A fan early fault warning and grading method based on multi-dimensional time-varying topology and sequence matching

Abstract

This invention discloses a method for early fault warning and classification of wind turbines based on multidimensional time-varying topology and sequence matching. First, multidimensional SCADA operating data of the wind turbine is acquired and preprocessed to form raw time-series samples. A double-layer sliding window reconstruction is used to integrate short-term fluctuations and long-term trend features. Second, the time-varying topological relationships of the sensors are determined, a temporal adjacency matrix is ​​constructed, and the dynamic features of the graph nodes are calculated. Next, the baseline prediction value for future times is obtained using a normal operating state baseline prediction model, and the comprehensive deviation between this prediction and the actual observed value is calculated. An early warning signal is output, and the early warning trigger time and early warning lead time are recorded. Then, the fault evolution features of the abnormal sequences that have triggered early warnings are extracted, and sequence matching is performed with a typical fault mode library to obtain the most similar fault mode and the matching distance. Finally, the fault hazard level is calculated based on the early warning lead time, the maximum deviation, and the minimum matching distance of the fault mode, and maintenance strategy recommendations are output.

Description

Technical Field

[0001] This invention relates to the field of wind power equipment operation and maintenance, specifically a method for early warning and classification of wind turbine faults based on multidimensional time-varying topology and sequence matching. Background Technology

[0002] Wind turbines operate under complex conditions of random wind speed fluctuations, alternating loads, and significant environmental disturbances. Key components such as the pitch system, gearbox, generator, and transmission system are prone to early signs of failure, including temperature increases, sudden current changes, transmission anomalies, and localized performance degradation. If these abnormal conditions are not identified in time during the initial stages of a fault, localized defects may continue to expand during operation, leading to unplanned outages, power generation losses, increased maintenance costs, and secondary damage to related components.

[0003] Currently, wind farms are generally equipped with Supervisory Control and Data Acquisition (SCADA) systems, which continuously record multi-dimensional operating parameters such as wind speed, power, rotational speed, pitch motor temperature, pitch motor current, gearbox temperature, and generator temperature. Utilizing SCADA data for wind turbine fault early warning has the advantages of requiring no additional hardware investment, wide coverage, and good engineering implementation foundation, and has become an important direction for online health management of wind turbine units.

[0004] However, existing wind turbine fault early warning methods still have the following shortcomings: First, most methods treat different sensor parameters as independent variables, making it difficult to characterize the conditional dependence between sensors over time, which can easily lead to false alarms when wind speed fluctuates and operating conditions switch frequently. Second, they rely solely on fixed thresholds or single residual indices for judgment, making it difficult to balance early warning sensitivity and false alarm suppression. Third, early warning results usually only indicate whether the fault is abnormal, lacking quantification of the degree of fault evolution and the level of danger, making it difficult to directly serve subsequent maintenance method selection, maintenance duration estimation, and maintenance cost accounting. Summary of the Invention

[0005] The purpose of this invention is to provide a method for early fault warning and classification of wind turbines based on multidimensional time-varying topology and sequence matching, comprising the following steps:

[0006] Step 1) Obtain multi-dimensional operation data of wind turbine SCADA, perform data preprocessing, and form raw time series samples;

[0007] Step 2) Reconstruct the original time series samples using a double-layer sliding window, extract the short-term fluctuation and long-term trend features of the original time series samples, and output the fused features;

[0008] Step 3) Determine the time-varying topological relationships of the sensors based on the fusion features and construct a temporal adjacency matrix;

[0009] Step 4) Calculate the dynamic features of graph nodes using a time-varying adjacency matrix and a dual attention mechanism;

[0010] Step 5) Input the dynamic features of the graph nodes into the pre-trained baseline prediction model for normal operation to obtain the baseline prediction value for the normal operation state in the next H steps;

[0011] Step 6) Calculate the overall deviation between the actual observed features and the baseline predicted values. If continuous All times satisfy If the warning signal is triggered, an early warning signal will be output, and the warning trigger time and the warning lead time will be recorded.

[0012] Step 7) Extract fault evolution features from the abnormal sequences that have triggered warnings, and perform sequence matching with the fault evolution features of typical fault modes pre-stored in the database to obtain the most similar fault modes and matching distances; Step 8) Calculate the fault hazard level based on the warning lead time, maximum deviation, and minimum matching distance of the fault modes, and output maintenance strategy recommendations.

[0013] Furthermore, in step 1), the SCADA multidimensional operating data of the wind turbine includes wind speed, wind direction, ambient temperature, active power, generator speed, pitch angle, gearbox temperature, bearing temperature, and main shaft speed.

[0014] Data preprocessing refers to aligning the SCADA multidimensional operating data of wind turbine units according to timestamps, removing data points that are shut down, have limited power, have communication anomalies, or do not meet physical constraints, and using interpolation of nearby times or window mean to fill in missing data, and unifying the units of measurement.

[0015] Furthermore, the original time series samples are shown below:

[0016] (1)

[0017] In the formula, T represents the total number of sampling times, and N represents the number of sensor nodes. Let t be the observation vector consisting of N operating parameters.

[0018] Furthermore, in step 2), the fusion features are as follows:

[0019] (2)

[0020] In the formula, This is a short-term window characteristic. This is a long-term window feature. For the concatenation operation, M is the gradient feature dimension;

[0021] The characteristics of a short-term window are as follows:

[0022] (3)

[0023] In the formula, The width of the inner sliding window. , The mean and standard deviation of the inner window are shown below:

[0024] (4)

[0025] The characteristics of a long-term window are as follows:

[0026] (5)

[0027] In the formula, The width of the outer sliding window. The trend gradient features extracted from the outer window, , The mean and standard deviation of the outer window are shown below:

[0028] (6)

[0029] (7)

[0030] In the formula, x i This represents the original time series sample or observed operating parameters at the i-th sampling time.

