sensor observation signals; Each initial sensor raw observation signal is scaled to the [0,1] interval to obtain normalized sensor data; Linear interpolation is performed on the missing values ​​in the normalized sensor data to obtain standard sensor data. Calculate the mean, standard deviation, and skewness of standard sensor data; Based on the mean, standard deviation, and skewness, a 5-dimensional feature vector is generated for each sensor within a sliding window, resulting in a standardized node feature matrix. A standardized time series matrix is ​​constructed based on the multidimensional time series vectors corresponding to the node feature matrices.

[0015] As a preferred technical measure: The method for creating a sensor graph structure based on the sensor's physical topology and statistical relationships is as follows: Step 11: Obtain the physical topology and statistical relationships of the sensors, including the sensor installation location relationships and signal variation patterns; Based on the sensor installation location relationship and signal change pattern, a weighted fusion formula is used to calculate the edge weight between any two nodes; Step 12: Based on the time-varying characteristics of the original sensor observation signal and the latest sliding window data, update the edge weights in real time to obtain new edge weights to reflect the latest correlations; at the same time, set a smoothing coefficient and introduce a smoothing formula to fuse the new edge weights and the old edge weights to obtain the fused edge weights corresponding to each edge. Step 13: Based on the edge weight threshold, judge the fusion edge weights and filter out the fusion edge weights that do not meet the edge weight threshold to obtain the fusion edge weights to be removed. Step 14: Based on the fusion edge weights to be removed, delete the corresponding edges to generate a sparse graph structure; Step 15: Based on the type of sensor, create one or more hierarchical subgraphs on the sparse graph structure, and create a sensor graph structure by connecting the key nodes of different hierarchical subgraphs across graph edges.

[0016] As a preferred technical measure: Step two, using a pre-created sensor drift simulation model, calculates the low-frequency drift index based on the relationship data between sensor nodes as follows: Acquire the relationship data between sensor nodes, which is a dynamic adjacency matrix representing the relationship between sensor nodes; Calculate the graph Laplacian matrix based on the dynamic adjacency matrix; The Laplacian matrix is ​​then subjected to eigenvalue decomposition to obtain the eigenvector matrix; The eigenvector matrix is ​​processed using a graph spectrum projection mechanism to obtain the frequency domain representation of the sensor, which includes low-frequency and high-frequency components. Low-frequency components correspond to small eigenvalues ​​and are used to reflect the global drift trend of the entire sensor network; high-frequency components correspond to large eigenvalues ​​and are used to reflect local rapid disturbances or noise. Then, the trend of low-frequency components over time is analyzed, and the low-frequency drift index is calculated.

[0017] When the low-frequency drift index exceeds the preset empirical threshold, it is determined that the sensor has drifted. Low-frequency drift indices can be used to distinguish between long-term slow drift dominated by low frequencies and short-term disturbances dominated by high frequencies, thereby identifying drift anomalies in oil temperature, oil quantity, engine room vibration, and gas-insulated switchgear. They can also be used to dynamically adjust the edge weights of graph structures and strengthen the weight of key nodes in graph signal analysis.

[0018] This invention analyzes the relationship data between sensor nodes and decomposes the signal into "low-frequency global drift" and "high-frequency local disturbance". It can quantitatively distinguish the zero-point drift and slow drift (such as ±2℃ drift of oil temperature sensor after long-term operation) caused by environmental temperature, electromagnetic interference and aging of various sensors such as oil temperature, oil quantity and engine room vibration, and the dynamic signal of normal equipment operation (such as instantaneous fluctuation of engine room vibration). It solves the problem that traditional methods cannot identify non-stationary drift and mistake drift for abnormality, and effectively improves the accuracy of drift judgment.

[0019] As a preferred technical measure: Step 3: Based on the pre-created sensor drift compensation model, adaptive compensation is performed on the original sensor observation signals according to the low-frequency drift index. The method for obtaining multiple sensor compensation signals is as follows: Step 31: Determine the low-frequency drift estimation component based on the original sensor observation signal; Based on the recursive estimation algorithm, the identity matrix, and the initial confidence constant, the covariance matrix is ​​calculated to reflect the uncertainty of parameter estimation; and the weight vector is determined based on the response of graph nodes to low-frequency modes. Based on the low-frequency drift index, the graph spectrum confidence weight is calculated and used to take a small value when the node drift is large in order to suppress over-update; Step 32: Calculate the adaptive gain vector based on the graph spectrum confidence weight, low-frequency drift estimation component, covariance matrix, and decay rate of historical data, to determine the parameter update magnitude at the current moment; Step 33: Based on the original observed signal, the low-frequency drift estimation component and the corresponding weight vector, calculate the prediction error of the node signal, which is used to represent the deviation between the actual signal and the low-frequency estimation. To suppress the impact of sudden noise, a Hubble function is introduced to robustly correct the prediction error, resulting in a robust error signal. The Hubble function is used to maintain sensitivity in the small residual range and limit gain in the large residual range. Step 34: Based on the robust error signal and adaptive gain vector, the weight vector is dynamically drift-modeled using robust residuals to obtain the recursively updated weight vector. Based on the low-frequency drift estimation components and the degree of weight decay of past samples, a recursively updated covariance matrix is ​​obtained to reflect the change in estimation confidence. Step 35: Calculate the low-frequency drift estimate based on the updated weight vector and the low-frequency drift estimation component; By subtracting the low-frequency drift estimate from the original observation signal, we finally obtain multiple sensor compensation signals after drift compensation. These signals are used to remove the drift trend that changes slowly over time and retain the high-frequency dynamic information of the real physical process.

