A method for early warning of minor faults in pitch systems based on SCADA data analysis

By decoupling using VMD and constructing a dynamic graph topology using DGNN, the problem of extracting minor fault features from SCADA data was solved, enabling high-accuracy early warning and low false alarm rate for pitch systems, supporting predictive maintenance, and reducing operation and maintenance costs.

CN123109968APending Publication Date: 2026-11-17BEIJING HUANENG XINRUI CONTROL TECH +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611356944.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-03
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively extract minute fault characteristics from SCADA data for pitch systems, and traditional methods neglect the complex spatial topological coupling relationships between multidimensional parameters, resulting in low accuracy and high false alarm rates in early warning.

Method used

Variational mode decomposition (VMD) is used to decouple the temporal features in SCADA data. Dynamic graph neural network (DGNN) is combined to construct a dynamic graph topology with multi-dimensional parameters. Fault warning is achieved through graph convolution and gated recurrent unit (GRU).

Benefits of technology

It significantly improves the accuracy of early warning of minor faults in the pitch system, reduces the false alarm rate, supports predictive maintenance, and reduces unplanned downtime and maintenance costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123109968A_ABST
    Figure CN123109968A_ABST
Patent Text Reader

Abstract

A method for early warning of minor faults in a pitch system based on SCADA data analysis includes the following steps: S1: Collect SCADA operating data of the wind turbine pitch system, perform missing value imputation, outlier removal, and standardization preprocessing to obtain a multi-dimensional time-series data matrix; S2: Use Variational Mode Decomposition (VMD) to adaptively decouple key time-series signals, calculate the permutation entropy and spectral centroid of each intrinsic mode function (IMF), and screen and reconstruct high-frequency mechanical characteristic signals; S3: Use the multi-dimensional monitoring parameters of the pitch system as graph nodes, dynamically calculate the edge weights between nodes based on a multi-head self-attention mechanism, and construct a dynamic graph topology and dynamic adjacency matrix; S4: Input the dynamic graph into a Dynamic Graph Neural Network (DGNN), complete spatial feature aggregation through graph convolution, complete time-series feature modeling through a gated recurrent unit (GRU), output the fault probability, and realize early warning of minor faults. This invention can improve the accuracy of early warning of minor faults and reduce the false alarm rate under complex operating conditions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of wind turbine generator condition monitoring and fault early warning, specifically to a method for early warning of minor faults in a pitch system based on SCADA data analysis. Background Technology

[0002] The pitch system of a wind turbine is a core component for controlling the rotor speed and ensuring the safe operation of the unit. Because the pitch system operates in a complex and variable natural environment and is subjected to alternating loads, it is highly susceptible to faults such as bearing wear, gear jamming, and motor overheating. Currently, wind farms mainly rely on operational data collected by SCADA (Supervisory Control and Data Acquisition) systems for condition monitoring. However, the operating conditions of the pitch system are complex, and SCADA data contains a large amount of non-stationary noise caused by wind speed fluctuations and grid disturbances. Early, minor jamming or wear fault characteristics are easily submerged in the fluctuating signals of normal operation. Traditional methods based on threshold alarms or simple machine learning models struggle to effectively extract minor fault characteristics from a background of strong noise and often ignore the complex spatial topological coupling relationships between multi-dimensional parameters of the pitch system (such as pitch angle, motor current, and temperature), resulting in low accuracy and high false alarm rates in early warning systems. Summary of the Invention

[0003] This invention provides a method for early warning of minor faults in a pitch system based on SCADA data analysis, in order to solve the technical problems existing in the prior art.

[0004] To achieve the above objectives, this invention provides a method for early warning of minor faults in a pitch system based on SCADA data analysis, which includes the following steps: S1: Collect SCADA operation data of wind turbine pitch system, perform missing value imputation, outlier removal and standardization preprocessing to obtain a multi-dimensional time series data matrix; S2: Variational Mode Decomposition (VMD) is used to adaptively decouple key time-series signals, calculate the permutation entropy and spectral centroid of each intrinsic mode function (IMF), and screen and reconstruct high-frequency mechanical characteristic signals. S3: Using the multi-dimensional monitoring parameters of the pitch system as graph nodes, dynamically calculate the edge weights between nodes based on the multi-head self-attention mechanism, and construct a dynamic graph topology and dynamic adjacency matrix; S4: Input the dynamic graph into the dynamic graph neural network (DGNN), complete the spatial feature aggregation through graph convolution, complete the temporal feature modeling through the gated recurrent unit (GRU), output the fault probability, and realize the early warning of minor faults.