[0031] Furthermore, in step 3), the steps for constructing the temporal adjacency matrix include:

[0032] Each sensor's operating parameter is treated as a graph node, and the conditional dependencies between sensors are treated as graph edges.

[0033] The sample covariance matrix is ​​calculated in each sliding time interval, and the sparse precision matrix is ​​estimated using the Toeplitz-GL algorithm.

[0034] The sample covariance matrix is ​​shown below:

[0035] (8)

[0036] In the formula, S k L is the sample covariance matrix within the k-th topology estimation window; c The length of the sliding time window used to estimate the sample covariance matrix; z i This is the fused feature vector corresponding to the i-th sampling time. For z i The transpose of ; the sparse precision matrix is ​​shown below:

[0037] (9)

[0038] In the formula, Let k be the sample covariance matrix within the time window. Let k be the accuracy matrix. For sparsity regularization parameters, For Toeplitz transformation operator;

[0039] Thresholding the precision matrix yields a time-varying adjacency matrix that reflects intra-layer and cross-layer associations.

[0040] The temporal adjacency matrix is ​​shown below:

[0041] (10)

[0042] In the formula, This is the adjacency threshold. Let be the adjacency matrix between sensor i and sensor j at time t. Let be the sparse precision matrix between sensor i and sensor j at time t.

[0043] Furthermore, in step 4), the steps for calculating the dynamic characteristics of graph nodes include:

[0044] Step 4.1) For the first Layer node feature matrix Performing a linear projection yields the query, key, and value vectors, as shown below:

[0045] (11)

[0046] In the formula, , , For learnable projection matrix;

[0047] Step 4.2) Calculate the spatial attention weights as follows:

[0048] (12)

[0049] In the formula, , Let i be the query vector of node i and j be the key vector of node j. For attention dimension;

[0050] Step 4.3) Fuse the spatial attention weights with the time-varying adjacency matrix to obtain the attention weights after fusing the topological prior, as shown below:

[0051] (13)

[0052] Step 4.4) After applying the activation function, the node features containing the spatial coupling relationship of the sensor are obtained, that is:

[0053] (14)

[0054] In the formula, For activation function, The longitude projection matrix of the j-th sensor node The value vector obtained by mapping;

[0055] Step 4.5) Perform sequence attention calculation on the time dimension to obtain the dynamic features of the graph nodes as shown below:

[0056] (15)

[0057] (16)

[0058] (17)

[0059] (18)

[0060] In the formula, The input is the node feature sequence within the time window. These are the query projection matrix, key projection matrix, and value projection matrix in the time attention module. Let t be the temporal attention weight of historical time s. For activation function, Let be the query vector corresponding to time t. This is the transpose of the key vector corresponding to historical time s. The length of the historical time window involved in the time attention calculation. Features resulting from enhanced time attention. For spatial attention modules, For time attention module, It is a feedforward network. This is a layer normalization operation.

[0061] Furthermore, in step 5), the baseline prediction model for normal operating conditions is a spatiotemporal graph neural network, which includes an encoder and a decoder;

[0062] The encoder is used to extract deep spatiotemporal features from the historical window, and the decoder generates predicted values ​​for the next H steps of normal operation in an autoregressive manner.

[0063] The baseline prediction model for normal operating conditions is trained using historical datasets;

[0064] Historical datasets include multiple input-output sample pairs;

[0065] The input is the dynamic features of the graph nodes; the output is the baseline prediction value under normal operating conditions in the next H steps.

[0066] The baseline prediction model for normal operating conditions is shown below:

[0067] (19)

[0068] In the formula, The semantic features output by the encoder, Given the input feature sequence, This corresponds to the time-varying adjacency matrix sequence;

[0069] The baseline predictions for the future H-step normal operation are shown below:

[0070] (20)

[0071] In the formula, The predicted value at time t+k. This is the hidden state of the decoder at step k-1;

[0072] The prediction error from the previous time step is introduced to correct the baseline prediction value for the next time step, that is:

[0073] (twenty one)

[0074] In the formula, This is the baseline prediction value for the next time step. For error feedback coefficients, This represents the prediction error at time t. The actual value at time t. Let t be the baseline prediction value at time t.

[0075] Furthermore, in step 6), the overall deviation between the actual observed values ​​and the baseline predicted values ​​is shown below:

[0076] (twenty two)

[0077] In the formula, For deviation fusion weights, The calculation is as follows:

[0078] (twenty three)

[0079] (twenty four)

[0080] (25)

[0081] In the formula, For residual covariance matrix estimation, The Mahalanobis distance;

[0082] The updated dynamic alert thresholds are shown below:

[0083] (26)

[0084] In the formula, , Nearly The mean and standard deviation of the deviation at each time point. This is the threshold adjustment coefficient;

[0085] The output warning signal is shown below:

[0086] (27)

[0087] In the formula, For continuous over-threshold window;

[0088] The advance warning time is shown below:

[0089] (28)

[0090] In the formula, The warning is triggered at the following time. This represents the actual time the fault occurred.

[0091] Furthermore, in step 7), the sequence matching step includes:

[0092] Step 7.1) Construct the multi-node residual matrix as follows:

[0093] (29)

[0094] In the formula, This is the abnormal window length;

[0095] Step 7.2) Extract the mean, variance, slope, and kurtosis statistical features of the deviation to form the current fault evolution feature sequence, i.e.:

[0096] (30)

[0097] In the formula, For vectorization operations, It is a statistical feature vector;

[0098] Step 7.3) Calculate the dynamic time-warped distance as shown below:

[0099] (31)

[0100] In the formula, For dynamic time-normalized distance, This is the reference feature sequence for the k-th typical failure mode. The length of the kth fault mode;

[0101] Step 7.4) Calculate the minimum matching distance and the starting position of the best matching subsequence for typical failure modes, i.e.:

[0102] (32)

[0103] Among all K typical fault modes, the fault mode with the smallest matching distance is selected to obtain the global minimum matching distance, as shown below:

[0104] (33)

[0105] In the formula, To find the minimum matching distance with the k-th typical failure mode, The starting position of the best matching subsequence. The minimum global matching distance is obtained by matching the abnormal sequence that triggers the early warning with all typical fault modes. Number the typical failure mode that is most similar to the current abnormal sequence;

[0106] Step 7.5) yields the following comprehensive measure of fault evolution:

[0107] (34)

[0108] In the formula, K is the number of reference failure modes in the typical failure mode library. This is the reference feature vector corresponding to the kth typical fault mode.

