Multi-view wind turbine generator anomaly detection method based on dynamic weight optimization

Abstract

The application provides a multi-view wind turbine generator abnormality detection method based on dynamic weight optimization, relates to the technical fields of wind turbine generator operation state monitoring, generator abnormality detection, multi-source heterogeneous sensor data fusion, dynamic weight optimization, deep learning, graph structure modeling, industrial artificial intelligence diagnosis and wind farm predictive maintenance, and aims at the abnormal state identification problem of wind turbine generators under complex working conditions, variable wind speed and variable load. A plurality of heterogeneous monitoring views are constructed by using vibration, temperature, current or electrical parameters, acoustic signals and SCADA operation data, and through view quality evaluation, dynamic weight distribution, shared latent space fusion and abnormality score calculation, the method can be applied to generator health state evaluation, abnormality detection, fault early warning, operation and maintenance auxiliary decision and predictive maintenance scene of onshore wind turbine generators, offshore wind turbine generators, double-fed wind turbine generators, direct-driven wind turbine generators and semi-direct-driven wind turbine generators.

Description

Technical Field

[0001] This invention relates to the field of wind power equipment condition monitoring and intelligent fault diagnosis technology, and more specifically, to a multi-view wind turbine generator anomaly detection method based on dynamic weight optimization. Background Technology

[0002] Wind turbines are key equipment for converting wind energy into electrical energy, and their long-term operational reliability directly affects the power generation efficiency, grid connection stability, and operation and maintenance costs of wind farms. As the core electrical equipment in the energy conversion chain of wind turbines, the generator is responsible for converting mechanical energy into electrical energy. Its operating status is closely related to the unit's power output, electromagnetic stability, mechanical transmission smoothness, and overall safe operation. Wind turbines typically operate in complex environments such as high altitude, low temperature, high humidity, salt spray, dust storms, strong turbulence, and frequent load changes. Generators are susceptible to factors such as wind speed fluctuations, load impacts, electromagnetic disturbances, changes in cooling conditions, insulation aging, bearing wear, and sensor drift, which can lead to problems such as winding overheating, bearing damage, rotor eccentricity, electromagnetic imbalance, inter-turn short circuits, abnormal slip ring temperature rise, cooling system malfunctions, and performance degradation.

[0003] Because wind turbine generators are typically installed inside high-altitude nacelles, on-site maintenance is difficult, downtime costs are high, and troubleshooting cycles are long. If generator anomalies are not detected early, they can lead to anything from decreased power generation efficiency, sustained temperature rise, and insulation degradation to generator burnout, bearing failure, forced shutdown, and even cascading failures. Therefore, establishing an online anomaly detection method for wind turbine generators to promptly identify early signs of a generator transitioning from a healthy to an abnormal state is crucial for reducing unplanned downtime risks in wind farms, improving equipment availability, and enabling predictive maintenance. Existing publicly available research shows that SCADA data has been widely used for predictive maintenance and anomaly detection in wind turbines, and can detect component anomaly trends in advance in some scenarios. Furthermore, research has also been conducted on detecting generator overheating faults using raw SCADA variables and structural features.

[0004] Currently, methods for detecting generator anomalies in wind turbines mainly include threshold-based or expert rule-based methods, single-source sensor data-based methods, traditional machine learning-based methods, and deep learning-based methods. Threshold-based or expert rule-based methods typically set fixed alarm thresholds based on operating variables such as generator winding temperature, bearing temperature, speed, current, voltage, and power. An alarm is triggered when the monitored value exceeds the preset range. This type of method is simple to implement and easy to deploy in engineering. However, wind turbine operating conditions exhibit significant non-stationarity; parameters such as generator temperature, current, vibration, and power fluctuate normally with wind speed, load, ambient temperature, speed, and control strategies. Fixed thresholds struggle to distinguish between normal fluctuations caused by changes in operating conditions and abnormal deviations caused by early faults, easily leading to false alarms, missed alarms, or alarm lag.

[0005] Anomaly detection methods based on single-source SCADA data typically utilize low-frequency operating parameters such as wind speed, power, engine speed, temperature, current, voltage, and pitch angle to establish a normal behavior model, and determine whether the generator is malfunctioning by analyzing the residuals between actual and predicted values. This type of method offers stable data sources and low deployment costs, making it suitable for engineering applications using existing monitoring systems in wind farms. However, SCADA data sampling periods are usually long, and it primarily represents a summary of operating conditions, making it difficult to fully reflect fine-grained anomalies such as early pitting of generator bearings, localized friction, electromagnetic imbalance, and localized insulation degradation in windings. When generator anomalies are in their early stages, fault symptoms may initially manifest as vibration spectra, current harmonics, acoustic characteristics, or localized temperature changes, while the SCADA residuals may not change significantly immediately. Therefore, relying solely on SCADA data can easily reduce the sensitivity of early anomaly detection.

[0006] Single-source detection methods based on vibration, acoustic, or current signals can reflect abnormal generator conditions from a specific physical perspective. For example, vibration signals are sensitive to bearing wear, rotor eccentricity, misalignment, and mechanical loosening; current and electrical parameter signals are sensitive to electromagnetic imbalance, inter-turn short circuits, negative sequence component abnormalities, and harmonic distortion; acoustic signals have a certain responsiveness to early bearing pitting, local friction, and abnormal noise. Related research also shows that methods such as high-frequency vibration data, acoustic data, and graph neural networks have been used for fault detection or condition identification of wind power equipment. However, single high-frequency signals are easily affected by sensor installation location, nacelle background noise, wind speed disturbances, structural resonance, and the condition of the acquisition equipment. When a certain type of sensor experiences noise pollution, data loss, abnormal sampling, or local inaccuracies, the judgment results of the single-source detection model may be significantly unstable, making it difficult to meet the reliability requirements of long-term online operation of wind turbines.

[0007] With the development of artificial intelligence and deep learning technologies, methods have begun to utilize models such as autoencoders, convolutional neural networks, recurrent neural networks, attention mechanisms, and graph neural networks to perform feature learning and anomaly detection on wind turbine operating data. These methods can automatically extract nonlinear features from historical data and have stronger data representation capabilities compared to traditional thresholding methods. However, most existing deep learning methods have the following shortcomings: First, most methods only model for a single data source and fail to fully integrate anomaly evidence from multiple physical fields such as mechanical, electromagnetic, thermal, and operating conditions of the generator; second, although some methods use multi-source data input, they usually employ simple splicing, fixed weighting, or uniform feature fusion methods, making it difficult to dynamically adjust the contribution of different views according to their reliability under the current operating conditions; third, models usually assume that the quality of various data types is stable, while actual wind farms often have problems such as missing sensors, communication delays, clock skew, noise pollution, data drift, and inconsistent sampling frequencies; fourth, the model output is mostly anomaly labels or anomaly scores, lacking explanations of key operational information such as "which view contributes the most, which variable is most obviously abnormal, and whether the current alarm may be caused by sensor anomalies."

[0008] For wind turbine generators, anomalies are often not caused by a single variable, but rather by the combined changes of multiple physical quantities. For example, localized overheating of the windings may simultaneously manifest as increased winding temperature, three-phase current imbalance, increased negative sequence components, and abnormal power residuals; early bearing damage may initially manifest as changes in vibration and acoustic characteristics, followed by temperature rise and changes in operating efficiency; cooling system anomalies may be obvious in the temperature view but not significant in the vibration or acoustic view; sensor malfunctions may manifest as an anomaly in one view while other views remain consistent. Therefore, generator anomaly detection requires simultaneous consideration of the complementarity, consistency, and reliability of different views, and cannot simply rely on a single signal or a fixed fusion strategy.

[0009] Existing multi-source fusion methods still have significant shortcomings in practical applications. First, the physical meaning, sampling frequency, and noise characteristics of different data sources vary greatly, making it difficult to directly unify the modeling of vibration, acoustic, current, temperature, and SCADA data. Second, different views have varying sensitivities to different fault types, and fixed weights cannot adapt to changes in various faults such as winding overheating, bearing damage, electromagnetic anomalies, and cooling anomalies. Third, the importance of each view changes under different operating conditions. For example, temperature and current views are more critical under high load conditions, while vibration and acoustic views are more sensitive under mechanical shock or bearing deterioration scenarios. Fourth, existing methods lack explicit evaluation of view quality and cannot dynamically adjust fusion weights based on missing rate, signal-to-noise ratio, distribution drift, operating condition adaptability, and time synchronization error. Fifth, when a view fails temporarily or a sensor drifts, traditional fusion models are prone to sudden changes in abnormal scores, increased false alarms, or distorted detection results.