[0005] In one embodiment of the present invention, optionally, in step S1, the SCADA operating data includes: the given pitch angle of blades 1 / 2 / 3, the actual pitch angle, the pitch motor current, the pitch motor temperature, the pitch inverter temperature, the wind speed, and the generator speed.

[0006] In one embodiment of the present invention, optionally, in step S1, the preprocessing adopts Z-score normalization to obtain a multidimensional time series data matrix with dimension N×T. , where N is the parameter dimension and T is the time step.

[0007] In one embodiment of the present invention, optionally, in step S2, variational mode decomposition (VMD) constructs a constrained variational problem, introduces a quadratic penalty factor and Lagrange multipliers, and uses the alternating direction method of multipliers (ADMM) to iteratively solve in the frequency domain to obtain K IMF components.

[0008] In one embodiment of the present invention, optionally, in step S2, the selection rule for the high-frequency mechanical characteristic signal is: select the IMF component with large arrangement entropy and high center frequency that can reflect minor jamming and wear impact for reconstruction, and filter out the low-frequency trend term caused by wind speed fluctuation.

[0009] In one embodiment of the present invention, optionally, in step S3, the dynamic graph topology... , where the set of nodes Dynamic adjacency matrix for monitoring parameters of pitch control system It is calculated using a multi-head self-attention mechanism and normalized by the function SoftMax.

[0010] In one embodiment of the present invention, optionally, in step S4, the dynamic graph neural network DGNN includes a graph convolutional network GCN and a gated recurrent unit GRU; wherein the GCN is used for spatial topological feature aggregation, and the GRU is used for temporal dependency modeling.

[0011] In one embodiment of the present invention, optionally, in step S4, the model training adopts a binary cross-entropy loss BCELoss joint graph structure regularization term, which is constructed based on the F-norm difference between the prior physical adjacency matrix and the dynamic adjacency matrix.

[0012] In one embodiment of the present invention, optionally, in step S4, the fault probability is output by the fully connected layer and the Sigmoid activation function; when the fault probability continuously exceeds a set threshold or the probability slope is abnormal, a minor fault warning for the pitch system is triggered.

[0013] In one embodiment of the present invention, optionally, in step S4, spatial feature aggregation employs graph convolution operation, with the following formula:

[0014] in, To incorporate a dynamic adjacency matrix with self-loops, For degree matrix, The node feature matrix, For graph convolution weights, This is the activation function.

[0015] The pitch system micro-fault early warning method based on SCADA data analysis provided by this invention has the following beneficial technical effects: (1) Improve the accuracy of early warning of minor faults: By extracting high-frequency mechanical features through VMD decoupling and combining DGNN to capture spatial topological changes, it can accurately identify abnormalities in the early stage of minor jamming or wear in the pitch system (usually several weeks earlier than traditional threshold alarms), and improve the warning accuracy to over 90%.

[0016] (2) Reduce false alarm rate under complex working conditions: effectively filter out non-stationary noise interference caused by drastic wind speed fluctuations and power grid disturbances, avoid false alarms that are easy to be generated under changing working conditions by traditional methods, and reduce the false alarm rate by more than 60%.

