Fault early warning method and system for operation and maintenance equipment of wind power plant

Abstract

The invention discloses a fault early warning method and system for operation and maintenance equipment of a wind power plant, and relates to the technical field of fan monitoring, and the method comprises the steps: deploying a multi-modal sensor on a fan, collecting and processing data, extracting a data vector, carrying out the vector fusion through a deep branch encoder, and outputting the fan abnormal probability; constructing scattering invariants and defining a third-order coupling amount based on the data vector, synchronously constructing a third-order amplitude through fusion vector decomposition, calculating a third-order residual error by combining the third-order coupling amount and the third-order amplitude, and outputting the third-order residual error; mapping topology according to fan energy transfer and calculating a propagation kernel of the topology, calculating a topological weight according to the propagation kernel, outputting an eikonal core factor and shape correction, and finally outputting a fan eikonal risk index; and constructing a fan topological graph, defining initial features of fan nodes, and analyzing fan abnormal node sources through a graph convolution algorithm. According to the method, false alarm and missing alarm are remarkably reduced, and high sensitivity and high robustness are still achieved under small sample abnormity and complex working condition disturbance.

Description

Technical Field

[0001] This invention relates to the field of wind turbine monitoring technology, and in particular to a fault early warning method and system for wind farm operation and maintenance equipment. Background Technology

[0002] Wind farm operation and maintenance (O&M) condition monitoring and fault early warning technology has undergone three evolutions: from single-sensor monitoring to multimodal perception, from threshold discrimination to intelligent diagnosis, and from offline analysis to edge computing. Early technologies relied on SCADA and accelerometers, using empirical threshold methods based on features such as root mean square, kurtosis, envelope spectrum, and gear meshing sidebands to achieve coarse-grained early warning. Subsequently, data-driven methods such as HMM, SVM, and RF emerged, along with deep learning methods that input vibration / acoustic spectra into CNNs and long-term operating conditions into LSTMs or CRNNs, enabling automatic identification of typical faults such as bearing pitting, tooth surface wear, and blade icing. In recent years, multimodal fusion (vibration / acoustics / vision / strain / environment), digital twins, and edge deployment have been introduced to explore coupling physical and data models of equipment to improve timeliness and interpretability. Simultaneously, graph networks (GNNs) have begun to show promise in complex system topology modeling, attempting to characterize fault propagation trajectories from the perspective of "component-connection-energy flow." Overall, the industry has moved from "point-based threshold monitoring" to "system-level spatiotemporal modeling," but existing technologies still have shortcomings. Common early / late-stage fusion methods are mostly based on experience-weighted or black-box splicing, lacking verifiable cross-modal consistency metrics. They are prone to generating false alarms in strong wind and turbulent scenarios and cannot meet the needs of high-precision wind turbine operation monitoring. Summary of the Invention

[0003] In view of the aforementioned existing problems, the present invention is proposed.

[0004] Therefore, this invention provides a fault early warning method and system for wind farm operation and maintenance equipment, which solves the problem that the existing technology lacks a verifiable cross-modal consistency metric, is prone to generating false alarms in strong wind and turbulent scenarios, and cannot meet the needs of high-precision wind turbine operation monitoring.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a fault early warning method for wind farm operation and maintenance equipment, comprising, Multimodal sensors are deployed on the wind turbine to collect and process data. Data vectors are extracted, and vector fusion is performed through a deep branch encoder to output the wind turbine anomaly probability. Based on the data vector, a scattering invariant is constructed and a third-order coupling quantity is defined. Simultaneously, a third-order amplitude is constructed through fusion vector decomposition. The third-order residual is calculated and output by combining the third-order coupling quantity and the third-order amplitude. Based on the wind turbine energy transfer mapping topology, the propagation kernel of the topology is calculated. Based on the propagation kernel, the topology weight is calculated and the eikonal core factor and shape correction are output. Finally, the wind turbine eikonal risk index is output. Construct a wind turbine topology graph and define the initial characteristics of wind turbine nodes. Analyze the sources of abnormal wind turbine nodes using a graph convolution algorithm.

[0006] As a preferred embodiment of the fault early warning method for wind farm operation and maintenance equipment described in this invention, the step of deploying multimodal sensors on the wind turbine to collect and process data refers to installing vibration sensors, acoustic sensors, visual sensors, and environmental sensors on the wind turbine to collect vibration, sound, image, and environmental data of the wind turbine operation, synchronizing the data collected by the multimodal sensors in time, segmenting the collected data by defining a time window, and using a Hamming window combined with a linear phase FIR low-pass filter for filtering and denoising, and then standardizing the denoised data.