[0020] As a preferred technical measure: Step four involves using a pre-created sensor fusion model and introducing dynamic weighting coefficients to adaptively fuse multiple sensor compensation signals, generating a multi-sensor fused signal sequence. The method is as follows: Step 41, based on the low-frequency drift index, calculate the dynamic confidence level, which includes the following: When a sensor drifts, the low-frequency drift index increases, and the corresponding dynamic confidence level automatically decreases to reduce the sensor's impact on the fusion results; conversely, when the signal is stable, the dynamic confidence level increases, giving it a high weight in the global prediction. Step 42: Introduce a graph Laplace smoothing term and set spatial correlation constraints to constrain the confidence distribution, thereby obtaining the smoothed and corrected dynamic confidence. Setting spatial correlation constraints is used to reduce the fluctuation of confidence of adjacent sensors in space, and can stabilize the weight distribution through the neighborhood propagation mechanism when local abnormal sensors drift. Step 43: Adaptively fuse the smoothed and corrected dynamic confidence and multiple sensor compensation signals to generate a multi-sensor fused signal sequence.

[0021] As a preferred technical measure: Step 5: Using a pre-created wind power state prediction model, the spatial dependence and dynamic evolution characteristics of the time series in the multi-sensor fusion signal sequence are captured to obtain the wind turbine state prediction results. The method is as follows: Step 51: Obtain a multi-sensor fusion signal sequence at several time steps, which includes the spatial feature distribution information of sensor nodes and their changes over time. Step 52, using the Laplacian matrix of the sensor network graph The multi-sensor fusion signal sequence is mapped to the graph frequency domain to model spatial correlation. By filtering the graph spectrum, global and local features corresponding to different frequency modes are extracted to obtain the feature vector of each node, so as to capture the dependency relationship between key sensors. Then, based on the Laplacian matrix eigenvector matrix, eigenvalue matrix, spectral filter kernel, and nonlinear activation function, the eigenvectors of the nodes are updated to obtain new eigenvectors for each graph node. Step 53: In the time dimension, perform one-dimensional convolution on the feature vector of each graph node to extract short-term dynamic features, including trend information, mutation information and local fluctuation information, and filter them through the convolution kernel to form the feature representation of the node's evolution over time. Step 54 introduces a time attention mechanism to capture long-term dependencies by calculating the correlation weights between historical time steps; and obtains a weighted output by weighted summation of the time series using a weight matrix. Step 55: Based on the feature representation and weighted output, obtain the state prediction result for future time moments through a fully connected mapping.

[0022] As a preferred technical measure: It also includes a closed-loop self-learning update model, which includes the following: S1, based on the state prediction results and the frequency domain representation of the node signals, calculate the prediction residual to obtain the prediction bias of the sensor; S2, adaptively adjusts the parameters, adjusting the forgetting factor and confidence level in real time according to the prediction deviation; S3. Based on the forgetting factor and confidence level, the sensor graph structure is re-estimated. When the prediction deviation deviates from the threshold, the edge weights are re-estimated using the graph re-estimation step size and the latest signal correlation coefficient between graph nodes, so that the sensor graph structure can adaptively reflect the dynamic dependence changes between sensors. S4 retrains one or more models and performs drift re-detection, updates the prediction model parameters, and recalculates the drift index to automatically correct the model parameters and topological weights, thereby achieving continuous adaptive learning of drift and abnormal states.

[0023] To achieve one of the above objectives, the second technical solution of the present invention is as follows: A wind turbine state prediction system with sensor drift self-compensation, comprising: One or more processing units; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processing units, the one or more processing units implement the above-described method for predicting the state of a wind turbine with sensor drift self-compensation.

[0024] Compared with existing technical solutions, the present invention has the following beneficial effects: This invention constructs a sensor data processing model, a sensor drift simulation model, a sensor drift compensation model, a sensor fusion model, and a wind power state prediction model. It fully considers the spatial correlation and non-stationary drift characteristics of multiple types of sensors. Based on the relationship data between sensor nodes, a low-frequency drift index is calculated to quantify the global drift degree of the original sensor observation signals. Adaptive compensation is then performed on the original sensor observation signals according to the low-frequency drift index, resulting in multiple sensor compensation signals. Simultaneously, dynamic weighting coefficients are introduced to adaptively fuse the multiple sensor compensation signals, generating a multi-sensor fused signal sequence. Finally, the spatial dependencies and dynamic evolution characteristics of the time series in the multi-sensor fused signal sequence are captured to obtain the wind turbine state prediction results. This allows for accurate identification of drift amounts and generation of highly reliable prediction inputs in multi-sensor scenarios, thus effectively improving the accuracy of the prediction model.

[0025] Furthermore, by constructing a wind power condition prediction model, this invention can simultaneously mine the "spatial dependence" (such as the physical correlation between oil temperature and oil pressure sensors) and "temporal dynamic evolution characteristics" (such as the periodic changes of vibration signals) of multi-sensor fusion signals. Compared with traditional neural network models that only focus on the time dimension, this invention can effectively reduce the prediction error of critical states and facilitate troubleshooting by operation and maintenance personnel.

[0026] Furthermore, this invention, through multi-sensor joint modeling, drift identification, and adaptive compensation, can identify various sensor drift types and amplitudes, including zero-point drift, range drift, and sudden drift; and can adaptively correct drift signals to eliminate the effects of environmental disturbances and aging; at the same time, it can generate highly robust multi-sensor fusion signals for predicting key conditions such as oil temperature, oil quantity, engine room vibration, partial discharge of gas-insulated switchgear (GIS), and transformer oil chromatography; thus, applying this invention can provide real-time prediction capabilities for online deployment in industrial settings, improving prediction lead time and accuracy. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the wind turbine state prediction method of the present invention. Figure 2 This is a schematic diagram of a model framework for the wind turbine state prediction method of the present invention; Figure 3 This is another flowchart illustrating the wind turbine state prediction method of the present invention. Detailed Implementation

[0028] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application. This invention covers any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of the invention as defined by the claims.