[0010] Furthermore, the operating status of wind turbine generators exhibits long-term evolutionary characteristics. Affected by factors such as seasonal variations, ambient temperature changes, unit aging, maintenance and replacement, control strategy adjustments, and sensor performance degradation, the distribution of generator health data drifts slowly over time. Existing offline training models are typically based on historical data; once the field operating distribution changes, the model's generalization ability tends to decline. Although some research has begun to focus on wind power fault diagnosis under conditions of cross-unit migration and insufficient new unit samples, in generator multi-view anomaly detection, how to perform online updates while ensuring safety, and how to avoid mislearning real faults as new normal states, remain problems that need to be solved in practical engineering applications.

[0011] Therefore, the existing technology has at least the following technical defects: First, single SCADA, vibration, current, temperature or acoustic data are insufficient to fully characterize the multi-physical coupling features of generator anomalies; Second, fixed thresholds, fixed weights, or simple splicing and fusion methods are difficult to adapt to the complex operating conditions and changes in fault types of wind turbines; Third, existing methods generally lack dynamic evaluation of the quality and reliability of different view data, making it difficult to handle practical problems such as view missingness, noise pollution, asynchronous sampling, and sensor drift. Fourth, existing deep learning anomaly detection models typically lack view contribution explanations for generator subsystems and readable outputs for operation and maintenance, which is not conducive to field personnel quickly determining the source of anomalies; Fifth, existing offline models are difficult to adapt to data distribution drift during long-term operation, and there is a risk of the model being contaminated by abnormal samples during online updates.

[0012] In summary, a method for anomaly detection in wind turbine generators is needed. This method should be able to comprehensively utilize heterogeneous information from multiple sources, such as vibration, temperature, current or electrical parameters, acoustics, and SCADA operational data, to construct multiple monitoring views reflecting different physical states of the generator. It should be able to dynamically adjust the weights of each view based on view quality, operating conditions, historical stability, and the strength of anomaly evidence. It should maintain stable detection capabilities even under conditions of missing views, increased noise, or sensor drift. Furthermore, it should be able to output anomaly scores, anomaly levels, key views, key variables, and anomaly evidence explanations, thereby improving the accuracy, robustness, early warning capability, and engineering applicability of wind turbine generator anomaly detection. Therefore, we propose a multi-view wind turbine generator anomaly detection method based on dynamic weight optimization. Summary of the Invention

[0013] The purpose of this invention is to provide a multi-view wind turbine generator anomaly detection method based on dynamic weight optimization, which can realize early anomaly identification, reliable alarm and interpretable diagnosis of wind turbine generators under complex operating conditions, multi-source noise, sensor absence and dynamic changes in operating status.

[0014] This invention addresses the problem of multi-physics coupling anomaly detection during wind turbine generator operation. It comprehensively utilizes vibration signals, temperature signals, current or electrical parameter signals, acoustic signals, and SCADA operation data to construct a multi-view monitoring model that reflects the mechanical, thermal, electromagnetic, acoustic, and operating conditions of the generator. By evaluating the data quality, operating condition adaptability, anomaly evidence strength, and historical stability of different views, the weight of each view in the anomaly detection process is dynamically adjusted, enabling the model to adaptively select more reliable anomaly evidence under different fault types, operating conditions, and data quality conditions.

[0015] The present invention also aims to establish a generator anomaly detection mechanism with fault tolerance for missing views and online adaptive capability. When a certain type of sensor data is missing, noise is enhanced, sampling is out of sync, or sensor drift occurs, the method can automatically reduce the fusion weight of abnormal or failed views based on view reliability and use other valid views for compensation, thereby avoiding false alarms, missed alarms, or distorted detection results caused by a single view anomaly or data quality degradation.

[0016] The purpose of this invention is also to output anomaly detection results that are understandable to wind farm operation and maintenance personnel. The method not only outputs generator anomaly scores and alarm levels, but also outputs dynamic weights for each view, key anomaly variables, main evidence of anomalies, and suggested review directions. This enables operation and maintenance personnel to determine whether the current anomaly mainly originates from mechanical vibration, electromagnetic parameters, temperature changes, acoustic response, or deviation from operating conditions, thereby improving the reliability, interpretability, and engineering application value of the generator anomaly detection results.

[0017] To achieve the above objectives, this invention can improve the accuracy, robustness, and early warning capability of wind turbine generator anomaly detection, reduce the risk of false alarms and missed alarms caused by fixed threshold detection, single-source data detection, and static fusion detection under complex operating conditions, and provide technical support for wind turbine generator health status assessment, fault early warning, and predictive maintenance.

[0018] A multi-view wind turbine generator anomaly detection method based on dynamic weight optimization is proposed. This method takes the wind turbine generator as the detection object and comprehensively utilizes vibration data, temperature data, current or electrical parameter data, acoustic data, and SCADA operation data during generator operation to construct multiple heterogeneous monitoring views that can respectively characterize the mechanical state, thermal state, electromagnetic state, acoustic state, and operating condition of the generator. Furthermore, through view quality assessment, dynamic weight optimization, shared potential space fusion, anomaly score calculation, missing view compensation, and online update mechanisms, the method can achieve early identification, reliable alarm, and interpretable output of wind turbine generator anomalies.

[0019] Unlike anomaly detection methods that rely solely on single SCADA data, single vibration data, or simple multi-source data stitching, this invention divides various sensor data into multiple views with clear physical meanings. Furthermore, it assesses each view based on its data integrity, signal-to-noise ratio, operational condition adaptability, distribution drift, and cross-view characteristics within the current time window. Figure 1 Consistency and historical stability are ensured by adaptively adjusting the contribution weight of each view in the final anomaly detection. Therefore, when a view experiences noise contamination, missing data, asynchronous sampling, or sensor drift, the system can automatically reduce the impact of that view and compensate using other reliable views, thereby improving the robustness and engineering applicability of wind turbine generator anomaly detection.

[0020] The multi-view wind turbine generator anomaly detection method based on dynamic weight optimization includes the following steps: S1, collect multi-source heterogeneous operation data of wind turbine generators: Collect multi-source monitoring data during the operation of wind turbine generators. The multi-source monitoring data includes at least two or more of the following: vibration data, temperature data, current or electrical parameter data, acoustic data, and SCADA operation data.

[0021] Vibration data is used to reflect mechanical anomalies such as generator bearing wear, rotor eccentricity, misalignment, mechanical loosening, and local impact; temperature data is used to reflect thermal anomalies such as stator winding temperature rise, bearing temperature rise, slip ring temperature rise, cooling system anomalies, and insulation degradation; current or electrical parameter data is used to reflect electromagnetic anomalies such as three-phase current imbalance, increased negative sequence components, harmonic distortion, inter-turn short circuits, and electromagnetic disturbances; acoustic data is used to reflect early bearing pitting, abnormal friction, rubbing, and local abnormal noise in the nacelle; SCADA operation data is used to reflect the overall operating status such as wind speed, engine speed, power, pitch angle, yaw state, ambient temperature, and overall engine operating conditions.

[0022] Set at time t The collected multi-source dataset is represented as follows: in, V Indicates the number of views; Indicates the first v Each view at time t The corresponding raw monitoring data. Preferably, V=5, corresponding to the vibration view, temperature view, current or electrical parameter view, acoustic view, and SCADA operating condition view, respectively.

[0023] To facilitate subsequent modeling, the data for each view is sliced ​​according to a sliding time window. Let the length of the time window be... L If the sliding step size is 1, then the first... t Within the first time window v The data for each view is represented as follows: in, Indicates the first v The view in the first t Input data within a time window; Indicates the first v Each view at time t The sampled values ​​or feature vectors.

[0024] As a preferred technical solution of the present invention, S2 involves performing time synchronization, missing data marking, and standardization processing on multi-source data. Since the sampling frequencies of vibration, acoustic, current, temperature and SCADA data are different, and there may be communication delays, data loss and clock skew during field operation, this invention first performs time synchronization processing on the data of each view, mapping data with different sampling frequencies to a unified time window scale.

[0025] For the v-th view, the original data is transformed into features within a unified time window using methods such as resampling, interpolation, statistical aggregation, or front-end feature compression. in, Indicates the first v Preprocessing functions corresponding to each view; Indicates the first preprocessed... v Individual view features.

[0026] To preserve data quality information, this invention does not directly delete missing views or missing channels, but instead constructs a missing marker matrix: in, Indicates the first v The number of variables or features in each view; when the i-th variable is valid within the current time window. When the variable is missing or invalid, .

[0027] Perform normalization on each view based on a health baseline: in, Represents the standardized first v Individual view features; and These represent the first [item] in a healthy state. v Mean and standard deviation of each view; To prevent the use of tiny constants with a denominator of zero, the above processing allows for the creation of multi-view inputs that can be used for unified modeling, even when different views have different sampling frequencies, units, and data integrity.

[0028] As a preferred technical solution of the present invention, S3, constructing a multi-view feature representation of the generator. Based on the anomaly mechanism of wind turbine generators, this invention constructs multiple monitoring views with clear physical meanings from multi-source data. Features are extracted independently from each view, and its corresponding physical interpretation is retained.