[0109] Furthermore, in step 8), the fault hazard level is as follows:

[0110] (35)

[0111] (36)

[0112] In the formula, Score the severity level of the fault; For the grade score weight, ; The deviation threshold; The maximum deviation is continuous. The overall deviation at each moment D t The maximum value in the range; Level I is the early abnormal development period; Level II is the mid-term accelerated evolution period; Level III is the critical emergency deterioration period.

[0113] The technical effects of this invention are undeniable, and are as follows:

[0114] (1) SCADA time series data is reconstructed by a double sliding window, while retaining short-term local fluctuations and long-term gradual trends, avoiding the problem that traditional single window cannot take into account both short-term anomalies and long-term degradation.

[0115] (2) By inferring the time-varying conditional dependence of sensors through the Toeplitz-GL algorithm, the intra-layer and cross-layer topological relationships between the pitch system, transmission system, generator system and environmental parameters can be characterized, thereby improving the ability to identify anomalies under complex working conditions.

[0116] (3) By integrating spatial attention and temporal attention, the ST-GNN baseline prediction model can adaptively learn key sensor nodes and key historical moments in fault propagation, thereby reducing the false alarm rate under strong noise and multiple operating conditions.

[0117] (4) The comprehensive deviation is constructed by the absolute value of the residual, the standardized residual and the Mahalanobis distance, and the adaptive threshold and the continuous over-threshold condition are introduced to make the early warning trigger both sensitive and stable.

[0118] (5) By matching the multidimensional matrix profile with the DTW sequence, the warning result of "whether it is abnormal" is further transformed into the fault risk level, providing quantitative input for maintenance method selection, maintenance time estimation and maintenance cost accounting. Attached Figure Description

[0119] Figure 1 This is a flowchart illustrating the overall method of the present invention;

[0120] Figure 2 This is a schematic diagram of the SCADA parameter double-layer sliding window and time-varying topology construction of the present invention. Detailed Implementation

[0121] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.

[0122] Example 1:

[0123] See Figures 1 to 2 A method for early fault warning and classification of wind turbines based on multidimensional time-varying topology and sequence matching includes the following steps:

[0124] Step 1) Obtain multi-dimensional operating data from the wind turbine's SCADA (Supervisory Control And Data Acquisition) system, perform data preprocessing, and form raw time-series samples;

[0125] Step 2) Reconstruct the original time series samples using a double-layer sliding window, extract the short-term fluctuation and long-term trend features of the original time series samples, and output the fused features;

[0126] Step 3) Determine the time-varying topological relationships of the sensors based on the fusion features and construct a temporal adjacency matrix;

[0127] Step 4) Calculate the dynamic features of graph nodes using a time-varying adjacency matrix and a dual attention mechanism;

[0128] Step 5) Input the dynamic features of the graph nodes into the pre-trained baseline prediction model for normal operation to obtain the baseline prediction value for the normal operation state in the next H steps;

[0129] Step 6) Calculate the overall deviation between the actual observed features and the baseline predicted values. If continuous All times satisfy If the warning signal is triggered, an early warning signal will be output, and the warning trigger time and the warning lead time will be recorded.

[0130] Actual observation feature acquisition process: After the future prediction time arrives, the actual operation data at the corresponding time is obtained from the wind turbine SCADA system, and the actual observation features with the same dimension as the baseline prediction value are obtained by following the same data preprocessing and two-layer sliding window reconstruction method as in steps 1)-2).

[0131] Step 7) Extract fault evolution features from the abnormal sequences that have triggered warnings, and perform sequence matching with the fault evolution features of typical fault modes pre-stored in the database to obtain the most similar fault modes and matching distances.

[0132] Typical failure modes include failure evolution patterns constructed from historical failure logs, alarm records, and maintenance records. They include at least one or more of the following: pitch system blade emergency shutdown activation, encoder malfunction, position asynchrony, pitch motor current over-limit, PMC failure, abnormal battery test interval, abnormal gearbox temperature, abnormal bearing temperature, and abnormal generator temperature.

[0133] Typical failure modes can also be classified according to the failure evolution process into early slow deviation type, mid-term continuous degradation type and critical rapid deterioration type. Each type of failure mode includes a corresponding reference characteristic sequence, failure occurrence time, failure type label and hazard level label.

[0134] Step 8) Calculate the fault hazard level based on the early warning lead time, maximum deviation, and minimum matching distance of the fault mode, and output maintenance strategy recommendations.

[0135] Example 2:

[0136] The main structure of this embodiment is the same as that of embodiment 1. Further, in step 1), the SCADA multidimensional operation data of the wind turbine includes wind speed, wind direction, ambient temperature, active power, generator speed, pitch angle, gearbox temperature, bearing temperature and main shaft speed.

[0137] Data preprocessing refers to aligning the SCADA multidimensional operation data of wind turbine units according to timestamps, removing data points that are shut down, have limited power, have communication anomalies, or do not meet physical constraints, and using interpolation of nearby times or window mean to fill in missing data, and unifying the units of measurement.

[0138] Power curtailment refers to the operating state of a wind turbine where, under normal power generation conditions, the actual active power is lower than the theoretical power or rated control curve power at the corresponding wind speed due to grid dispatching, power control at the wind farm, curtailment, turbine control strategies, or safety protection strategies. Under power curtailment conditions, the normal mapping relationship between wind speed, power, speed, and pitch angle is artificially altered, and therefore it is removed or marked separately before normal baseline modeling.

