Two-stage fan equipment quality abnormity tracing method based on hierarchical information representation

Abstract

The invention provides a two-stage fan equipment quality abnormity tracing method based on hierarchical information characterization, and relates to the technical field of fan operation and maintenance quality management and control, and the method comprises the steps: obtaining SCADA monitoring data, operation logs and operation and maintenance records of fan equipment, carrying out data preprocessing through overlapping sampling, and constructing a time sequence database; calculating the maximum information coefficient, related to the fault, of each feature in the time sequence database, and screening key features to form a feature subset; constructing a double-end Transform model, taking the feature subset as the input of the model, introducing a classification mark, and outputting a feature prediction value and fault category probability distribution; the difference between the fault sample and the normal sample is calculated based on the mahalanobis distance, the tracing key features are screened according to the contribution degree of the difference, and multi-dimensional time domain statistical indexes are extracted from the tracing key features to construct a high-dimensional feature matrix; and inputting the high-dimensional feature matrix into a tracing model represented by hierarchical information, judging an abnormal reason through a feature matrix threshold value, and outputting a tracing result.

Description

Technical Field

[0001] This invention relates to the field of wind turbine operation and maintenance quality control technology, and in particular to a two-stage wind turbine equipment quality anomaly tracing method based on hierarchical information representation. Background Technology

[0002] Wind energy, as a core component of clean energy, plays a vital role in the global energy transition. Wind turbine equipment, especially megawatt-class and above units, is the core equipment of wind power systems, and its operational stability directly affects power generation efficiency and operation and maintenance costs. However, wind turbine equipment operates in a complex environment, such as strong winds, temperature differences, and dust storms. It also has many components with close coupling between them, leading to quality anomalies characterized by complex causes, insidious propagation, and severe consequences.

[0003] The current management and control of wind turbine quality anomalies faces the following problems: First, there is the challenge of data processing. The operational data collected by SCADA systems is high-dimensional, typically 20-50 dimensions, and contains missing data and noise. Furthermore, the proportion of fault samples is low, resulting in low data utilization. Second, there is insufficient accuracy in early warning. Traditional early warning models, such as SVM and LSTM, struggle to capture the nonlinear coupling relationships in high-dimensional time-series data, leading to low recognition rates for early, subtle anomalies. Finally, the tracing mechanism is ambiguous. Existing methods mostly stop at the stage of issuing early warnings, lacking quantitative tracing of the causes of anomalies. Maintenance personnel must rely on experience for troubleshooting, which is time-consuming and prone to misjudgment. Summary of the Invention

[0004] The purpose of this invention is to provide a two-stage wind turbine equipment quality anomaly tracing method based on hierarchical information representation, so as to solve the problems of low data utilization, feature redundancy, insufficient early warning accuracy and difficulty in anomaly tracing in the prior art mentioned in the background.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a two-stage wind turbine equipment quality anomaly tracing method based on hierarchical information representation, comprising the following steps: acquiring SCADA monitoring data, operation logs, and maintenance records of the wind turbine equipment; performing data preprocessing using overlapping sampling to construct a time-series database; calculating the maximum information coefficient of each feature in the time-series database related to the fault, and selecting key features to form a feature subset; constructing a dual-head Transformer model including a prediction head and a classification head, using the feature subset as input to the model, introducing classification labels, and outputting feature prediction values ​​and fault category probability distributions; calculating the difference between fault samples and normal samples identified by the dual-head Transformer model based on Mahalanobis distance, and selecting key tracing features according to the contribution of the difference, extracting multi-dimensional time-domain statistical indicators from the key tracing features to construct a high-dimensional feature matrix; inputting the high-dimensional feature matrix into the hierarchical information representation tracing model, determining the cause of the anomaly through a feature matrix threshold, and outputting the root cause tracing result of the quality anomaly.

[0006] Optionally, key information is extracted from the operation logs and maintenance records to obtain fault status labels aligned with timestamps, and these labels are integrated with the SCADA monitoring data to form time-series data. Assuming a time window length of T and an overlap rate of 40%-60%, samples of T consecutive time steps are extracted from the time-series data, with adjacent samples overlapping by T / 2 time steps, to construct a time-series database. For missing values ​​in the time-series data, the mean of the feature within the sliding window is used for filling. The filled features are then standardized to eliminate the influence of units.

[0007] Optionally, the step of calculating the maximum information coefficients of each feature in the time-series database related to the fault and screening key features specifically includes: calculating the maximum information coefficients of the preprocessed features and the fault state labels, and performing grid division in a two-dimensional plane based on the sample size. The calculation formula is as follows: In the formula: For the mutual information between feature X and label Y, To find the minimum grid dimension, the grid dimension (a, b) satisfies , To control the upper limit parameter of grid partitioning complexity; N is the sample size; sort by MIC value from high to low, and select features with MIC values ​​greater than a preset threshold; construct a simple early warning model and compare the accuracy improvement to verify the effectiveness of the features and determine the key features.

[0008] Optionally, the step of using the feature subset as input to the model, introducing classification labels, and outputting feature prediction values ​​and fault category probability distributions specifically includes: constructing an input sequence from the key feature subset according to time windows, generating query, key, and value matrices through linear mapping, calculating attention weights using a self-attention mechanism, and capturing feature relationships of different dimensions through a multi-head attention mechanism; adding classification labels at the beginning of the input sequence to aggregate global information, and using them as input to the classification head after Transformer encoding; outputting feature prediction values ​​for the next time step through the prediction head, and outputting the fault category probability distribution through the classification head.