[0029] S3.1, Constructing a vibration view Vibration views are used to characterize the dynamic response of mechanical components such as generator bearings, rotors, and frames. Vibration view features are obtained by extracting time-domain, frequency-domain, or time-frequency-domain features from the vibration signal. in, Represents the root mean square value; Indicates kurtosis; Indicates the peak factor; Indicates spectral power; Indicates the energy of the envelope spectrum; This view indicates the order of characteristics. It is primarily used to capture bearing pitting, rotor eccentricity, misalignment, mechanical loosening, and localized impact anomalies.

[0030] S3.2, Construct a temperature view Temperature views are used to characterize the thermal state of the generator. The temperature data includes one or more of the following: stator winding temperature, bearing temperature, slip ring temperature, cooling system temperature, and ambient temperature. Temperature rise rate and interphase temperature difference characteristics are constructed: in, Indicates the first i At time 1, temperature measuring points t Temperature; This indicates the rate of temperature change at that measuring point; This view represents the temperature difference between two phase windings or two temperature measuring points. It is primarily used to detect winding overheating, poor cooling, bearing friction rise, and abnormal slip ring temperature.

[0031] S3.3, Construct a current or electrical parameter view Current or electrical parameter views are used to characterize the electromagnetic state of a generator. This view includes one or more of the following: three-phase current, three-phase voltage, active power, reactive power, power factor, harmonic components, negative-sequence components, and zero-sequence components. Taking current harmonic distortion rate as an example, its calculation method is as follows: in, Indicates time t Total harmonic distortion; Represents the fundamental current amplitude; (I_{k,t}) represents the amplitude of the first wave. k The amplitude of the second harmonic current; K This indicates the highest harmonic order involved in the calculation.

[0032] The degree of three-phase current imbalance can be expressed as: in, These represent the currents of phases A, B, and C, respectively; (U_t^I) represents the degree of three-phase current imbalance. This view is mainly used to capture electromagnetic modulation and converter-side anomalies caused by inter-turn short circuits, electromagnetic imbalance, rotor eccentricity, etc.

[0033] S3.4, Constructing an acoustic view Acoustic views are used to characterize the sound response during generator operation. Short-time Fourier transform, Mel-spectrum, or cepstral feature extraction are performed on the acoustic signals. in, Represents the acoustic feature extraction function; The acoustic view features can include Mel spectrum, MFCC features, spectral centroid, spectral entropy, and modulation spectral energy. This view is primarily used to capture abnormal acoustic responses caused by early pitting, friction noise, localized rubbing, and structural loosening in bearings.

[0034] S3.5, Construct SCADA operating condition view The SCADA operating condition view is used to characterize the global operating status of the generator. This view includes variables such as wind speed, engine speed, active power, reactive power, pitch angle, yaw angle, ambient temperature, nacelle temperature, generator power, and generator torque. To reduce the interference of operating condition changes on anomaly detection, this invention constructs normal behavior residuals: in, Indicates the current SCADA observation variable; This represents the SCADA variables predicted by the normal behavior model under current wind speed, speed of rotation, pitch angle, and environmental conditions. This view represents the SCADA operating residuals. It is primarily used to identify overall generator performance degradation, power deviations, and abnormal operating conditions.

[0035] As a preferred technical solution of the present invention, S4, constructing the graph structure or association structure of each view. To express the relationships between different sensors, variables, or features within the same view, this invention constructs a corresponding graph structure or association structure for each view. For the (v)th view, the sensor measurement points, feature channels, or variables are used as graph nodes, and the physical connections, statistical correlations, or similarities between nodes are used as graph edges.

[0036] No. v The diagram structure of a single view is represented as follows: in, Indicates the first v A collection of nodes for each view; Represents the set of edges; This represents the adjacency matrix. The adjacency matrix can be constructed from both physical topological adjacencies and data-driven adjacencies. in, This represents the physical adjacency matrix obtained based on the generator sensor arrangement, phase relationship, winding connection relationship, bearing position, or cooling circuit relationship; This represents the data-driven adjacency matrix calculated based on the data from the current time window. This represents the fusion coefficient between physical adjacency and data-driven adjacency. Data-driven adjacency can be represented as: in, and Let i and j represent the features of the i-th and j-th nodes in the (v)-th view, respectively; This represents the scale parameter. To reduce noisy and redundant edges, the top edges with the largest weights for each node can be retained. k The edges form a sparse adjacency matrix.

[0037] It should be noted that the "multiple views" in this invention do not refer to multiple graph construction methods for the same set of sensor variables, but rather to heterogeneous monitoring views from different physical modes and with different anomaly sensitivities; graph structure modeling is only used to enhance the ability to express the correlation between variables within each view.

[0038] As a preferred technical solution of the present invention, S5, a multi-view encoder is constructed and latent representations are extracted. A corresponding view encoder is constructed for each view. The view encoder can employ a convolutional neural network, a temporal convolutional network, a recurrent neural network, a Transformer encoder, a graph convolutional network, a graph attention network, or a combination thereof. The latent representation of the v-th view is: in, This represents the latent feature representation of the v-th view within the t-th time window; This represents the v-th view encoder; This represents the learnable parameters of the encoder; This represents the standardized view features; This represents the adjacency matrix corresponding to the view.

[0039] To enable the encoder to learn normal patterns in a healthy state, this invention sets up a reconstruction branch and a prediction branch for each view. The reconstruction branch is used to reconstruct the current input: in, Indicates the first v Decoder for each view; Indicates decoder parameters; This represents the reconstructed view features. The prediction branch is used to predict the view features for the next time window or a short future time window: in, Represents the predictor for the v-th view; Indicates the predictor parameters; This represents the predicted features for the next time window. The reconstruction error and prediction error are respectively: in, This represents the reconstruction error of the v-th view; Indicates the first v The prediction error of each view. Under normal conditions, the model can reconstruct and predict view features well; when the generator malfunctions, the reconstruction error or prediction error of the corresponding view will increase.

[0040] As a preferred technical solution of the present invention, S6, calculate the data quality and reliability index of each view. To avoid noisy views, missing views, or drifting views having an excessive impact on the final detection results, this invention constructs a view quality evaluation mechanism. For the first... v The view, in the first t Construct a quality vector within each time window: in, Indicates data integrity metrics; Indicates the signal-to-noise quality index; Indicates the distribution drift index; Indicates time alignment reliability metrics; indicates cross-view reliability metrics. Figure 1 Consistency indicators; This represents a historical stability indicator. Data integrity indicators can be represented as: in, Missing marker; This represents the number of variables in the v-th view. A larger value indicates more complete data for that view. The distribution drift metric can be expressed as: in, This represents the data distribution of the v-th view within the current time window; Indicates the distribution of healthy baseline; This represents the distribution distance function, which can be expressed as KL divergence, maximum mean difference, or Wasserstein distance. A larger value indicates that the current view is closer to a healthy baseline distribution; a smaller value indicates a more significant drift in the view. (Cross-view) Figure 1 Consistency indicators can be expressed as: in, and They represent the first v The first view and the firstu The projection matrix of each view onto the shared latent space. The larger this value, the better the projection of the first view onto the shared latent space. v The more consistent a view is with other views in the shared space, the better. The first view is calculated based on the mass vector. v Reliability priors for each view: in, This represents the reliability prior of the v-th view within the current time window; This represents a learnable or preset quality evaluation parameter. This reliability prior is used to constrain subsequent dynamic weight allocation, preventing noisy or missing views from receiving excessive weight.

[0041] As a preferred technical solution of the present invention, S7, calculates the fusion weights of each view based on the dynamic weight optimizer. This invention constructs a dynamic weight optimizer that calculates the dynamic weight of each view within the current time window based on the latent representation, quality vector, and runtime condition vector of each view. Let the current runtime condition vector be... It can be composed of variables such as wind speed, engine speed, power, blade pitch angle, ambient temperature, and load level. First, calculate the... v Weighted scoring of each view: in, Indicates the first v Weighted scores for each view; , , Represents the learnable weight matrix; Represents the score vector; Indicates the bias term; This represents the hyperbolic tangent activation function. The initial dynamic weights are obtained through normalization: in, Indicates the first v The initial weights of each view; T t This represents a temperature coefficient used to control the smoothness of the weight distribution. To ensure that the dynamic weights are simultaneously constrained by both the model learning results and the prior reliability of the view, this invention fuses the initial dynamic weights with the reliability prior: in, Indicates the view weights after reliability constraints; This represents the balance coefficient between the model score and the reliability prior. To prevent drastic fluctuations in weights between adjacent time windows, a time smoothing mechanism is introduced: in, Indicates the final dynamic weight; This indicates the weight of the view corresponding to the previous time window; This represents the weight update coefficient. The dynamic weights satisfy the following constraints: in, Indicates the lower bound of the weight, used to prevent a view from being completely ignored for an extended period of time; This indicates the upper bound of the weight, used to prevent a single view from monopolizing the detection results; This represents the threshold for weight change constraints; This represents the view weight vector for the current time window.

