Wind turbine generator fault early warning method and system based on multi-modal data fusion

Abstract

The invention relates to the technical field of wind turbine generator fault early warning, and discloses a wind turbine generator fault early warning method and system based on multi-modal data fusion, and the method comprises the steps: collecting the data of a multi-modal sensor, and carrying out the time-space alignment preprocessing; multi-modal features are extracted through variational mode decomposition, STL decomposition and other methods, and cross-modal fusion is achieved through dimension adaptive projection and a multi-head attention mechanism; calculating a dynamic weight based on three factors of data quality, fault type correlation and information gain, and carrying out weighted fusion; constructing a dynamic unit topological graph, and capturing cross-unit association features by using a space-time diagram convolutional network; long-time early warning with confidence is realized through double-branch gating fusion in combination with a Bayesian neural network; a multi-label classification identification multi-fault mode is adopted, and an operation and maintenance decision is optimized through an adaptive large neighborhood search algorithm. According to the method, the long early warning window of the offshore wind turbine generator can be realized, and uncertainty quantification and intelligent operation and maintenance decision support are provided.

Description

Technical Field

[0001] This invention relates to the field of wind turbine fault early warning technology, and more specifically, to a wind turbine fault early warning method and system based on multimodal data fusion. Background Technology

[0002] Offshore wind power, as a crucial component of clean energy, plays a key role in the global energy transition. However, offshore wind turbines operate in harsh environments such as high humidity, high salt spray, and strong winds, making critical components like gearboxes, generators, and blades prone to failure. Due to the high cost of offshore operation and maintenance and the strict limitations imposed by weather conditions, traditional scheduled maintenance and reactive repair methods are insufficient to meet practical needs. There is an urgent need for accurate fault early warning technology to enable preventative maintenance, reduce operation and maintenance costs, and improve power generation efficiency.

[0003] Existing wind turbine fault early warning methods mainly suffer from the following technical problems: single-modal data cannot fully reflect the health status of the unit; when fusing multi-modal data, there is a lack of adaptive processing capability for data quality fluctuations and fault type differences, resulting in insufficient early warning accuracy; the wake effect and fault propagation relationship between units within the wind farm are ignored, making it impossible to identify the risk of cascading faults across units; the early warning time window is short, usually only a few hours to a few days in advance, which cannot provide sufficient preparation time for offshore operation and maintenance, and lacks the quantification of uncertainty of prediction results and intelligent operation and maintenance decision support.

[0004] Therefore, there is an urgent need to develop a wind turbine fault early warning method that can integrate multimodal data, consider the correlation between units, realize long-term early warning windows, and provide intelligent operation and maintenance decision-making, so as to meet the actual needs of intelligent operation and maintenance management of offshore wind farms. Summary of the Invention

[0005] This invention provides a method and system for early warning of wind turbine faults based on multimodal data fusion, which solves the technical problems in related technologies such as lack of adaptive processing capability, neglect of inter-unit correlation, and short early warning time window.

[0006] This invention provides a wind turbine fault early warning method based on multimodal data fusion, comprising the following steps:

[0007] S1: Collect raw data from multimodal sensors, use basic preprocessing methods, and output a spatiotemporally aligned multimodal dataset;

[0008] S2: Obtain a spatiotemporally aligned multimodal dataset, extract features using mode decomposition and spectral analysis, and output a cross-modal fusion feature set;

[0009] S3, based on a cross-modal fusion feature set, employs a three-factor dynamic weighted fusion process to output the final fusion feature;

[0010] S4 receives the final fused features, and uses dynamic topology construction, wake coefficient calculation and spatiotemporal graph convolution processing to output cross-unit correlation features;

[0011] S5 constructs comprehensive features based on the final fusion features and cross-unit correlation features, and uses dual-branch gated fusion and Bayesian network processing to output early warning results with confidence.

[0012] S6 obtains warning results with confidence levels, uses multi-label classification to identify faults, and outputs warning results and operation and maintenance decision-making solutions.

[0013] In a preferred embodiment, S1 includes:

[0014] Deploy a multimodal sensor network at key locations to collect vibration data, temperature data, acoustic data, electrical parameter data, SCADA data, and environmental meteorological data; perform basic preprocessing on the collected raw multimodal data;

[0015] Timestamp alignment is performed on data with different sampling frequencies, and a spatiotemporally aligned multimodal dataset is obtained with the sampling time of SCADA data as the reference time axis.

[0016] In a preferred embodiment, S2 includes:

[0017] Variational mode decomposition is applied to the vibration data to decompose the signal into multiple intrinsic mode functions. For each intrinsic mode function, multi-scale permutation entropy features, sample entropy, approximate entropy and time-domain statistical features are extracted and concatenated to form a vibration feature vector.

[0018] The STL decomposition method is used to decompose the temperature data into trend components, periodic components and residual components. Features are extracted for each component and concatenated to form a temperature feature vector.

[0019] Mel frequency cepstral coefficients and frequency domain features are extracted from acoustic data, and wavelet packet decomposition is used to obtain wavelet packet energy spectra. These are then spliced ​​together to form acoustic feature vectors.

[0020] Power curve deviation analysis and current harmonic analysis are performed on electrical parameter data, and the data are spliced ​​together to form an electrical feature vector.

[0021] In a preferred embodiment, S2 further includes:

[0022] To address the differences in feature dimensionality across different modalities, an adaptive projection method is employed to unify them to the target dimension; principal component analysis is used for dimensionality reduction of high-dimensional features, while nonlinear mapping is used for dimensionality increase of low-dimensional features.

[0023] Based on a feature set with unified dimensions, a cross-modal attention mechanism is constructed. A multi-head attention mechanism is adopted, in which each head independently calculates the attention weight and then the two heads are spliced ​​and fused.

[0024] Calculate the attention weights between modalities, obtain the similarity score by the inner product of the query vector and the key vector, and obtain the attention weight distribution after normalization; perform feature interaction based on the attention weights, add the interaction features to the original projected features, and output the cross-modal fusion feature set.

[0025] In a preferred embodiment, S3 includes:

[0026] Based on the cross-modal fusion characteristics, the dynamic fusion weights of each mode for different fault types are calculated. The dynamic weights comprehensively consider three factors: data quality, fault type correlation and information gain.

[0027] The correlation coefficient of fault type is obtained by statistical analysis of historical fault data, and the correlation between mode and fault type is quantified by mutual information method.

[0028] The original weight score is obtained by multiplying the three factors, scaling it by the temperature parameter, applying an exponential function and normalizing it to obtain the final weight; the information gain of each mode is calculated.

[0029] A weighted fusion strategy is adopted to sum the cross-modal fusion features of each modality with their corresponding dynamic weights to obtain the final fusion feature vector.

[0030] In a preferred embodiment, S4 includes:

[0031] Real-time wind speed and direction data of wind farms are acquired, and the wake relationship between units is determined based on real-time wind direction and unit geographical location to construct a dynamic unit spatial topology map;

[0032] The wake influence coefficient between computer groups takes into account three factors: distance between units, wind direction angle, and wind speed.

[0033] Based on the dynamic topology graph and the wake influence coefficient, a spatiotemporal graph convolutional network is constructed. Each layer contains two operations: spatial convolution and temporal convolution. Spatial convolution aggregates the features of spatial neighbor nodes, and temporal convolution uses gated recurrent units to capture temporal dependencies.

[0034] By employing bilinear transformation to determine the fault propagation probability among computer groups, a cascading fault risk map of the unit group is constructed, and cross-unit correlation characteristics are output.

[0035] In a preferred embodiment, S5 includes:

[0036] The final fused features and cross-unit related features are concatenated to obtain a comprehensive feature vector, which is then used to construct a historical feature sequence.

[0037] Historical data is decomposed into multiple time scales, and time aggregation is performed on each level. For each time scale level, an LSTM network is used to model the temporal evolution pattern, and an in-layer attention mechanism is introduced to obtain the context vector.

[0038] An inter-layer attention mechanism is used to fuse information from different time scales to obtain a multi-scale fused context vector. The fused context vector is then input into a fully connected layer for feature transformation to obtain a deep feature representation. An extreme gradient boosting tree is used to process the structured features.

[0039] In a preferred embodiment, S5 further includes:

[0040] A gating fusion mechanism is used to integrate the outputs of LSTM and XGBoost branches, construct the gating weights of the gating network output, and calculate the final warning score based on the gating weights.

[0041] A Bayesian neural network is introduced to quantify uncertainty, transforming the weights of the fully connected layer in the early warning model from deterministic parameters to probabilistic distribution parameters, and approximate weight posterior distributions are inferred through variational inference.

[0042] During the prediction phase, multiple samples are taken from the weighted posterior distribution. Each sample is forward-propagated to obtain a warning score, and the multiple scores constitute the prediction distribution.

[0043] The mean of the predicted distribution is calculated as the final warning score, and the standard deviation is calculated as a measure of uncertainty. Based on the predicted mean and standard deviation, the confidence interval is calculated, and the warning result with confidence level is output.

[0044] In a preferred embodiment, S6 includes:

[0045] A multi-label classification strategy is adopted to identify multiple fault modes simultaneously. A multi-output neural network is constructed, which includes a shared feature layer and multiple fault type-specific output layers, and independently calculates the occurrence probability of each fault type.