[0009] Optionally, the dual-head Transformer model is optimized based on multi-task weighted loss, specifically including: for imbalanced data of fault samples, weighted cross-entropy loss is used to process the data and assign higher weights to minority class samples; mean squared error loss is used to calculate the difference between the predicted value and the true value; classification loss and prediction loss are fused through a weighted joint loss function, and backpropagation is used to optimize the model parameters.

[0010] Optionally, the step of calculating the difference between the faulty samples and normal samples identified by the dual-head Transformer model based on Mahalanobis distance, and selecting key traceability features based on the contribution of the difference, specifically includes: calculating the Mahalanobis distance between the faulty samples and normal samples output by the dual-head Transformer model; constructing a Mahalanobis distance matrix, and ranking the contribution of each feature to the distance matrix, selecting the features with the highest contribution as key traceability features, wherein the contribution is evaluated by the change in sample distance after removing specific features.

[0011] Optionally, the step of extracting multi-dimensional time-domain statistical indicators from the key traceability features to construct a high-dimensional feature matrix specifically includes: extracting trend indicators, dispersion indicators, extreme value indicators, distribution pattern indicators, and dynamic change indicators from the key traceability features; constructing a feature matrix based on the number of key features and statistical indicators, with each row corresponding to the statistical features of a sample, used to characterize the statistical characteristics of abnormal states.

[0012] Optionally, the step of inputting the high-dimensional feature matrix into the tracing model of hierarchical information representation and judging the cause of the anomaly through the feature matrix threshold specifically includes: predicting the real-time data of the wind turbine based on the dual-head Transformer model, and judging it as a potential anomaly when the failure probability exceeds the threshold; inputting the high-dimensional feature matrix of the potential anomaly sample into the tracing model of hierarchical information representation, setting the threshold range of the normal state feature matrix based on the statistics of historical normal samples, calculating the deviation between the anomaly sample matrix and the normal threshold, and locating the component, subsystem and / or whole machine with the largest deviation.

[0013] Optionally, the construction steps of the traceability model for the hierarchical information representation specifically include: establishing a three-level physical hierarchy relationship of component level, subsystem level, and whole machine level based on the physical structure and functional association of wind turbine equipment. Component level corresponds to the time-domain statistical indicators of monitored parts, subsystem level is obtained by fusing functionally associated component-level features, and whole machine level is generated by weighted aggregation of subsystem-level features; establishing a mapping relationship between each level and the feature matrix, where component-level features are mapped to time-domain statistical indicators in the feature matrix, subsystem-level features are obtained through correlation calculation of component-level features, and whole machine-level features are obtained through anomaly propagation results of subsystem-level features; constructing an inter-level association rule base based on historical fault data and quantifying the association strength; and, based on the hierarchical relationship and association rule base, calculating the deviation between the abnormal sample feature matrix and the normal threshold, and verifying the deviation propagation path to achieve root cause localization from feature matrix deviation to physical component anomalies.

[0014] On the other hand, the present invention also provides a two-stage wind turbine equipment quality anomaly tracing system based on hierarchical information representation, comprising: an acquisition module for acquiring SCADA monitoring data, operation logs, and maintenance records of wind turbine equipment, performing data preprocessing using overlapping sampling, and constructing a time-series database; an early warning feature screening module for calculating the maximum information coefficient of each feature in the time-series database related to the fault, and screening key features to form a feature subset; a model building module for constructing a dual-head Transformer model including a prediction head and a classification head, using the feature subset as input to the model, introducing classification labels, and outputting feature prediction values ​​and fault category probability distributions; a tracing feature screening module for calculating the difference between fault samples and normal samples identified by the dual-head Transformer model based on Mahalanobis distance, and screening tracing key features according to the contribution of the difference, extracting multi-dimensional time-domain statistical indicators from the tracing key features to construct a high-dimensional feature matrix; and a tracing result output module for inputting the high-dimensional feature matrix into the hierarchical information representation tracing model, judging the cause of the anomaly through the feature matrix threshold, and outputting the root cause tracing result of the quality anomaly.

[0015] Compared with the prior art, the beneficial effects of the present invention are: The first phase of this application, based on the maximum information coefficient and the dual-head Transformer model, enables intelligent early warning, achieving high-precision capture and classification of early, subtle fault signs in wind turbine equipment, significantly improving the sensitivity and timeliness of fault identification. The second phase, based on feature extraction using Mahalanobis distance and multidimensional time-domain statistical indicators, combined with a tracing model representing hierarchical information, accurately maps data anomalies to specific physical levels of components, subsystems, and the entire unit. This achieves clear and interpretable localization from data alarms to the physical root cause, providing a clear direction for operation and maintenance decisions and greatly improving the intelligence and precision of wind turbine equipment quality risk management. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method steps of the present invention.

[0017] Figure 2 This is a block diagram of the feature engineering technology of the present invention.

[0018] Figure 3 This is a block diagram of the dual-head Transformer encoder of the present invention.

[0019] Figure 4 This is a flowchart of the multi-source data acquisition process of the present invention.

[0020] Figure 5 This is a flowchart of the data preprocessing process of the present invention.

[0021] Figure 6 This is a flowchart of the early warning model identification process of the present invention.

[0022] Figure 7 This is a flowchart of the traceability model for this invention.

[0023] Figure 8 This is a flowchart of the traceability model output of the present invention.