[0042] Through this step, the present invention can dynamically adjust the contribution of each view in anomaly detection according to different fault types, different operating conditions, and different data quality states.

[0043] As a preferred technical solution of the present invention, S8, constructing a shared potential space and performing multi-view fusion. To integrate complementary information from different views, this invention maps the latent representations of each view to a unified shared latent space. v The shared space of a view is represented as: in, Indicates the first v The shared space representation after view mapping; Indicates the first v The projection matrix of each view. The shared representation of each view is fused based on dynamic weights: in, This represents the shared potential representation of the merged multi-view system. This representation simultaneously includes generator mechanical state, thermal state, electromagnetic state, acoustic state, and operating condition information.

[0044] To enhance consistency constraints between different views, cross-view constraints are introduced. Figure 1 Induced loss: in, Indicates cross-view Figure 1 Induced loss; u and vThis represents the different view numbers. This loss is used to constrain different views to map to similar shared latent spaces under healthy conditions, thereby enhancing the stability of multi-view fusion.

[0045] As a preferred technical solution of the present invention, S9, calculate the generator anomaly score and output the anomaly detection result. For each view, construct a single-view anomaly evidence score: in, This represents the score of the abnormal evidence for the v-th view; Indicates reconstruction error; Indicates prediction error; This indicates the consistency error of the diagram structure. , , This represents the weighting coefficient. The graph structure consistency error can be expressed as: in, Represents the original adjacency matrix; This represents the adjacency matrix reconstructed from the latent representation; Let Frobenius norm be represented. The final generator anomaly score is: in, S t This represents the overall anomaly score of the generator within the current time window; the first item represents the dynamic weighted result of anomaly evidence for each view; the second item represents inconsistent anomaly evidence between different views. u This represents the cross-view inconsistency adjustment coefficient.

[0046] Based on the statistical distribution of abnormal scores in a healthy state, an adaptive abnormal threshold is set: in, This represents the anomaly threshold for the current time window; This represents the dynamic mean of abnormal scores in a healthy state. The dynamic standard deviation of abnormal scores in a healthy state; k This represents the threshold adjustment coefficient. A generator is deemed to have an anomaly when the overall anomaly score satisfies the following formula: in, y t =1 indicates a generator malfunction; y t =0 indicates that no abnormality was detected in the generator.

[0047] To avoid false alarms caused by momentary disturbances in a single time window, this invention can also set continuous alarm criteria: Where H represents the number of consecutive detection windows; N a Indicates the minimum number of abnormal windows required to trigger an alarm; Indicates an indicator function. When there are no fewer than H consecutive windows... N a When a window malfunctions, a formal alarm will be output.

[0048] As a preferred technical solution of the present invention, S10, generating an interpretable anomaly detection result. To improve on-site operation and maintenance availability, this invention not only outputs anomaly tags but also anomaly explanation information. v The contribution of each view to the overall anomaly score can be expressed as: in, Indicates the first v The abnormal contribution of each view. The magnitude of the value can be used to determine whether the current anomaly originates from the vibration view, temperature view, current or electrical parameter view, acoustic view, or SCADA operating condition view.

[0049] For the v The first view in the i The contribution of each variable at the variable level can be expressed as: in, Indicates the first v The first view i Abnormal contribution of each variable; This indicates the input characteristics of the variable; This indicates the reconstruction characteristics of the variable.

[0050] The final output includes: generator anomaly score, anomaly level, anomaly identification result, dynamic weight of each view, anomaly contribution of each view, key anomaly variables, data quality indicators, possible anomaly types, and suggested review directions. For example, when the temperature and current views both have high contributions, the output suggests winding overheating, electromagnetic imbalance, or cooling anomalies; when the vibration and acoustic views have high contributions, the output suggests bearing wear, rotor eccentricity, or mechanical loosening; when a single view is anomaly but other views are not, the output suggests... Figure 1 When the consistency is good, the output will indicate sensor abnormality or data quality abnormality.

[0051] As a preferred technical solution of the present invention, S11, a fault tolerance and compensation mechanism for missing views is constructed. When a view experiences data loss, communication interruption, or low quality within the current time window, this invention does not directly stop detection. Instead, it reduces the weight of that view through missing data flags and reliability priors. If the... v If the missing rate of a view exceeds a preset threshold, then: in, This represents the data integrity threshold. At this point, the dynamic weight optimizer automatically lowers the reliability prior for that view. This reduces the final weight. .

[0052] At the same time, the missing view is compensated by utilizing the shared potential space: in, This represents the shared representation of the v-th view obtained through compensation; This represents the compensation weight of the u-th view for the v-th missing view. This mechanism enables the system to maintain its anomaly detection capability even when some sensors fail or some data sources are missing.

[0053] As a preferred technical solution of the present invention, S12, an online drift detection and security update mechanism is constructed. During long-term operation of wind turbine generators, the distribution of health data may drift slowly due to seasonal changes, unit aging, maintenance and replacement, and adjustments to control strategies. To improve the long-term online operation capability of the model, this invention establishes an online drift detection and safety update mechanism.

[0054] The degree of drift between the current distribution and the healthy baseline distribution is expressed as: in, This indicates the distribution of operational data over a recent period. Indicates the historical baseline distribution of health; This represents the distribution distance function. A security update is triggered when the following conditions are met: in, Indicates the drift threshold; This represents the anomaly proportion threshold. This condition means that the model is only allowed to update online when the data distribution undergoes a stable drift and the anomaly proportion within the most recent window is below the safety threshold. This prevents real-world fault samples from being mistakenly learned as new normal states.

[0055] Online updates prioritize updating the dynamic weight optimizer and standardized statistics, rather than directly and significantly updating the main encoders of each view. The update method is as follows: in, Indicates the parameters of the dynamic weight optimizer; Indicates the online learning rate; This indicates that the loss function is updated online. For the main encoder, a shadow model can be used for slow updates, and then replaced with the main model after stability has been verified.

[0056] As a preferred technical solution of the present invention, S13, model training objective function The overall training objective of this invention consists of view reconstruction loss, view prediction loss, and cross-view loss. Figure 1 The loss consists of consistency loss, dynamic gating constraint loss, missing view compensation loss, and weight smoothing loss. in, Indicates the overall training loss; , , , , , These represent the weighting coefficients of different loss terms. The dynamic gating constraint loss is defined as: in, Indicates KL divergence; This represents the dynamic weight distribution generated by the model; This represents the prior distribution of reliability obtained from the view quality assessment. This loss is used to ensure that the dynamic weights are adaptively learned by the model without deviating from the reliability constraints reflected by the view quality. The missing view compensation loss is defined as: in, This represents the set of views that are randomly occluded or actually missing during the training process. This represents a shared representation of the real view; This represents the shared representation obtained through compensation from other views. This loss is used to improve the model's robustness in scenarios where views are missing.

[0057] As a preferred technical solution of the present invention, S14, technical effects and implementation method Through the above steps, the present invention forms a complete closed loop of "multi-source data acquisition - multi-view construction - view encoding - quality assessment - dynamic weight optimization - shared potential space fusion - anomaly scoring - interpretation output - online update".

[0058] This method addresses the following technical issues: First, by jointly modeling vibration, temperature, current or electrical parameters, acoustics, and SCADA operating condition views, it solves the problem that single-source data cannot fully characterize multi-physical coupling anomalies in generators; Second, through view quality assessment and reliability-constrained dynamic weight optimization, it addresses the difficulty of adapting to complex operating conditions using fixed weighting, simple stitching, and single-time attention fusion; Third, through missing view markers, automatic weight reduction, and shared latent space compensation, it solves the detection instability caused by sensor deficiencies, noise contamination, and asynchronous sampling; Fourth, through cross-view... Figure 1 Fifth, by using consistency constraints and adaptive anomaly scoring, the early anomaly detection capability is enhanced; sixth, by using view weights, variable contribution, and anomaly evidence output, the interpretability and operational availability of the detection results are improved; and seventh, by using online drift detection and a secure update mechanism, the long-term deployment adaptability of the model is improved.

[0059] This invention can be deployed on the edge-side industrial control computer of a wind turbine, the server of a wind farm central control center, or a cloud-edge collaborative platform. The method is not limited to a specific deep learning framework, a specific sensor brand, or a specific model, and is applicable to anomaly detection, health assessment, and predictive maintenance of generators in doubly-fed, direct-drive, and semi-direct-drive wind turbines.