[0029] like Figure 1 , Figure 2 As shown, this is the first specific embodiment of the wind turbine state prediction method with sensor drift self-compensation of the present invention: A wind turbine state prediction method with sensor drift self-compensation includes the following: Step 1: Collect raw observation signals from multiple sensors using a pre-created sensor data processing model, and process the raw observation signals to obtain data on the relationships between sensor nodes. Step 2: Using a pre-created sensor drift simulation model, a low-frequency drift index is calculated based on the relationship data between sensor nodes. This index is used to quantify the global drift degree of the original sensor observation signal. Step 3: Based on the pre-created sensor drift compensation model, adaptive compensation is performed on the original sensor observation signal according to the low-frequency drift index to obtain multiple sensor compensation signals. Step 4: Using a pre-created sensor fusion model, dynamic weight coefficients are introduced to adaptively fuse multiple sensor compensation signals to generate a multi-sensor fusion signal sequence. Step 5: Using a pre-created wind power state prediction model, capture the spatial dependence and dynamic evolution characteristics of the time series in the multi-sensor fusion signal sequence to obtain the wind turbine state prediction results.

[0030] like Figure 3 As shown, this is the second specific embodiment of the wind turbine state prediction method with sensor drift self-compensation of the present invention: A method for predicting the state of a wind turbine with sensor drift self-compensation includes the following steps: The first step is to collect data and preprocess it, which includes the following: The system uses the SCADA interface of the wind turbine monitoring and data acquisition system to collect signals from multiple sensors in real time, including oil temperature. Oil quantity cabin vibration Gas-insulated switchgear (GIS) partial discharge and transformer oil chromatography The sampling frequency can be set to once per minute to generate a multi-dimensional time series vector. Its expression is as follows: in The total number of sensors, This represents the signal data of the i-th sensor, each... Includes timestamp .

[0031] The preprocessing steps are as follows: Outlier removal: An improved Z-score method is used to handle outliers. If so, then remove it.

[0032] Normalization: The data from each sensor is scaled to the [0,1] interval, and the calculation formula is as follows: Missing value imputation: linear interpolation is used for values ​​less than 5 minutes, and historical averages are used for values ​​greater than 5 minutes.

[0033] Feature creation: Generate a 5-dimensional feature vector for each sensor within a 1-minute sliding window. The calculation formula is as follows: in , These are the mean, standard deviation, and skewness, respectively.

[0034] The final output is standardized data. ,in The length of the sliding window. For feature dimensions.

[0035] The second step is to create a sensor network graph and update it dynamically, which includes the following: Treat each sensor as a graph node , forming a set of nodes Each node is accompanied by a feature vector. This includes standardized signal values, time differences, and statistical characteristics. Based on this, a graph structure is created according to the sensor's physical topology and historical statistical relationships. For any two nodes and , border rights It is calculated by the normalized inner product of the physical connection weights and the signals within the sliding window, and the calculation formula is as follows: in: Represents a node and In the sliding window length The standardized signal vector within; This refers to the vector dot product operation. Representing vectors The Euclidean norm; The physical connection weights between nodes are determined by prior engineering considerations; , is the weighting adjustment coefficient, which controls the contribution ratio of physical prior to statistical correlation, and is set to 0.6.

[0036] The edge weights are updated based on the latest sliding window data, and exponential weighted smoothing is applied to obtain the new edge weights. The calculation formula is as follows: in The smoothness level is controlled by a value of 0.7. for The right to control at any moment.

[0037] For below the threshold Prune the edges to generate a sparse adjacency matrix. This serves as input for graph signal spectrum analysis and sensor drift compensation algorithms. This invention can create hierarchical subgraphs based on sensor type and achieve multi-sensor joint modeling through cross-graph edges, ensuring that key nodes maintain sufficient weight in the graph structure to adapt to the complex multi-dimensional data environment of wind turbines.

[0038] The third step is to analyze the spectrum of the graph signal, and to analyze the node signal at each time step. Mapping to the graph spectral domain to extract global drift and local perturbation information, which includes the following: First, based on the dynamic adjacency matrix Calculate the graph Laplacian matrix The calculation formula is as follows: in: The adjacency matrix generated in the second step; For a degree matrix, the diagonal elements ; It is the unnormalized graph Laplace matrix.

[0039] The Laplacian matrix is ​​then subjected to eigenvalue decomposition, calculated using the following formula: in: For each column of the eigenvector matrix, ... Corresponding to a frequency pattern in a graph; This is an eigenvalue matrix, with the eigenvalues ​​arranged in ascending order. .

[0040] Using graph spectrum projection to project node signals Mapped to the frequency domain, the calculation formula is as follows: in In the frequency domain representation, the low-frequency components (corresponding to smaller eigenvalues) The high-frequency component (corresponding to larger eigenvalues) mainly reflects local rapid disturbances or noise, while the high-frequency component reflects the global drift trend of the entire sensor network.

[0041] To achieve sensor drift self-compensation, the system analyzes the trend of low-frequency components over time and calculates the drift index. The calculation formula is as follows: in: For nodes in the past Low-frequency components at each time step; Used to perform drift correction on node signals, resulting in corrected node signals. The calculation formula is as follows: Through the above compensation, the global drift in the node signal is effectively removed, while high-frequency information is preserved, providing more stable and drift-free data input for subsequent multi-sensor predictions of oil temperature, oil quantity, engine room vibration, etc.

[0042] In addition, the system can set frequency segmentation thresholds according to different sensor types to adaptively divide low-frequency and high-frequency components to adapt to the dynamic characteristics of wind turbines under different operating conditions.

[0043] The fourth step is to detect sensor drift, which includes the following: Based on the graph spectrum representation obtained in the third step Define nodes Low-frequency drift index The formula used to quantify the global drift of a sensor signal is as follows: in: For nodes The Each graph frequency component; This is a set of low-frequency component indices, corresponding to the portion of the graph with smaller Laplacian eigenvalues, used to reflect the global drift trend; This represents the total number of nodes or the total frequency. This indicates that the absolute value is taken to ensure that the drift amplitude calculation is not affected by the positive or negative direction.

[0044] When low-frequency drift index Exceeding the preset experience threshold At that time, determine the sensor The drift phenomenon occurs, and its calculation formula is as follows: threshold Set it within the range of 0.3 to 0.5.