[0024] Figure 9 This is a structural block diagram of the traceability model for hierarchical information representation in this invention.

[0025] Figure 10 This is a schematic diagram of the system structure of the present invention.

[0026] In the diagram: 10 - Acquisition module, 20 - Early warning feature selection module, 30 - Model building module, 40 - Traceability feature selection module, 50 - Traceability result output module. Detailed Implementation

[0027] The present invention will now be clearly and completely described in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0028] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be used interchangeably where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0029] Those skilled in the art will understand that, unless explicitly stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in the specification of this application means the presence of features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0030] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0031] It should be understood that the sequence number and size of each step in this embodiment do not imply the order of execution. The execution order of each process is determined by its function and internal logic, and should not constitute any limitation on the implementation process of this application embodiment.

[0032] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0033] Please refer to Figures 1-9This invention discloses a two-stage wind turbine equipment quality anomaly tracing method based on hierarchical information representation. The steps include: acquiring SCADA monitoring data, operation logs, and maintenance records of the wind turbine equipment; performing data preprocessing using overlapping sampling to construct a time-series database; calculating the maximum information coefficient related to faults for each feature in the time-series database; selecting key features to form a feature subset; constructing a dual-head Transformer model including a prediction head and a classification head; using the feature subset as input to the model and introducing classification labels to output feature prediction values ​​and fault category probability distributions; calculating the difference between fault samples and normal samples identified by the dual-head Transformer model based on Mahalanobis distance; selecting key tracing features based on the contribution of the difference; extracting multi-dimensional time-domain statistical indicators from the key tracing features to construct a high-dimensional feature matrix; inputting the high-dimensional feature matrix into the hierarchical information representation tracing model; determining the cause of the anomaly through a feature matrix threshold; and outputting the root cause tracing result of the quality anomaly.

[0034] The first phase of this application, based on the maximum information coefficient and the dual-head Transformer model, enables intelligent early warning, achieving high-precision capture and classification of early, subtle fault signs in wind turbine equipment, significantly improving the sensitivity and timeliness of fault identification. The second phase, based on feature extraction using Mahalanobis distance and multidimensional time-domain statistical indicators, combined with a tracing model representing hierarchical information, accurately maps data anomalies to specific physical levels of components, subsystems, and the entire unit. This achieves clear and interpretable localization from data alarms to the physical root cause, providing a clear direction for operation and maintenance decisions and greatly improving the intelligence and precision of wind turbine equipment quality risk management.

[0035] In some embodiments, key information is extracted from the operation logs and maintenance records to obtain fault status labels aligned with timestamps, and these labels are integrated with the SCADA monitoring data to form time-series data. Assuming a time window length of T and an overlap rate of 40%-60%, samples of T consecutive time steps are extracted from the time-series data, with adjacent samples overlapping by T / 2 time steps, to construct a time-series database. For missing values ​​in the time-series data, the mean of the feature within the sliding window is used for filling. The filled features are then standardized to eliminate the influence of dimensions.

[0036] Specifically, firstly, overlapping sampling is performed: based on the wind turbine operating cycle, such as sampling every 10 minutes, 18 time steps are taken as the window width {T=18}, and with a sliding step size of {T / 2=9}, the continuous SCADA data is divided into overlapping windows, and the sample format is as follows: ,in This is the feature vector of the i-th window at time j, which improves data utilization.

[0037] Missing value handling: handling missing data points The average of the three valid data points before and after it is used to fill the gap, and the formula is as follows: ; If there are fewer than 3 data points at the boundary, the average of the existing valid data will be taken. Z-score standardization: Maps data to a distribution with a mean of 0 and a standard deviation of 1. The formula is: ; In the formula: Let j be the global mean of the j-th feature. Let be the global standard deviation of the j-th feature.

[0038] This application employs a sliding window sampling method with an overlap rate of 40%-60%, and in this embodiment, a 50% overlap rate. This generates a sufficient number of samples with limited data and ensures adequate temporal continuity between adjacent samples. Since the changes in the wind turbine's operating state are continuous and gradual, a 40%-60% overlap rate ensures sufficient temporal correlation information between adjacent samples, avoiding the loss of local features due to fragmented sampling windows. If the overlap rate is too low, the number of samples will be insufficient and the temporal correlation will be weak; if the overlap rate is too high, a large number of redundant samples will be generated, increasing the computational burden and diminishing marginal benefits. Wind turbine SCADA data typically has a fixed sampling period, such as 10 minutes / time. The window length T is set according to its operating cycle, for example, 3 hours corresponds to 18 time steps; 50% overlap, i.e., 9 time steps, corresponds to 1.5 hours, which can effectively cover local fluctuations, such as parameter fluctuations caused by short-term wind speed changes, and medium-term trends, such as the equipment temperature rise process, matching the dynamic characteristics of the actual wind turbine operation and improving the effectiveness of the preprocessed data.

[0039] In some embodiments, the step of calculating the maximum information coefficients of each feature in the time-series database related to the fault and screening key features specifically includes: calculating the maximum information coefficients of the preprocessed features and the fault state labels, and performing grid division in a two-dimensional plane based on the sample size. The calculation formula is as follows: In the formula: For the mutual information between feature X and label Y, To find the minimum grid dimension, the grid dimension (a, b) satisfies , To control the upper limit parameter of grid partitioning complexity; N is the sample size; sort by MIC value from high to low, and select features with MIC values ​​greater than a preset threshold; construct a simple early warning model and compare the accuracy improvement to verify the effectiveness of the features and determine the key features.