[0139] The physical constraints are as follows:

[0140] (1) The wind speed shall not be less than zero and shall not exceed the allowable wind speed range of the unit design;

[0141] (2) The active power is not less than zero and does not exceed the allowable margin of the rated power;

[0142] (3) The generator speed, main shaft speed and pitch angle are within the range specified in the unit operation manual;

[0143] (4) The gearbox temperature, bearing temperature and ambient temperature are within the effective measurement range of the sensor;

[0144] (5) When the wind speed is lower than the cut-in wind speed, the active power should be close to zero. When the wind speed is higher than the rated wind speed, the active power should not exceed the allowable margin of the rated power.

[0145] If a data point violates any of the above constraints, it is determined to be a data point that does not meet the physical constraints.

[0146] Example 3:

[0147] The main structure of this embodiment is the same as any one of embodiments 1-2. Furthermore, the original time series sample is shown below:

[0148] (1)

[0149] In the formula, T represents the total number of sampling times, and N represents the number of sensor nodes. Let t be the observation vector consisting of N operating parameters.

[0150] Example 4:

[0151] The main structure of this embodiment is the same as any one of embodiments 1-3. Further, in step 2), the fusion features are as follows:

[0152] (2)

[0153] In the formula, This is a short-term window characteristic. This is a long-term window feature. For the concatenation operation, M is the gradient feature dimension;

[0154] The characteristics of a short-term window are as follows:

[0155] (3)

[0156] In the formula, The width of the inner sliding window. , The mean and standard deviation of the inner window are shown below:

[0157] (4)

[0158] The characteristics of a long-term window are as follows:

[0159] (5)

[0160] In the formula, The width of the outer sliding window. The trend gradient features extracted from the outer window. , The mean and standard deviation of the outer window are shown below:

[0161] (6)

[0162] (7)

[0163] In the formula, x i This represents the original time series sample or observed operating parameters at the i-th sampling time.

[0164] Example 5:

[0165] The main structure of this embodiment is the same as any one of embodiments 1 to 4. Further, in step 3), the step of constructing the temporal adjacency matrix includes:

[0166] Each sensor's operating parameter is treated as a graph node, and the conditional dependencies between sensors are treated as graph edges.

[0167] The sample covariance matrix is ​​calculated in each sliding time interval, and the sparse precision matrix is ​​estimated using the Toeplitz-GL algorithm.

[0168] The sample covariance matrix is ​​shown below:

[0169] (8)

[0170] In the formula, S k L is the sample covariance matrix within the k-th topology estimation window; c The length of the sliding time window used to estimate the sample covariance matrix, i.e., the number of time samples contained within the window; z i This is the fused feature vector corresponding to the i-th sampling time. For z i Transpose of;

[0171] The sparse precision matrix is ​​shown below:

[0172] (9)

[0173] In the formula, Let k be the sample covariance matrix within the time window. Let k be the accuracy matrix. For sparsity regularization parameters, For Toeplitz transformation operator;

[0174] Thresholding the precision matrix yields a time-varying adjacency matrix that reflects intra-layer and cross-layer associations.

[0175] The temporal adjacency matrix is ​​shown below:

[0176] (10)

[0177] In the formula, This is the adjacency threshold. Let be the adjacency matrix between sensor i and sensor j at time t. Let be the sparse precision matrix between sensor i and sensor j at time t.

[0178] Example 6:

[0179] The main structure of this embodiment is the same as any one of embodiments 1-5. Further, in step 4), the step of calculating the dynamic features of graph nodes includes:

[0180] Step 4.1) For the first Layer node feature matrix Performing a linear projection yields the query, key, and value vectors, as shown below:

[0181] (11)

[0182] In the formula, , , For learnable projection matrix;

[0183] Step 4.2) Calculate the spatial attention weights as follows:

[0184] (12)

[0185] In the formula, , Let i be the query vector of node i and j be the key vector of node j. For the attention dimension;

[0186] Step 4.3) Fuse the spatial attention weights with the time-varying adjacency matrix to obtain the attention weights after fusing the topological prior, as shown below:

[0187] (13)

[0188] Step 4.4) After applying the activation function, the node features containing the spatial coupling relationship of the sensor are obtained, that is:

[0189] (14)

[0190] In the formula, For activation function, The longitude projection matrix of the j-th sensor node The value vector obtained from the mapping is used to represent the feature information of node j in the current layer;

[0191] Step 4.5) Perform sequence attention calculation on the time dimension to obtain the dynamic features of the graph nodes as shown below:

[0192] (15)

[0193] (16)

[0194] (17)

[0195] (18)

[0196] In the formula, The input is the node feature sequence within the time window. These are the query projection matrix, key projection matrix, and value projection matrix in the time attention module. Let t be the temporal attention weight of historical time s. For activation function, Let be the query vector corresponding to time t. This is the transpose of the key vector corresponding to historical time s. The length of the historical time window involved in the time attention calculation. Features resulting from enhanced time attention. For spatial attention modules, For time attention module, It is a feedforward network. This is a layer normalization operation.

[0197] Example 7:

[0198] The main structure of this embodiment is the same as any one of embodiments 1 to 6. Further, in step 5), the normal operation state baseline prediction model is a spatiotemporal graph neural network, including an encoder and a decoder.

[0199] The encoder is used to extract deep spatiotemporal features from the historical window, and the decoder generates predicted values ​​for the next H steps of normal operation in an autoregressive manner.

[0200] The baseline prediction model for normal operating conditions is trained using historical datasets;

[0201] Historical datasets include multiple input-output sample pairs;

[0202] The input is the dynamic features of the graph nodes; the output is the baseline prediction value under normal operating conditions in the next H steps.

[0203] The baseline prediction model for normal operating conditions is shown below:

[0204] (19)

[0205] In the formula, The semantic features output by the encoder, Given the input feature sequence, This corresponds to the time-varying adjacency matrix sequence;

[0206] The baseline predictions for the future H-step normal operation are shown below:

[0207] (20)

[0208] In the formula, The predicted value at time t+k. This is the hidden state of the decoder at step k-1;

[0209] The prediction error from the previous time step is introduced to correct the baseline prediction value for the next time step, that is:

[0210] (twenty one)

[0211] In the formula, This is the baseline prediction value for the next time step. For error feedback coefficients, This represents the prediction error at time t. The actual value at time t. This represents the baseline prediction value at time t.