[0045] This indicator can distinguish between long-term slow drift (dominated by low frequency) and short-term disturbances (dominated by high frequency), thus effectively identifying drift anomalies in multi-dimensional data such as oil temperature, oil quantity, engine room vibration, and GIS sensors of gas-insulated switchgear. This provides a reliable basis for subsequent drift self-compensation and multi-sensor prediction. Simultaneously, the drift detection results can be used to dynamically adjust the edge weights of the graph structure, strengthening the weight of key nodes in graph signal analysis and enhancing the system's sensitivity to abnormal signals and predictive robustness.

[0046] The fifth step, based on the GSAF algorithm, is to compensate for the adaptive drift, which includes the following: To suppress low-frequency drift components in sensor node signals, this invention proposes a drift compensation algorithm based on graph spectral adaptive filtering (GSAF). This algorithm introduces a graph spectral confidence adjustment mechanism and the Huber robust estimation criterion within the recursive least squares (RLS) filtering framework, achieving joint suppression of non-Gaussian noise and structural drift, thereby significantly improving the long-term stability and anomaly resistance of multi-sensor systems. Specifically, it includes the following processes: Step (1): Input the original observation signal into the algorithm and perform initialization operations at time... ,node The original observation signal is Its low-frequency drift estimation component is: in, :node Low-frequency drift estimation components in the graph frequency domain.

[0047] The algorithm aims to output the drift-compensated signal. To achieve recursive estimation, the following state variables and parameters are introduced, and their calculation formulas are as follows: in, : An identity matrix of order 1; Initial confidence constant, used to control the initial response amplitude; : Number of low-frequency bases; Weight vector, representing the node Response coefficients to low-frequency modes; The covariance matrix reflects the uncertainty of parameter estimation.

[0048] Spectrum Confidence Weights The calculation formula is as follows: in, Low-frequency drift index.

[0049] Step (2), at each time point The adaptive gain vector is calculated using the following formula: in, Gain vector, used to determine the parameter update magnitude at the current moment; : Spectral confidence Forgetting factor: controls the rate of decay of historical data; a smaller value is taken when node drift is large to suppress over-update; denominator term Used to prevent matrix singularities and control update balance.

[0050] This formula introduces a spectral confidence coefficient on the basis of traditional RLS gain, which enables nodes to automatically reduce their update rate when drift is significant, and enhance their response capability when the node is stable, thereby realizing adaptive gain adjustment based on graph structure.

[0051] Step (3) Calculate the prediction error of the node signal, and the calculation formula is as follows: in, Prediction error represents the deviation between the actual signal and the low-frequency estimate.

[0052] To suppress the impact of sudden noise, the Huber function is introduced for robust correction, and its calculation formula is as follows: in, Error signals that have undergone robust processing; : Sign function, used to maintain the direction of error; Huber threshold: used to identify abnormal residuals.

[0053] when When, maintain linear response; when At the same time, limit the error range to prevent outliers from dominating parameter updates.

[0054] This robust mechanism maintains sensitivity in the small residual range and limits gain in the large residual range, thus balancing sensitivity and stability, and is suitable for non-Gaussian noise environments.

[0055] Step (4) Based on the robust error signal, recursively update the weight vector and covariance matrix, and the calculation formula is as follows: in, Updated low-frequency drift estimation weights; The updated covariance matrix reflects the change in the estimate confidence level; Forgetting factor: Used to control the weight decay of past samples.

[0056] Weight updates utilize robust residuals to model dynamic drift, while covariance updates balance historical memory and rapid response through a forgetting factor. To avoid matrix singularities, numerical corrections can be performed after the update, calculated using the following formula: in, Tiny positive numbers (typically taken as...) This ensures the positive definiteness of the matrix.

[0057] Step (5) Output drift compensation. The final drift compensation signal is calculated as follows: in, : Signal output after drift compensation; :node The low-frequency drift estimate.

[0058] By subtracting the low-frequency estimated components from the original signal, the slowly changing drift trend over time can be effectively removed, while preserving the high-frequency dynamic information of the actual physical process. The compensated signal can be directly input into subsequent anomaly detection or prediction models, significantly improving the stability and reliability of the overall monitoring system.

[0059] Compared to traditional RLS or Kalman filters, the GSAF algorithm of this invention achieves spatial adaptive compensation between nodes by combining a graph spectrum confidence mechanism; it also incorporates the Huber robustness criterion to improve resistance to non-Gaussian noise and sudden drift; at the same time, it has low computational complexity, is easy to deploy online, and is suitable for real-time drift correction in distributed sensor networks.

[0060] The sixth step, to further improve the global consistency and prediction reliability of the signal after drift compensation, proposes a multi-sensor fusion method based on graph confidence weighting. This method introduces dynamic weight coefficients. This enables adaptive fusion of information between various sensor nodes, thereby obtaining a robust, low-drift multi-sensor fusion signal sequence.

[0061] The method for fusing multiple sensors is as follows: Step (1) Set up the fusion model, setting the time as follows: There are a total of The expression for a sensor signal after GSAF drift compensation is as follows: The fused signal is obtained by weighting based on confidence level. The calculation formula is as follows: in, :sensor The signal after drift compensation; The global signal after multi-sensor fusion; Sensor confidence weights reflect the reliability of nodes in a graph structure.

[0062] Step (2) Calculate the dynamic confidence level. The node confidence level is reflected by the low-frequency drift index, which is defined as follows: in: :node The low-frequency drift index represents the percentage of global drift of the sensor in the frequency domain. : To prevent small constants from having a denominator of zero (typically taken as...) ).

[0063] When a node exhibits significant drift Increase, corresponding confidence level Automatically lowering the threshold reduces the node's impact on the fusion result; conversely, when the signal is stable, Increase its weight so that it receives a higher weight in the global prediction.

[0064] Step (3) Set spatial correlation constraints. To avoid drastic fluctuations in the confidence levels of adjacent sensors in space, this invention introduces a graph Laplace smoothing term to constrain the confidence level distribution, thus obtaining the constrained confidence level. The calculation formula is as follows: in: : Confidence vector of all nodes; The normalized Laplace matrix of the graph (derived from the adjacency matrix in step two) create); Smoothing coefficient, used to control the strength of spatial consistency; : Identity matrix.