[0040] Specifically, based on the selection of key features with the maximum information coefficient, the maximum information coefficient between each monitoring indicator and the wind turbine operating status is calculated, key features with strong correlation are screened, the input dimension is reduced, and nonlinear correlation information is retained.

[0041] Dataset construction: using standardized feature data as input Where {D} is the original feature dimension, and the running state label S is the output, where, , This indicates that it is normal. Indicates different fault types.

[0042] MIC calculation: for features With tags Divide in a two-dimensional plane Grid in Calculate the mesh mutual information and take the maximum value. The formula is: ; In the formula: For grid Mutual information under the following For the maximum mutual information under the same dimension, To find the minimum grid dimension, the grid dimension (a, b) satisfies , The parameter is used to control the upper limit of the complexity of grid partitioning; N is the sample size.

[0043] Feature filtering: Setting the MIC threshold Select Key feature set It retains the linear and nonlinear characteristics that are strongly correlated with the fault state.

[0044] Verify the effectiveness of the feature subset: Construct a simple early warning model, such as logistic regression, using the feature subset. If the model accuracy is more than 10% higher than that of the full feature set, then it is determined to be the final input feature.

[0045] This application employs the Maximum Information Coefficient (MIC) as the feature selection criterion, which can efficiently uncover the complex linear and nonlinear dependencies between features and fault states, overcoming the limitations of traditional linear methods. By calculating the maximum value of grid mutual information and setting a threshold, it can quickly select the most relevant subset of key features from dozens of high-dimensional features, significantly reducing the dimensionality of the model input, eliminating the interference of redundant features, and thus improving the training efficiency, generalization ability, and interpretability of the subsequent early warning model.

[0046] In some embodiments, the steps of using the feature subset as input to the model, introducing classification labels, and outputting feature prediction values ​​and fault category probability distributions specifically include: constructing an input sequence from the key feature subset according to time windows, generating a query, key, and value matrix through linear mapping, calculating attention weights using a self-attention mechanism, and capturing feature relationships of different dimensions through a multi-head attention mechanism; adding classification labels at the beginning of the input sequence to aggregate global information, and using them as input to the classification head after Transformer encoding; outputting feature prediction values ​​for the next time step through the prediction head, and outputting the fault category probability distribution through the classification head.

[0047] Specifically, the feature subsets are used to construct the input sequence according to time windows. T is the time step, and d is the feature dimension, generated through linear mapping. matrix: In the formula: These are trainable parameters.

[0048] Key feature time series Mapped to model dimensions through the embedding layer The vector is given by the formula: ;; In the formula: This is the embedding matrix.

[0049] Perform position encoding: Add sine and cosine position encoding to impart timing information for each time step. and dimensions : ,like It is an even number; ,like It is an odd number; The embedded sequence is .

[0050] Adding a classification tag [CLS] token: Inserting a learnable vector at the beginning of the sequence. Used to aggregate global features, the new sequence is .

[0051] Bullish Self-Attention: Mapped to The Q, K, V vectors of the head, where each head has dimensions. : .

[0052] Single-head attention calculation: ; Multi-head fusion: ; in, This is the output matrix.

[0053] Constructing Transformer encoder layers: Each encoder layer contains a multi-head attention and feedforward neural network structure, optimized through residual connections and layer normalization. ; ; where FFN is a two-layer fully connected network with ReLU as the intermediate activation function.

[0054] Dual-head output: stacking After one encoder layer, the classification head is based on Vector output fault category probability , The prediction head outputs the state parameters for the next time step. Linear is a linear mapping layer.

[0055] This application constructs a dual-head Transformer model and introduces a classification token [CLS], enabling the model to simultaneously learn two tasks: time series prediction and fault classification. The self-attention mechanism effectively captures long-distance dependencies and complex coupling characteristics in high-dimensional time series data, while the classification token [CLS] provides rich contextual features for the classification head. This joint learning architecture for prediction and classification ensures that the model extracts representations that include both the data's own variation patterns and fault discrimination information. The two tasks mutually reinforce each other, jointly improving the model's overall early warning performance.

[0056] In some embodiments, the dual-head Transformer model is optimized based on multi-task weighted loss, specifically including: for imbalanced data of fault samples, weighted cross-entropy loss is used to process the data and assign higher weights to minority class samples; mean squared error loss is used to calculate the difference between the predicted value and the true value; classification loss and prediction loss are fused through a weighted joint loss function, and backpropagation is used to optimize the model parameters.

[0057] Specifically, the classification loss is calculated as follows: to address the problem of few faulty samples, weights are assigned to each category. ,in, For category Sample size The total sample size is given by the formula: In the formula: If the sample Category Otherwise, it is 0; Predict samples for the model Category The probability of.

[0058] Calculate regression loss: measure the difference between predicted and actual values ​​of state parameters. ; In the formula: For the first Sample No. The true value of the dimensional feature. This is a predicted value; Calculate the joint loss: by weights Balancing two types of losses: ; The Adam optimizer is used, where the initial learning rate is... The weight decays to Minimize the total loss to improve the model's sensitivity to minority class failures.

[0059] This application addresses the problem of highly imbalanced wind turbine fault samples by designing a multi-task weighted loss function. It assigns higher loss weights to fault categories with fewer samples, forcing the model to pay more attention to minority class samples during training. This mechanism effectively alleviates the prediction bias caused by the large differences in the number of samples in each category, significantly improves the recognition rate of rare fault types, and ensures the balanced sensitivity of the early warning model to various fault types and the overall accuracy of early warning.