[0212] Example 8:

[0213] The main structure of this embodiment is the same as any one of embodiments 1-7. Further, in step 6), the comprehensive deviation between the actual observed value and the baseline predicted value is as follows:

[0214] (twenty two)

[0215] In the formula, For deviation fusion weights, The calculation is as follows:

[0216] (twenty three)

[0217] (twenty four)

[0218] (25)

[0219] In the formula, For residual covariance matrix estimation, The Mahalanobis distance;

[0220] The updated dynamic alert thresholds are shown below:

[0221] (26)

[0222] In the formula, , Nearly The mean and standard deviation of the deviation at each time point. This is the threshold adjustment coefficient;

[0223] The output warning signal is shown below:

[0224] (27)

[0225] In the formula, For continuous over-threshold window;

[0226] The advance warning time is shown below:

[0227] (28)

[0228] In the formula, The warning trigger time, This represents the actual time when the fault occurred.

[0229] Example 9:

[0230] The main structure of this embodiment is the same as any one of embodiments 1-8. Further, in step 7), the sequence matching step includes:

[0231] Step 7.1) Construct the multi-node residual matrix as follows:

[0232] (29)

[0233] In the formula, This is the abnormal window length;

[0234] Step 7.2) Extract the mean, variance, slope, and kurtosis statistical features of the deviation to form the current fault evolution feature sequence, i.e.:

[0235] (30)

[0236] In the formula, This is a vectorization operation used to transform the multi-node residual matrix R f Expanded column-wise into a one-dimensional feature vector; It is a statistical feature vector;

[0237] Step 7.3) Calculate the dynamic time-warped distance as shown below:

[0238] (31)

[0239] In the formula, For dynamic time-normalized distance, This is the reference feature sequence for the k-th typical failure mode. The length of the kth fault mode;

[0240] Step 7.4) Calculate the minimum matching distance and the starting position of the best matching subsequence for typical failure modes, i.e.:

[0241] (32)

[0242] Among all K typical fault modes, the fault mode with the smallest matching distance is selected to obtain the global minimum matching distance, as shown below:

[0243] (33)

[0244] In the formula, To find the minimum matching distance with the k-th typical failure mode, The starting position of the best matching subsequence. The minimum global matching distance is obtained by matching the abnormal sequence that triggers the early warning with all typical fault modes. Number the typical failure mode that is most similar to the current abnormal sequence;

[0245] Step 7.5) yields the following comprehensive measure of fault evolution:

[0246] (34)

[0247] In the formula, K is the number of reference failure modes in the typical failure mode library. This is the reference feature vector corresponding to the kth typical fault mode, used to characterize the typical evolutionary features of this type of fault.

[0248] Example 10:

[0249] The main structure of this embodiment is the same as any one of embodiments 1 to 9. Further, in step 8), the fault hazard level is as follows:

[0250] (35)

[0251] (36)

[0252] In the formula, Score the severity level of the fault; For the grade score weight, ; The deviation threshold; The maximum deviation is continuous. The overall deviation at each moment D t The maximum value in the output is as follows: Level I is the early abnormal development period, and the output maintenance strategy recommendation is basic inspection and light intervention; Level II is the mid-term accelerated evolution period, and the output maintenance strategy recommendation is component repair or partial replacement; Level III is the critical emergency deterioration period, and the output maintenance strategy recommendation is overhaul or complete replacement of key components.

[0253] Example 11:

[0254] This invention proposes a wind turbine early fault warning and classification method based on multidimensional time-varying topology and sequence matching. Addressing the problems of existing wind turbine fault warning methods, such as the tendency to generate false alarms under complex operating conditions, difficulty in characterizing the time-varying coupling relationships between multiple sensor variables, and the difficulty in further converting warning results into fault hazard levels, this method utilizes existing SCADA multidimensional operating data of the wind turbine. Through time-series data reconstruction, sensor dynamic topology inference, normal state baseline prediction, continuous deviation warning, and abnormal sequence matching, it outputs early warning signals, warning trigger times, warning lead times, and fault hazard levels. This method does not rely on adding new high-frequency vibration sensors, can improve the stability of fault warnings under complex operating conditions under existing wind farm monitoring conditions, and provides a basis for subsequent maintenance method selection and cost accounting.

[0255] This method mainly includes the following steps:

[0256] Step 1: Obtain multi-dimensional operation data of wind turbine SCADA and form raw time series samples.

[0257] Wind speed, wind direction, ambient temperature, active power, generator speed, pitch angle, gearbox temperature, bearing temperature, spindle speed, and other available operating parameters are read from the wind farm's existing SCADA system. The collected data is aligned according to timestamps, and data points indicating shutdowns, power limitations, communication anomalies, or obvious non-compliance with physical constraints are removed. Missing data is filled using nearest-time interpolation or window averaging. Operating parameters of different dimensions are standardized to obtain a multidimensional time series for subsequent calculations.

[0258] (1)

[0259] Where T represents the total number of sampling times, N represents the number of sensor nodes, and x t This represents the observation vector at time t, consisting of N operating parameters. After this step, the output is SCADA time-series data X, which has a consistent time sequence, has had outliers removed, and has unified dimensions.

[0260] Step 2: Reconstruct SCADA time series data using a double-layer sliding window and extract short-term fluctuation and long-term trend features.

[0261] Set the width W of the inner sliding window s and the width W of the outer sliding window l The inner window is used to preserve short-term local fluctuations in the operating state, while the outer window is used to extract trend changes over a longer time scale. For each time t, the inner window features and outer window features are calculated separately, and then the two types of features are concatenated to form the fused features required for time-varying topological inference and graph model input.