[0065] This constraint can stabilize the weight distribution through a neighborhood propagation mechanism when local abnormal nodes drift, thereby avoiding overall prediction distortion caused by single-point drift.

[0066] Step (4) outputs the fused signal, the final fused signal. Represented as: in, : Node confidence after graph smoothing correction.

[0067] The weighted fused signal has the following characteristics: Anti-drift: Reduces the impact of highly drifted nodes on the global results; Spatial consistency: State changes among neighboring sensors remain coordinated; Prediction stability: Provides high-confidence input for subsequent Graph Spectrum Time Series Prediction Network (GFTT), reducing short-term oscillations and anomalous responses.

[0068] In summary, the method of this invention realizes confidence-adaptive signal fusion based on graph structure, which is not only robust to drift but also has physical topological consistency, providing a highly reliable data foundation for subsequent state prediction.

[0069] Step 7: For the signal after drift compensation and multi-sensor fusion, this invention proposes a Graph-Frequency Temporal Transformer (GFTT) model to achieve short-term dynamic prediction of key state parameters of wind turbines. This model can simultaneously capture the spatial dependencies between sensors and the dynamic evolution characteristics of the time series. The method for predicting the state using the GFTT model is as follows: (1) Obtain model input data, which is a multi-sensor fusion signal sequence of the past τ time steps. , denoted as: in, This represents the signal from the i-th sensor after drift compensation and fusion. This sequence contains spatial distribution information of node features and their changes over time.

[0070] (2) Create a graph spectrum convolution module and use the Laplacian matrix of the sensor network graph. By mapping node signals to the graph frequency domain, spatial correlation can be modeled. Filtering the graph spectrum allows for the extraction of global and local features corresponding to different frequency modes, thereby capturing the dependencies between key sensors.

[0071] In this module, the feature vector of each node It will be updated through graph spectrum convolution, and its calculation formula is as follows: in, The eigenvector matrix of the Laplacian matrix. For the eigenvalue matrix, For spectral filtering core, This is a non-linear activation function. This formula is used to help understand the convolution process, but the core technique lies in extracting the spatial dependencies between nodes through spectral transformation.

[0072] (3) Create a temporal convolution module. In the time dimension, perform one-dimensional convolution on the historical signals of each node to extract short-term dynamic features, including trends, abrupt changes, and local fluctuations. This is achieved through the convolution kernel. Filtering forms a feature representation of the node's evolution over time. This module ensures that the model remains sensitive to short-term anomalies, rapid changes, or periodic signals.

[0073] (4) Create a temporal attention module, introducing a temporal attention mechanism to capture long-term dependencies by calculating the correlation weights between historical time steps. The model uses a weight matrix... The time series is weighted and summed to obtain a weighted output. The calculation formula is as follows: in, The value matrix represents the time series, and the attention weights are... It automatically reflects the importance of a signal at different times. This mechanism enables the model to dynamically learn long-term trends and cross-time period effects.

[0074] (5) After processing by graph convolution, temporal convolution and attention module, the model obtains the state prediction result at the future time Δt through fully connected mapping. The calculation formula is as follows: in These are the trainable parameters for the output layer. Model training aims to minimize the prediction error, and its calculation formula is as follows: LOSS is the loss function.

[0075] Step 8: To achieve long-term stable operation of the algorithm, this invention establishes a closed-loop self-learning mechanism of drift detection—compensation—prediction—feedback. The system executes the following dynamic update process at each time step. The closed-loop self-learning update mechanism includes the following: (1) Calculate the prediction residual The calculation formula is as follows: in, : No. Each sensor at time Prediction bias.

[0076] (2) The parameters are adaptively adjusted. The forgetting factor and confidence level in the GSAF algorithm are adjusted in real time according to the residual statistics. The calculation formula is as follows: in, : Learning rate coefficient, used for smooth updates; : Represents the instantaneous prediction residual of the i-th sensor at time t, whose value is the actual observation value of the sensor at that time. Fusion signal with model prediction The difference; and : These represent the adaptive forgetting factors of the i-th sensor node at time t and the next time t+1, respectively. As the drift error continues to increase, the system automatically reduces... Enhance the weight of historical drift memory to improve compensation stability.

[0077] (3) The graph structure is re-estimated. When the residual deviates from the threshold for a long period of time, the edge weight is re-estimated. The calculation formula is as follows: in, : The edge weights after re-estimation; : Edge weights before re-estimation; : Step size for graph weight estimation; The latest signal correlation coefficient between nodes. This mechanism ensures that the graph topology can adaptively reflect the dynamic dependencies between sensors.

[0078] (4) Retrain the model and perform drift re-detection, which includes the following: During the offline phase, incremental training is performed periodically to update the prediction model parameters. At the same time, the drift index is recalculated. To maintain the accuracy of the test.

[0079] Through the aforementioned closed-loop feedback mechanism, the system can automatically correct model parameters and topological weights during long-term operation, achieving continuous adaptive learning of drift and abnormal states.

[0080] This invention achieves real-time compensation for low-frequency drift of sensors through graph spectral adaptive filtering (GSAF), and effectively enhances signal robustness and spatial consistency by combining graph-structured multi-sensor fusion and a closed-loop self-learning mechanism. The drift-compensated and fused multi-sensor signals can provide stable input for the GFTT model based on graph spectral convolution and temporal attention, enabling high-precision short-term prediction of key state parameters of wind turbines, with a significantly reduced prediction error compared to traditional methods. Furthermore, this invention is not dependent on specific sensor types or layouts, and has advantages such as high computational efficiency, ease of online deployment, and strong engineering versatility, providing a reliable technical solution for state monitoring and anomaly early warning of wind turbines and similar industrial systems.

[0081] A specific embodiment of using the method of the present invention to predict the state of a coastal wind farm: This embodiment takes a coastal wind farm in southern China as the research object. The wind farm has a total of 70 2.5MW wind turbine generators. Each unit is equipped with approximately 40 key monitoring points, covering sensors for oil temperature, oil pressure, gearbox vibration, wind speed, power generation, main shaft torque, converter temperature, and blade strain, with a total of over 2000 sensors. The data from all sensors are uniformly sampled by the SCADA system, with a sampling period of 1 minute, and uploaded to the edge computing node for real-time analysis via the on-site fiber optic communication network.