[0046] A fault interaction graph neural network is constructed, with fault types as graph nodes, and a graph attention network is used to learn the interaction influence coefficients between faults; the comprehensive occurrence probability of each fault type is updated based on the fault interaction influence coefficients.

[0047] The overall risk score for each unit is calculated based on the updated overall probability of failure occurrence, prediction uncertainty, and failure severity.

[0048] An operation and maintenance plan optimization model is constructed by combining weather window forecasts and operation and maintenance resource constraints; an adaptive large neighborhood search algorithm is used to solve the problem. The algorithm includes three stages: initial solution construction, damage repair iteration, and adaptive weight adjustment, and outputs the optimal operation and maintenance plan and operation and maintenance decision report.

[0049] In a preferred embodiment, a wind turbine fault early warning system based on multimodal data fusion is used to perform the steps of the wind turbine fault early warning method based on multimodal data fusion described above, including:

[0050] The data acquisition and preprocessing module is used to acquire raw data from multimodal sensors, and outputs a spatiotemporally aligned multimodal dataset using basic preprocessing methods.

[0051] The feature extraction and fusion module is used to acquire spatiotemporally aligned multimodal datasets, extract features using mode decomposition and spectral analysis, and output a cross-modal fused feature set;

[0052] The dynamic weighted fusion module is used to perform three-factor dynamic weighted fusion processing based on cross-modal fusion feature sets and output the final fused features.

[0053] The cross-unit correlation analysis module is used to receive the final fused features, and output cross-unit correlation features by using dynamic topology construction, wake coefficient calculation and spatiotemporal graph convolution processing.

[0054] The long-term early warning module is used to construct comprehensive features based on the final fusion features and cross-unit correlation features. It adopts dual-branch gated fusion and Bayesian network processing to output early warning results with confidence.

[0055] The fault identification and decision-making module is used to obtain warning results with confidence levels, identify faults using multi-label classification, and output warning results and operation and maintenance decision-making solutions.

[0056] The beneficial effects of this invention are as follows:

[0057] By employing a three-factor dynamic weight fusion mechanism, which comprehensively considers data quality, fault type correlation, and information gain, adaptive fusion of multimodal data is achieved, effectively addressing data quality fluctuations in extreme marine environments. Simultaneously, a dynamic unit topology map is constructed, and a spatiotemporal graph convolutional network is used to capture cross-unit correlation features, enabling the identification of wake effects and cascading failure risks between units, thus improving the comprehensiveness and accuracy of early warnings.

[0058] By adopting a hierarchical timescale decomposition and dual-branch gating fusion architecture, combined with Bayesian neural networks for uncertainty quantification, a long-term early warning window of 10-15 days is achieved and confidence assessment is provided. By optimizing operation and maintenance decisions through an adaptive large neighborhood search algorithm, and comprehensively considering weather windows, resource constraints and geographical proximity, intelligent operation and maintenance decision support is provided for offshore wind farms, reducing operation and maintenance costs and improving operation and maintenance efficiency. Attached Figure Description

[0059] Figure 1 This is a flowchart of a wind turbine fault early warning method based on multimodal data fusion according to the present invention;

[0060] Figure 2 This is a flowchart of a wind turbine fault early warning method based on multimodal data fusion according to the present invention;

[0061] Figure 3 This is a block diagram of a wind turbine fault early warning system based on multimodal data fusion according to the present invention. Detailed Implementation

[0062] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0063] At least one embodiment of the present invention discloses a wind turbine fault early warning method based on multimodal data fusion, such as... Figures 1 to 2 As shown, it includes the following steps:

[0064] S1: Collect raw data from multimodal sensors, use basic preprocessing methods, and output a spatiotemporally aligned multimodal dataset;

[0065] Specifically, the following steps are included:

[0066] S11, Multimodal Sensor Network Deployment and Data Acquisition: A multimodal sensor network is deployed at key locations such as the gearbox, generator, and main bearings to collect raw data from the multimodal sensors. This raw data includes vibration data, temperature data, acoustic data, electrical parameter data, SCADA data, and environmental meteorological data. Specifically, vibration sensors use accelerometers to measure the vibration acceleration signals of the high-speed, medium-speed, and low-speed shafts of the gearbox, as well as the drive and non-drive ends of the generator. Temperature sensors use resistance temperature detectors (RTDs) to measure gearbox oil temperature, bearing temperature, generator winding temperature, and ambient temperature. Acoustic sensors use capacitive microphones installed inside the engine compartment to collect acoustic signals from the generator unit's operation. Electrical parameter sensors include current transformers and voltage transformers to measure the generator's three-phase current, three-phase voltage, and active power output. The SCADA system collects operating parameters such as generator speed, power, wind speed, wind direction, and blade pitch angle. Environmental meteorological sensors measure meteorological parameters such as ambient temperature, humidity, air pressure, and salt spray concentration. All sensors are connected to a data acquisition server via an industrial Ethernet network for real-time data transmission and storage.

[0067] S12, Basic Data Preprocessing: Basic preprocessing is performed on the collected raw multimodal data, including outlier detection, missing value imputation, and data standardization. Outlier detection uses the 3σ criterion; data points deviating from the mean by more than three times the standard deviation are marked as outliers and corrected using linear interpolation. Missing value imputation uses forward imputation, filling missing positions with the nearest valid data point. Data standardization uses Z-score standardization to convert each modality of data into a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminating differences in the units and ranges of data from different modalities.

[0068] Data from different sampling frequencies were timestamped, using the sampling time of the SCADA data as the baseline time axis, with a sampling interval of 10 seconds. For high-frequency data such as vibration and acoustic data, statistical features were extracted within each 10-second time window. These features included the mean, standard deviation, maximum, and minimum values, and were used as representative values ​​for that moment. For low-frequency data such as environmental meteorological data, linear interpolation was used to upsample the data to a 10-second interval. After time alignment, a spatiotemporally aligned multimodal dataset was obtained.

[0069] S2: Obtain a spatiotemporally aligned multimodal dataset, extract features using mode decomposition and spectral analysis, and output a cross-modal fusion feature set;

[0070] Fault features are extracted from spatiotemporally aligned multimodal data, and deep interactive fusion of features is achieved through dimensionality adaptation and cross-modal attention mechanisms; specifically, the following steps are included:

[0071] S21, Variational Mode Decomposition and Multi-Scale Entropy Feature Extraction of Vibration Data: Variational mode decomposition is applied to the vibration data, decomposing the signal into eight intrinsic mode functions (EMFs), each corresponding to a specific frequency range, thus separating the vibration characteristics of different fault sources. An EMF refers to a single frequency component obtained by decomposing a complex vibration signal. For example, after decomposing the gearbox vibration signal, the first EMF corresponds to the high-speed shaft meshing frequency of 1500Hz, the second to the medium-speed shaft meshing frequency of 600Hz, the third to the low-speed shaft meshing frequency of 150Hz, and the fourth to eighth to the bearing fault characteristic frequency, resonance frequency, and low-frequency modulation signal, respectively. This decomposition allows for the separation of fault characteristics of different components for independent analysis. A penalty parameter of 2000 and a convergence tolerance of 1e-7 are set, and the solution is obtained iteratively using the alternating direction multiplier method.

[0072] For each intrinsic mode function, multi-scale permutation entropy features, sample entropy, and approximate entropy are extracted to quantify signal complexity and irregularity. The embedding dimensions of the multi-scale permutation entropy features range from 3 to 7, with a total of 5 scales. At the same time, time-domain statistical features are extracted, including peak factor, impulse factor, margin factor, and waveform factor, to reflect the impulse and waveform characteristics of the signal. The features of the 8 mode functions are concatenated to form a vibration feature vector.

[0073] S22, Seasonal Trend Decomposition and Anomaly Indicator Construction of Temperature Data: The STL decomposition method is used to decompose the temperature series into trend components, periodic components, and residual components through local weighted regression. The period length is set to 144 time points, corresponding to a 24-hour daily cycle, to separate the long-term temperature variation trend, periodic fluctuations, and random disturbances.

[0074] In one embodiment of the present invention, the STL decomposition of temperature data specifically involves: extracting periodic components from the original temperature sequence; dividing the sequence into multiple period segments with a period length of 144; performing local weighted regression smoothing on the temperature value at the same time point within each period, such as 10 AM every day, to obtain the estimated value of the periodic component at that time; the local weighted regression uses cubic polynomial fitting, with a cubic weighting function and a window width of 7 periods; for example, for the temperature at 10 AM on the 5th day, selecting the temperature values ​​at 10 AM on the 3 days before and after, for a total of 7 days, calculating the weights based on the time distance (the closer the distance, the greater the weight), performing weighted polynomial fitting, and the fitted value is the periodic component at that time; traversing all time points to obtain the complete periodic component sequence;

[0075] The deperiodic sequence is obtained by subtracting the periodic component from the original sequence. The trend component is then extracted from the deperiodic sequence, and the entire sequence is smoothed using locally weighted regression with a window width of 1441 time points (approximately 10 days). The weighting function is also a cubic weighting function. For each time point, 720 time points before and after the selected time points are used, and weights are calculated based on the time distance. A weighted polynomial is then fitted, and the fitted value is the trend component at that time. For example, for a certain time point on day 5, locally weighted regression is performed on data from day 0.5 to day 9.5 to obtain the long-term trend value at that time.