[0060] In some embodiments, the step of calculating the difference between faulty samples and normal samples identified by the dual-head Transformer model based on Mahalanobis distance, and selecting key traceability features based on the contribution of the difference, specifically includes: calculating the Mahalanobis distance between faulty samples and normal samples output by the dual-head Transformer model; constructing a Mahalanobis distance matrix, and ranking the contribution of each feature to the distance matrix, selecting the feature with the highest contribution as the key traceability feature, wherein the contribution is evaluated by the change in sample distance after removing a specific feature.

[0061] Specifically, sample set construction: selecting from historical data A normal sample , To standardize the key feature vector, select from the early warning results one abnormal sample ; Statistical calculation: Normal sample mean vector Covariance matrix: ; In the formula: M is the total number of normal samples. This is an abnormal sample. This is the mean vector of the normal samples.

[0062] Mahalanobis distance calculation: outlier samples The Mahalanobis distance from the normal sample is: ; In the formula: Abnormal sample Mahalanobis distance from normal samples This is the mean vector of the normal samples.

[0063] Mahalanobis distance eliminates the influence of feature dimensions and correlation, and more accurately quantifies differences. The calculation formula is as follows: ; In the formula: These are the original eigenvalues. The characteristic mean, The characteristic standard deviation, This is the standardized value.

[0064] Screening for significant anomalies: Calculate the mean Mahalanobis distance of normal samples. with standard deviation Select The sample is a significantly abnormal sample.

[0065] Feature contribution calculation: Through covariance matrix decomposition, the contribution of the {k}th dimension feature to the Mahalanobis distance is: ; In the formula: It is the Kth diagonal element of the covariance matrix.

[0066] The top 6 features by contribution were selected as key features for source tracing. .

[0067] This application calculates sample differences and constructs a distance matrix based on Mahalanobis distance, which can fully consider the correlation between features and reflects the true differences between samples better than Euclidean distance. By calculating the change in contribution to the overall distance matrix after removing a single feature, the contribution of each feature to the fault can be accurately quantified. This allows for further screening of the most sensitive and contributing features to fault tracing from key features, providing the most crucial and direct input for subsequent accurate root cause analysis.

[0068] In some embodiments, the step of extracting multidimensional time-domain statistical indicators from key traceability features to construct a high-dimensional feature matrix specifically includes: extracting trend indicators, dispersion indicators, extreme value indicators, distribution pattern indicators, and dynamic change indicators from key traceability features; constructing a feature matrix based on the number of key features and the number of statistical indicators, with each row corresponding to the statistical features of a sample, used to characterize the statistical properties of abnormal states.

[0069] Specifically, multidimensional time-domain statistical indicator extraction: Nine types of statistical indicators are extracted from key features, including: Central trends: mean, median.

[0070] Dispersion: standard deviation, coefficient of variation.

[0071] Extreme values: maximum value and minimum value.

[0072] Distribution patterns: skewness, kurtosis.

[0073] Dynamic change: Mean of the absolute value of adjacent differences.

[0074] Mean: .

[0075] Standard deviation: .

[0076] Maximum value: .

[0077] Minimum value: .

[0078] Skewness: ; measures the symmetry of the distribution.

[0079] Kuroshi: ; measures the steepness of the distribution.

[0080] Coefficient of variation: The dispersion index that eliminates the influence of the mean.

[0081] Percentage of zero values: ; This is an indicator function.

[0082] Mean of the absolute values ​​of adjacent differences: ; to measure the degree of temporal fluctuation.

[0083] Feature matrix construction: Integrating the 9-dimensional indicators of 6 key features into matrix ,in Indicates the first The first feature The matrix contains index values, with rows representing features and columns representing statistical indicators.

[0084] This application extracts multi-category time-domain statistical indicators from key traceability features, enabling a comprehensive and three-dimensional characterization of behavioral patterns under fault conditions from multiple dimensions, including central trend, dispersion, extreme values, distribution patterns, and dynamic changes. The high-dimensional feature matrix constructed in this way contains richer and more refined state information compared to the original data or a single indicator, providing a powerful and comprehensive basis for the traceability model, allowing it to more accurately distinguish the subtle differences exhibited by different fault causes.

[0085] In some embodiments, the step of inputting the high-dimensional feature matrix into the tracing model of hierarchical information representation and determining the cause of the anomaly through the feature matrix threshold specifically includes: predicting the real-time data of the wind turbine based on the dual-head Transformer model, and determining it as a potential anomaly when the failure probability exceeds the threshold; inputting the high-dimensional feature matrix of the potential anomaly sample into the tracing model of hierarchical information representation, setting the threshold range of the normal state feature matrix based on the statistics of historical normal samples, calculating the deviation between the anomaly sample matrix and the normal threshold, and locating the component, subsystem and / or whole machine with the largest deviation.

[0086] Specifically, quality status prediction: Based on the early warning model, real-time data of the wind turbine is predicted. If the failure probability output by the classification head is greater than the threshold, it is judged as a potential anomaly.

[0087] Feature matrix generation: Potentially abnormal samples are fed into the tracing model, and multidimensional statistical indicators of their key features are extracted to form a feature matrix.