[0082] The system needs to automatically identify and compensate for low-frequency drift and sudden anomalies in sensor signals under long-term unattended operation conditions to ensure the stability and accuracy of the wind turbine key parameter prediction model. This specifically includes the following steps: Step 1: Obtain the raw signals from the SCADA monitoring and data acquisition system.

[0083] Step 2: After receiving the raw signals from the SCADA monitoring and data acquisition system, the system first performs data quality checks and preprocessing, which includes the following: Outlier detection and correction: Outliers are identified using a sliding window median method. When data at a certain moment deviates from the median value of the nearest 10-minute window by more than 3 times the median absolute deviation, it is marked as an outlier and corrected using time neighborhood interpolation.

[0084] Missing value handling: If the continuous missing time is less than 5 minutes, linear interpolation is used to fill in the missing value; if it exceeds 5 minutes, the historical average of similar time periods (such as similar wind speed intervals) is introduced to fill in the missing value.

[0085] Feature creation: For each sensor, the mean, standard deviation, skewness, kurtosis, and first-order difference features are calculated within a sliding window (30 minutes in length) to form node feature vectors, providing input for subsequent graph modeling.

[0086] Step 3: In the spatial relationship modeling stage, each sensor is treated as a node in the graph. The edge connections between nodes are determined by two parts: Physical connection relationships: such as setting high initial edge weights between temperature, vibration and power sensors in the same cabin.

[0087] Statistical correlation: Calculate the Pearson correlation coefficient based on historical data from the past 7 days. If the absolute value of the correlation coefficient is greater than 0.6, then a connection is established.

[0088] The edge weights are dynamically updated every 30 minutes. The system adaptively adjusts them based on the relevance of the latest data, and edges below a threshold are automatically pruned to prevent noise propagation. Through this dynamic graph structure, the model can simultaneously reflect the physical dependencies and statistical coupling characteristics within the unit.

[0089] Step 4: For each sensor node, the system quantifies the degree of drift by estimating the proportion of low-frequency energy through spectral decomposition. When the proportion of low-frequency energy exceeds a set threshold (0.25), the sensor is determined to have a drift trend.

[0090] The system then employs an adaptive drift compensation algorithm, using a recursive filtering structure to update the drift estimate in real time. The update of the filter weights is combined with robust residual estimation (Huber function) to suppress the effects of sudden noise, thereby achieving smooth and continuous drift correction.

[0091] The system identifies low-frequency drift components in sensor signals based on graph spectrum decomposition. When the proportion of low-frequency energy exceeds a threshold of 0.25, the system automatically determines that the node has a drift trend. Taking the lubricating oil temperature sensor of Unit 12 as an example, after approximately 90 days of continuous operation, its signal drifted by an average of approximately +2.3℃. After processing by the graph spectrum adaptive filtering (GSAF) algorithm proposed in this invention, the drift was reduced to within ±0.25℃, and the mean square error (MSE) decreased by 88.6% compared to the uncompensated original signal.

[0092] Meanwhile, the compensation effect of the traditional first-order Kalman filter (KF) method only reduced the MSE by about 41.2%, indicating that the GSAF algorithm is significantly better than the traditional method in long-term slow drift correction. For details, please refer to the table below: Table 1 Step 5: To avoid the impact of single sensor drift on the overall prediction accuracy of the system, the system adopts a confidence-weighted multi-sensor fusion strategy. Each sensor is automatically assigned a confidence weight based on its drift index, with sensors exhibiting greater drift having lower weights. The state changes of adjacent sensors are smoothed and constrained through a graph structure, ensuring that the fusion result retains key features while suppressing anomalies.

[0093] Table 2 Comparison of signal stability of typical nodes before and after fusion. Table 2 shows the signal stability comparison results of typical nodes before and after fusion. After confidence-weighted fusion, the average signal noise amplitude decreased by about 38%, and the correlation between key parameters increased by about 0.2, indicating that the fused signal is smoother and has stronger physical consistency.

[0094] Step Six: The fused multi-sensor data is input into the GFTT model for short-term state prediction. This model consists of a graph spectral convolution module, a temporal convolution module, and an attention mechanism module, responsible for capturing spatial dependence, short-term dynamic changes, and long-term temporal correlations, respectively. The model is trained using 60 consecutive days of data from the Nantong pneumatic field, and the prediction targets are the changing trends of nacelle vibration amplitude and main shaft torque over the next 10 minutes.

[0095] Table 3 Comparison of Predictive Performance of Different Models Table 3 shows the comparison results of the prediction performance of different models. It can be seen that the GFTT model of the present invention improves the prediction accuracy by about 38.6% compared with LSTM and improves the prediction stability by about 22% under extreme wind conditions, which can effectively achieve high-precision and robust prediction of the state of wind turbine units.

[0096] Step 7: After six months of actual operation at the Nantong Power Plant, the prediction system designed based on this invention achieved an overall data effectiveness rate of 98.7%, reduced the number of sensor drift alarms by approximately 60%, and improved the response time of the prediction results to actual unit power fluctuations by approximately 3 minutes. Therefore, through this method, the operation and maintenance center can detect abnormal trends in the units in advance, rationally arrange maintenance plans, and significantly reduce the rate of unplanned outages. No false alarms or misreporting were observed during system operation, proving that this method has good engineering practicality and stability.

[0097] A server embodiment applying the method of the present invention: A server comprising: One or more processing units; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processing units, the one or more processing units implement the above-described method for predicting the state of a wind turbine with sensor drift self-compensation.

[0098] The storage device can be internal memory, external memory, cache memory, or other special memory. The processing unit has signal processing capabilities and can be a general-purpose processor, a digital signal processor, an application-specific integrated circuit, an off-the-shelf programmable gate array, or other programmable logic device.