[0076] Subtracting the trend component from the deperiodic sequence yields the residual component; the residual component contains random fluctuations and anomalous events that cannot be explained by trends and periods; through iterative optimization, the above period extraction and trend extraction processes are repeated twice to make each component more stable and accurate; the three components obtained finally satisfy: original temperature sequence = trend component + period component + residual component.

[0077] Taking the gearbox bearing temperature at a certain moment as an example, the original observed value is 78°C. After STL decomposition, we get: a trend component of 70°C, indicating a gradual temperature rise due to long-term bearing wear; a periodic component of +8°C, indicating a periodic temperature rise during high-load periods in the daytime; and a residual component of 0°C, indicating no abnormal random fluctuations at present. Verification: 70 + 8 + 0 = 78°C. If the residual component suddenly reaches +5°C at a certain moment, it indicates that there may be a sudden anomaly.

[0078] For the trend component, slope and curvature features are extracted to reflect the long-term temperature change trend and acceleration. A continuous rise or abrupt curvature changes may indicate fault development. For the periodic component, amplitude and frequency features are extracted. An abnormal increase in amplitude may indicate poor heat dissipation, and a frequency shift may indicate abnormal operating conditions. For the residual component, statistical features such as mean, standard deviation, skewness, and kurtosis are calculated to quantify the characteristics of random disturbances. The rate of temperature change and acceleration within a sliding window are calculated. The sliding window length is 6 time points. The instantaneous rate of temperature change and acceleration are captured by first-order and second-order differences. Abrupt changes may indicate sudden faults. The trend features, periodic features, residual features, and dynamic features are concatenated to form a temperature feature vector.

[0079] S23, Mel frequency cepstral coefficients and frequency domain feature extraction from acoustic data; Mel frequency cepstral coefficients and frequency domain features are extracted from acoustic data to capture acoustic feature changes caused by loosening, wear, and abnormal vibration of mechanical parts; a Mel filter bank is used to process the acoustic spectrum, with 40 triangular Mel filters designed to cover the frequency range of 20Hz to 20kHz. A short-time Fourier transform is performed on the acoustic signal with a window length of 2048 sampling points and an overlap rate of 50%. Energy is calculated using the Mel filter bank to obtain a 40-dimensional Mel spectrum. After taking the logarithm of the Mel spectrum, a discrete cosine transform is performed, and the first 13 coefficients are extracted as Mel frequency cepstral coefficients. Their first and second order differences are calculated and concatenated to obtain a 39-dimensional Mel frequency cepstral feature.

[0080] A three-layer wavelet packet decomposition method using Daubechies wavelets was employed to decompose the acoustic signal into eight frequency bands. The energy of each band was calculated and normalized to obtain the wavelet packet energy spectrum, reflecting the energy distribution of different frequency bands. Frequency domain statistical features were extracted, including the spectral centroid, spectral spread, spectral roll-off point, and spectral flatness. The spectral roll-off point is the frequency point where the accumulated energy reaches 85%, reflecting the center position, dispersion, high-frequency components, and flatness of the spectrum, respectively. The Mel frequency cepstral features, wavelet packet energy spectrum, and frequency domain statistical features were concatenated to form an acoustic feature vector.

[0081] S24, Power curve analysis and harmonic feature extraction of electrical parameter data; power curve deviation analysis and current harmonic analysis are performed on electrical parameter data to extract electrical fault characteristics; a theoretical power curve model is established, using a piecewise function to describe the relationship between wind speed and output power: power is zero below the cut-in wind speed; power is proportional to the cube of the wind speed between the cut-in and rated wind speeds; rated power is maintained between the rated and cut-out wind speeds; and the unit shuts down above the cut-out wind speed. The theoretical power is calculated based on measured wind speeds, and the power residual is obtained by comparing it with the actual power. The mean, standard deviation, maximum, and minimum values ​​of the residuals are extracted to reflect the degree of deviation of the power curve and systematic bias.

[0082] A Fast Fourier Transform (FFT) is performed on the three-phase current signal to identify the fundamental frequency and harmonics from the 2nd to the 20th harmonics, with the fundamental frequency set at 50Hz. The amplitude of each harmonic is extracted. The total harmonic distortion rate (THD) is calculated to quantify the degree of current waveform distortion. The amplitude and phase of characteristic harmonics such as the 5th, 7th, 11th, and 13th harmonics are extracted, as these harmonics are related to the converter switching frequency. The power factor and voltage-current phase difference are calculated to reflect energy utilization efficiency and reactive power compensation status. The power curve characteristics, harmonic characteristics, and phase characteristics are concatenated to form an electrical characteristic vector.

[0083] S25, Dimension Adaptive Projection Achieves Feature Space Unification; To address the differences in feature dimensions across different modalities, a dimension adaptive projection method is used to unify them to the target dimension, balancing feature information preservation and computational complexity;

[0084] For high-dimensional features, including vibration and acoustic features, principal component analysis (PCA) is used for dimensionality reduction. The feature covariance matrix is ​​calculated and eigenvalue decomposition is performed. The first 32 principal components are selected to form a projection matrix, projecting the original features onto the target dimension space. For low-dimensional features, including temperature and electrical features, nonlinear mapping is used for dimensionality increase. A single-layer neural network is constructed. After standardizing the original features, a linear transformation is performed using a weight matrix and a bias vector. The weight matrix dimension is 32 times the original dimension. Then, the ReLU activation function is applied to obtain the target dimension feature vector. The weight matrix and bias vector are obtained through pre-training with an autoencoder. The encoder maps the original features to the target dimension, and the decoder reconstructs them back to the original dimension, minimizing the reconstruction error during training. Through dimensional adaptive projection, a dimensionally unified target dimension feature set is obtained.

[0085] S26, Cross-modal feature interaction fusion based on attention mechanism; Based on a feature set with unified dimension, a cross-modal attention mechanism is constructed to realize deep interaction and correlation learning of features of different modalities, and automatically learn the contribution weight of each modality to fault warning.

[0086] A multi-head attention mechanism is employed, comprising four attention heads, each independently calculating attention weights before concatenation and fusion. A query, key, and value transformation matrix is ​​defined, with its dimension being the target dimension multiplied by 8, mapping the target dimension features to an 8-dimensional space. Inter-modal attention weights are calculated: for modality i and modality j, the query vector and key vector are calculated separately, and a similarity score is obtained through the inner product. This score is then scaled by the square root of 8 to prevent gradient vanishing. Softmax normalization is applied to the scores of all modalities to obtain the attention weight distribution. The attention weights reflect the importance of modality j to modality i, automatically identifying key modal combinations under different fault types.

[0087] Feature interaction based on attention weights: The value vectors of all modalities are calculated and weighted by the corresponding attention weights to obtain the interaction features. The interaction features are added to the original projected features, and a residual connection is used to obtain the fused modal features. The outputs of the four attention heads are concatenated to obtain the cross-modal fused features. This captures diverse correlation patterns between modalities from different representation subspaces, enhancing the feature expressive power.

[0088] The output cross-modal fusion feature set includes fused vibration, temperature, acoustic, and electrical features, which retains the unique information of each mode and incorporates cross-modal correlation information.

[0089] In some embodiments, a graph attention network is used instead of a multi-head attention mechanism. Each modal feature is treated as a graph node, and the physical relationships between modalities are treated as graph edges. The features of neighboring nodes are aggregated through a message passing mechanism. This method can better utilize prior knowledge of intermodal relationships, such as the causal relationship between vibration and temperature, and improve the physical interpretability of feature fusion and the generalization ability of the model.

[0090] S3, based on a cross-modal fusion feature set, employs a three-factor dynamic weighted fusion process to output the final fusion feature;

[0091] Based on the correlation between data quality and fault type, the fusion weights of each modality are dynamically adjusted to achieve robust fusion that adapts to the environment; specifically, this includes the following steps:

[0092] S31, dynamic weight calculation based on quality score and correlation; based on the quality score output by S1 and the cross-modal fusion features output by S2, calculate the dynamic fusion weight of each modality for different fault types; the dynamic weight comprehensively considers three factors: data quality, fault type correlation and information gain.

[0093] The correlation coefficient of fault types is obtained through statistical analysis of historical fault data. Fault cases from the past three years are collected, including gearbox faults, generator bearing faults, blade cracks, electrical system faults, etc. For each fault type, the abnormal performance of each mode data before the fault occurs is statistically analyzed. The mutual information method is used to quantify the correlation between the mode and the fault type. The larger the mutual information value, the stronger the early warning capability of the mode for the fault type.

[0094] The mutual information values ​​are normalized to the range of 0 to 1 to obtain the fault type correlation coefficients. For example, the correlation coefficient of vibration mode with gearbox faults is 0.9, and the correlation coefficient with electrical faults is 0.3; the correlation coefficient of electrical parameter mode with electrical faults is 0.95, and the correlation coefficient with gearbox faults is 0.2. The correlation coefficient matrix stores the correlation relationships between all modes and all fault types.