[0262] (2)

[0263] (3)

[0264] (4)

[0265] (5)

[0266] in, Indicates short-term window characteristics, Indicates long-term window characteristics, and These represent the mean and standard deviation of the inner window, respectively. and These represent the mean and standard deviation of the outer window, respectively. This represents the trend gradient features extracted from the outer window, Concat(·) represents the concatenation operation, and M represents the gradient feature dimension. Through this step, the original SCADA data is transformed into a fused feature z that simultaneously contains both short-term local changes and long-term evolutionary trends. t .

[0267] Step 3: Based on the fusion features, infer the time-varying topological relationships of the sensors to obtain the dynamic adjacency matrix.

[0268] Each sensor's operating parameter is treated as a graph node, and the conditional dependencies between sensors are treated as graph edges. To address the issue that sensor relationships change over time under different wind turbine operating conditions, the sample covariance matrix is ​​calculated within each sliding time interval. The Toeplitz-GL algorithm is then used to estimate the sparse precision matrix, and the precision matrix is ​​thresholded to obtain a time-varying adjacency matrix that reflects intra-layer and cross-layer correlations.

[0269] (6)

[0270] (7)

[0271] (8)

[0272] Among them, S k Θ represents the sample covariance matrix within the window at time k. k Let represent the accuracy matrix at time k, λ represent the sparsity regularization parameter, T(·) represent the Toeplitz operator, and τ represent the adjacency threshold. This represents the adjacency weight between sensor i and sensor j at time t. This step outputs the dynamic graph structure G. t =(V,E t A t ), which serves as the structural input for subsequent baseline prediction models.

[0273] Step 4: Calculate the dynamic features of graph nodes using a time-varying adjacency matrix and a dual attention mechanism.

[0274] The fusion feature z obtained in step two t And the time-varying adjacency matrix A obtained in step three t Input a spatiotemporal graph neural network. First, process the feature matrix of the l-th layer node. Linear projection is performed to obtain query, key, and value vectors; then spatial attention weights are calculated and fused with the time-varying adjacency matrix, so that the model can utilize data-driven adaptive weights while retaining sensor topological priors.

[0275] (9)

[0276] (10)

[0277] (11)

[0278] (12)

[0279] in, , , For learnable projection matrices, and Let d represent the query vector of node i and the key vector of node j, respectively. k Represents the attention dimension. Represents spatial attention weights. Let σ represent the attention weights after fusing topological priors, and σ represent the activation function. Through the above calculations, node features containing sensor spatial coupling relationships are obtained.

[0280] Subsequently, sequence attention is performed on the time dimension, enabling the model to use abnormal evolution information within the historical window to correct the features of the current node.

[0281] (13)

[0282] (14)

[0283] (15)

[0284] (16)

[0285] Where, β ts This represents the temporal attention weight of time t to historical time s. The output represents the features enhanced by temporal attention. SpatialAttn(·) represents the spatial attention module, TemporalAttn(·) represents the temporal attention module, FFN(·) represents the feedforward network, and LayerNorm(·) represents the layer normalization operation. After this step, the output is the dynamic features of the graph nodes enhanced by dual attention.

[0286] Step 5: Construct a baseline prediction model for normal operation and obtain the baseline prediction values ​​for future times.

[0287] The dynamic features of the graph nodes obtained in step four are input into a spatiotemporal graph neural network with an encoder-decoder structure. The encoder extracts deep spatiotemporal features from the historical window, and the decoder generates predicted values ​​for the next H steps of normal operation in an autoregressive manner. To reduce the accumulation of errors in the autoregressive prediction, the prediction error from the previous time step is introduced to robustly correct the prediction result for the next time step.

[0288] (17)

[0289] (18)

[0290] (19)

[0291] Among them, H enc This represents the semantic features output by the encoder. Indicates the input feature sequence. This represents the corresponding time-varying adjacency matrix sequence. This represents the predicted value at time t+k. Let γ represent the hidden state of the decoder at step k-1, and let e represent the error feedback coefficient. t This represents the prediction error at time t. This step yields the baseline prediction value for the wind turbine under normal operating conditions.

[0292] Step 6: Calculate the overall deviation between the actual observed value and the baseline predicted value, and trigger an early warning.

[0293] The actual fusion feature z t Compared with baseline forecast By comparing the results, the absolute residuals, covariance-weighted residuals, and Mahalanobis distance are calculated separately, and then combined according to their weights to form the overall deviation D. t The warning threshold is then dynamically updated based on the mean and standard deviation of the recent deviation sequence, and a continuous over-threshold window is set to avoid false alarms caused by single-point fluctuations.

[0294] (20)

[0295] (twenty one)

[0296] ;(twenty two)

[0297] (twenty three)

[0298] (twenty four)

[0299] Where Σ represents the estimation of the residual covariance matrix, d M(·) Let w1, w2, and w3 represent the deviation fusion weights, and δ represent the Mahalanobis distance. t This represents the dynamic warning threshold at time t. and They represent near L respectively θ The mean and standard deviation of the deviation at each time point, where η represents the threshold adjustment coefficient.

[0300] (25)

[0301] (26)

[0302] Among them, L w Indicates a continuous out-of-threshold window, t w Indicates the time when the warning is triggered, t f The value represents the actual time of the fault occurrence, and Δt represents the lead time for the warning. If L is continuous... w At each time point, D satisfies D. t >δ t If the warning signal is triggered, an early warning signal will be output and the trigger time will be recorded. Based on the magnitude of Δt, the early warning signal can be further divided into early warning, mid-term warning, and emergency warning.

[0303] Step 7: Extract fault evolution features from the abnormal sequences that have triggered warnings and perform sequence matching with typical fault modes.

[0304] After the warning is triggered in step six, the abnormal window starting from the warning trigger time is captured, a multi-node residual matrix is ​​constructed, and statistical features such as mean deviation, variance, slope, and kurtosis are extracted to form the current fault evolution feature sequence. Subsequently, using the multi-dimensional matrix profile and dynamic time warping distance, the current abnormal sequence is matched one by one with the reference sequences in the typical fault mode library to obtain the most similar fault mode and its matching distance.