[0099] An embodiment of a device applying the method of the present invention: An electronic device is provided with a computer-readable storage medium on which a computer program is stored. When the program is executed by a processing unit, it implements the above-described method for predicting the state of a wind turbine with sensor drift self-compensation.

[0100] Computer-readable storage media refers to physical carriers capable of storing computer-recognizable data, instructions, or programs. These media must meet the core characteristic of being "readable by a computer" (i.e., the data exists in the form of electrical, magnetic, or optical signals and can be converted into binary information that a computer can process through appropriate devices). The physical carrier can be a magnetic storage medium, optical storage medium, semiconductor storage medium, or other storage media.

[0101] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, optical storage, etc.) containing computer-usable program code.

[0102] The model or module in this application is an object that constitutes an objective description of form and structure through physical or virtual representation. The object is not the same as a physical object, and is not limited to physical or virtual. It can be a data processing function, software program, processing mode, usage method, operation mode, workflow, application process, electronic hardware, circuit module, processing system, system imitation or simulation object.

[0103] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features within the scope of the technology disclosed in the present invention; and these modifications or substitutions will not cause the substance of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. Any modifications or equivalent substitutions that do not deviate from the spirit and scope of the present invention should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A wind turbine state prediction method with sensor drift self-compensation, characterized in that: It comprises the following steps: Step 1: Collect multiple sensor original observation signals through a pre-created sensor data processing model, and process the sensor original observation signals to obtain sensor node relationship data; Step 2: Use a pre-created sensor drift simulation model to calculate a low-frequency drift index based on the sensor node relationship data, which is used to quantify the global drift degree of the sensor original observation signals; Step 3: Based on the pre-created sensor drift compensation model, the low-frequency drift index is used to adaptively compensate the sensor original observation signals to obtain multiple sensor compensation signals; Step 4: Use a pre-created sensor fusion model to adaptively fuse the multiple sensor compensation signals by introducing a dynamic weight coefficient to generate a multi-sensor fusion signal sequence; Step 5: Use a pre-created wind power state prediction model to capture the spatial dependence relationship and dynamic evolution characteristics of the time sequence in the multi-sensor fusion signal sequence to obtain the wind turbine state prediction result. 2.The wind turbine state prediction method with sensor drift self-compensation according to claim 1, characterized in that: Step 1: Collect multiple sensor original observation signals through a pre-created sensor data processing model, and process the sensor original observation signals to obtain sensor node relationship data by the following method: Collect one or more sensor original observation signals and pre-process the sensor original observation signals to obtain a standardized time series matrix and a node feature matrix; Map all monitoring sensors of the wind turbine to graph nodes one by one to ensure that each sensor corresponds to a unique graph node; Create a sensor graph structure according to the physical topology and statistical relationship of the sensors; Map the time series matrix and the node feature matrix to the sensor graph structure to generate the sensor node relationship data. 3.The wind turbine state prediction method with sensor drift self-compensation according to claim 2, characterized in that: The method for collecting one or more sensor original observation signals and pre-processing the sensor original observation signals to obtain a standardized time series matrix and a node feature matrix is as follows: Real-time collect multiple sensor original observation signals through the interface of the wind turbine monitoring and data acquisition system, which at least includes oil temperature signal, oil quantity signal, cabin vibration signal, partial discharge signal of gas insulated switchgear and transformer oil chromatogram signal; Set the sampling frequency to Y minutes once to generate a multi-dimensional time series vector; Remove outliers from the multiple sensor original observation signals to obtain several preliminary sensor original observation signals; Scale each preliminary sensor original observation signal to the interval [0, 1] to obtain normalized sensor data; Linearly interpolate the missing values in the normalized sensor data to obtain standard sensor data; Calculate the mean, standard deviation and skewness of the standard sensor data; Generate a 5-dimensional feature vector for each sensor within a sliding window according to the mean, standard deviation and skewness to obtain a standardized node feature matrix; Based on the multi-dimensional time series vector corresponding to the node feature matrix, a standardized time series matrix is constructed.

4. The wind turbine state prediction method of claim 2, wherein: According to the sensor physical topology and statistical relationship, the method for creating the sensor graph structure is as follows: Step 11: Obtain the sensor physical topology and statistical relationship, which includes the sensor installation position relationship and signal change law; Based on the sensor installation position relationship and signal change law, the edge weight between any two nodes is calculated using a weighted fusion formula; Step 12: According to the time-varying characteristics of the sensor original observation signal and the latest sliding window data, the edge weight is updated in real time to obtain a new edge weight to reflect the latest correlation relationship; At the same time, a smoothing coefficient is set to introduce a smoothing formula to fuse the new edge weight and the old edge weight to obtain a fused edge weight corresponding to each edge; Step 13: According to the edge weight threshold, the fused edge weight is judged to screen out the fused edge weight that does not meet the edge weight threshold to obtain the to-be-removed fused edge weight; Step 14: Based on the to-be-removed fused edge weight, the corresponding edge is deleted to generate a sparse graph structure; Step 15: According to the type of the sensor, one or more layered subgraphs are created on the sparse graph structure, and key nodes of different layered subgraphs are connected through cross-graph edges to create a sensor graph structure.

5. The wind turbine state prediction method of claim 1, wherein: Step two, using the pre-created sensor drift simulation model, based on the sensor node relationship data, the method for calculating the low-frequency drift index is as follows: Obtain the sensor node relationship data, which is a dynamic adjacency matrix representing the relationship between sensor nodes; Calculate the graph Laplacian matrix according to the dynamic adjacency matrix; Then, the Laplacian matrix is decomposed to obtain an eigenvector matrix; The eigenvector matrix is processed using a graph spectrum projection mechanism to obtain a frequency domain representation of the sensor, which includes low-frequency components and high-frequency components; The low-frequency component corresponds to a small eigenvalue, which is used to reflect the global drift trend of the entire sensor network; The high-frequency component corresponds to a large eigenvalue, which is used to reflect local rapid disturbance or noise; Then analyze the trend of the low-frequency component over time to calculate the low-frequency drift index; When the low-frequency drift index exceeds the pre-set empirical threshold, it is determined that the sensor has drifted; The low-frequency drift index can be used to distinguish between long-term slow drift dominated by low frequency and short-term disturbance dominated by high frequency to identify drift anomalies of oil temperature, oil quantity, cabin vibration and gas insulated switchgear, and can also be used to dynamically adjust the edge weight of the graph structure to strengthen the weight of the key nodes in the graph signal analysis.