[0095] The dynamic fusion weights are calculated by multiplying the quality score, fault type correlation coefficient, and information gain to obtain the original weight score. Since all three factors have been normalized to the range of 0 to 1, the multiplication result is also between 0 and 1. The original weight score is then scaled by dividing it by a temperature parameter of 0.5. The temperature parameter controls the smoothness of the weight distribution; a smaller temperature parameter makes the weight distribution more concentrated on high-quality, high-relevance modes. An exponential function is applied to the scaled score to obtain the exponential weights. The exponential weights of all modes are then normalized. The final weight of each mode is equal to the exponential weight of that mode divided by the sum of the exponential weights of all modes, ensuring that the sum of the weights of all modes is 1.

[0096] When the data quality of a certain modality deteriorates, for example, when the signal-to-noise ratio decreases due to salt spray corrosion of the sensor, the quality score of that modality decreases, and the dynamic weight decreases accordingly to reduce the negative impact of low-quality data on the fusion results. When the warning target is a specific fault type, the weight of the modality that is highly correlated with that fault type is automatically increased to achieve targeted feature fusion.

[0097] S32, Information gain calculation based on feature change rate; calculate the information gain of each mode, quantify the dynamic changes and anomalies of modal features; information gain comprehensively considers both feature change rate and deviation from baseline features;

[0098] The feature change rate is calculated by dividing the norm of the difference between the fused features at the current time and the fused features at the previous time by the norm of the features at the previous time. The feature change rate reflects the dynamic change speed of the modal features. The larger the change rate, the more significant the mode is undergoing, which may contain important fault information.

[0099] The deviation from the baseline features is calculated, and the difference between the current feature distribution and the normal baseline feature distribution is quantified by KL divergence. The baseline features are obtained by statistical modeling of historical data under normal operating conditions, and the feature distribution is fitted by Gaussian mixture model. KL divergence is obtained by calculating the relative entropy of the two probability distributions. The larger the KL divergence, the further the current feature deviates from the normal state and the higher the degree of abnormality.

[0100] When calculating information gain, the feature change rate and KL divergence are normalized to ensure they are within the same numerical range. The feature change rate is normalized by dividing by the historical maximum change rate, and the KL divergence is normalized by dividing by a preset maximum divergence threshold. The normalized values ​​are both between 0 and 1. Information gain equals the normalized feature change rate multiplied by a weighting coefficient of 0.4, plus the normalized KL divergence multiplied by a weighting coefficient of 0.6, with the sum of the two weighting coefficients being 1. A slightly higher weight for KL divergence indicates that the deviation from the baseline more directly reflects the fault risk. A larger information gain value indicates that the modality contains more fault-related information at the current moment and should be assigned a higher fusion weight. Through the information gain mechanism, key modalities at the current moment can be dynamically identified. When a modality undergoes an abnormal change, its information gain increases, and the fusion weight increases accordingly, enabling the early warning system to capture fault signs in a timely manner.

[0101] S33, a weighted fusion strategy is used to generate the final fusion features. Based on the calculated dynamic fusion weights, a weighted fusion strategy is employed to generate the final fusion features. For a specific fault type, the cross-modal fusion features of each mode are multiplied by their corresponding dynamic weights and then summed to obtain the final fusion feature vector for that fault type.

[0102] The final fused feature vector is 32-dimensional, taking the target dimension as the dimension, which integrates information from all modalities. It also undergoes adaptive weighting based on data quality, fault correlation, and information gain. When the data quality of a modality decreases or its information gain is low, its weight is automatically reduced to avoid interference from low-quality or low-information data. When a modality is highly correlated with the target fault type and has a high information gain, its weight is automatically increased to enhance the contribution of key information. Through environmentally adaptive dynamic weight fusion, a robust fused feature is obtained. This feature can adapt to data quality fluctuations in extreme maritime environments, ensuring the stability and reliability of the early warning system.

[0103] In some embodiments, a meta-learning method is used to replace the fixed weight calculation formula for dynamic weight learning; a meta-learning network is constructed, with the input being the quality score, feature statistics and fault type label of each modality, and the output being the fusion weight of each modality; by training the meta-learning network under multiple fault scenarios, it can automatically learn the optimal weight allocation strategy under different scenarios; the meta-learning method can capture more complex weight allocation rules and adapt to more diverse fault modes and environmental conditions.

[0104] S4 receives the final fused features, and uses dynamic topology construction, wake coefficient calculation and spatiotemporal graph convolution processing to output cross-unit correlation features;

[0105] Constructing a dynamic unit topology diagram to capture wake effects and coupling relationships between units, and identifying the risk of cascading failures across units; specifically including the following steps:

[0106] S41, real-time wind direction sensing and dynamic topology map construction; based on the final fusion features output by S3, wind speed and wind direction data of the wind farm are acquired in real time to construct a dynamic unit spatial topology map; the wind farm units are distributed in the sea area, the wind farm units are preset to 100 units, the relative positions between the units are fixed, but the wake effect changes dynamically with the wind direction.

[0107] Real-time wind direction data from the wind farm's meteorological station is obtained, with due north as 0 degrees and rotating clockwise, east wind is 90 degrees, south wind is 180 degrees, and west wind is 270 degrees; the 10-minute average wind direction is obtained by averaging through a sliding window to eliminate the influence of short-term wind direction fluctuations.

[0108] Based on real-time wind direction and unit geographical location, determine the wake relationship between units; for any two units i and j, calculate the azimuth angle from unit i to unit j, which is calculated using the latitude and longitude coordinates of the two units; determine the angle between the azimuth angle and the prevailing wind direction. When the angle is less than 90 degrees, unit i is located upwind of unit j and may have a wake effect on unit j; when the angle is greater than 90 degrees, unit i is located downwind or crosswind of unit j, and the wake effect can be ignored.

[0109] A dynamic unit topology graph is constructed, where nodes represent units and edges are dynamically established based on wake relationships. When unit i has a wake influence on unit j, a directed edge is established between node i and node j, with the edge pointing from i to j, indicating the direction of wake propagation. As the wind direction changes, the edge set of the topology graph is dynamically updated to reflect the real-time unit relationships.

[0110] S42, wind speed adaptive wake influence coefficient calculation; based on dynamic topology graph, wake influence coefficient between computer groups, quantifying the intensity of wake effect; the wake influence coefficient comprehensively considers three factors: distance between units, wind direction angle, and wind speed.

[0111] The distance attenuation factor is calculated using a Gaussian attenuation model. The attenuation factor is equal to the square of the negative inter-unit distance divided by an exponential function of twice the square of the wake diffusion parameter. The wake diffusion parameter is related to the wind speed. The higher the wind speed, the faster the wake diffusion and the larger the diffusion parameter. The wake diffusion parameter is equal to the basic diffusion parameter plus the wind speed coefficient multiplied by the wind speed. The basic diffusion parameter is set to 200 meters and the wind speed coefficient is set to 10 meters per second.

[0112] The wind alignment factor is calculated using a cosine function. The alignment factor is equal to the cosine of the angle between the azimuth of the line connecting the units and the prevailing wind direction. When the angle is 0 degrees, unit j is located directly downwind of unit i, the alignment factor is 1, and the wake effect is strongest. As the angle increases, the alignment factor decreases, and the wake effect weakens. When the angle is greater than 90 degrees, the alignment factor is negative, and the wake effect coefficient is set to 0.

[0113] The wake influence coefficient is calculated by multiplying the distance attenuation factor and the wind direction alignment factor to obtain the preliminary influence coefficient. Since the wind direction alignment factor is negative when the angle is greater than 90 degrees, it needs to be corrected. The larger value between the preliminary influence coefficient and 0 is taken as the final wake influence coefficient to ensure that the coefficient is non-negative. Specifically, the wake influence coefficient is equal to the maximum value between the distance attenuation factor multiplied by the wind direction alignment factor and 0. Through this calculation method, the wake influence coefficient is between 0 and 1, and the larger the coefficient, the stronger the wake influence. For turbine pairs that are close to each other and located directly downwind, the wake influence coefficient is close to 1. For turbine pairs that are far apart or deviate from the wind direction, the wake influence coefficient is close to 0. The wake influence coefficient is updated in real time with wind speed and wind direction. When the wind direction changes, the wake relationship between the turbines is recalculated, and the edge weights of the topology graph are dynamically adjusted to achieve accurate modeling of the real-time operating status of the wind farm.

[0114] S43, Spatiotemporal Graph Convolutional Network for Unit Feature Propagation: Based on a dynamic topology graph and wake influence coefficient, a spatiotemporal graph convolutional network is constructed to achieve the propagation and aggregation of unit features in both spatial and temporal dimensions. The spatiotemporal graph convolutional network can simultaneously capture the spatial correlations and temporal evolution patterns between units.

[0115] The final fused features of each unit are used as the initial feature representation of the graph nodes; a three-layer spatiotemporal graph convolutional network is constructed, with each layer containing two operations: spatial convolution and temporal convolution.

[0116] In the spatial convolution operation, each unit node aggregates the features of its spatial neighbor nodes; a spatial neighbor is defined as a node that is connected to the node by an edge in the dynamic topology graph, that is, a unit that is affected by the wake of the node or has a wake effect on the node; the aggregation operation adopts a weighted summation, with the weight being the normalized wake effect coefficient, and the normalization is achieved by dividing by the product of the square root of the number of neighbor nodes and the square root of the number of neighbors of the current node.