[0088] Anomaly cause tracing: Based on historical normal sample statistics, a threshold range for the normal state feature matrix is ​​set, the deviation between the abnormal sample matrix and the normal threshold is calculated, and combined with hierarchical information representation, such as the component corresponding to the feature with the largest hierarchical positioning deviation of component-subsystem-whole machine, the possible causes of the anomaly are output, such as gearbox cooling system failure, insufficient lubricating oil, etc.

[0089] This application compares the fault samples output by the early warning model with the statistical thresholds of historical normal samples, transforming abstract faults into quantifiable deviations from specific thresholds, thus making the judgment process objective and standardized. By combining hierarchical information representation of components, subsystems, and the entire machine, the deviation values ​​at the data layer can be mapped to specific physical components or subsystems, ultimately outputting location results that engineers can directly understand, greatly improving the efficiency of operation and maintenance decisions.

[0090] In some embodiments, the construction steps of the traceability model for hierarchical information representation specifically include: establishing a three-level physical hierarchy relationship of component level, subsystem level, and whole machine level based on the physical structure and functional association of wind turbine equipment, wherein the component level corresponds to the time-domain statistical indicators of monitored components, the subsystem level is obtained by fusing functionally associated component-level features, and the whole machine level is generated by weighted aggregation of subsystem-level features; establishing a mapping relationship between each level and the feature matrix, wherein component-level features are mapped to time-domain statistical indicators in the feature matrix, subsystem-level features are obtained through correlation calculation of component-level features, and whole machine-level features are obtained through anomaly transmission results of subsystem-level features; constructing an inter-level association rule base based on historical fault data and quantifying the association strength; and, based on the hierarchical relationship and association rule base, realizing the root cause localization from feature matrix deviation to physical component anomaly by calculating the deviation between the abnormal sample feature matrix and the normal threshold, combined with deviation transmission path verification.

[0091] Specifically, the hierarchy is first defined by its basis and content. Based on the physical structure and functional relationships of the wind turbine equipment, the hierarchy is divided into three levels, each corresponding to the core elements of a particular fault: The component level includes key parts directly involved in the generation and transmission of faults, with each component corresponding to directly monitorable raw indicators in the feature matrix. Next is the subsystem level, composed of functionally related components; subsystem-level features are obtained through the fusion of component-level features. Finally, the overall turbine level reflects the overall operating status of the wind turbine, incorporating comprehensive indicators such as power generation; its features are generated by weighted aggregation of subsystem-level features.

[0092] Secondly, there's the mapping relationship between hierarchical levels and feature matrices. Each level is directly correlated through indicators in the feature matrix. Taking a high gearbox oil temperature fault as an example: component-level features directly map to the time-domain statistical indicators of the corresponding components in the feature matrix, such as the mean, maximum, and MAE of gearbox oil temperature for the gearbox, and the standard deviation and skewness of bearing temperatures for the drive / non-drive ends. Subsystem-level features are calculated through the correlation of component-level features. For example, the temperature coordination of the transmission subsystem is characterized by the correlation coefficient between gearbox oil temperature and bearing temperature, and the degree of heat dissipation failure of the cooling subsystem is characterized by the ratio of the mean gearbox oil temperature to the mean cooling pump flow rate. Finally, system-level features are reflected through the abnormal transmission results of subsystem-level features. For example, the system power loss rate is calculated by multiplying the efficiency decrease of the transmission subsystem by the generator speed decrease.

[0093] Next, the hierarchical association rules are constructed. Based on the propagation path of the high gearbox oil temperature fault, causal association rules between levels are mined: frequent patterns of component anomaly - subsystem performance degradation - overall system index deviation are extracted from historical fault data. For example, insufficient cooling pump flow at the component level - decreased heat dissipation efficiency of the cooling subsystem at the subsystem level - gearbox oil temperature exceeding the threshold at the component level - decreased stability of the transmission subsystem at the subsystem level - reduced generator speed at the overall system level. The association strength is quantified, and effective rules are selected based on confidence and support to form a hierarchical association rule library.

[0094] Finally, the anomaly localization logic is based on hierarchical relationships. In tracing the cause of anomalies, the root cause is located through the following steps: Calculate the deviation between the anomaly sample feature matrix and the normal threshold, and filter out the component-level features with the largest deviations. Based on the hierarchical association rule base, verify whether the component-level deviation has propagated to the corresponding subsystem. If the subsystem-level deviation further propagates to the entire unit level, then the component is confirmed as the root cause of the anomaly; if only the component-level deviation does not trigger a subsystem-level anomaly, then it is determined to be a local component anomaly. Through the above hierarchical relationship construction, a precise mapping from feature matrix deviation to physical component anomalies is achieved, providing a clear logical path for tracing the root cause of wind turbine failures.

[0095] Low-level information: feature matrix The contribution of Mahalanobis distance reflects the statistical differences in the original features.

[0096] Intermediate information: Multi-head attention weight matrix of the Transformer model Positions with higher weights are more critical for anomaly identification and can pinpoint the starting time step of an anomaly.

[0097] Advanced information: Anomaly propagation path pattern library, built based on historical fault cases, such as high gearbox oil temperature - increased drive end bearing temperature - pitch angle lock.