6. The wind turbine state prediction method of claim 1, wherein: Step three, based on the pre-created sensor drift compensation model, the sensor original observation signal is adaptively compensated according to the low-frequency drift index to obtain a plurality of sensor compensation signals, the method being as follows: Step 31, determine the low-frequency drift estimation component according to the sensor original observation signal; According to the recursive estimation algorithm, the unit matrix and the initial confidence constant, the covariance matrix is calculated to reflect the uncertainty of parameter estimation; And based on the response degree of the low-frequency mode to the graph node, the weight vector is determined; According to the low-frequency drift index, the graph spectrum confidence weight is calculated, which takes a small value when the node drift is large to suppress excessive update; Step 32, according to the graph spectrum confidence weight, the low-frequency drift estimation component, the covariance matrix, and the decay rate of the historical data, the adaptive gain vector is calculated to determine the parameter update amplitude at the current time; Step 33, based on the original observation signal, the low-frequency drift estimation component and the corresponding weight vector, the prediction error of the node signal is calculated to represent the deviation between the actual signal and the low-frequency estimation; In order to suppress the influence of burst noise, Huber function is introduced to robustly correct the prediction error to obtain the robust error signal, and Huber function is used to maintain sensitivity in small residual interval and limit gain in large residual interval; Step 34, based on the robust error signal and the adaptive gain vector, the robust residual is used to dynamically model the weight vector to obtain the recursively updated weight vector; According to the low-frequency drift estimation component and the weight decay degree of the past samples, the recursively updated covariance matrix is obtained to reflect the estimation confidence change; Step 35, according to the updated weight vector and the low-frequency drift estimation component, the low-frequency drift estimation value is calculated; The original observation signal is subtracted from the low-frequency drift estimation value to finally obtain the drift-compensated multiple sensor compensation signals, which are used to remove the slowly changing drift trend over time and retain the high-frequency dynamic information of the real physical process.

7. The wind turbine state prediction method of claim 1, wherein: Step four, using the pre-created sensor fusion model, the multiple sensor compensation signals are adaptively fused by introducing a dynamic weight coefficient to generate a multi-sensor fusion signal sequence, and the method is as follows: Step 41, based on the low-frequency drift index, the dynamic confidence is calculated, which includes the following contents: When the sensor drifts, the low-frequency drift index increases, and the corresponding dynamic confidence automatically decreases to reduce the influence of the sensor on the fusion result; on the contrary, when the signal is stable, the dynamic confidence increases, so that it obtains a high weight in the global prediction; Step 42, the graph Laplace smoothing term is introduced to set the spatial correlation constraint to constrain the confidence distribution and obtain the smoothed dynamic confidence; the spatial correlation constraint is set to reduce the fluctuation of the confidence of adjacent sensors in space, and can stabilize the weight distribution through the neighborhood propagation mechanism when the local abnormal sensor drifts; Step 43, the smoothed dynamic confidence and the multiple sensor compensation signals are adaptively fused to generate a multi-sensor fusion signal sequence.

8. The wind turbine state prediction method of claim 1, wherein: Step five, using the pre-created wind power state prediction model, the spatial dependence relationship and the dynamic evolution characteristics of the time sequence in the multi-sensor fusion signal sequence are captured to obtain the wind turbine state prediction result, and the method is as follows: Step 51, obtain a multi-sensor fusion signal sequence of several time steps, which includes spatial feature distribution information of the sensor nodes and its change over time; Step 52, using the Laplacian matrix of the sensor network graph The multi-sensor fusion signal sequence is mapped to a graph frequency domain, modeling of spatial correlation is realized, and by filtering the graph frequency spectrum, global features and local features corresponding to different frequency modes are extracted, a feature vector of each node is obtained, and the dependence relationship between key sensors is captured. Then update the feature vector of the node based on the Laplacian matrix eigenvector matrix, eigenvalue matrix, spectral filtering kernel and nonlinear activation function, to obtain the new feature vector of each graph node; Step 53, one-dimensional convolution is performed on the feature vector of each graph node in the time dimension to extract short-term dynamic features including trend information, mutation information and local fluctuation information, and filtering is performed through the convolution kernel to form the feature representation of the node evolution over time; Step 54, introduce a time attention mechanism, capture long-term dependencies by calculating the correlation weights between historical time steps; obtain the weighted output by weighting and summing the time series through the weight matrix; Step 55, according to the feature representation and the weighted output, and through full connection mapping, the state prediction result of the future time is obtained.

9. The wind turbine state prediction method for self-compensation of sensor drift according to claim 1, characterized in that: Further comprising a closed-loop self-learning update model, which Comprises the following contents: S1, based on the state prediction result and the frequency domain representation of the node signal, calculate the prediction residual to obtain the prediction deviation of the sensor; S2, adaptively adjust the parameters, and adjust the forgetting factor and the confidence in real time according to the prediction deviation; S3, according to the forgetting factor and the confidence, re-estimate the sensor graph structure, when the prediction deviation deviates from the threshold, use the graph re-estimation step and the latest signal correlation coefficient between the graph nodes to re-estimate the edge weight, obtain the new edge weight, so that the sensor graph structure can adaptively reflect the dynamic dependence change between the sensors; S4, retrain one or more models, and detect drift again, update the prediction model parameters, and recalculate the drift index, so as to automatically correct the model parameters and the topology weight, realize the continuous self-adaptive learning of the drift and the abnormal state.

10. A wind turbine state prediction system for self-compensation of sensor drift, characterized in that: It comprises: One or more processing units; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processing units, the one or more processing units realize the wind turbine state prediction method for self-compensation of sensor drift according to any one of claims 1-9.

Images