[0007] As a preferred embodiment of the fault early warning method for wind farm operation and maintenance equipment described in this invention, the extracted data vectors are fused by a deep branch encoder and the wind turbine abnormality probability index is output. The corresponding data vectors are extracted by feature engineering methods for the normalized vibration, sound, image and environmental data respectively. A branch encoder is used to process the data vector, and the data vector is input into each branch encoder to obtain a low-dimensional embedding; Calculate the normalized branch attention weights of the data vectors respectively. And obtain the fusion vector F; The fused vector is input into the logistic regression head, which outputs the wind turbine anomaly probability. .

[0008] As a preferred embodiment of the fault early warning method for wind farm operation and maintenance equipment described in this invention, the method involves: constructing scattering invariants based on data vectors and defining third-order coupling quantities; simultaneously constructing third-order amplitudes through fusion vector decomposition; combining third-order coupling quantities and third-order amplitudes to calculate third-order residuals and outputting exponents based on the square of the sum norm of the extracted data vectors in Euclidean space to construct scattering invariants. The first-order coupling quantity is constructed by aligning the scattering invariant with the cross-modal coupling of the data vector; The consistency quantity of the three-mode multiplication is calculated based on the modal correlation in the first-order coupling quantity, and the second-order coupling quantity is constructed. Define the consistency of the four-mode “cluster” and map it to six third-order primitives to construct a third-order coupling quantity; The fusion vector F is constructed using the tensor kernel of low-rank CP decomposition to construct first to third order amplitudes; Calculate and output the third-order residual by combining the third-order coupling quantity and the third-order amplitude; For each time window m, output the fusion vector, anomaly probability, third-order coupling quantity, third-order magnitude, and third-order residual.

[0009] As a preferred embodiment of the fault early warning method for wind farm operation and maintenance equipment described in this invention, the method involves mapping the wind turbine energy transfer topology and calculating the propagation kernel of the topology, calculating the topology weight based on the propagation kernel and outputting the eikonal core factor and shape correction, and finally outputting the wind turbine eikonal risk index, which is mapped to four topology routes based on the wind turbine energy transfer route. For each wind turbine component in each topology route, during the healthy period, the average spectral peak of the sensor for that component is statistically analyzed, and the mode index with the largest average spectral peak is taken as... And calculate the Green's function of this component in the topology propagation kernel. ; Green's function based on wind turbine components calculates the topology propagation kernel for four types of topology routes; The topological average kernel strength is defined based on the propagation kernel of each topology; Generating graded coupling scores based on third-order coupling quantities; The order weights are calculated by combining the order coupling scores and then normalized. Calculate the full-order expansion factor for the current time window to obtain the eikonal core factor. With shape correction ; The Eikonal risk index of the wind turbine is calculated by combining the Eikonal core factor and shape correction. .

[0010] As a preferred embodiment of the fault early warning method for wind farm operation and maintenance equipment described in this invention, the following steps are described: constructing a wind turbine topology graph and defining initial features of wind turbine nodes; analyzing the sources of abnormal wind turbine nodes using a graph convolution algorithm; constructing a wind turbine component topology graph based on the components of the wind turbine; establishing connecting edges based on the relationships between nodes and performing undirected processing in a graph convolutional network (GCN); and assigning initial node features to each node. ; A two-layer GCN combined with residuals is used to perform forward propagation of node features; The node suspicion level is calculated based on the final output of GCN, and the node with the highest suspicion level is identified as the source of the abnormal node and an early warning is issued.

[0011] As a preferred embodiment of the fault early warning method for wind farm operation and maintenance equipment described in this invention, the abnormal node source of the wind turbine is obtained, and the abnormal node is marked and highlighted in the three-dimensional digital model of the wind turbine.

[0012] Secondly, the present invention provides a fault early warning system for wind farm operation and maintenance equipment, comprising, The anomaly analysis module is used to deploy multimodal sensors on the wind turbine to collect and process data, extract data vectors, perform vector fusion through a deep branch encoder, and output the wind turbine anomaly probability. The structure mapping module is used to construct scattering invariants and define third-order coupling quantities based on data vectors, simultaneously construct third-order amplitudes through fusion vector decomposition, and calculate and output third-order residuals by combining third-order coupling quantities and third-order amplitudes. The structural disturbance analysis module is used to map the topology based on the energy transfer of the wind turbine and calculate the propagation kernel of the topology. Based on the propagation kernel, the topology weights are calculated and the eikonal core factor and shape correction are output. Finally, the wind turbine eikonal risk index is output. The topology analysis module is used to construct the wind turbine topology graph and define the initial characteristics of the wind turbine nodes, and analyze the sources of abnormal nodes in the wind turbine through graph convolution algorithms.

[0013] Thirdly, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the fault early warning method for wind farm operation and maintenance equipment as described in the first aspect of the present invention.