[0017] (3) Achieve predictive maintenance and reduce operation and maintenance costs: By outputting continuous dynamic curves of failure probability, wind farm operation and maintenance personnel are provided with intuitive trends of equipment health status deterioration, supporting the transformation from "post-event maintenance" to "predictive maintenance", significantly reducing unplanned downtime and lowering the cost of spare parts replacement and manual inspection. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is the complete process from SCADA data acquisition, VMD time-series feature decoupling, dynamic map construction to DGNN early warning; Figure 2 The flowchart for decoupling VMD time series features shows the detailed process of signal input, variational problem construction, multiplier alternating direction method solution and high-frequency IMF extraction. Figure 3This is a diagram of the Dynamic Graph Neural Network (DGNN) architecture, illustrating the internal network structure of node feature input, dynamic adjacency matrix calculation, GCN spatial aggregation, GRU temporal modeling, and probability output. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This invention mainly solves the following two technical problems: (1) Difficulty in extracting minor fault features: SCADA data is non-stationary and contains strong noise. The abnormal fluctuation amplitude caused by minor jamming or wear in the early stage of the pitch system is extremely small and difficult to be effectively extracted by conventional time domain or frequency domain analysis methods.

[0022] (2) Complex multidimensional parameter coupling relationship: The pitch system involves multiple physical fields such as mechanical, electrical and thermal. There is a dynamic nonlinear coupling relationship between multidimensional parameters such as pitch angle, motor current and inverter temperature. Traditional methods are difficult to capture such abnormal changes in spatial topology.

[0023] The main inventive points of this invention are as follows: (1) SCADA timing feature decoupling method based on VMD: In view of the problem that non-stationary noise is strong and minor fault features are easily submerged in the SCADA data of the pitch system, the VMD algorithm is introduced for the first time to adaptively decouple key signals such as pitch motor current. By calculating the permutation entropy and spectral centroid, the high-frequency intrinsic mode function (IMF) reflecting mechanical jamming or wear is accurately extracted, effectively filtering out low-frequency interference such as wind speed fluctuation.

[0024] (2) Capturing spatial anomalies of multidimensional parameters based on dynamic graph topology: This method breaks through the limitation of traditional methods that treat multidimensional parameters as independent time series. It defines parameters such as pitch angle, motor current, and temperature of the pitch system as graph nodes and uses a multi-head self-attention mechanism to dynamically calculate the edge weights between nodes to construct a dynamic graph topology structure, which can capture abnormal changes in the nonlinear coupling relationship between multi-physics parameters in real time.

[0025] (3) DGNN micro-fault early warning model integrating spatiotemporal features: By combining graph convolutional network (GCN) and gated recurrent unit (GRU), an end-to-end dynamic graph neural network (DGNN) early warning model was constructed. At the same time, spatial topological feature aggregation and temporal dependency modeling were realized, which significantly improved the early warning accuracy of micro-faults in the pitch system.

[0026] First, let's explain the parameters and functions involved in this case: 1.

[0027] A multidimensional time-series data matrix; where N = the dimension of the monitoring parameters, and T = the time step.

[0028] 2.

[0029] Key timing signals of the pitch system (such as motor current and pitch angle).

[0030] 3. Z-score standardization Data normalization methods are used to eliminate the influence of dimensions.

[0031] 4.

[0032] The total number of intrinsic mode functions (IMFs) in VMD decomposition.

[0033] 5.

[0034] The center frequency of the k-th IMF component.

[0035] 6.

[0036] The k-th IMF component (intrinsic mode function).

[0037] 7.

[0038] The second-order penalty factor controls the decomposition accuracy and convergence speed.

[0039] 8.

[0040] Lagrange multipliers are used to solve constrained variational problems.

[0041] 9.

[0042] Dirac impulse function.

[0043] 10.j Imaginary unit (used in frequency domain calculations).

[0044] 11.

[0045] The first-order partial derivative with respect to time t.

[0046] 12.

[0047] L2 norm (vector / signal energy).

[0048] 13.

[0049] Inner product operation.

[0050] 14.min{...} Objective: To minimize the sum of the estimated bandwidths of all IMFs.

[0051] 15. st Constraint: The sum of all IMFs equals the original signal. .

[0052] 16.

[0053] Augmented Lagrangian function (unconstrained variational problem).

[0054] 17. ADMM Alternating direction method of multipliers; used for frequency domain iterative solution of VMD.

[0055] 18.PE (Permutation Entropy) Permutation entropy; measures the degree of non-stationarity and irregularity of a signal.

[0056] 19.

[0057] The raw attention scores for nodes i and j.

[0058] 20. ,

[0059] Learnable weight matrices for queries and keys.

[0060] twenty one.