[0305] (27)

[0306] (28)

[0307] (29)

[0308] (30)

[0309] (31)

[0310] Among them, R f L represents a multi-node residual matrix. f Indicates the length of the exception window, s f f represents the fault evolution feature vector. stat Represents a statistical eigenvector. L represents the reference feature sequence for the k-th typical failure mode. k The length of the k-th fault mode is represented by DTW(·), and DTW(·) represents the dynamic time warping distance, m. k This represents the minimum matching distance to the k-th typical failure mode. M represents the starting position of the best matching subsequence. f This represents a comprehensive measure of fault evolution.

[0311] Step 8: Calculate the fault hazard level based on the early warning lead time, maximum deviation, and fault mode matching distance, and output maintenance strategy recommendations.

[0312] In obtaining the early warning lead time Δt and the maximum deviation D within the observation window. max and the minimum DTW distance m from the typical failure mode min Then, the fault hazard level score S is calculated. The shorter the warning lead time, the greater the maximum deviation, and the smaller the fault mode matching distance, the higher the hazard level score. Based on the score range, the fault status is divided into Level I, Level II, or Level III, and mapped to the corresponding maintenance strategy.

[0313] (32)

[0314] (33)

[0315] Where α, β, and γ represent the weights of the grade scores, and α + β + γ = 1; D th The threshold represents the deviation; ε represents a small positive number to avoid division by zero; S1 and S2 represent the level classification thresholds. Level I represents the early abnormal development period, corresponding to basic inspection and mild intervention; Level II represents the mid-term accelerated evolution period, corresponding to component maintenance or partial replacement; Level III represents the critical emergency deterioration period, corresponding to major overhaul or complete replacement of key components.

[0316] Through the above steps, this invention can identify early faults and quantify hazard levels under complex operating conditions using existing SCADA data without adding extra high-frequency monitoring hardware. Compared with early warning methods that rely solely on single-variable thresholds or static correlations, this invention can simultaneously utilize short-term fluctuations, long-term trends, time-varying sensor topology relationships, and abnormal sequence matching results to reduce false alarms caused by operating condition fluctuations, and further enable the early warning results to serve in the selection of maintenance methods, determination of maintenance timing, and assessment of maintenance costs.

Claims

1. A method for early fault warning and classification of wind turbines based on multidimensional time-varying topology and sequence matching, characterized in that, Includes the following steps: Step 1) Obtain multi-dimensional operation data of wind turbine SCADA, perform data preprocessing, and form raw time series samples; Step 2) Reconstruct the original time series samples using a double-layer sliding window, extract the short-term fluctuation and long-term trend features of the original time series samples, and output the fused features; Step 3) Determine the time-varying topological relationships of the sensors based on the fusion features and construct a temporal adjacency matrix; Step 4) Calculate the dynamic features of graph nodes using a time-varying adjacency matrix and a dual attention mechanism; Step 5) Input the dynamic features of the graph nodes into the pre-trained baseline prediction model for normal operation to obtain the baseline prediction value for the normal operation state in the next H steps; Step 6) Calculate the overall deviation between the actual observed features and the baseline predicted values. If continuous All times satisfy If the warning signal is triggered, an early warning signal will be output, and the warning trigger time and the warning lead time will be recorded. Step 7) Extract fault evolution features from the abnormal sequences that have triggered warnings, and perform sequence matching with the fault evolution features of typical fault modes pre-stored in the database to obtain the most similar fault modes and matching distances. Step 8) Calculate the fault hazard level based on the early warning lead time, maximum deviation, and minimum matching distance of the fault mode, and output maintenance strategy recommendations.

2. The method for early fault warning and classification of wind turbines based on multidimensional time-varying topology and sequence matching according to claim 1, characterized in that, In step 1), the SCADA multidimensional operating data of the wind turbine includes wind speed, wind direction, ambient temperature, active power, generator speed, pitch angle, gearbox temperature, bearing temperature and main shaft speed. Data preprocessing refers to aligning the SCADA multidimensional operating data of wind turbine units according to timestamps, removing data points that are shut down, have limited power, have communication anomalies, or do not meet physical constraints, and using interpolation of nearby times or window mean to fill in missing data, and unifying the units of measurement.

3. The method for early fault warning and classification of wind turbines based on multidimensional time-varying topology and sequence matching according to claim 2, characterized in that, The original time series samples are shown below: (1) In the formula, T represents the total number of sampling times, and N represents the number of sensor nodes. Let t be the observation vector consisting of N operating parameters.

4. The method for early fault warning and classification of wind turbines based on multidimensional time-varying topology and sequence matching according to claim 3, characterized in that, In step 2), the fusion features are as follows: (2) In the formula, This is a short-term window characteristic. This is a long-term window feature. For the concatenation operation, M is the gradient feature dimension; The characteristics of a short-term window are as follows: (3) In the formula, The width of the inner sliding window. , The mean and standard deviation of the inner window are shown below: (4) The characteristics of a long-term window are as follows: (5) In the formula, The width of the outer sliding window. The trend gradient features extracted from the outer window. , The mean and standard deviation of the outer window are shown below: (6) (7) In the formula, x i This represents the original time series sample or observed operating parameters at the i-th sampling time.