[0014] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the fault early warning method for wind farm operation and maintenance equipment as described in the first aspect of the present invention.

[0015] The beneficial effects of this invention are as follows: By deploying multimodal sensors on the wind turbine, a standardized data vector is first formed, which is then fused early by a dedicated branch encoder to output the anomaly probability. Further, a "scattering invariant" is constructed based on the data vector, and a third-order coupling quantity is defined. Simultaneously, a third-order amplitude decomposition is performed on the fused vector, and the third-order residual is calculated through the "coupling-amplitude" correspondence for cross-modal self-consistency verification. Then, based on the wind turbine energy transfer topology, a propagation kernel and topological weights are constructed, and an eikonal core factor and shape correction are introduced to construct a closed-loop aggregated eikonal risk index to quantify early risks. Finally, a wind turbine component graph is established, and initial node features are defined. Graph convolution is used to locate anomaly sources on the topology, significantly reducing false alarms and false negatives. It maintains high sensitivity and robustness even under small-sample anomalies and complex operating conditions, shortening fault location and troubleshooting time. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the fault early warning method for wind farm operation and maintenance equipment in Example 1.

[0018] Figure 2 This is a structural diagram of the fault early warning system for wind farm operation and maintenance equipment in Example 1. Detailed Implementation

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0022] Example 1, referring to Figure 1 and Figure 2 This is the first embodiment of the present invention, which provides a fault early warning method for wind farm operation and maintenance equipment, including the following steps: S1. Deploy multimodal sensors on the wind turbine to collect and process data, extract data vectors, perform vector fusion through a deep branch encoder, and output the wind turbine anomaly probability. Specifically, deploying multimodal sensors on the wind turbine to collect and process data involves installing vibration sensors, acoustic sensors, visual sensors, and environmental sensors on the wind turbine to collect vibration, sound, images, and environmental data during wind turbine operation. The environmental data includes temperature, humidity, wind speed, wind direction, and air pressure. The data collected by the multimodal sensors is synchronized in time, time windows are defined to segment the collected data, and Hamming windows combined with linear phase FIR low-pass filters are used for filtering and denoising. The denoised data is then standardized.

[0023] Furthermore, the extracted data vectors are fused using a deep branch encoder, and the wind turbine anomaly probability index is output. Feature engineering methods are then used to extract corresponding data vectors from the normalized vibration, sound, image, and environmental data, including: The normalized vibration data were subjected to Fourier transform and frequency domain features were extracted. The frequency domain amplitude values ​​in the frequency domain features were then concatenated into a matrix within a fixed frequency band (0-2.5kHz). Where M is the channel multiplication point and T is the time frame, the matrix is ​​decomposed using the PCA dimensionality reduction algorithm within the time window m to obtain the covariance:

[0024] in Let covariance matrix be the variance matrix. The eigenvector matrix, For the eigenvalue matrix, For transpose; By performing principal component dimensionality reduction within a window using the vibration frequency domain feature matrix to obtain the vibration vector, a concentrated expression of the principal modal energy of the structure and compression of the noise dimension are achieved. Its function is to significantly reduce the instability of indicators caused by wind noise and slow drift. Its application is to obtain a high signal-to-noise ratio structural characterization at the edge with a fixed amount of computation. Thus, it can maintain detection sensitivity and controllable data transmission and storage load even in the early stage of weak anomalies. Select the first K eigenvectors from the eigenvector matrix, and then use the matrix of the time window m. Projection to obtain the vibration vector of the m-th window :

[0025] Fourier transform is used to extract the frequency domain features of the sound data, and the sound frequency domain features are mapped using the mel frequency band and the logarithmic energy is taken to extract the sound vector. ; By using Mel mapping and logarithmic energy expression of acoustic features, robust capture of meshing harmonics and bearing overtones is achieved; its function is to reduce the masking of spectral peaks by broadband wind shearing noise; its purpose is to provide a second independent evidence when lightly loaded or in poor vibration sensing position; thereby achieving the effect of improving the detection rate and reducing reliance on a single mode in low excitation scenarios. Based on the analysis of wind turbine image data using frozen ResNet-18, for the m-th window... Extracting image vectors from frame images :

[0026] in The avgpool activation for frame f is obtained from each frame image; By freezing the image network and obtaining image vectors by frame averaging within a window, low-variance quantization of gradually changing appearances such as cracks, erosion, and icing is achieved. Its purpose is to avoid drift and overfitting caused by online fine-tuning. Its application is to maintain feature stability under varying lighting and angle conditions, thereby achieving the effect of being usable and reproducible under small sample conditions. Environmental vectors are constructed by extracting wind speed, temperature, relative humidity, and gust indices from environmental data. Among them, the gust index for:

[0027] in Let be the wind speed at time t; By introducing gust indicators and incorporating them into the environmental vector, a clear characterization of short-term operating condition changes is achieved. Its function is to provide exogenous explanatory variables for the probability of anomalies, distinguishing between load disturbances and actual degradation. Its purpose is to perform adaptive correction in the alarm pre-processing stage, thereby reducing false positives and unnecessary maintenance triggers. The data vector is processed using a branch encoder. The data vector is input into each branch encoder to obtain a low-dimensional embedding, including: Vibration vector The input vibration is extracted using LSTM branches, and the hidden states are then normalized by average pooling to obtain the normalized vibration embedding. , ; sound vector The audio is input to a CRNN branch, which performs 1D convolution and then uses an RNN to extract the hidden states before normalization to obtain the sound embedding. , v; Image vector The input image is linearly mapped and normalized by a CNN branch to obtain the image embedding. , ; environment vector The input environment MLP branch is processed using a two-layer perceptron, followed by up-dimensional projection and normalization to obtain the environment embedding. , ; Calculate the normalized branch attention weights of the data vectors respectively. And obtain the fusion vector F:

[0028]

[0029]

[0030] in Let be the attention weights for each branch. The normalized attention weights for the k-th branch are... and These are the attention weight parameters; By outputting low-dimensional embeddings through a four-branch encoder and completing early fusion with attention, adaptive amplification of the most relevant modes is achieved under different fault conditions. Its function is to automatically weight the fusion results according to the context, without relying on a manual weight table. Its purpose is to uniformly output a usable risk feature for upper-level discrimination, thereby achieving the effect of generalization improvement across working conditions and machine types and traceable explanatory clues. The fused vector is input into the logistic regression head, which outputs the wind turbine anomaly probability. :

[0031] in As weight, For bias, This is the activation function.

[0032] By outputting anomaly probabilities through logistic regression headers, lightweight inference and auditable thresholds are achieved. Its function is to facilitate direct integration with the alarm policies of existing operation and maintenance platforms. Its purpose is to enable rapid deployment and unified interface management, thereby achieving the effect of implementation and large-scale operation and maintenance without modifying existing systems.

[0033] S2. Construct scattering invariants based on data vectors and define third-order coupling quantities. Simultaneously construct third-order amplitudes through fusion vector decomposition. Calculate and output third-order residuals by combining third-order coupling quantities and third-order amplitudes. Specifically, scattering invariants are constructed based on data vectors, and a third-order coupling quantity is defined. Simultaneously, a third-order amplitude is constructed through fused vector decomposition. The third-order residual is calculated by combining the third-order coupling quantity and the third-order amplitude, and the output is the sum norm squared of the extracted data vectors in Euclidean space.

[0034] in For the structural response-acoustic joint excitation intensity, For structure-environment coupling deviation, This is a structural-visual defect coupling bias. By constructing scattering invariants and aligning them with multimodal data within the same time window, observations with different noise characteristics and dimensions are converted into a unified and stable scalar basis. Its function is to measure three types of cross-modal coupling with a fixed caliber, avoiding bias diffusion caused by operating condition disturbances; its purpose is to provide external anchor points for subsequent consistency checks that are independent of training weights; ultimately, it achieves the effect of maintaining measurement stability even under conditions of weak anomalies and wind noise coexisting. Construct a first-order coupling quantity by aligning the scattering invariant with the cross-modal coupling of the data vector:

[0035]

[0036] in For modal correlation, It is a first-order coupling quantity; Aligning scattering invariants with cross-modal correlations using first-order coupling quantities achieves a projection from geometric similarity to physical correlation. Its function is to express intermodal synchronicity with comparable strength, unaffected by misleading peaks in single-modal anomalies; its purpose is to provide a second, independent chain of evidence for early anomalies; ultimately, it improves detection reliability without increasing sensing hardware. Calculate the consistency quantity of the three-mode multiplication based on the modal correlation in the first-order coupling quantity, and construct the second-order coupling quantity:

[0037]

[0038] in To measure the consistency of the three models, It is a second-order coupling quantity; By constructing trimodal consistency and second-order coupling quantities, the co-occurrence of multi-source anomalies is characterized. Its function is to distinguish between single-point noise and cross-modal resonance, especially to identify hidden faults caused by combined structural-acoustic-visual changes; its application is as an important escalation trigger condition, used in parallel with probabilistic outputs; ultimately achieving a stronger discriminative effect for complex coupled anomalies. Define the four-mode "cluster" consistency and map it to six third-order primitives to construct a third-order coupling quantity:

[0039]

[0040]

[0041]