[0061] Attention vector (transposed).

[0062] 22.|| Feature splicing operation.

[0063] twenty three.

[0064] Activation function of linear unit with leakage correction.

[0065] twenty four.

[0066] Apply Softmax normalization to dimension j and output weights in the range [0,1].

[0067] 25.

[0068] The set of adjacent nodes of node i.

[0069] 26.

[0070] Spatial aggregation features of graph convolution output It is an identity matrix.

[0071] 27.

[0072] Candidate hidden state.

[0073] 28.

[0074] The GRU is currently in a hidden state.

[0075] 29.

[0076] The state that was hidden in the previous moment.

[0077] 30. , ,

[0078] GRU learnable weight matrix.

[0079] 31. , ,

[0080] GRU bias term.

[0081] 32.tanh Hyperbolic tangent activation function.

[0082] 33.⊙ Multiplying elements together (Hadamard product).

[0083] 34.

[0084] The model predicts the probability of failure.

[0085] 35. ,

[0086] Weights and biases of fully connected layers.

[0087] 36. Sigmoid S-type activation function; outputs probabilities [0,1].

[0088] 37.

[0089] Given X, the failure probability of the pitch system at time t.

[0090] 38.

[0091] True label (0 = normal, 1 = fault).

[0092] 39.L Total loss function.

[0093] 40. BCE Loss Binary cross-entropy loss (classification task loss).

[0094] 41.

[0095] Graph structure regularization coefficients.

[0096] 42.

[0097] Prior adjacency matrix based on physical mechanisms.

[0098] 43. normalize() Data standardization function.

[0099] 44.VMD() Variational mode decomposition execution function.

[0100] 45.select_high_frequency_imfs( ) High-frequency IMF screening function.

[0101] 46. ​​concatenate() Original features + high-frequency features are spliced ​​together.

[0102] 47.compute_dynamic_adjacency() Dynamic adjacency matrix calculation.

[0103] 48. GCN_layer() Figure: convolutional layer.

[0104] 49. GRU_cell() GRU cell timing modeling.

[0105] 50. Linear() Fully connected layer.

[0106] 51. Sigmoid() Probability-based activation.

[0107] First embodiment: Figure 1 For the complete process from SCADA data acquisition, VMD time-series feature decoupling, dynamic map construction to DGNN early warning, such as Figure 1 As shown, this invention provides a method for early warning of minor faults in a pitch system based on SCADA data analysis, which includes the following steps: S1: Collect SCADA operation data of wind turbine pitch system, perform missing value imputation, outlier removal and standardization preprocessing to obtain a multi-dimensional time series data matrix; In this embodiment, the SCADA operating data includes, for example, the given pitch angle values ​​and actual pitch angle values ​​for blades 1 / 2 / 3, the pitch motor current, the pitch motor temperature, the pitch inverter temperature, the wind speed, and the generator speed. In other embodiments, the SCADA operating data can be added to or subtracted from the above data, and this invention is not limited to the SCADA operating data listed above.

[0108] In this embodiment, preprocessing can, for example, employ Z-score normalization to obtain a multidimensional time-series data matrix of dimension N×T. , where N is the parameter dimension and T is the time step.

[0109] S2: Variational Mode Decomposition (VMD) is used to adaptively decouple key time-series signals, calculate the permutation entropy and spectral centroid of each intrinsic mode function (IMF), and screen and reconstruct high-frequency mechanical characteristic signals. Figure 2 The flowchart for decoupling VMD time series features shows the detailed process of signal input, variational problem construction, multiplier alternating direction method solution, and high-frequency IMF extraction.

[0110] In this embodiment, variational mode decomposition (VMD) can be solved iteratively in the frequency domain by constructing a constrained variational problem, introducing a quadratic penalty factor and Lagrange multipliers, and using the alternating direction method of multipliers (ADMM) to obtain K IMF components.

[0111] In this embodiment, the selection rule for high-frequency mechanical characteristic signals is, for example, to select IMF components with high arrangement entropy and high center frequency that can reflect minor jamming and wear impacts for reconstruction, while filtering out low-frequency trend terms caused by wind speed fluctuations. In other embodiments, other selection rules can be set, and this invention is not limited to the selection rules listed above.