[0117] The aggregated features are linearly transformed using a spatial weight matrix; after the linear transformation, the ReLU activation function is applied to introduce nonlinear transformation capability; in the temporal convolution operation, each unit node fuses its features from historical moments; a gated recurrent unit is used to capture temporal dependencies, which includes update gates and reset gates, and can selectively retain or forget historical information; the features of the current moment and the hidden state of the previous moment are input into the gated recurrent unit, and the new hidden state of the current moment is output;

[0118] The outputs of spatial convolution and temporal convolution are added together to obtain the output features of this layer. By stacking three layers of spatiotemporal graph convolution, the feature representation of each unit node incorporates the spatial information of three-hop neighbors and the temporal information of multiple historical moments, which can capture a wider range of unit associations and a longer time evolution trend.

[0119] S44, Fault propagation probability calculation and cascading risk identification; Based on the unit feature representation output by the spatiotemporal graph convolutional network, the fault propagation probability between computer groups is calculated to identify potential cascading fault risks; the fault propagation probability reflects the possibility of a fault propagating from one unit to another.

[0120] A bilinear transformation is used to score the fault propagation between computer groups. Specifically, for group i and group j, the feature representation vector of group i is multiplied by the fault propagation transformation matrix to obtain an intermediate vector. The inner product of this intermediate vector and the feature representation vector of group j is calculated. The inner product is obtained by multiplying the corresponding elements of the two vectors and then summing them to obtain a scalar value as the propagation score. The fault propagation transformation matrix is ​​obtained by training through historical fault propagation cases and can capture the pattern features of fault propagation.

[0121] A weighted term for the wake effect coefficient is added to the propagation score. Specifically, the propagation score obtained by bilinear transformation is normalized by mapping the score to the range of 0 to 1 using the sigmoid function; the wake effect coefficient is multiplied by the weighting coefficient to obtain the wake contribution term; finally, the normalized propagation score is added to the wake contribution term to obtain the comprehensive propagation score; the weighting coefficient reflects the degree of physical influence of the wake effect on fault propagation. The wake effect will cause downstream units to experience greater load fluctuations and turbulence intensity, increasing the risk of faults.

[0122] Applying the sigmoid activation function to the propagation score maps the score to a probability range of 0 to 1, thus obtaining the fault propagation probability. The closer the probability value is to 1, the greater the probability that the fault will propagate from unit i to unit j; the closer the probability value is to 0, the smaller the probability of propagation.

[0123] Based on the failure propagation probability, a cascading failure risk graph for the unit group is constructed. A propagation probability threshold of 0.7 is set. When the propagation probability between units exceeds the threshold, directed edges are established in the risk graph to indicate a significant failure propagation risk. By analyzing the topology of the risk graph, high-risk propagation paths and key nodes are identified.

[0124] Calculate the impact range of each unit as a fault source. The impact range is equal to the sum of the products of the fault propagation probability from that unit to all other units and the fault severity of the target unit. Units with a large impact range may trigger a large-scale chain reaction once a fault occurs, and should be the key monitoring targets. Obtain cross-unit correlation features, which include the spatiotemporal graph convolution features of each unit and the fault impact range index.

[0125] In some embodiments, causal inference methods are used instead of correlation analysis for fault propagation modeling; Granger causality tests are used to identify causal relationships between units and determine whether a state change in one unit is a Granger cause of a state change in another unit; a causal graph is constructed, where the direction of the edges represents the direction of the causal relationship and the weight of the edges represents the causal strength; fault propagation analysis based on the causal graph can more accurately identify the root cause and propagation path of the fault and avoid interference from spurious associations.

[0126] S5 constructs comprehensive features based on the final fusion features and cross-unit correlation features, and uses dual-branch gated fusion and Bayesian network processing to output early warning results with confidence.

[0127] Construct an early warning model capable of achieving a 10-15 day long warning window and provide quantification of the uncertainty of the prediction results; specifically including the following steps:

[0128] S51, Construction of comprehensive feature vector; The final fused feature and cross-unit correlation feature are concatenated to obtain the comprehensive feature vector; The final fused feature contains multimodal fusion information of a single unit, and the cross-unit correlation feature contains spatial correlation and fault propagation information of the unit group;

[0129] Construct a historical feature sequence and collect comprehensive feature vectors from the past 30 days. The original sequence length is 4320 time points. The calculation method is 30 days multiplied by 144 points per day. The sampling interval is 10 seconds. Downsampling to a 10-minute interval yields 2160 time points, reducing computational complexity.

[0130] S52, hierarchical timescale decomposition and feature aggregation; decomposes 30 days of historical data into 5 timescale levels to capture the evolutionary patterns of different time spans: the first level is 1 day, the second level is 2 days, the third level is 4 days, the fourth level is 8 days, and the fifth level is 16 days, forming a multi-scale time system;

[0131] For each level, the data is divided into non-overlapping time periods. A combination of average pooling and max pooling is used for time aggregation, where average pooling captures the average state and max pooling captures the peak state. The aggregated features are then concatenated. The first layer yields a feature sequence of length 30, the second layer has a length of 15, the third layer has a length of 7, the fourth layer has a length of 3, and the fifth layer has a length of 1. Through hierarchical decomposition, long time series data are transformed into multiple short time series data, with each level focusing on the change patterns at different time scales.

[0132] S53, In-layer LSTM modeling and attention mechanism; For each time scale layer, an LSTM network is used to model the temporal evolution law. The LSTM network contains 2 layers and the number of hidden units is 128; Information is selectively retained or forgotten through a gating mechanism to handle long-term dependencies. The gating mechanism includes a forget gate, an input gate, and an output gate.

[0133] Taking the first layer as an example, a feature sequence of length 30 is input into an LSTM and processed step by step to obtain 30 hidden state sequences, each with 128 dimensions. An intra-layer attention mechanism is introduced, where the hidden states are input into a single-layer neural network to calculate attention scores, which are then normalized using softmax to obtain the attention weight distribution. The hidden state sequences are weighted and summed based on the attention weights to obtain a context vector. By integrating information from all time steps in this layer, key moments are automatically identified, including the moments when fault symptoms appear. The same method is used for other layers, with the second to fifth layers each obtaining their own context vectors.

[0134] S54, Inter-layer attention fusion and multi-scale information integration; Inter-layer attention mechanism is used to fuse information at different time scales; The context vectors of the five layers are input into a single-layer neural network to calculate the inter-layer attention score, and the weight distribution is obtained by softmax normalization, with the sum of the weights being 1;

[0135] The five context vectors are weighted and summed based on inter-layer attention weights to obtain a multi-scale fused context vector, which integrates multi-scale information from 1 day to 16 days. The inter-layer attention mechanism adaptively adjusts the weights according to the fault mode: rapid development faults focus on the short-term scale, and slow degradation faults focus on the long-term scale.

[0136] The fused context vector is input into two fully connected layers (256 neurons in the first layer and 128 neurons in the second layer, ReLU activation) for feature transformation to obtain a deep feature representation for early warning.

[0137] S55, Construction of auxiliary branches using extreme gradient boosting trees (XGBoost); An auxiliary branch is constructed using extreme gradient boosting trees (XGBoost) to handle structured features, complementing the LSTM branch; XGBoost excels at handling tabular data and nonlinear relationships; Statistical features, time-series statistics, and domain knowledge features are extracted from the comprehensive feature vector and concatenated to form a structured feature vector; Statistical features include the mean, standard deviation, maximum, minimum, trend, and periodicity of the past 30 days; Time-series statistics include autocorrelation coefficient, partial autocorrelation coefficient, trend strength, and seasonality strength; Domain knowledge features include the correlation between temperature and vibration, cumulative deviation of power curves, and rate of change of harmonic distortion.

[0138] Structured features are input into the XGBoost model, which contains decision trees with a maximum depth of 6 and a learning rate of 0.1. XGBoost iteratively builds decision trees using a gradient boosting framework, fitting residuals to each new tree. The objective function includes log loss and a regularization term, which includes the number of leaf nodes and the L2 norm of the weights to prevent overfitting. The output is a warning score based on structured features, ranging from 0 to 1. XGBoost automatically performs feature selection and interaction to identify the most important feature combinations.

[0139] S56, a gated fusion mechanism integrates the outputs of two branches; a gated fusion mechanism is used to integrate the outputs of LSTM branch and XGBoost branch, adaptively adjusting the contribution weights to achieve complementary advantages; the deep feature representation of LSTM branch is input into the fully connected layer, and the activation function is sigmoid to obtain the LSTM branch warning score, which ranges from 0 to 1;

[0140] A gated network is constructed, with the input being a concatenated vector of LSTM deep features and XGBoost warning scores. The gated network consists of two fully connected layers: the first layer contains 64 neurons activated by ReLU, and the second layer contains one neuron activated by sigmoid. The output is a gate weight, ranging from 0 to 1, representing the contribution ratio of the LSTM branch. The final warning score is equal to the gate weight multiplied by the LSTM score, plus 1, minus the gate weight, and then multiplied by the XGBoost score. The gated fusion learns the optimal fusion strategy through end-to-end training, adaptively adjusting the weights according to data characteristics and fault modes to improve warning accuracy and robustness.

[0141] S57, Uncertainty Quantification of Bayesian Neural Networks: Bayesian neural networks are introduced to quantify uncertainty, providing confidence assessment for early warning results and helping maintenance personnel judge the reliability of early warnings.