[0042] in This is a triple consistency product. It is a third-order coupling quantity; By defining four-mode clusters and mapping third-order coupling quantities, closed-loop coupling identification across components and channels is achieved. Its function is to capture collaborative anomalies caused by chain propagation within a single window, providing prior evidence for localization and risk ranking. Its purpose is to provide interpretable input for path weighting in subsequent risk aggregation, ultimately achieving a more sensitive response to chain propagation and early propagation. The fusion vector F is used to construct first to third order amplitudes through the tensor kernel of low-rank CP decomposition:

[0043]

[0044]

[0045]

[0046] in , , as well as These represent the zeroth, first, second, and third order amplitudes, respectively. It is a first-order linear kernel. and The direction vector of the second-order CP low-rank kernel. , as well as The direction vector of the third-order CP low-rank kernel is obtained through offline training; By applying linear and low-rank tensor kernel decomposition to the fusion vector to obtain multi-order amplitudes, the high-dimensional embedding is decomposed into additive and comparable order contributions. Its purpose is to organize information with a fixed order, avoiding the uninterpretability caused by black box splicing. Its purpose is to provide a reference quantity for the same order alignment of the consistency residual. Ultimately, it achieves the effect of completing the decomposition at the edge with controllable complexity and maintaining real-time performance. Calculate and output the third-order residual by combining the third-order coupling quantity and the third-order amplitude:

[0047]

[0048]

[0049] in , as well as These are the first-order, second-order, and third-order consistent residuals, respectively. and These are engineering coupling constants, obtained through offline training and calibration. As the base width, ; By comparing the coupling quantities and amplitudes of the same order in a differentiated manner and outputting consistent residuals, a two-way verification of model evidence and physical evidence is achieved. Its function is to provide a robust triggering reference from the residuals even when the probability output is affected by sudden changes in operating conditions; its purpose is to introduce a mutual verification mechanism in early warning decision-making to reduce false triggers; ultimately, it achieves the effect of significantly reducing false alarms and false negatives and improving alarm credibility in volatile environments. For each time window m, output the fusion vector, anomaly probability, third-order coupling quantity, third-order magnitude, and third-order residual.

[0050] S3. Based on the wind turbine energy transfer mapping topology, calculate the propagation kernel of the topology, calculate the topology weights based on the propagation kernel, and output the eikonal core factor and shape correction. Finally, output the wind turbine eikonal risk index. Specifically, based on the wind turbine energy transfer mapping topology and calculating the topology propagation kernel, the topology weights are calculated based on the propagation kernel, and the eikonal core factor and shape correction are output. Finally, the wind turbine eikonal risk index is output, which is mapped to four topology routes based on the wind turbine energy transfer path, including: X1, Trapezoidal 1: Blade → Main shaft → High-speed gearbox → Generator rotor; X2, Trapezoidal 2: Blade → Main shaft → Low-speed gearbox → Generator stator; X3, Cross 1: Blade → Main shaft → (Cross-stage coupling) → Generator stator; X4, Cross 2: Blade → Main shaft → (Cross-stage coupling) → Generator rotor; For each wind turbine component in each topology route, during the healthy period, the average spectral peak of the sensor for that component is statistically analyzed, and the mode index with the largest average spectral peak is taken as... And calculate the Green's function of this component in the topology propagation kernel. :

[0051] in Maximum peak value discrete frequency points, , The peak value of the maximum vibration PSD. The damping ratio is obtained by dividing the half-power bandwidth by twice the peak value of the vibration PSD. Green's functions based on wind turbine components are used to calculate the topology propagation kernel for four types of topology routes, including: For trapezoidal type 1: blade → main shaft → high-speed gearbox → generator rotor:

[0052] in For the propagation kernel of topology route X1, Let m be the set of discrete frequencies within a time window. Let Green's function be the blade. Green's function with the main axis, Let i be the Green's function for the high-speed gearbox, and i be the imaginary unit. For trapezoidal type 2: blade → main shaft → low-speed gearbox → generator stator:

[0053] in For the propagation kernel of topology route X2, For the Green's function of the low-speed gearbox; For cross-type 1: Blade → Main shaft → (cross-stage coupling) → Generator stator:

[0054] in For the propagation kernel of topology route X3, The Green's function of the generator stator. The frequency domain consistency factor of the main shaft and generator stator; For cross 2: Blade → Main shaft → (cross-stage coupling) → Generator rotor:

[0055] in For the propagation kernel of topology route X4, Let Green's function be the generator rotor. The frequency domain consistency factor of the main shaft and generator rotor; The amplitude coherence is obtained through STFT estimation calculation:

[0056] in The amplitudes of components x and y are coherent, and x and y include the main shaft, generator rotor, and generator stator; By mapping energy transfer to two trapezoidal paths and two intersecting paths, and fixing a unique dominant mode for each component based on the spectral peak during the healthy period, sparse modeling of complex dynamic networks is achieved. Its function is to exclude redundant modal noise and retain only the channels with the most energy explanatory power. It can significantly reduce volatility and computational cost in online kernel calculations, thereby maintaining the stability of path discrimination in the early stages of small signals. By introducing a frequency domain consistency factor in the intersecting paths, it can quantitatively characterize the cross-level direct coupling strength. Its function is to distinguish between conventional coupling of "passing along the main chain level by level" and abnormal coordination of "cross-level short circuit". It can be used to identify atypical energy jumps that may occur in the early stages of faults, thereby providing higher sensitivity to cross-coupled faults without increasing hardware overhead. The topological average kernel strength is defined based on the propagation kernel of each topology:

[0057] in and These represent the average core strengths of the trapezoidal and cross-trapezoidal structures, respectively, and indicate the structural transfer gain of the trapezoidal and cross-trapezoidal structures. Generating order-wise coupling scores based on third-order coupling quantities:

[0058]

[0059]

[0060] in , as well as Scoring is given for first, second, and third order trapezoidal coupling. , as well as Scoring is given for first, second, and third order cross-trapezoidal couplings; Calculate the order weights by combining the order coupling scores and then normalize them:

[0061] in For order, , and The order weights for trapezoids and cross trapezoids; By characterizing path propagation gain with average kernel intensity and jointly normalizing the order coupling score with it into order weights, the same-scale fusion of data consistency evidence and structure propagation evidence is achieved. Its function is to make different orders and different paths comparable on the same scale. It is used to dynamically emphasize the most credible path-order combination in the window, thereby reducing misjudgments caused by random spectral peaks. Calculate the full-order expansion factor for the current time window to obtain the eikonal core factor. With shape correction :

[0062]

[0063] in For the Gamma function, To minimize the time, preventing singularities caused by excessively small time t. The eikonal engineering coupling constant is calibrated in one step based on the health dataset. Dimensional regularization; By calculating the core factors and shape correction of the full-order expansion, and outputting a unified risk index under factorial decay and order-based weight constraints, it realizes the encapsulation of multi-source, multi-order, and cross-path evidence into a single threshold-manageable indicator. Its function is to avoid the contradictions and maintenance difficulties caused by the stacking of multiple thresholds. It is used for risk ranking and operation and maintenance decisions across units and time periods, thereby improving the consistency and portability of early warning caliber. The Eikonal risk index of the wind turbine is calculated by combining the Eikonal core factor and shape correction. :

[0064] in For factorial, and These are the normalized weights.

[0065] S4. Construct a wind turbine topology graph and define the initial features of wind turbine nodes. Analyze the sources of abnormal wind turbine nodes using a graph convolution algorithm. Specifically, a wind turbine topology graph is constructed and initial features of the wind turbine nodes are defined. Anomaly node sources are analyzed using a graph convolution algorithm. A wind turbine component topology graph is constructed based on the turbine's components, including three blade nodes, one main shaft node, one high-speed gearbox node, one low-speed gearbox node, one generator rotor node, and one generator stator node. Connection edges are established based on the relationships between nodes and processed undirected in a graph convolutional network (GCN). These relationships include the direction of force or energy flow. Initial node features are assigned to each node. :

[0066] in , as well as This is a global variable for the same user, copied to each node. For the aggregation of node sensor data vectors, For the route strength alignment term, use max-min normalization; By constructing a minimal and sufficient topology graph with components as nodes, the physical constraints of the wind turbine's mechanical links and electromagnetic terminals are explicitly encoded. This transforms the localization problem from pure feature clustering to propagation discrimination constrained by topology, avoiding "hot zone drift" in complex coupling scenarios. In practical applications, this graph can propagate at a fixed scale and constant at the edge, keeping computational burden controllable. This achieves stable convergence to reasonable candidates even with small samples and noise. By injecting global variables into each node and combining them with local sensor statistics to form initial features, "global risk background - local anomaly clues" are aligned within the same domain. This prevents nodes from mistaking global disturbances for local faults when the overall load changes drastically. In practical applications, this design reduces reliance on complex threshold compensation, thus significantly reducing false alarms in situations such as strong winds and grid disturbances. A two-layer GCN combined with residuals is used for forward propagation of node features:

[0067]

[0068] in and The outputs of the two GCN layers respectively, , , as well as These are the trainable parameters of GCN. This is the final output of GCN. The residual coefficient; By employing two layers of graph convolution supplemented with residuals, the physical radius of information propagation is limited while preserving in-situ semantics, overcoming common oversmoothing issues. Its purpose is to balance the two-hop impact of cascading failures with source-end resolution; practically, it reduces the number of layers and parameter scale, facilitating real-time operation on edge devices; thus achieving the effect of neither losing neighborhood evidence nor sacrificing real-time performance. The node suspicion level is calculated based on the final output of GCN, and the node with the highest suspicion level is identified as the source of the abnormal node and an alert is issued.