[0112] S3: Using the multi-dimensional monitoring parameters of the pitch system as graph nodes, dynamically calculate the edge weights between nodes based on the multi-head self-attention mechanism, and construct a dynamic graph topology and dynamic adjacency matrix; In this embodiment, the dynamic graph topology , where the set of nodes Dynamic adjacency matrix for monitoring parameters of pitch control system It is calculated using a multi-head self-attention mechanism and normalized by the function SoftMax.

[0113] S4: Input the dynamic graph into the dynamic graph neural network (DGNN), complete the spatial feature aggregation through graph convolution, complete the temporal feature modeling through the gated recurrent unit (GRU), output the fault probability, and realize the early warning of minor faults. Figure 3 This is a diagram of the Dynamic Graph Neural Network (DGNN) architecture, illustrating the internal network structure of node feature input, dynamic adjacency matrix calculation, GCN spatial aggregation, GRU temporal modeling, and probability output.

[0114] In this embodiment, the dynamic graph neural network (DGNN) may include, for example, a graph convolutional network (GCN) and a gated recurrent unit (GRU); wherein the GCN is used for spatial topological feature aggregation, and the GRU is used for temporal dependency modeling.

[0115] In this embodiment, model training can, for example, employ a binary cross-entropy loss (BCE) combined with a graph structure regularization term, which is constructed based on the F-norm difference between the prior physical adjacency matrix and the dynamic adjacency matrix.

[0116] In this embodiment, the fault probability can be output by a fully connected layer and a sigmoid activation function, for example; when the fault probability continuously exceeds a set threshold or the probability slope is abnormal, a minor fault warning for the pitch system is triggered.

[0117] Furthermore, in step S4, spatial feature aggregation employs graph convolution operation, with the following formula:

[0118] in, To incorporate a dynamic adjacency matrix with self-loops, For degree matrix, The node feature matrix, For graph convolution weights, This is the activation function.

[0119] Second embodiment: The pitch system minor fault early warning method based on SCADA data analysis provided in this embodiment is as follows: Step 1: SCADA Data Acquisition and Preprocessing Collect relevant operational data of the pitch system from the wind turbine's SCADA system, including but not limited to: the setpoint and actual pitch angle values ​​for blades 1 / 2 / 3, pitch motor current, pitch motor temperature, pitch inverter temperature, wind speed, and generator speed. Perform missing value imputation, outlier removal, and Z-score normalization on the collected data to obtain a multidimensional time-series data matrix. , where N is the parameter dimension and T is the time step.

[0120] Step 2: Decoupling of SCADA temporal features based on VMD For key signals such as pitch motor current that contain rich mechanical state information The variational mode decomposition (VMD) algorithm is used to decouple the temporal features.

[0121] 1. Constructing a variational problem: Assume the signal... K with specific center frequencies The intrinsic mode functions (IMF) Composition. To minimize the sum of the estimated bandwidths for each mode, the following constrained variational problem is constructed:

[0122]

[0123] 2. Solving the variational problem: Introducing a quadratic penalty factor. and Lagrange multipliers The constrained variational problem is transformed into an unconstrained variational problem:

[0124] Iterative updates in the frequency domain are performed using the alternating direction of multipliers (ADMM) method. , and Continue until the convergence condition is met.

[0125] 3. High-Frequency Mechanical Characteristic IMF Extraction: After decoupling to obtain K IMF components, the permutation entropy (PE) and spectral centroid of each component are calculated. IMF components with larger permutation entropy and higher center frequency (reflecting non-stationary high-frequency impacts caused by minor jamming / wear) are selected as features of the reconstructed signal, denoted as . It effectively filters out low-frequency trend terms caused by wind speed changes.

[0126] Step 3: Dynamic Graph Topology Construction Define the multidimensional parameters of the pitch system as nodes of a graph to construct a dynamic graph. .

[0127] 1. Node Feature Matrix: Node Set Represents N monitoring parameters. At time step t, the node feature matrix... It is composed of the original features and the high-frequency features after VMD decoupling.