[0060] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention significantly improves fault identification sensitivity, with five types of views covering all physical characteristics of generator faults. Minor faults such as early bearing pitting and localized winding overheating can be captured in advance through vibration, acoustic, and temperature single views, making up for the lag defect of single SCADA low-frequency data. This invention features adaptive fusion under complex operating conditions, with dynamic weights adjusted in real time according to fault type, load, and data quality. Under high load conditions, the weights of temperature and electrical parameters are strengthened, while under mechanical fault scenarios, the weights of vibration and acoustics are increased. Compared with fixed weight fusion, this invention can effectively reduce the false alarm rate and false alarm rate under complex operating conditions. This invention has strong fault tolerance and engineering applicability. It automatically reduces the weight of the failed view and performs feature compensation in scenarios such as sensor communication interruption, measurement point drift, and data loss. It can continuously and stably monitor without stopping the sensor for maintenance, and is suitable for the complex field environment of wind farms. The diagnostic results of this invention are interpretable, improving operation and maintenance efficiency: it outputs the contribution of each view and each measuring point to anomalies, directly locates faulty subsystems and failed measuring points, distinguishes between actual equipment faults and sensor data anomalies, and significantly shortens the time for on-site fault diagnosis. The model of this invention can run stably online for a long time. The online update mechanism with safety threshold constraints updates the model baseline only when there is no concentrated failure, which solves the problem of data distribution drift caused by unit aging and seasonal temperature changes, and eliminates the need for frequent offline retraining of the model. This invention has strong deployment versatility, and the method is compatible with various wind turbine units such as doubly fed, direct drive, and semi-direct drive. It can be deployed on edge industrial control, centralized control servers, and cloud platforms, and is compatible with mainstream vibration, temperature, and electrical parameter acquisition sensors and SCADA systems on the market. The engineering implementation cost is low. Attached image description: Figure 1 The overall flowchart of the algorithm provided in this paper is for this invention; Figure 2 A schematic diagram of the multi-view branch coding and view internal association modeling structure provided by the present invention; Figure 3 A schematic diagram of the view quality-driven dynamic weight optimization and shared latent space fusion structure provided by the present invention; Figure 4 This is a schematic diagram of the multi-task anomaly detection and interpretable diagnostic output structure provided by the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are specific implementations of the present invention and are not limited to all embodiments.

[0062] Therefore, the following detailed description of embodiments of the present invention is not intended to limit the scope of the claimed invention, but merely illustrates some embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0063] The implementation of the present invention will be described below with reference to specific embodiments. These embodiments are used to explain the technical solution of the present invention and do not limit the scope of protection of the present invention. Without departing from the core concept of the present invention, the type of wind turbine, the type of generator, the number of sensors, the sampling frequency, the model structure, and the parameter values ​​can all be adjusted according to the actual wind farm conditions.

[0064] The core implementation process of this invention includes: multi-source data acquisition of wind turbine generators, construction of five types of heterogeneous monitoring views, feature extraction of each view, modeling of internal associations within the view, view quality assessment, dynamic weight optimization, fusion of anomaly evidence from multiple views, comprehensive anomaly determination, fault-tolerant handling of missing views, online drift update, and output of interpretable detection results.

[0065] This embodiment focuses on the generator of an operational wind turbine in a wind farm. This wind turbine can be a doubly-fed induction generator (DFIG), a direct-drive wind turbine, or a semi-direct-drive wind turbine. The generator is equipped with vibration sensors, temperature sensors, current or electrical parameter acquisition devices, and acoustic sensors, and is connected to the wind turbine's SCADA system. Each acquisition device is connected to an edge computing unit via a data acquisition module. The edge computing unit is connected to the wind farm's central control server via Ethernet, an industrial bus, or the wind farm monitoring network. The edge computing unit performs data synchronization, feature extraction, model inference, and alarm output; the wind farm server stores historical data, trains models, and manages model versions.

[0066] Example 1: Detection of Generator Winding Overheating 1. Generator multi-source data acquisition and access In this embodiment, the system performs online monitoring of the wind turbine generator. The sensors installed on the generator and their data acquisition targets are as follows.

[0067] Vibration sensors are installed in the front bearing housing, rear bearing housing, or frame of the generator to collect vibration acceleration signals during generator operation. Temperature sensors are installed in the U-phase, V-phase, W-phase, front bearing, rear bearing, slip rings, and cooling circuit of the generator stator windings to collect generator thermal status data. Current or electrical parameter acquisition devices are connected to the generator's three-phase current and three-phase voltage measurement circuits to collect A-phase current, B-phase current, C-phase current, three-phase voltage, active power, reactive power, power factor, and harmonic components. Acoustic sensors are installed inside or near the generator nacelle to collect operating noise. The SCADA system provides operating data such as wind speed, generator speed, generator power, pitch angle, yaw status, ambient temperature, and nacelle temperature.

[0068] The system collected the following five types of data within a certain detection window: in, This represents the vibration data within the current window, which actually corresponds to the vibration signal of the generator bearing or base. This indicates the temperature data within the current window, which actually corresponds to the winding temperature, bearing temperature, slip ring temperature, and cooling circuit temperature. This indicates the current or electrical parameter data within the current window, which actually corresponds to the three-phase current, voltage, power, and harmonic data. This represents the acoustic data within the current window, which actually corresponds to the acoustic signal in the generator nacelle. This indicates the SCADA operating data within the current window, which actually corresponds to operating condition variables such as wind speed, engine speed, power, and pitch angle.

[0069] In this embodiment, the system divides the above data into five views: vibration view, temperature view, current or electrical parameter view, acoustic view, and SCADA operating condition view. Winding overheating anomalies are usually first reflected in the temperature view, and may be accompanied by electromagnetic and operating condition anomalies such as current imbalance, harmonic distortion, and increased power residual. Therefore, this embodiment focuses on explaining how the temperature view, current view, and SCADA operating condition view work together to support the detection of winding overheating anomalies.

[0070] 2. Data synchronization and window slicing Because vibration and acoustic signals have high sampling frequencies, while temperature and SCADA signals have low sampling frequencies, the system first timestamps the various data types. Let the detection window length be... L , No. t The first window v The data for each view is: in, v This represents a specific view; when v = temp, The corresponding winding, bearing, and slip ring temperature sequences within the current window; when v=elec, The corresponding three-phase current, voltage, and harmonic sequences within the current window; when v=sca, This corresponds to the wind speed, engine speed, power, and pitch angle data within the current window.

[0071] For high-frequency vibration and acoustic data, the system extracts frequency domain or time-frequency domain features within each detection window and then aligns them with low-frequency data. For temperature, current, and SCADA data, the system uses interpolation, mean aggregation, or last effective value preservation to form a unified window feature. Through this step, the five types of heterogeneous data are converted into multi-view inputs under the same detection window.

[0072] 3. Construction of temperature view, current view and SCADA operating condition view In winding overheating anomaly detection, the system first constructs a temperature view. The nodes in the temperature view include... U Phase winding temperature, V Phase winding temperature, W Phase winding temperature, front bearing temperature, rear bearing temperature, slip ring temperature, cooling inlet temperature, and cooling outlet temperature.

[0073] The system calculates the temperature rise rate at each temperature measuring point: in, Indicates the first i Temperature values ​​of each temperature measuring point in the current window; This indicates the temperature value of the same measuring point in the previous window; This indicates the rate of temperature rise at that measuring point. For example, if... i Corresponding to the U-phase winding temperature, then This indicates the increase in temperature of the U-phase winding in the current window relative to the previous window.

[0074] The system also calculates the temperature difference of the three-phase windings: in, , , These represent the winding temperatures of phase U, phase V, and phase W respectively within the current window; , , These represent the temperature differences between the three phase windings. If a phase winding becomes locally overheated, the corresponding temperature difference will increase significantly.

[0075] The system then constructs a current or electrical parameter view. The nodes in the current view include phase A current, phase B current, phase C current, fundamental current, harmonic current, negative sequence component, zero sequence component, active power, reactive power, and power factor. The system calculates the three-phase current imbalance: in, , , These represent the effective values ​​of the currents in phases A, B, and C within the current window, respectively. Indicates the degree of three-phase current imbalance; To prevent the use of tiny constants with a denominator of zero, the three-phase current imbalance typically increases if there is localized overheating, inter-turn short circuits, or electromagnetic imbalance in the windings. The system also calculates the total harmonic distortion (THD). in, This indicates the amplitude of the fundamental current within the current window; Indicates the first k The amplitude of the second harmonic current; K represents the highest harmonic order involved in the calculation; This indicates the total harmonic distortion rate of the current. This indicator is used to reflect whether there are abnormal disturbances in the electromagnetic state of the generator.

[0076] The system then constructs a SCADA operating condition view. This view is used to determine whether the temperature rise is caused by normal high-load operating conditions. The system combines wind speed, generator speed, generator power, pitch angle, ambient temperature, and nacelle temperature into an operating condition vector. in, This indicates the average wind speed at the current window. Indicates the generator speed; Indicates generator power; Indicates the pitch angle; Indicates ambient temperature; This indicates the cabin temperature. Using this operating condition vector, the system can determine whether the current temperature rise meets the thermal response requirements under normal load conditions.