[0069] in Let i be the degree of suspicion of node i. These are the weighting coefficients. For bias.

[0070] By embedding nodes into a lightweight input readout layer to obtain suspiciousness, and using the maximum value as the anomaly source output, a single criterion closed loop is achieved from "whether it is abnormal" to "where the anomaly is". Its function is to transform complex multimodal and multipath evidence into auditable and traceable numerical outputs, facilitating direct integration with existing operation and maintenance platforms. In terms of application, it supports the automatic issuance of handling strategies based on fixed thresholds or sorting policies, thereby reducing manual review costs and shortening location time.

[0071] Furthermore, after obtaining the source of the abnormal node in the wind turbine, the abnormal node is marked and highlighted in the three-dimensional digital model of the wind turbine.

[0072] This embodiment also provides a fault early warning system for wind farm operation and maintenance equipment, including: The anomaly analysis module is used to deploy multimodal sensors on the wind turbine to collect and process data, extract data vectors, perform vector fusion through a deep branch encoder, and output the wind turbine anomaly probability. The structure mapping module is used to construct scattering invariants and define third-order coupling quantities based on data vectors, simultaneously construct third-order amplitudes through fusion vector decomposition, and calculate and output third-order residuals by combining third-order coupling quantities and third-order amplitudes. The structural disturbance analysis module is used to map the topology based on the energy transfer of the wind turbine and calculate the propagation kernel of the topology. Based on the propagation kernel, the topology weights are calculated and the eikonal core factor and shape correction are output. Finally, the wind turbine eikonal risk index is output. The topology analysis module is used to construct the wind turbine topology graph and define the initial characteristics of the wind turbine nodes, and analyze the sources of abnormal nodes in the wind turbine through graph convolution algorithms.

[0073] This embodiment also provides a computer device applicable to the fault early warning method for wind farm operation and maintenance equipment, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the fault early warning method for wind farm operation and maintenance equipment as proposed in the above embodiment.

[0074] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0075] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the fault early warning method for wind farm operation and maintenance equipment as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0076] In summary, this invention deploys multimodal sensors on wind turbines to first form standardized data vectors, which are then fused early by a dedicated branch encoder to output anomaly probabilities. Further, it constructs "scattering invariants" based on the data vectors and defines third-order coupling quantities. Simultaneously, it performs third-order amplitude decomposition on the fused vectors and calculates third-order residuals through the "coupling-amplitude" correspondence for cross-modal self-consistency verification. Then, based on the wind turbine energy transfer topology, it constructs propagation kernels and topological weights, introduces eikonal core factors and shape corrections, and constructs a closed-loop aggregated eikonal risk index to quantify early risks. Finally, it establishes a wind turbine component graph and defines initial node features, using graph convolution to locate anomaly sources on the topology. This significantly reduces false alarms and false negatives, maintaining high sensitivity and robustness even under small-sample anomalies and complex operating conditions, thus shortening fault location and troubleshooting time.

[0077] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A fault early warning method for wind farm operation and maintenance equipment, characterized in that: include, Multimodal sensors are deployed on the wind turbine to collect and process data. Data vectors are extracted, and vector fusion is performed through a deep branch encoder to output the wind turbine anomaly probability. Based on the data vector, a scattering invariant is constructed and a third-order coupling quantity is defined. Simultaneously, a third-order amplitude is constructed through fusion vector decomposition. The third-order residual is calculated and output by combining the third-order coupling quantity and the third-order amplitude. Based on the wind turbine energy transfer mapping topology, the propagation kernel of the topology is calculated. Based on the propagation kernel, the topology weight is calculated and the eikonal core factor and shape correction are output. Finally, the wind turbine eikonal risk index is output. Construct a wind turbine topology graph and define the initial characteristics of wind turbine nodes. Analyze the sources of abnormal wind turbine nodes using a graph convolution algorithm.

2. The fault early warning method for wind farm operation and maintenance equipment as described in claim 1, characterized in that: The process of deploying multimodal sensors on the wind turbine to collect and process data refers to installing vibration sensors, acoustic sensors, vision sensors, and environmental sensors on the wind turbine to collect vibration, sound, image, and environmental data during wind turbine operation. The data collected by the multimodal sensors is synchronized in time, time windows are defined to segment the collected data, and Hamming windows combined with linear phase FIR low-pass filters are used for filtering and denoising. The denoised data is then standardized.