[0128] 2. Dynamic Adjacency Matrix: Due to the dynamic changes in the coupling relationship between parameters with the operating conditions, a multi-head self-attention mechanism is used to dynamically calculate the edge weights between nodes and generate a dynamic adjacency matrix. :

[0129]

[0130] in, The weight matrix is ​​a learnable matrix. This indicates a splicing operation. This is the attention vector.

[0131] Step 4: Feature Aggregation and Fault Prediction using Dynamic Graph Neural Network (DGNN) Spatial Feature Aggregation: Utilizing Graph Convolutional Networks (GCNs) to aggregate spatial features on dynamic graphs, capturing topological variations between multidimensional parameters:

[0132] in, , As a unit array, For degree matrix, The graph convolution weight matrix, This is the activation function.

[0133] Temporal feature modeling: spatially aggregated feature sequences The input is fed into a gated recurrent unit (GRU) to capture time dependencies:

[0134]

[0135]

[0136]

[0137] Fault probability output: hides the GRU's state. Input a fully connected layer, and output the failure probability of the pitch system at time step t via a sigmoid function. :

[0138] Loss function: The model is trained using binary cross-entropy loss (BCE Loss) combined with a graph structure regularization term.

[0139] in, This is a priori adjacency matrix constructed based on physical mechanisms. is the regularization coefficient.

[0140] The pseudocode implementation of this invention is as follows: # Pseudocode: Early Warning Algorithm for Pitch System Based on VMD and DGNN def VMD_DGNN_Warning(SCADA_data, K, alpha): # 1. Data Preprocessing X_norm = normalize(SCADA_data) # 2. VMD Decoupling for High-Frequency Feature Extraction X_high = [] for signal in X_norm.key_signals: # For key signals such as motor current u, omega, _ = VMD(signal, K, alpha) # Calculate permutation entropy and reconstruct IMFs with high frequency and large entropy values. imf_selected = select_high_frequency_imfs(u) X_high.append(sum(imf_selected)) # Segmenting original features and high-frequency features H = concatenate(X_norm, X_high) # 3. DGNN Forward Propagation P_fault = [] h_prev = initialize_hidden_state() for t in range(T): # Dynamic graph topology construction (attention mechanism) A_t = compute_dynamic_adjacency(H[t]) # Graph Convolution Spatial Aggregation Z_t = GCN_layer(A_t, H[t]) # GRU Temporal Modeling h_t = GRU_cell(Z_t, h_prev) h_prev = h_t # Fault probability output p_t = Sigmoid(Linear(h_t)) P_fault.append(p_t) return P_fault User scenarios for this invention: This invention is mainly applied to wind farm control centers, remote operation and maintenance diagnostic centers of wind turbine manufacturers, and status monitoring platforms of third-party wind power operation and maintenance service providers. Operation and maintenance engineers can view the health status and fault probability prediction curves of the pitch system of each wind turbine in real time through the system's large screen or mobile device. After receiving early warning alerts, they can arrange targeted maintenance plans.

[0141] Applications of this invention include: a wind turbine intelligent health management system (PHM), a wind farm SCADA advanced data analysis module, and dedicated intelligent monitoring and diagnostic software for pitch control systems.

[0142] The detection method of the present invention: 1. Acquire the pitch system operation data of the wind turbine SCADA system in real time / near real time through OPC protocol or database interface.

[0143] 2. Input the data into the VMD-DGNN early warning model deployed on the edge computing gateway or cloud server for online inference.

[0144] 3. The model outputs the failure probability of the pitch system in real time. When the probability continuously exceeds the set threshold (such as 0.75) or the probability rises at an abnormal rate, a minor fault warning is triggered, and a diagnostic report containing abnormal parameter nodes is generated.

[0145] The pitch system micro-fault early warning method based on SCADA data analysis provided by this invention has the following beneficial technical effects: (1) Improve the accuracy of early warning of minor faults: By extracting high-frequency mechanical features through VMD decoupling and combining DGNN to capture spatial topological changes, it can accurately identify abnormalities in the early stage of minor jamming or wear in the pitch system (usually several weeks earlier than traditional threshold alarms), and improve the warning accuracy to over 90%.