[0077] 4. View Internal Association Modeling The system constructs the temperature view, current view, and SCADA operating condition view as graph structures. Taking the temperature view as an example, its graph nodes include three-phase winding temperature, bearing temperature, slip ring temperature, and cooling circuit temperature. The system constructs an adjacency matrix based on the generator's physical structure and temperature correlation. in, This represents the adjacency matrix of the temperature views within the current window. This represents the physical adjacency matrix obtained based on the generator's thermal conduction relationship and the arrangement of measuring points. For example, there is an adjacency relationship between the U / V / W three-phase windings, and a local thermal coupling relationship between the bearing temperature and the slip ring temperature. This represents the data adjacency matrix calculated based on the current window temperature change trend; This represents the fusion coefficient between physical adjacency and data adjacency.

[0078] For example, if the U-phase winding temperature and the V-phase winding temperature have similar trends in a healthy state over a long period of time, the edge weight between them will be larger; if the current U-phase winding temperature suddenly deviates from the V-phase and W-phase winding temperatures, the data association between the U-phase node and other winding nodes will change, and this change will serve as an important basis for subsequent anomaly judgment.

[0079] 5. View Encoding and Anomaly Evidence Calculation The system inputs the temperature view, current view, and SCADA operating condition view into the corresponding view encoders to obtain the latent representation: in, This represents the potential representation of the temperature view within the current window, but actually reflects the thermal state of the generator. The current view represents the potential representation of the generator's electromagnetic state; This indicates the potential representation of the SCADA operating conditions view, which actually reflects the current operating load and environmental conditions; , , These represent the temperature, current, and SCADA view encoders, respectively. , , This represents the standardized view features.

[0080] Each view encoder simultaneously sets up a reconstruction branch. Taking the temperature view as an example, the system reconstructs temperature features based on the potential temperature representation: in, This represents the reconstructed temperature view features; (g{temp}) represents the temperature view decoder. The temperature view reconstruction error is: in, This indicates evidence of anomalies in the current temperature view. If the generator is in a healthy state and the temperature changes conform to historical healthy patterns, the error is small; if localized overheating of the windings or abnormal cooling causes the temperature pattern to deviate from healthy patterns, the error increases.

[0081] Similarly, the system calculates the current view reconstruction error separately. and SCADA condition view reconstruction error .

[0082] 6. View quality assessment and dynamic weight calculation In this embodiment, the system does not consistently assume that the temperature is apparent. Figure 1 The most important factor is not fixed values, but rather the weights are dynamically calculated based on the data quality and evidence of anomalies for each view within the current window.

[0083] The system first calculates the mass vector of the temperature view: in, This indicates the integrity of the temperature view data; Indicates the stability of temperature data or signal-to-noise quality; This indicates the degree of deviation in the distribution of temperature data relative to a healthy baseline; This indicates the consistency between the temperature view, current view, and SCADA operating condition view.

[0084] If the temperature of the U-phase winding in the current window increases, and the current imbalance and SCADA residual also increase, then the temperature view is highly consistent with other views; if only one temperature sensor changes abruptly, while the current and SCADA are normal, then the system tends to judge that there may be a sensor malfunction or local data quality problem at that temperature measurement point.

[0085] The dynamic weight optimizer calculates weight scores based on the view's latent representation, view quality, and runtime conditions. in, Indicates the temperature view rating; This indicates a potential representation of the temperature view; Represents the mass vector of the temperature view; This represents the current operating condition vector; , , 、(u) These are the model parameters.

[0086] After normalizing the scores of the five views, the view weight of the current window is obtained: For example, in the early overheating window of a certain winding, the system obtains: Wherein, 0.12 represents the vibration view weight, 0.36 represents the temperature view weight, 0.31 represents the current view weight, 0.08 represents the acoustic view weight, and 0.13 represents the SCADA operating condition view weight. This result indicates that the current anomaly is mainly supported by the temperature and current views, while vibration and acoustic evidence are weaker, making it more consistent with anomalies such as winding overheating or electromagnetic imbalance.

[0087] 7. Comprehensive Anomaly Score Calculation and Alarm Judgment The system calculates the anomaly evidence score for each view. Taking this embodiment as an example: If prediction error and graph structure error are used, they can also be added to the corresponding view anomaly evidence score.

[0088] The overall abnormality score is: in, This indicates the overall anomaly score of the generators within the current window; , , , , These represent the scores for abnormal evidence in vibration, temperature, current, acoustics, and SCADA views, respectively. Indicates cross-view inconsistency exceptions; This represents the cross-view anomaly adjustment coefficient. The system calculates dynamic thresholds based on historical health data. in, Indicates the current window's abnormal threshold; This represents the dynamic mean of the comprehensive abnormality score in a healthy state. The dynamic standard deviation of the overall abnormality score in a healthy state; k This represents the threshold adjustment coefficient.

[0089] If multiple consecutive windows satisfy If the overall anomaly score exceeds the threshold in four out of five consecutive detection windows, the system will output a formal alarm.

[0090] 8. Output of test results For a given detection window, the system outputs the following structured results: Among them, "Generator G01" indicates the object being monitored; "Winding overheating abnormality" indicates the type of abnormality indicated by the system. t Indicates the detection time; Indicates the overall abnormal score; Indicates the current anomaly threshold; w represents the dynamic weights of the five views; C t This indicates the contribution of each of the five views to the current anomaly.

[0091] This output indicates that the current generator's overall anomaly score is above the threshold. The temperature view contributes 0.43, and the current view contributes 0.34, both constituting primary evidence of anomaly. The system further outputs key anomaly variables: U-phase winding temperature, three-phase current imbalance, total harmonic distortion rate, and generator power residual.

[0092] The final diagnostic results are as follows: The current wind turbine generator G01 is in the detection window t An anomaly was detected, with a comprehensive anomaly score of 0.87, exceeding the dynamic threshold of 0.62. Evidence of the anomaly primarily came from temperature and current views, with the U-phase winding temperature rise rate, three-phase current imbalance, and current harmonic distortion rate contributing significantly. The system determined that the anomaly might be related to localized winding overheating, electromagnetic imbalance, or abnormal cooling. It is recommended to prioritize checking the generator winding temperature measurement points, cooling airflow, three-phase current balance, and winding insulation condition.

[0093] Example 2: Early Damage Anomaly Detection of Generator Bearings 1. Data Acquisition and View Construction In this embodiment, the front bearing of the wind turbine generator exhibits early pitting or slight wear. Such anomalies may not necessarily cause a significant temperature rise or immediate SCADA power deviation in the early stages, but they typically manifest as periodic impacts, increased energy in abnormal frequency bands, or enhanced modulation characteristics in vibration and acoustic signals.

[0094] The system collects vibration signals from the generator front bearing housing, acoustic signals from the nacelle, bearing temperature, three-phase current, and SCADA operating condition variables. The system constructs a vibration view from the vibration signals, an acoustic view from the acoustic signals, a temperature view from the bearing temperature, a current view from the three-phase current, and a SCADA operating condition view from the wind speed, engine speed, and power.

[0095] 2. Extraction of abnormal features from vibration views For the vibration signal of the generator front bearing, the system extracts the envelope spectrum energy: in, This represents the envelope spectrum energy of the bearing fault-related frequency band within the current window; Indicates the envelope spectrum at frequency f Energy at the location; This represents the frequency bands near the characteristic frequency of bearing failure and its harmonics.

[0096] When early pitting occurs in a bearing, the impact component causes an increase in the envelope spectrum energy near the fault characteristic band. Therefore, This is important vibration evidence of early bearing damage. The system also calculates the vibration kurtosis: in, Indicates the number of elements in the current window. i One vibration sampling point; This represents the average vibration signal value within the current window. This indicates the kurtosis of the vibration. The kurtosis typically increases when the bearing experiences increased impact.

[0097] 3. Extraction of anomaly features from acoustic view For acoustic signals, the system extracts Mel-spectral features and anomalous frequency band energy. The acoustic anomalous frequency band energy is represented as: in, This indicates the energy of the acoustic anomaly frequency band within the current window; Indicates the acoustic signal at frequency f Energy at the location; This represents the set of frequency bands associated with bearing friction, impact, or abnormal noise.

[0098] If early pitting occurs in the generator bearing, the envelope spectrum energy in the vibration view and the abnormal frequency band energy in the acoustic view will increase simultaneously. If the acoustic energy increases but the vibration view does not respond, the system tends to determine that it may be due to nacelle background noise or acoustic sensor malfunction, rather than directly determining that the bearing is faulty.

[0099] 4. Dynamic weight optimization and anomaly fusion In the current window, the system calculates the weights of five views: Wherein, 0.38 represents the vibration view weight, 0.12 represents the temperature view weight, 0.10 represents the current view weight, 0.30 represents the acoustic view weight, and 0.10 represents the SCADA operating condition view weight. This weight distribution indicates that the current anomaly is mainly supported by the vibration and acoustic views, while no obvious anomalies have been found in the temperature, current, and SCADA views. System output structured detection results: in, The vibration view contributes 0.46 and the acoustic view contributes 0.34, both of which are the main evidence of anomalies.