5. The wind turbine early fault warning and classification method based on multidimensional time-varying topology and sequence matching according to claim 4, characterized in that, Step 3) involves constructing the temporal adjacency matrix, including: Each sensor's operating parameter is treated as a graph node, and the conditional dependencies between sensors are treated as graph edges. The sample covariance matrix is ​​calculated in each sliding time interval, and the sparse precision matrix is ​​estimated using the Toeplitz-GL algorithm. The sample covariance matrix is ​​shown below: (8) In the formula, Sk is the sample covariance matrix within the k-th topology estimation window; Lc is the length of the sliding time window used to estimate the sample covariance matrix; and zi is the fusion feature vector corresponding to the i-th sampling time. This is the transpose of zi; The sparse precision matrix is ​​shown below: (9) In the formula, Let k be the sample covariance matrix within the time window. Let k be the accuracy matrix. For sparsity regularization parameters, For Toeplitz transformation operator; Thresholding the precision matrix yields a time-varying adjacency matrix that reflects intra-layer and cross-layer associations. The temporal adjacency matrix is ​​shown below: (10) In the formula, This is the adjacency threshold. Let be the adjacency matrix between sensor i and sensor j at time t. Let be the sparse precision matrix between sensor i and sensor j at time t.

6. The method for early fault warning and classification of wind turbines based on multidimensional time-varying topology and sequence matching according to claim 5, characterized in that, Step 4) involves calculating the dynamic characteristics of graph nodes, including: Step 4.1) For the first Layer node feature matrix Performing a linear projection yields the query, key, and value vectors, as shown below: (11) In the formula, , , For learnable projection matrix; Step 4.2) Calculate the spatial attention weights as follows: (12) In the formula, , Let i be the query vector of node i and j be the key vector of node j. For the attention dimension; Step 4.3) Fuse the spatial attention weights with the time-varying adjacency matrix to obtain the attention weights after fusing the topological prior, as shown below: (13) Step 4.4) After applying the activation function, the node features containing the spatial coupling relationship of the sensor are obtained, that is: (14) In the formula, For activation function, The longitude projection matrix of the j-th sensor node The value vector obtained by mapping; Step 4.5) Perform sequence attention calculation on the time dimension to obtain the dynamic features of the graph nodes as shown below: (15) (16) (17) (18) In the formula, The input is the node feature sequence within the time window. These are the query projection matrix, key projection matrix, and value projection matrix in the time attention module. Let t be the temporal attention weight of historical time s. For activation function, Let be the query vector corresponding to time t. This is the transpose of the key vector corresponding to historical time s. The length of the historical time window involved in the time attention calculation. Features resulting from enhanced time attention. For spatial attention modules, For time attention module, It is a feedforward network. This is a layer normalization operation.

7. The method for early fault warning and classification of wind turbines based on multidimensional time-varying topology and sequence matching according to claim 6, characterized in that, In step 5), the baseline prediction model for normal operation is a spatiotemporal graph neural network, which includes an encoder and a decoder; The encoder is used to extract deep spatiotemporal features from the historical window, and the decoder generates predicted values ​​for the next H steps of normal operation in an autoregressive manner. The baseline prediction model for normal operating conditions is trained using historical datasets; Historical datasets include multiple input-output sample pairs; The input is the dynamic features of the graph nodes; the output is the baseline prediction value under normal operating conditions in the next H steps. The baseline prediction model for normal operating conditions is shown below: (19) In the formula, The semantic features output by the encoder, Given the input feature sequence, This corresponds to the time-varying adjacency matrix sequence; The baseline predictions for the future H-step normal operation are shown below: (20) In the formula, The predicted value at time t+k. This is the hidden state of the decoder at step k-1; The prediction error from the previous time step is introduced to correct the baseline prediction value for the next time step, that is: (21) In the formula, This is the baseline prediction value for the next time step. For error feedback coefficients, This represents the prediction error at time t. The actual value at time t. This represents the baseline prediction value at time t.

8. The method for early fault warning and classification of wind turbines based on multidimensional time-varying topology and sequence matching according to claim 7, characterized in that, In step 6), the combined deviation between the actual observed values ​​and the baseline predicted values ​​is shown below: (22) In the formula, For deviation fusion weights, The calculation is as follows: (23) (24) (25) In the formula, For residual covariance matrix estimation, The Mahalanobis distance; The updated dynamic alert thresholds are shown below: (26) In the formula, , Nearly The mean and standard deviation of the deviation at each time point. This is the threshold adjustment coefficient; The output warning signal is shown below: (27) In the formula, For continuous over-threshold window; The advance warning time is shown below: (28) In the formula, The warning trigger time, This represents the actual time when the fault occurred.

9. A method for early fault warning and classification of wind turbines based on multidimensional time-varying topology and sequence matching according to claim 8, characterized in that, In step 7), the sequence matching step includes: Step 7.1) Construct the multi-node residual matrix as follows: (29) In the formula, This is the abnormal window length; Step 7.2) Extract the mean, variance, slope, and kurtosis statistical features of the deviation to form the current fault evolution feature sequence, i.e.: (30) In the formula, For vectorization operations, It is a statistical feature vector; Step 7.3) Calculate the dynamic time-warped distance as shown below: (31) In the formula, For dynamic time-normalized distance, This is the reference feature sequence for the k-th typical failure mode. The length of the kth fault mode; Step 7.4) Calculate the minimum matching distance and the starting position of the best matching subsequence for typical failure modes, i.e.: (32) Among all K typical fault modes, the fault mode with the smallest matching distance is selected to obtain the global minimum matching distance, as shown below: (33) In the formula, To find the minimum matching distance with the k-th typical failure mode, The starting position of the best matching subsequence. The minimum global matching distance is obtained by matching the abnormal sequence that triggers the early warning with all typical fault modes. Number the typical failure mode that is most similar to the current abnormal sequence; Step 7.5) yields the following comprehensive measure of fault evolution: (34) In the formula, K is the number of reference failure modes in the typical failure mode library. This is the reference feature vector corresponding to the kth typical fault mode.

10. A method for early warning and classification of wind turbine faults based on multidimensional time-varying topology and sequence matching according to claim 9, characterized in that, In step 8), the fault hazard level is as follows: (35) (36) In the formula, Score the severity level of the fault; For the grade score weight, ; The deviation threshold; The maximum deviation is continuous. The overall deviation at each moment D t The maximum value in the range; Level I is the early abnormal development period; Level II is the mid-term accelerated evolution period; Level III is the critical emergency deterioration period.

Images