[0098] Anomaly path identification: First, time localization is performed, and attention weights are used in intermediate information. No. Head Time Step and The strength of the association, the time step of the sudden increase in screening weight As an abnormal initiation point; among them, a sudden increase in the judgment weight is considered in the multi-head attention weight matrix of the Transformer model. The Middle Time step correlation strength of the head The following quantitative conditions are used for joint determination: Absolute weight threshold condition: Based on historical normal sample statistics, set an upper limit threshold for attention weights under normal conditions. ,when When the time comes, mark it as a potential candidate for a sudden increase; Dynamic rate of change condition: Calculate the relative rate of change of the weights at the current time step compared to the weights at the previous time step. ,like Then take the absolute difference and set a threshold for the rate of change. ,when At that time, it was determined that the weight showed a sudden increase.

[0099] Multi-head consistency verification: Since different heads focus on different feature dimensions in the multi-head attention mechanism, at least 50% of the heads must meet the above two conditions to avoid misjudgment caused by noise from a single head.

[0100] When time step When all the above conditions are met, it is determined to be a sudden increase in weight, and this time step... It was identified as the abnormal start time.

[0101] Feature diffusion: Analyzing the temporal changes of the feature matrix in low-level information, if the features... For example, the gearbox oil temperature is The mean value at any given time increases sharply, and the characteristics are... If the bearing temperature is If the mean value increases over time, then it is determined that... The transmission relationship.

[0102] Root cause determination: Match the identified propagation path with the pattern library in the advanced information, and combine it with threshold conditions, such as the average value of gearbox oil temperature failure state being greater than 1.5 times the normal average value, to output the abnormal root cause, such as the gearbox cooling system failure causing excessive oil temperature and key influencing indicators.

[0103] This application establishes a clear knowledge system and causal logic for the tracing process by defining a well-defined three-level physical hierarchy and its mapping rules to the feature matrix, and by mining fault propagation paths and association rules based on historical fault data. This enables the tracing model not only to discover data anomalies, but also to understand the propagation chain of anomalies in the physical system, such as component anomaly - subsystem performance degradation - overall system performance deviation, thereby achieving intelligent reasoning and precise localization from data deviation to the physical root cause.

[0104] Please refer to Figure 10On the other hand, the present invention also provides a two-stage wind turbine equipment quality anomaly tracing system based on hierarchical information representation, comprising: an acquisition module for acquiring SCADA monitoring data, operation logs, and maintenance records of wind turbine equipment, performing data preprocessing using overlapping sampling, and constructing a time-series database; an early warning feature screening module for calculating the maximum information coefficient of each feature in the time-series database related to the fault, and screening key features to form a feature subset; a model building module for constructing a dual-head Transformer model including a prediction head and a classification head, using the feature subset as input to the model, introducing classification labels, and outputting feature prediction values ​​and fault category probability distributions; a tracing feature screening module for calculating the difference between fault samples and normal samples identified by the dual-head Transformer model based on Mahalanobis distance, and screening tracing key features according to the contribution of the difference, extracting multi-dimensional time-domain statistical indicators from the tracing key features to construct a high-dimensional feature matrix; and a tracing result output module for inputting the high-dimensional feature matrix into the hierarchical information representation tracing model, judging the cause of the anomaly through the feature matrix threshold, and outputting the root cause tracing result of the quality anomaly.

[0105] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0106] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, database, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0107] The above are merely embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention's specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A two-stage method for tracing quality anomalies in wind turbine equipment based on hierarchical information representation, characterized in that the steps include... include: Acquire SCADA monitoring data, operation logs, and maintenance records of wind turbine equipment, perform data preprocessing using overlap sampling, and construct a time-series database; Calculate the maximum information coefficient of each feature in the time series database that is related to the fault, and select key features to form a feature subset; A two-head Transformer model including a prediction head and a classification head is constructed. The feature subset is used as the input of the model, and classification labels are introduced to output the feature prediction values ​​and the probability distribution of the fault category. The difference between fault samples and normal samples identified by the dual-head Transformer model is calculated based on Mahalanobis distance, and key traceability features are selected according to the contribution of the difference. Multidimensional time-domain statistical indicators are extracted from the key traceability features to construct a high-dimensional feature matrix. The high-dimensional feature matrix is ​​input into the hierarchical information representation traceability model. The cause of the anomaly is determined by the feature matrix threshold, and the root cause of the quality anomaly is output.

2. The two-stage wind turbine equipment quality anomaly tracing method based on hierarchical information representation according to claim 1, characterized in that, The steps for data preprocessing using overlap sampling specifically include: Key information is extracted from the operation logs and maintenance records to obtain fault status tags aligned with timestamps, and these tags are integrated with the SCADA monitoring data to form time-series data. Let the time window length be T, and the overlap rate be 40%-60%. Extract samples from the time series data for T consecutive time steps, with adjacent samples overlapping by T / 2 time steps, and construct a time series database. For missing values ​​in the time series data, the mean of this feature within a sliding window is used to fill them; The filled features are standardized to eliminate the influence of dimensions.

3. The two-stage wind turbine equipment quality anomaly tracing method based on hierarchical information representation according to claim 1, characterized in that, The step of calculating the maximum information coefficient of each feature in the time series database related to the fault and screening key features specifically includes: The maximum information coefficient between the preprocessed features and the fault state labels is calculated. Based on the sample size, a two-dimensional plane is divided into meshes. The calculation formula is as follows: ； In the formula: For the mutual information between feature X and label Y, To find the minimum grid dimension, the grid dimension (a, b) satisfies , The parameter that controls the upper limit of the complexity of grid partitioning; N is the sample size; Sort features by MIC value from high to low and filter features with MIC value greater than a preset threshold. A simplified early warning model was constructed, and the effectiveness of the features was verified by comparing the accuracy improvement. Key features were then identified.