[0142] The weights of the fully connected layer in the early warning model are transformed from deterministic parameters to probabilistic distribution parameters. The prior distribution of the weights is defined as a Gaussian distribution with a mean of 0 and a standard deviation of 0.1. An approximate posterior distribution of the weights is obtained through variational inference, and this posterior distribution also adopts a Gaussian distribution, including the mean and standard deviation parameters. A variational inference objective function is constructed, containing two terms: the first term is the expected prediction likelihood, approximated by Monte Carlo sampling with N equal to 10 sampling times. Weights are sampled from the posterior distribution and forward propagated to calculate the log-likelihood between the predicted output and the true label, and then averaged. The second term is the KL divergence between the posterior and prior distributions, calculated analytically using a Gaussian distribution. The objective function equals the expected prediction likelihood minus the KL divergence. The Adam optimizer is trained for 100 epochs with a learning rate of 0.001, maximizing the objective function to obtain the mean and standard deviation parameters of the posterior weight distribution.

[0143] S58, Prediction distribution sampling and confidence interval calculation; In the prediction phase, the weighted posterior distribution is sampled 100 times, and each sampling yields a set of weight parameters for forward propagation to obtain a warning score; The 100 scores constitute the prediction distribution.

[0144] The mean of the predicted distribution is calculated as the final warning score, which is the arithmetic mean of 100 scores. The standard deviation is calculated as a measure of uncertainty; a larger standard deviation indicates higher prediction uncertainty and insufficient model confidence, while a smaller standard deviation indicates lower uncertainty and higher model confidence. Uncertainty sources include cognitive uncertainty and random uncertainty. Cognitive uncertainty stems from model parameter uncertainty, while random uncertainty arises from inherent data noise. Based on the predicted mean and standard deviation, a 95% confidence interval is calculated: the lower bound equals the mean minus 1.96 times the standard deviation, and the upper bound equals the mean plus 1.96 times the standard deviation. The confidence interval reflects the possible range of warning scores, providing a risk assessment basis for operational decisions.

[0145] Output the warning results with confidence, including the warning score, prediction standard deviation and confidence interval, and pass them to step seven for multi-fault identification and operation and maintenance decision optimization;

[0146] In some embodiments, an ensemble learning method is used instead of a Bayesian neural network; multiple early warning models are trained using different random initialization and training data subsets; all models predict independently during prediction; the mean and variance are calculated as a measure of uncertainty, with the variance reflecting the prediction discrepancy between models; ensemble learning has higher computational efficiency and is suitable for real-time early warning scenarios, reducing computational complexity while ensuring the effectiveness of uncertainty quantification.

[0147] S6 obtains warning results with confidence levels, uses multi-label classification to identify faults, and outputs warning results and operation and maintenance decision-making solutions;

[0148] Identify multiple failure modes, analyze the interactions between failures, and optimize operational decisions based on risk assessment; specifically, this includes the following steps:

[0149] S61, Fault Type Identification of Multi-Label Classification Network; Employing a multi-label classification strategy to simultaneously identify multiple fault modes, including gearbox wear, generator bearing fatigue, blade corrosion cracks, electrical system moisture, yaw system failure, and hydraulic system leakage, identifying multiple concurrent fault scenarios;

[0150] A multi-output neural network is constructed, with the input being the deep feature representation of an LSTM branch. The network includes a shared feature layer and six dedicated output layers for each fault type. The shared feature layer consists of two fully connected layers: the first layer contains 256 neurons, and the second layer contains 128 neurons, with ReLU activation function, extracting common features for all fault types. The dedicated output layers for each fault type are single fully connected layers with a 128-dimensional input dimension and a 1-dimensional output dimension, using sigmoid activation function to independently calculate the occurrence probability of each fault type, ranging from 0 to 1. Through multi-label classification, the network simultaneously outputs the occurrence probabilities of the six fault types, identifying concurrent multi-fault scenarios.

[0151] S62, Fault Interaction Graph Neural Network Modeling: Construct a fault interaction graph neural network to model the mutual influence relationships between faults. Complex interactions exist between faults; one fault may accelerate the development of another, or multiple faults may work together to produce more severe consequences.

[0152] Six fault types are used as graph nodes. A fault interaction graph is constructed based on historical fault co-occurrence data, which is obtained by statistically analyzing fault records from the past three years. The co-occurrence frequency of any two fault types is calculated, and when it exceeds a threshold, an undirected edge is created to represent the association.

[0153] Node features are defined, including attributes such as failure probability, historical frequency of occurrence, and failure severity. A graph attention network is used to learn the interaction influence coefficients between failures: the feature vectors of node k and node l are linearly transformed through the weight matrix W, and the dot product with the attention parameter vector a yields the attention score. Softmax normalization is then applied to all neighboring nodes of node k to obtain the failure interaction influence coefficient. The failure interaction influence coefficient reflects the strength of the influence of failure k on failure l; a larger coefficient indicates a greater influence. The graph attention network automatically learns the interaction patterns between failures from the data without manual specification.

[0154] S63 considers the comprehensive probability update of fault interaction; based on the fault interaction influence coefficient, the comprehensive occurrence probability of each fault type is updated, taking into account the mutual promotion effect between faults; the initial occurrence probability is output by the multi-label classification network, without considering the interaction effect;

[0155] For fault type k, calculate the interaction effect of other fault types on it: the effect value is equal to the probability of occurrence of fault l multiplied by the interaction effect coefficient of fault l on fault k, and the total interaction effect is obtained by summing over all other fault types; the updated comprehensive probability of occurrence is equal to the initial probability of occurrence multiplied by 1 plus the total interaction effect; when the probability of occurrence of other faults is high and the interaction effect coefficient is large, the comprehensive probability of occurrence of fault k increases, reflecting the risk amplification effect of multiple fault concurrency.

[0156] By modeling fault interactions, we can more accurately assess the risks of concurrent multi-fault scenarios and avoid underestimating synergistic effects. For example, gearbox wear leads to increased vibration, which in turn accelerates generator bearing fatigue. The two faults reinforce each other, and the combined risk is higher than the sum of the risks considered individually.

[0157] S64, Calculation of Unit Comprehensive Risk Score: Based on the updated comprehensive failure probability, prediction uncertainty and failure severity, calculate the comprehensive risk score for each unit, taking into account the probability of failure, prediction reliability and severity of consequences.

[0158] For unit i, iterate through all fault types k and calculate the risk contribution: normalize the prediction standard deviation, adding a small constant of 0.01 during normalization to avoid division by zero, and calculate the prediction confidence weight equal to 1 divided by the normalized standard deviation, with a higher weight for smaller standard deviations; the risk contribution value equals the overall probability of fault k multiplied by the severity coefficient of fault k multiplied by the normalized confidence weight; the confidence weight is scaled by dividing by the maximum value to ensure it is within a reasonable range.

[0159] The severity coefficient of the fault is set according to the degree of impact: 1.0 for gearbox and generator faults, because they may lead to long downtime and high maintenance costs; 0.8 for blade and electrical faults; and 0.6 for yaw and hydraulic faults.

[0160] Sum the risk contributions of all failure types to obtain the comprehensive risk score of unit i; calculate the scores of 100 units and sort them from high to low to generate a unit risk ranking list;

[0161] S65, Weather Window Forecasting and O&M Resource Constraint Modeling: Combining weather window forecasting and O&M resource constraints, an O&M plan optimization model is constructed. Maritime O&M is strictly limited by weather conditions, requiring activities to be scheduled within suitable weather windows while considering limited resources.

[0162] Acquire 15-day weather forecast data, including wind speed, wave height, and visibility. Define a weather suitability index: a suitability score of 1 indicates perfect suitability when wind speed is less than 12 meters per second, wave height is less than 2 meters, and visibility is greater than 5 kilometers; a score of 0 indicates unsuitability when any parameter exceeds the safe range; linear interpolation is used to obtain a suitability value between 0 and 1 for the critical range. Divide the next 15 days into 15 time windows, each lasting one day, and calculate the weather suitability for each window. Window with high suitability is the preferred time for operation and maintenance.

[0163] Define the constraints of operation and maintenance resources: 2 operation and maintenance ships, each ship can serve a maximum of 5 units per day, 20 operation and maintenance personnel are divided into 4 groups, considering operation and maintenance time and ship round trip, each time window can support a maximum of 10 units, as a hard constraint condition for the optimization model;

[0164] S66, Construction and solution of multi-objective optimization model; Construct a multi-objective optimization model to simultaneously optimize operation and maintenance efficiency and resource utilization, including maximizing risk coverage and minimizing operation and maintenance costs;

[0165] Define decision variables: For unit i and time window t, define binary decision variables, where 1 indicates that unit i is maintained in time window t and 0 indicates that it is not maintained. The decision variable matrix has a dimension of 100 multiplied by 15, for a total of 1500 decision variables.

[0166] The optimization objectives are structured as follows: The first objective is to maximize risk coverage. The objective function is equal to the sum of decision variables multiplied by the unit risk score and then multiplied by the weather suitability, encouraging the scheduling of operations and maintenance (O&M) for high-risk units and high-suitability windows. The second objective is to minimize O&M costs. O&M costs include vessel voyage costs and O&M operation costs. Vessel voyage costs are fixed costs for each window with O&M tasks, while O&M operation costs are determined based on the fault type and severity. The total cost equals the voyage cost plus the operation cost. The two objectives are weighted and combined to form a comprehensive objective function.