[0100] 5. Final output of diagnostic results The system outputs the following diagnostic results: The current wind turbine generator G01 shows signs of early bearing damage, with a comprehensive anomaly score of 0.74, higher than the current threshold of 0.58. The anomaly is primarily supported by vibration and acoustic views, with the vibration envelope spectrum energy, vibration kurtosis, and acoustic anomaly frequency band energy of the generator's front bearing contributing significantly. The temperature view has not yet shown a significant increase, indicating that the anomaly may be in its early stages. It is recommended to schedule a review of the generator bearing condition, focusing on the lubrication status of the front bearing, bearing clearance, installation alignment, and operating noise trends.

[0101] This embodiment illustrates that the present invention can capture early bearing anomalies using vibration and acoustic views when temperature and SCADA are not yet obviously abnormal, thereby improving early warning capabilities.

[0102] Example 3: Generator Anomaly Detection under Conditions of Missing Acoustic Views 1. Acoustic sensor communication anomaly In this embodiment, a brief communication interruption occurs in the acoustic sensor during wind turbine operation, resulting in missing acoustic data within the current window. Traditional multi-source stitching models typically require complete input; if any type of input is missing, the model may be unable to calculate or may exhibit abnormal score distortion. This invention achieves stable detection through a missing view fault-tolerance mechanism.

[0103] The system first calculates data integrity based on missing markers in the acoustic view: in, This indicates the integrity of the acoustic view data; n acu Indicates the number of acoustic view features; This indicates whether the i-th acoustic feature is missing. If the acoustic view is completely missing, then... ,thereby Approaching 0. When the system detects: in, If the threshold for data integrity is indicated, the acoustic view is determined to be currently not entirely reliable.

[0104] 2. Missing views are automatically downgraded in weight. The dynamic weight optimizer incorporates the data integrity of the acoustic view into the quality vector, thus reducing the prior reliability of the acoustic view. The system then obtains the following weights: Here, 0.03 represents the acoustic view weight. This result indicates that the system did not stop detection due to the lack of acoustic views, but instead automatically reduced the influence of acoustic views on the final judgment and increased the relative weights of vibration, temperature, current, and SCADA views.

[0105] 3. Shared potential space compensation To avoid incomplete fusion representation due to missing acoustic views, the system compensates for the acoustic views using other valid views: in, This represents the shared space representation of the compensated acoustic view. , , , Shared space representations for vibration, temperature, current, and SCADA views, respectively; , , , These represent the compensation weights of other views on the acoustic view.

[0106] If the vibration, temperature, current, and SCADA view do not show any abnormalities at this time, the system will not output a generator fault alarm due to the lack of acoustic view, but will instead output a data quality prompt.

[0107] 4. Output of test results The system outputs the following structured results: in, This indicates the current overall abnormal score; Indicates the abnormal threshold; This indicates that the acoustic view quality is low. The final output diagnostic result is: The current wind turbine generator G01 has not met the equipment malfunction alarm conditions, but the acoustic view data quality has significantly degraded. The system has reduced the weight of the acoustic view and is performing this window's detection based on vibration, temperature, current, and SCADA views. It is recommended to check the acoustic sensor power supply, communication link, installation location, and acquisition module status.

[0108] This embodiment illustrates that the present invention can not only detect generator abnormalities, but also distinguish between "equipment abnormalities" and "data view abnormalities", thereby reducing false alarms caused by the failure of a single sensor.

[0109] Example 4: Detection of Generator Cooling Anomalies 1. Operating Scenario In this embodiment, the wind turbine generator is operating under high load conditions. SCADA data shows high wind speed and generator power, and the temperature view shows a continuous increase in stator winding temperature and cooling outlet temperature, but the three-phase current imbalance has not increased significantly, and the vibration and acoustic views do not show obvious impact characteristics. This scenario may correspond to a decrease in generator cooling system efficiency, blockage of cooling ducts, or abnormal cooling fan operation.

[0110] 2. Joint judgment based on temperature and operating conditions The system predicts the expected normal temperature response under the current load conditions based on the SCADA operating condition vector: in, This indicates that under the current wind speed, power, rotational speed, and ambient temperature conditions, the [missing information] is... i Predicted normal temperature at each temperature measuring point; Indicates the first i A normal behavior model for a temperature measuring point; This represents the current operating condition vector. The temperature residual is: in, This indicates the deviation between the actual temperature and the predicted temperature under current operating conditions. If the generator is under high load, a temperature rise may be normal; however, if the actual temperature is significantly higher than the predicted temperature under current operating conditions, it indicates an abnormal thermal response.

[0111] 3. Dynamic weighting results In this scenario, the system obtains the following weights: The temperature view has a weight of 0.41, and the SCADA operating condition view has a weight of 0.29. This weighting indicates that the current anomaly assessment primarily relies on the temperature and operating condition views, rather than solely on absolute temperature thresholds. The low weighting of the current view suggests that the current anomaly is not typical of electromagnetic imbalance; the low weightings of vibration and acoustics indicate insufficient evidence of mechanical impact.

[0112] 4. Output of test results The system outputs structured results: D4={Generator GO1, suspected cooling anomaly,} The final diagnostic results are as follows: The current wind turbine generator G01 is suspected of having a cooling anomaly. The current temperature level exceeds the healthy baseline under the same wind speed, load, and similar ambient temperature, and the cooling outlet temperature and winding temperature are rising synchronously. The anomaly is mainly supported by the temperature view and SCADA operating condition view; the evidence of anomalies in the current, vibration, and acoustic views is weaker. It is recommended to check the generator cooling duct, cooling fan, filter, cooling medium circulation status, and temperature sensor installation status.

[0113] This embodiment illustrates that the present invention does not simply determine an anomaly by the temperature exceeding a fixed threshold, but rather combines the operating conditions to determine "whether the temperature is abnormal under the current load conditions", thereby reducing false alarms caused by normal temperature rise under high load.

[0114] Example 5: Online Drift Update Scenario 1. Long-term data drift During the long-term operation of wind turbines, seasonal temperature changes, turbine aging, maintenance and replacement, sensor calibration, and control strategy adjustments can cause slow changes in the distribution of generator health data. If the model is not updated for a long time, normal data may be misjudged as abnormal; if the model is updated without constraints, it may mislearn real faults as new normal states.

[0115] In this embodiment, the system calculates the distance between the recent data distribution and the historical health baseline at preset intervals: in, This indicates the distribution of generator multi-view health data over a recent period. Indicates the historical baseline distribution of health; This represents the distribution distance function.

[0116] 2. Security Update Determination The system will only perform an online update when the following conditions are met simultaneously: in, Indicates the drift threshold; Indicates the threshold for the proportion of abnormalities; H This indicates the number of historical windows used for statistical analysis; Indicates the first i Is the window abnormal?

[0117] The first formula indicates that the data distribution has undergone a stable drift; the second formula indicates that the proportion of abnormal windows has been low recently. Only when both conditions are met does the system consider the drift to be more likely caused by environmental changes, seasonal changes, or normal equipment aging, rather than a malfunction.

[0118] 3. Online update strategy The system prioritizes updating the dynamic weight optimizer and standardized statistics, rather than directly and significantly updating the main encoder for each view. in, Indicates the parameters of the dynamic weight optimizer; Indicates the online learning rate; This represents the loss during online updates. For the view encoder, the system uses a shadow model for background updates, replacing the current main model only when the shadow model performs stably in the most recent validation window.

[0119] 4. Output Results The system outputs the following information: The current multi-view health data distribution of wind turbine generator G01 has shown a stable drift relative to the historical baseline, but the anomaly rate in the most recent statistical period is below the safety threshold. The system has updated the dynamic weight optimizer and standardized statistics; the main encoder will not be replaced for now, and the previous model version will be retained. If the anomaly score continues to rise, online updates will be suspended and a manual review process will commence.

[0120] This embodiment illustrates that the present invention can adapt to changes in long-term operating data while avoiding mislearning real faults as new normal states.

[0121] Overall output format of the embodiments of the present invention In practical engineering deployments, this invention can output unified structured detection results: in, Indicates the generator number; Type indicates the exception type or data quality warning; t Indicates the detection time; Indicates the overall abnormal score; Indicates the current abnormal threshold; Indicates the dynamic weights of each view; Indicates the contribution of each view to anomalies; Indicates the quality indicators for each view; Indicates key outlier variables; This indicates the suggested direction for review.

[0122] For example: D={G01, winding overheating abnormal, 2026-07-11 14:30, 0.87, 0.62 [0.12, 0.36, 0.31, 0.08, 0.13], [0.09, 0.43, 0.34, 0.04, 0.10], normal. Check U-phase winding temperature, three-phase current imbalance, THD, winding insulation, and cooling system.} The structured results can be displayed in the wind farm control center, written into the wind turbine status monitoring database, and used as a data source for subsequent maintenance work order generation, fault verification, and model retraining.