[0146] (2) Reduce false alarm rate under complex working conditions: effectively filter out non-stationary noise interference caused by drastic wind speed fluctuations and power grid disturbances, avoid false alarms that are easy to be generated under changing working conditions by traditional methods, and reduce the false alarm rate by more than 60%.

[0147] (3) Achieve predictive maintenance and reduce operation and maintenance costs: By outputting continuous dynamic curves of failure probability, wind farm operation and maintenance personnel are provided with intuitive trends of equipment health status deterioration, supporting the transformation from "post-event maintenance" to "predictive maintenance", significantly reducing unplanned downtime and lowering the cost of spare parts replacement and manual inspection.

[0148] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0149] Those skilled in the art will understand that the modules in the apparatus of the embodiments can be distributed in the apparatus of the embodiments as described in the embodiments, or they can be located in one or more devices different from this embodiment with corresponding changes. The modules of the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

[0150] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for early warning of micro-failure of a variable pitch system based on SCADA data analysis, characterized in that, Includes the following steps: S1: Collect SCADA operation data of wind turbine pitch system, perform missing value imputation, outlier removal and standardization preprocessing to obtain a multi-dimensional time series data matrix; S2: Variational Mode Decomposition (VMD) is used to adaptively decouple key time-series signals, calculate the permutation entropy and spectral centroid of each intrinsic mode function (IMF), and screen and reconstruct high-frequency mechanical characteristic signals. S3: Using the multi-dimensional monitoring parameters of the pitch system as graph nodes, dynamically calculate the edge weights between nodes based on the multi-head self-attention mechanism, and construct a dynamic graph topology and dynamic adjacency matrix; S4: Input the dynamic graph into the dynamic graph neural network (DGNN), complete the spatial feature aggregation through graph convolution, complete the temporal feature modeling through the gated recurrent unit (GRU), output the fault probability, and realize the early warning of minor faults.

2. The method of claim 1, wherein, In step S1, the SCADA operating data includes: the given pitch angle of blades 1 / 2 / 3, the actual pitch angle, the pitch motor current, the pitch motor temperature, the pitch inverter temperature, the wind speed, and the generator speed.

3. The method of claim 1, wherein, In step S1, the preprocessing adopts Z-score standardization to obtain a multi-dimensional time series data matrix with a dimension of N x T wherein N is a parameter dimension, and T is a time step.

4. The method of claim 1, wherein, In step S2, variational mode decomposition (VMD) constructs a constrained variational problem, introduces a quadratic penalty factor and Lagrange multipliers, and uses the alternating direction method of multipliers (ADMM) to iteratively solve the problem in the frequency domain, obtaining K IMF components.

5. The method of claim 1, wherein, In step S2, the selection rule for high-frequency mechanical characteristic signal is as follows: select IMF components with large arrangement entropy and high center frequency that can reflect minor jamming and wear impact for reconstruction, and filter out low-frequency trend terms caused by wind speed fluctuations.

6. The method of claim 1, wherein, In step S3, the dynamic graph topology where the set of nodes For the pitch system monitoring parameters, the dynamic adjacency matrix is calculated by the multi-head self-attention mechanism and normalized by the function SoftMax.

7. The method of claim 1, wherein, In step S4, the Dynamic Graph Neural Network (DGNN) includes a Graph Convolutional Network (GCN) and a Gated Recurrent Unit (GRU); where the GCN is used for spatial topological feature aggregation and the GRU is used for temporal dependency modeling.

8. The method of claim 1, wherein, In step S4, the model training uses a binary cross-entropy loss (BCE) combined with a graph structure regularization term, which is constructed based on the difference in F norm between the prior physical adjacency matrix and the dynamic adjacency matrix.

9. The method of claim 1, wherein, In step S4, the fault probability is output by the fully connected layer and the Sigmoid activation function; when the fault probability continuously exceeds the set threshold or the probability slope is abnormal, a minor fault warning for the pitch system is triggered.

10. The method of claim 1, wherein, In step S4, spatial feature aggregation uses graph convolution operation, with the following formula: wherein, is a dynamic adjacency matrix with self-loops, is a degree matrix, is a node feature matrix, is a graph convolution weight, is an activation function.