[0167] S67, Constraint Setting and Adaptive Large Neighborhood Search Solution; Constraints are set to ensure the feasibility of the operation and maintenance plan: The first constraint is the resource capacity constraint, the number of units to be operated and maintained in each time window is less than or equal to 10; the second constraint is the unit uniqueness constraint, each unit is operated and maintained at most once within the planning period to avoid duplicate operation and maintenance; the third constraint is the geographical proximity constraint, the units to be operated and maintained in the same time window should be geographically close as much as possible to reduce ship travel time; the fourth constraint is the fault urgency constraint, high-risk units must complete operation and maintenance before their predicted fault occurrence time;

[0168] An adaptive large neighborhood search algorithm is used to solve the multi-objective optimization problem. This algorithm is particularly suitable for the dynamic constraints and real-time adjustment requirements of offshore wind power operation and maintenance scenarios. The algorithm consists of three stages: initial solution construction, damage repair iteration, and adaptive weight adjustment.

[0169] Specifically, in the initial solution construction phase, a greedy heuristic strategy is used to generate high-quality initial solutions. Units are sorted from highest to lowest based on their comprehensive risk score, and maintenance time windows are allocated to each unit accordingly. For unit i, its comprehensive score in each time window t is calculated. The comprehensive score equals the unit risk score multiplied by the weather suitability score divided by 1, plus the number of units already allocated in that window. The time window with the highest comprehensive score that satisfies the resource capacity constraint is prioritized. Geographical proximity is also considered; if a unit is already allocated in that window, the average distance between unit i and the allocated units is calculated. If the distance is less than 5 kilometers, a proximity bonus of 1.2 is given to encourage geographical concentration. All units requiring maintenance are traversed to obtain the initial feasible solution.

[0170] In the damage and repair iteration phase, four damage operators and three repair operators are designed to achieve efficient exploration of the solution space through adaptive selection. Damage operators include: a random removal operator that randomly selects 10% to 30% of the assigned units to remove their maintenance schedules; a worst-case removal operator that removes the 20% of units with the lowest contribution to the objective function; a time window clearing operator that randomly selects 2 to 3 time windows to clear all maintenance schedules; and a correlation removal operator that identifies groups of units with similar fault types and removes the entire group's schedules to re-optimize collaborative maintenance. Repair operators include: a greedy insertion operator that selects the optimal time window for inserting removed units in descending order of risk score; a regret value insertion operator that calculates the regret value of each unit in each window, where the regret value equals the score of the second-best window minus the score of the best window, prioritizing the insertion of units with higher regret values ​​to avoid subsequent scheduling difficulties; and a clustering repair operator that clusters removed units by geographical location, assigning each cluster to the same time window to maximize vessel utilization efficiency.

[0171] In each iteration, a destruction operator and a repair operator are randomly selected. The current solution is destroyed, and a new solution is generated through repair. The objective function value of the new solution is calculated, and the simulated annealing acceptance criterion is used to determine whether to accept the new solution. The acceptance probability is 1 when the new solution is better than the current solution, and equal to the negative objective function difference divided by the exponential function of the current temperature when the new solution is worse than the current solution. The initial temperature is set to 100, and every 50 iterations, the temperature is multiplied by a cooling coefficient of 0.95, allowing the algorithm to accept worse solutions in the early stages to escape local optima, and gradually converge to a high-quality solution in the later stages.

[0172] In the adaptive weight adjustment phase, the selection probabilities of each destruction and repair operator are dynamically adjusted. Initially, all operators have equal probabilities. Every 50 iterations, the performance of each operator is statistically analyzed, with scores calculated based on the number of times the operator is used and the number of times it generates an improved solution. An operator that generates a new optimal solution receives a score increase of 30, an operator that generates an acceptable solution receives a score increase of 10, and an operator that generates a rejected solution receives a score increase of 1. The operator selection probabilities are updated based on these scores, using a roulette wheel selection mechanism where the probability equals the operator's score divided by the sum of all operator scores. Operators with better performance receive higher selection probabilities, achieving adaptive learning for the algorithm.

[0173] The iteration termination condition is set as 100 consecutive iterations without improvement or reaching the maximum number of iterations (1000). The algorithm records historical optimal solutions during the process and outputs the optimal operation and maintenance plan after the iteration ends. Compared to traditional genetic algorithms, adaptive large neighborhood search converges faster and provides higher quality solutions when dealing with complex constraints and large-scale problems. It can also flexibly adapt to the dynamic adjustment needs of offshore wind power operation and maintenance, and can quickly re-optimize the operation and maintenance plan when weather forecasts are updated or the unit status changes.

[0174] The best historical solution is selected as the optimal operation and maintenance plan. The operation and maintenance time window and geographical grouping of each unit are specified to achieve a multi-objective balance of risk coverage, cost control and efficient resource utilization.

[0175] S68, Operation and Maintenance Decision Report Generation; Based on the optimal operation and maintenance solution, a detailed operation and maintenance decision report is generated to provide actionable guidance for the operation and maintenance team, including the following:

[0176] Unit Maintenance Prioritization: Units requiring maintenance are listed from highest to lowest based on their comprehensive risk score, with key fault types, probability of occurrence, prediction confidence level, and risk score clearly indicated to help quickly identify the most urgent tasks. Recommended Maintenance Time Window: An optimal maintenance time window is assigned to each unit, noting weather suitability, expected wind speed, wave height, and other meteorological conditions, comprehensively considering unit risk, weather conditions, and resource availability. Estimated Maintenance Man-Hours and Spare Parts List: Maintenance man-hours are estimated based on fault type and severity, including diagnostic time, repair time, and testing time. A spare parts list is provided, including name, specifications, and quantity, ensuring sufficient spare parts are carried to avoid maintenance delays. Maintenance Resource Allocation Plan: Vessel arrangements and personnel configurations are specified for each time window to ensure efficient resource utilization. Multiple units within the same time window are grouped by geographical location to optimize vessel routes and reduce round-trip time. Risk Mitigation Effect Prediction: The risk reduction after implementing the maintenance plan is calculated, and the comprehensive risk score before and after maintenance is compared to quantify maintenance benefits, providing a quantitative basis for decision-making rationality.

[0177] Outputs final early warning results and operation and maintenance decision recommendations, including fault early warning information, unit risk assessment, operation and maintenance priority ranking, and detailed operation and maintenance plan, supporting intelligent operation and maintenance management of offshore wind farms.

[0178] In one embodiment of the present invention, an application example of the present invention is provided;

[0179] This invention focuses on the application of intelligent operation and maintenance in offshore wind farms. Based on the actual operation data of a large offshore wind farm, it demonstrates the application effect of the method of this invention.

[0180] Specific description of application scenarios:

[0181] Located in the East China Sea, approximately 50 kilometers from the coastline, this offshore wind farm comprises 100 5MW wind turbines with a total installed capacity of 500MW. The wind farm operates in harsh environments, with an average annual wind speed of 8.5 meters per second and a maximum wind speed of 35 meters per second, an average annual humidity of 85%, and high salt spray concentrations. Maintenance costs are exorbitant, with a single offshore maintenance trip costing approximately 150,000 yuan, and only 180 days of offshore maintenance allowed per year.

[0182] Wind turbine unit WT-025 was selected as an application example. This unit was put into operation in 2020 and has accumulated more than 30,000 hours of operation. The fault early warning system of this invention was used to monitor and issue early warnings for this unit continuously for 30 days.

[0183] Data collection example:

[0184] During the monitoring period, multimodal data of the unit were collected; Table 1 shows an example of some sensor data collected at a certain moment.

[0185] Table 1: Examples of Multimodal Sensor Data

[0186]

[0187] As shown in Table 1, the quality scores of the data from each modal sensor are all above 0.85, indicating good data quality. The gearbox oil temperature and generator winding temperature are within the normal operating range, but the vibration peak of the high-speed shaft of the gearbox is slightly higher than the historical average, suggesting a possible early sign of failure.

[0188] Example of feature extraction results:

[0189] Based on the collected multimodal data, feature vectors for each modality are obtained using the feature extraction method of this invention. Table 2 shows some examples of the extracted features.

[0190] Table 2: Examples of Multimodal Feature Extraction Results

[0191]

[0192] As shown in Table 2, slight anomalies were observed in the IMF2 arrangement entropy, peak factor, and impulse factor in the vibration characteristics, and slight anomalies were also observed in the spectral centroid and wavelet packet energy ratio in the acoustic characteristics. These anomalies corroborate each other, suggesting that the gearbox may have an early wear failure. The temperature and electrical characteristics were basically normal, indicating that the failure was still in its early stage and had not yet caused significant temperature increases or electrical parameter anomalies.

[0193] Fault warning results:

[0194] The long-term early warning model of this invention is used to predict the failure risk of the unit over the next 15 days. The early warning results show that the probability of gearbox failure of the unit within the next 12 days is 0.78, with a confidence level of 0.85 and a prediction standard deviation of 0.12; at the same time, a potential risk of generator bearing failure is identified, with a failure probability of 0.45 and a confidence level of 0.72.