4. The two-stage wind turbine equipment quality anomaly tracing method based on hierarchical information representation according to claim 1, characterized in that, The steps of using the feature subset as input to the model, introducing classification labels, and outputting feature prediction values ​​and fault category probability distributions specifically include: The key feature subset is constructed into an input sequence according to a time window. A query, key, and value matrix is ​​generated through linear mapping. Attention weights are calculated using a self-attention mechanism, and feature relationships of different dimensions are captured through a multi-head attention mechanism. A classification marker is added at the beginning of the input sequence to aggregate global information. After being encoded by Transformer, it is used as the input to the classification head. The prediction head outputs the feature prediction value for the next time step, and the classification head outputs the probability distribution of the fault category.

5. The two-stage wind turbine equipment quality anomaly tracing method based on hierarchical information representation according to claim 1, characterized in that, The dual-head Transformer model is optimized based on multi-task weighted loss, specifically including: For imbalanced data with faulty samples, a weighted cross-entropy loss is used to assign higher weights to minority class samples. The difference between the predicted and actual values ​​is calculated using the mean squared error loss. The classification loss and prediction loss are fused by a weighted joint loss function, and the model parameters are optimized by backpropagation.

6. The two-stage wind turbine equipment quality anomaly tracing method based on hierarchical information representation according to claim 1, characterized in that, The steps of calculating the difference between faulty samples and normal samples identified by the dual-head Transformer model based on Mahalanobis distance, and screening key traceability features based on the contribution of the difference, specifically include: Calculate the Mahalanobis distance between the faulty samples and normal samples output by the dual-head Transformer model; A Mahalanobis distance matrix is ​​constructed, and the contributions of each feature to the distance matrix are ranked. The features with the highest contribution ranking are selected as key features for tracing. The contribution is evaluated by the change in sample distance after removing a specific feature.

7. The two-stage wind turbine equipment quality anomaly tracing method based on hierarchical information representation according to claim 1, characterized in that, The steps of extracting multidimensional time-domain statistical indicators from key traceability features to construct a high-dimensional feature matrix specifically include: Extract trend indicators, dispersion indicators, extreme value indicators, distribution pattern indicators, and dynamic change indicators from key traceability features; A feature matrix is ​​constructed based on the number of key features and statistical indicators, with each row corresponding to the statistical features of a sample, used to characterize the statistical properties of abnormal states.

8. The two-stage wind turbine equipment quality anomaly tracing method based on hierarchical information representation according to claim 1, characterized in that, The step of inputting the high-dimensional feature matrix into the tracing model representing hierarchical information and determining the cause of the anomaly through a feature matrix threshold specifically includes: Based on the dual-head Transformer model, real-time data of wind turbines is predicted, and when the failure probability exceeds the threshold, it is judged as a potential anomaly. The high-dimensional feature matrix of potential abnormal samples is input into the tracing model of hierarchical information representation. Based on the statistics of historical normal samples, the threshold range of the normal state feature matrix is ​​set, the deviation between the abnormal sample matrix and the normal threshold is calculated, and the component, subsystem and / or whole machine with the largest deviation is located.

9. The two-stage wind turbine equipment quality anomaly tracing method based on hierarchical information representation according to claim 1, characterized in that, The specific steps for constructing the tracing model of the hierarchical information representation include: Based on the physical structure and functional relationship of wind turbine equipment, a three-level physical hierarchy relationship is established: component level, subsystem level, and whole machine level. Among them, the component level corresponds to the time-domain statistical indicators of monitored parts, the subsystem level is obtained by fusing the functionally related component level features, and the whole machine level is generated by weighted aggregation of subsystem level features. Establish the mapping relationship between each level and the feature matrix. The component-level features are mapped to the time-domain statistical indicators in the feature matrix. The subsystem-level features are obtained by calculating the correlation of the component-level features. The whole machine-level features are obtained by calculating the abnormal transmission results of the subsystem-level features. A hierarchical association rule base is constructed based on historical fault data, and the association strength is quantified; Based on the hierarchical relationship and association rule base, by calculating the deviation between the abnormal sample feature matrix and the normal threshold, and combining the deviation propagation path verification, the root cause of the abnormality of the physical component can be located from the deviation of the feature matrix.

10. A two-stage wind turbine equipment quality anomaly traceability system based on hierarchical information representation, characterized in that, include: The acquisition module is used to acquire SCADA monitoring data, operation logs and maintenance records of wind turbine equipment, and uses overlap sampling for data preprocessing to build a time series database; The early warning feature filtering module is used to calculate the maximum information coefficient of each feature in the time series database related to the fault, and filter key features to form a feature subset; The model building module is used to build a two-head Transformer model including a prediction head and a classification head. It takes the feature subset as the input of the model, introduces classification labels, and outputs feature prediction values ​​and fault category probability distribution. The traceability feature filtering module is used to calculate the difference between the fault samples and normal samples identified by the dual-head Transformer model based on Mahalanobis distance, and to filter key traceability features according to the contribution of the difference. Multidimensional time-domain statistical indicators are extracted from the key traceability features to construct a high-dimensional feature matrix. The traceability result output module is used to input the high-dimensional feature matrix into the traceability model represented by hierarchical information, determine the cause of the anomaly through the feature matrix threshold, and output the root cause traceability result of the quality anomaly.

Images