3. The fault early warning method for wind farm operation and maintenance equipment as described in claim 2, characterized in that: The extracted data vectors are fused using a deep branch encoder and the wind turbine anomaly probability index is output. Feature engineering methods are used to extract corresponding data vectors for the normalized vibration, sound, image and environmental data respectively. A branch encoder is used to process the data vector, and the data vector is input into each branch encoder to obtain a low-dimensional embedding; Calculate the normalized branch attention weights of the data vectors respectively. And obtain the fusion vector F; The fused vector is input into the logistic regression head, which outputs the wind turbine anomaly probability. .

4. The fault early warning method for wind farm operation and maintenance equipment as described in claim 3, characterized in that: The process involves constructing scattering invariants based on data vectors and defining a third-order coupling quantity. Simultaneously, a third-order amplitude is constructed through fusion vector decomposition. The third-order residual is calculated by combining the third-order coupling quantity and the third-order amplitude, and the output is based on the square of the sum norm of the extracted data vector in Euclidean space to construct scattering invariants. The first-order coupling quantity is constructed by aligning the scattering invariant with the cross-modal coupling of the data vector; The consistency quantity of the three-mode multiplication is calculated based on the modal correlation in the first-order coupling quantity, and the second-order coupling quantity is constructed. Define the consistency of the four-mode "cluster" and map it to six third-order primitives to construct a third-order coupling quantity; The fusion vector F is constructed using the tensor kernel of low-rank CP decomposition to construct first to third order amplitudes; Calculate and output the third-order residual by combining the third-order coupling quantity and the third-order amplitude; For each time window m, output the fusion vector, anomaly probability, third-order coupling quantity, third-order magnitude, and third-order residual.

5. The fault early warning method for wind farm operation and maintenance equipment as described in claim 4, characterized in that: The process involves mapping the topology based on the wind turbine's energy transfer and calculating the propagation kernel of the topology. Based on the propagation kernel, the topology weights are calculated, and the eikonal core factor and shape correction are output. Finally, the wind turbine eikonal risk index is output, which is mapped to four topology routes based on the wind turbine's energy transfer route. For each wind turbine component in each topology route, during the healthy period, the average spectral peak of the sensor for that component is statistically analyzed, and the mode index with the largest average spectral peak is taken as... And calculate the Green's function of this component in the topology propagation kernel. ; Green's function based on wind turbine components calculates the topology propagation kernel for four types of topology routes; The topological average kernel strength is defined based on the propagation kernel of each topology; Generating graded coupling scores based on third-order coupling quantities; The order weights are calculated by combining the order coupling scores and then normalized. Calculate the full-order expansion factor for the current time window to obtain the eikonal core factor. With shape correction ; The Eikonal risk index of the wind turbine is calculated by combining the Eikonal core factor and shape correction. .

6. The fault early warning method for wind farm operation and maintenance equipment as described in claim 5, characterized in that: The process involves constructing a wind turbine topology graph and defining initial features for each node. Anomaly node sources are analyzed using a graph convolution algorithm. A wind turbine component topology graph is constructed based on the turbine's components. Connection edges are established based on the relationships between nodes and processed undirected in a graph convolutional network (GCN). Each node is then assigned initial features. ; A two-layer GCN combined with residuals is used to perform forward propagation of node features; The node suspicion level is calculated based on the final output of GCN, and the node with the highest suspicion level is identified as the source of the abnormal node and an early warning is issued.

7. The fault early warning method for wind farm operation and maintenance equipment as described in claim 6, characterized in that: After obtaining the source of the abnormal node in the wind turbine, the abnormal node is marked and highlighted in the three-dimensional digital model of the wind turbine.

8. A fault early warning system for wind farm operation and maintenance equipment, based on the fault early warning method for wind farm operation and maintenance equipment according to any one of claims 1 to 7, characterized in that: include, The anomaly analysis module is used to deploy multimodal sensors on the wind turbine to collect and process data, extract data vectors, perform vector fusion through a deep branch encoder, and output the wind turbine anomaly probability. The structure mapping module is used to construct scattering invariants and define third-order coupling quantities based on data vectors, simultaneously construct third-order amplitudes through fusion vector decomposition, and calculate and output third-order residuals by combining third-order coupling quantities and third-order amplitudes. The structural disturbance analysis module is used to map the topology based on the energy transfer of the wind turbine and calculate the propagation kernel of the topology. Based on the propagation kernel, the topology weights are calculated and the eikonal core factor and shape correction are output. Finally, the wind turbine eikonal risk index is output. The topology analysis module is used to construct the wind turbine topology graph and define the initial characteristics of the wind turbine nodes, and analyze the sources of abnormal nodes in the wind turbine through graph convolution algorithms.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the fault early warning method for wind farm operation and maintenance equipment as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the fault early warning method for wind farm operation and maintenance equipment as described in any one of claims 1 to 7.

Images