[0195] Based on the early warning results, the system calculated the unit's comprehensive risk score to be 8.6 out of 10, ranking it 3rd among the 100 units in the wind farm. It is considered a high-risk unit and it is recommended to prioritize its operation and maintenance.

[0196] The system is expected to require 8 hours of maintenance, which will involve replacing the high-speed shaft bearings and lubricating oil in the gearbox. The required spare parts include 2 sets of high-speed shaft bearings and 200 liters of lubricating oil.

[0197] Actual verification results:

[0198] Following system recommendations, the maintenance team performed preventative maintenance on the unit. Disassembly and inspection revealed early pitting and wear marks on the high-speed shaft bearing of the gearbox, with a wear depth of approximately 0.3 mm. If not addressed promptly, this was expected to develop into a serious failure within 1-2 weeks, leading to unit shutdown. After replacing the bearing and lubricating oil, the unit returned to normal operation, and its vibration and acoustic characteristics returned to normal levels.

[0199] This case demonstrates that the method of the present invention can accurately predict early gearbox failures, giving the maintenance team sufficient preparation time and avoiding sudden downtime and greater equipment damage.

[0200] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for early warning of wind turbine faults based on multimodal data fusion, characterized in that, Includes the following steps: S1: Collect raw data from multimodal sensors, use basic preprocessing methods, and output a spatiotemporally aligned multimodal dataset; S2: Obtain a spatiotemporally aligned multimodal dataset, extract features using mode decomposition and spectral analysis, and output a cross-modal fusion feature set; S3, based on a cross-modal fusion feature set, employs a three-factor dynamic weighted fusion process to output the final fusion feature; S4 receives the final fused features, and uses dynamic topology construction, wake coefficient calculation and spatiotemporal graph convolution processing to output cross-unit correlation features; S5 constructs comprehensive features based on the final fusion features and cross-unit correlation features, and uses dual-branch gated fusion and Bayesian network processing to output early warning results with confidence. S6 obtains warning results with confidence levels, uses multi-label classification to identify faults, and outputs warning results and operation and maintenance decision-making solutions.

2. The wind turbine fault early warning method based on multimodal data fusion according to claim 1, characterized in that, S1 includes: Deploy a multimodal sensor network at key locations to collect vibration data, temperature data, acoustic data, electrical parameter data, SCADA data, and environmental meteorological data; perform basic preprocessing on the collected raw multimodal data; Timestamp alignment is performed on data with different sampling frequencies, and a spatiotemporally aligned multimodal dataset is obtained with the sampling time of SCADA data as the reference time axis.

3. The wind turbine fault early warning method based on multimodal data fusion according to claim 1, characterized in that, S2 includes: Variational mode decomposition is applied to the vibration data to decompose the signal into multiple intrinsic mode functions. For each intrinsic mode function, multi-scale permutation entropy features, sample entropy, approximate entropy and time-domain statistical features are extracted and concatenated to form a vibration feature vector. The STL decomposition method is used to decompose the temperature data into trend components, periodic components and residual components. Features are extracted for each component and concatenated to form a temperature feature vector. Mel frequency cepstral coefficients and frequency domain features are extracted from acoustic data, and wavelet packet decomposition is used to obtain wavelet packet energy spectra. These are then spliced ​​together to form acoustic feature vectors. Power curve deviation analysis and current harmonic analysis are performed on electrical parameter data, and the data are spliced ​​together to form an electrical feature vector.

4. The wind turbine fault early warning method based on multimodal data fusion according to claim 1, characterized in that, S2 further includes: To address the differences in feature dimensionality across different modalities, an adaptive projection method is employed to unify them to the target dimension; principal component analysis is used for dimensionality reduction of high-dimensional features, while nonlinear mapping is used for dimensionality increase of low-dimensional features. Based on a feature set with unified dimensions, a cross-modal attention mechanism is constructed. A multi-head attention mechanism is adopted, in which each head independently calculates the attention weight and then the two heads are spliced ​​and fused. Calculate the attention weights between modalities, obtain the similarity score by the inner product of the query vector and the key vector, and obtain the attention weight distribution after normalization; perform feature interaction based on the attention weights, add the interaction features to the original projected features, and output the cross-modal fusion feature set.

5. The wind turbine fault early warning method based on multimodal data fusion according to claim 1, characterized in that, S3 includes: Based on the cross-modal fusion characteristics, the dynamic fusion weights of each mode for different fault types are calculated. The dynamic weights comprehensively consider three factors: data quality, fault type correlation and information gain. The correlation coefficient of fault type is obtained by statistical analysis of historical fault data, and the correlation between mode and fault type is quantified by mutual information method. The original weight score is obtained by multiplying the three factors, scaling it by the temperature parameter, applying an exponential function and normalizing it to obtain the final weight; the information gain of each mode is calculated. A weighted fusion strategy is adopted to sum the cross-modal fusion features of each modality with their corresponding dynamic weights to obtain the final fusion feature vector.

6. The wind turbine fault early warning method based on multimodal data fusion according to claim 1, characterized in that, S4 includes: Real-time wind speed and direction data of wind farms are acquired, and the wake relationship between units is determined based on real-time wind direction and unit geographical location to construct a dynamic unit spatial topology map; The wake influence coefficient between computer groups takes into account three factors: distance between units, wind direction angle, and wind speed. Based on the dynamic topology graph and the wake influence coefficient, a spatiotemporal graph convolutional network is constructed. Each layer contains two operations: spatial convolution and temporal convolution. Spatial convolution aggregates the features of spatial neighbor nodes, and temporal convolution uses gated recurrent units to capture temporal dependencies. By employing bilinear transformation to determine the fault propagation probability among computer groups, a cascading fault risk map of the unit group is constructed, and cross-unit correlation characteristics are output.

7. The wind turbine fault early warning method based on multimodal data fusion according to claim 1, characterized in that, S5 includes: The final fused features and cross-unit related features are concatenated to obtain a comprehensive feature vector, which is then used to construct a historical feature sequence. Historical data is decomposed into multiple time scales, and time aggregation is performed on each level. For each time scale level, an LSTM network is used to model the temporal evolution pattern, and an in-layer attention mechanism is introduced to obtain the context vector. An inter-layer attention mechanism is used to fuse information from different time scales to obtain a multi-scale fused context vector. The fused context vector is then input into a fully connected layer for feature transformation to obtain a deep feature representation. An extreme gradient boosting tree is used to process the structured features.

8. The wind turbine fault early warning method based on multimodal data fusion according to claim 1, characterized in that, The S5 also includes: A gating fusion mechanism is used to integrate the outputs of LSTM and XGBoost branches, construct the gating weights of the gating network output, and calculate the final warning score based on the gating weights. A Bayesian neural network is introduced to quantify uncertainty, transforming the weights of the fully connected layer in the early warning model from deterministic parameters to probabilistic distribution parameters, and approximate weight posterior distributions are inferred through variational inference. During the prediction phase, multiple samples are taken from the weighted posterior distribution. Each sample is forward-propagated to obtain a warning score, and the multiple scores constitute the prediction distribution. The mean of the predicted distribution is calculated as the final warning score, and the standard deviation is calculated as a measure of uncertainty. Based on the predicted mean and standard deviation, the confidence interval is calculated, and the warning result with confidence level is output.

9. A wind turbine fault early warning method based on multimodal data fusion according to claim 1, characterized in that, S6 includes: A multi-label classification strategy is adopted to identify multiple fault modes simultaneously. A multi-output neural network is constructed, which includes a shared feature layer and multiple fault type-specific output layers, and independently calculates the occurrence probability of each fault type. A fault interaction graph neural network is constructed, with fault types as graph nodes, and a graph attention network is used to learn the interaction influence coefficients between faults; the comprehensive occurrence probability of each fault type is updated based on the fault interaction influence coefficients. The overall risk score for each unit is calculated based on the updated overall probability of failure occurrence, prediction uncertainty, and failure severity. An operation and maintenance plan optimization model is constructed by combining weather window forecasts and operation and maintenance resource constraints; an adaptive large neighborhood search algorithm is used to solve the problem. The algorithm includes three stages: initial solution construction, damage repair iteration, and adaptive weight adjustment, and outputs the optimal operation and maintenance plan and operation and maintenance decision report.

10. A wind turbine fault early warning system based on multimodal data fusion, used to execute the steps in the wind turbine fault early warning method based on multimodal data fusion as described in any one of claims 1-9, characterized in that, include: The data acquisition and preprocessing module is used to acquire raw data from multimodal sensors, and outputs a spatiotemporally aligned multimodal dataset using basic preprocessing methods. The feature extraction and fusion module is used to acquire spatiotemporally aligned multimodal datasets, extract features using mode decomposition and spectral analysis, and output a cross-modal fused feature set; The dynamic weighted fusion module is used to perform three-factor dynamic weighted fusion processing based on cross-modal fusion feature sets and output the final fused features. The cross-unit correlation analysis module is used to receive the final fused features, and output cross-unit correlation features by using dynamic topology construction, wake coefficient calculation and spatiotemporal graph convolution processing. The long-term early warning module is used to construct comprehensive features based on the final fusion features and cross-unit correlation features. It adopts dual-branch gated fusion and Bayesian network processing to output early warning results with confidence. The fault identification and decision-making module is used to obtain warning results with confidence levels, identify faults using multi-label classification, and output warning results and operation and maintenance decision-making solutions.

Images