[0123] In this embodiment, "generator multi-source data acquisition" corresponds to the multi-source heterogeneous data acquisition step in the technical solution; "construction of five types of views: vibration, temperature, current, acoustics, and SCADA" corresponds to the multi-view construction step; "view encoding and reconstruction error calculation" corresponds to the multi-view encoding and anomaly evidence extraction step; "view quality assessment and dynamic weight calculation" corresponds to the dynamic weight optimization step; "comprehensive anomaly score and threshold determination" corresponds to the anomaly detection step; "automatic weight reduction for missing views and shared space compensation" corresponds to the missing view fault tolerance step; "online drift safety update" corresponds to the online model update step; and "structured detection result output" corresponds to the interpretable anomaly detection result output step.

[0124] Through the above specific implementation methods, the present invention can achieve early identification, robust detection, and interpretable output of abnormal generator states in wind turbine units for actual wind farm scenarios such as winding overheating, early bearing damage, abnormal cooling, missing views, and long-term data drift.

[0125] Although the invention has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made therein without departing from the spirit and scope of the invention. Accordingly, this specification and drawings are merely exemplary descriptions of the invention as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if such modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include such modifications and modifications.

Claims

1. A multi-view wind turbine generator anomaly detection method based on dynamic weight optimization, characterized in that, Includes the following steps: S1. Collect multi-source heterogeneous operation data of wind turbine generator, wherein the multi-source heterogeneous operation data includes at least two of the following: vibration data, temperature data, current electrical parameter data, acoustic data, and SCADA operation condition data; S2. Perform time synchronization, missing data marking and standardization preprocessing on all data sources, unify the time window scale of each data source, retain missing data marking information that represents data integrity, and complete the standardization of each data source based on the generator health baseline. S3. Based on the physical fault mechanisms corresponding to different data, five independent heterogeneous monitoring views are constructed, namely, a vibration view for characterizing mechanical faults, a temperature view for characterizing thermal faults, a current parameter view for characterizing electromagnetic faults, an acoustic view for characterizing abnormal noise faults, and a SCADA condition view for characterizing operating load status. Each type of view independently extracts the corresponding fault-sensitive features. S4. Build an internal association graph structure for each type of monitoring view, and use the sensor measurement points and feature variables in the view as graph nodes. Combine the physical installation topology of the generator with the correlation of data statistics to build the graph edge adjacency relationship. S5. Configure a dedicated encoder and decoder for each type of view to form a view encoding branch. The encoder outputs a view-specific latent feature representation. The decoder completes the view feature reconstruction and combines it with the temporal prediction module to generate view prediction features. Based on the reconstruction deviation and prediction deviation, obtain single-view anomaly evidence. S6. Perform multi-dimensional quality assessment on each type of view for each detection window, generate a view quality vector that includes data integrity, signal-to-noise ratio, data distribution drift, temporal alignment accuracy, cross-view matching degree, and long-term operational stability, and calculate the view reliability prior based on the quality vector; S7. Build a dynamic weight optimizer, input the potential features of each view, the view quality vector, and the real-time operating condition vector to calculate the initial weight of the view, and then adjust the initial weight by fusing the view reliability prior. Finally, apply the upper and lower limit constraints of the weight and the time series smoothing constraints to obtain the final dynamic fusion weight of each view. S8. Set up a unified shared potential space, map the exclusive potential features of each type of view to the shared space, combine the final dynamic fusion weights of each view to weighted fusion of all view shared features, and add cross-view consistency constraints to optimize the fusion effect. S9. The generator's comprehensive anomaly score is calculated by integrating multi-dimensional anomaly evidence from a single view and inconsistent feature evidence across views. An adaptive anomaly judgment threshold is dynamically generated based on historical health data. The generator's preliminary anomaly identification result is output by combining continuous multi-window joint judgment rules. S10. Calculate the contribution of each type of view anomaly based on the dynamic weight of each view and the evidence of anomalies in a single view, locate the key anomaly variables inside the view step by step, and output an interpretable diagnostic result that includes anomaly level, view weight, source of anomaly contribution, key anomaly variables, and fault review suggestions. S11. Configure the fault tolerance compensation logic for missing views. When the data integrity of any view is lower than the preset threshold, the dynamic weight of the view will be automatically reduced, and the missing view will be compensated for by relying on the shared space characteristics of the other valid views. This will distinguish between device body abnormalities and sensor data quality abnormalities. S12. Monitor the distribution drift of multi-view data relative to the healthy baseline in real time during long-term operation. Only when the data drifts stably and the proportion of recent abnormal samples is lower than the safety threshold, perform online safety updates of the hierarchical model. Prioritize updating the weight optimizer and the standardized baseline, and use the shadow model to update the encoder asynchronously and slowly.

2. The multi-view wind turbine generator anomaly detection method based on dynamic weight optimization according to claim 1, characterized in that, The fault-sensitive features extracted for each type of view in step S3 are as follows: time-domain statistical features, frequency-domain spectral energy features, envelope spectrum features, and order features are extracted from the vibration view. The temperature view extracts the temperature rise rate at each measuring point and the temperature difference characteristics between different measuring points; The current parameter view extracts three-phase current imbalance index, harmonic distortion index, and positive and negative sequence component index; the acoustic view extracts Mel spectrum, acoustic anomalous frequency band energy, and cepstral characteristics. The SCADA operating condition view constructs the residual characteristics of each operating parameter relative to the healthy operating condition.

3. The multi-view wind turbine generator anomaly detection method based on dynamic weight optimization according to claim 1, characterized in that, In step S4, the graph adjacency relationship of each type of view is generated by fusing two parts. The first part is a fixed physical adjacency matrix determined based on the generator winding arrangement, bearing installation, and cooling circuit structure. The second part is a data-driven adjacency matrix calculated based on the correlation of data of each variable in the current window.

4. The multi-view wind turbine generator anomaly detection method based on dynamic weight optimization according to claim 1, characterized in that, In step S5, the view encoder adopts any one or more combinations of graph convolutional networks, graph attention networks, temporal convolutional networks, recurrent neural networks, and Transformer encoders. The decoder, predictor, and encoder are set up in a one-to-one correspondence. The single-view anomaly evidence is obtained by weighted calculation of feature reconstruction error, temporal prediction error, and graph structure adjacency reconstruction error.

5. The multi-view wind turbine generator anomaly detection method based on dynamic weight optimization according to claim 1, characterized in that, The dynamic weight constraint rules in step S7 include: setting a minimum lower limit for the weight of a single type of view to avoid complete loss of view information; setting a maximum upper limit for the weight of a single type of view to prevent a single view from dominating the detection results; and setting a maximum fluctuation threshold for the view weights of adjacent time windows to suppress drastic weight jumps.

6. The multi-view wind turbine generator anomaly detection method based on dynamic weight optimization according to claim 1, characterized in that, The implementation of the cross-view consistency constraint in step S8 is as follows: under the constraint health state, the feature distance after different views are mapped to the shared latent space is kept within a small range, and the cross-view feature distance loss is incorporated into the overall model loss function during the training phase.

7. The multi-view wind turbine generator anomaly detection method based on dynamic weight optimization according to claim 1, characterized in that, In step S9, the adaptive abnormal threshold is updated in real time based on the mean and standard deviation of the comprehensive abnormal scores of historical health windows. The joint judgment rule for continuous windows is: set the number of continuous detection windows and the number of abnormal windows required to trigger an alarm. Only when the number of abnormal windows in the continuous window reaches the standard will a formal device alarm be output.

8. The multi-view wind turbine generator anomaly detection method based on dynamic weight optimization according to claim 1, characterized in that, When the missing view fault tolerance compensation logic is executed in step S11, a data quality alarm is output synchronously. Only the remaining complete views are retained to participate in the subsequent comprehensive anomaly score calculation. The overall detection process will not be interrupted due to the missing of a single type of view.

9. The multi-view wind turbine generator anomaly detection method based on dynamic weight optimization according to claim 1, characterized in that, The hierarchical online security update strategy in step S12 is as follows: directly updating the main encoder based on data containing a large number of real fault samples is prohibited; the standardized statistical baseline and dynamic weight optimizer are only updated when the dual conditions of drift and low anomaly ratio are met; the main encoder uses a shadow model for background iterative training, and the stability of the shadow model is verified before replacing the online main model.

10. The multi-view wind turbine generator anomaly detection method based on dynamic weight optimization according to claim 1, characterized in that, The overall approach is deployed on edge control equipment of wind turbine units, wind farm centralized control servers, or cloud-edge collaborative platforms, and is suitable for health monitoring, fault early warning, and predictive maintenance scenarios of various wind turbine generators, including doubly-fed, direct-drive, and semi-direct-drive types.

Images