Offshore wind turbine multi-source heterogeneous data fusion and state evaluation method and device and medium

Abstract

The application discloses a kind of offshore wind turbine multi-source heterogeneous data fusion and state evaluation method, device and medium, the method includes: constructing offshore wind turbine multi-source heterogeneous database;And data preprocessing is carried out;The multi-source heterogeneous data after data preprocessing is mapped to uniform embedding representation space using physical modeling and deep learning technology;In uniform embedding representation space, construct the feature interaction operator based on multi-head cross attention mechanism, execute feature decoupling, and extract panoramic service state vector;Panoramic service state vector is introduced as global observation item into integrated physical information neural network variational bayes calibration framework, realizes the dynamic calibration of digital twin physical parameters, generates offshore wind turbine service state evaluation report with uncertainty quantification support.The application effectively solves the problems of data island, conservative evaluation and decision lag, and generates a service state evaluation report with physical interpretability and uncertainty support.

Description

Technical Field

[0001] This invention relates to a method, device, and medium for multi-source heterogeneous data fusion and condition assessment of offshore wind turbines, belonging to the field of offshore wind power big data processing and structural health monitoring. Background Technology

[0002] As offshore wind power expands into deeper waters and on a larger scale, wind turbines operate for extended periods in extreme environments such as strong winds, giant waves, salt spray, and complex seabed erosion. Their structural safety is crucial to the efficiency of the wind farm. Currently, offshore wind farms have accumulated massive amounts of diverse and heterogeneous data, including high-frequency waveforms generated by CMS vibration sensors, low-frequency environmental scalars generated by weather stations, visual information such as monitoring videos and inspection images, and unstructured text such as operation and maintenance logs and fault reports.

[0003] However, the existing monitoring system faces severe challenges: First, there is a lack of effective semantic alignment and spatiotemporal correlation mechanisms among various types of data, resulting in severe "information silos" and making it difficult to achieve deep coupling of multi-dimensional features. Second, traditional assessment models are mostly designed for single-modal data and cannot automatically extract and integrate expert experience and macroscopic features contained in images and text, leading to insufficient data utilization depth. Finally, because it is impossible to establish an accurate mapping between real-time monitored heterogeneous information and the actual structural bearing capacity, operation and maintenance decisions still rely on delayed periodic inspections, lacking predictive capabilities based on multi-modal risk perception, and making it difficult to achieve a balance between power generation efficiency and scientific operation and maintenance in complex environments. Summary of the Invention

[0004] In view of this, the present invention provides a method, device, computer equipment and storage medium for multi-source heterogeneous data fusion and status assessment of offshore wind turbines, which effectively solves the problems of data silos, conservative assessment and decision lag, and generates a service status assessment report with physical interpretability and uncertainty support.

[0005] The first objective of this invention is to provide a method for fusing and assessing the condition of multi-source heterogeneous data from offshore wind turbines.

[0006] The second objective of this invention is to provide a device for multi-source heterogeneous data fusion and condition assessment of offshore wind turbines.

[0007] A third objective of this invention is to provide a computer device.

[0008] A fourth objective of this invention is to provide a storage medium.

[0009] The first objective of this invention can be achieved by adopting the following technical solution:

[0010] A method for fusing multi-source heterogeneous data and assessing the condition of offshore wind turbines, the method comprising:

[0011] Acquire offshore wind turbine operating status data, structural dynamic response data, external environmental meteorological data, visual morphology of key components, and full life cycle operation and maintenance records to construct a multi-source heterogeneous database for offshore wind turbines;

[0012] Data preprocessing is performed on a multi-source heterogeneous database of offshore wind turbines, including data cleaning and spatiotemporal alignment.

[0013] Using physical modeling and deep learning techniques, multi-source heterogeneous data that has undergone data preprocessing is mapped to a unified embedded representation space;

[0014] Within the unified embedding representation space, a feature interaction operator based on a multi-head cross-attention mechanism is constructed to perform feature decoupling and extract a fusion feature vector that reflects the panoramic operation status of offshore wind turbines, which serves as the panoramic service status vector.

[0015] By introducing the panoramic service status vector as a global observation term into the variational Bayesian calibration framework of the integrated physical information neural network, dynamic calibration of the physical parameters of the digital twin is achieved, generating an offshore wind turbine service status assessment report with uncertainty quantification support.

[0016] Furthermore, the acquisition of offshore wind turbine operating status data, structural dynamic response data, external environmental meteorological data, visual morphology of key components, and full life-cycle operation and maintenance records includes:

[0017] The offshore wind turbine data acquisition system synchronously collects offshore wind turbine operating status data, structural dynamic response data, external environmental meteorological data, visual morphology of key components, and full life cycle operation and maintenance records. It constructs a multi-source heterogeneous database of offshore wind turbines that integrates low-frequency time series, high-frequency waveforms, continuous video streams, discrete high-definition images, and semi-structured text. The offshore wind turbine data acquisition system includes an offshore wind turbine SCADA system, a CMS monitoring system, a nacelle wind measurement system, sea surface meteorological buoys, industrial-grade visual monitoring equipment, intelligent inspection equipment for drones and underwater robots, and an operation and maintenance management system.

[0018] Furthermore, the offshore wind turbine data acquisition system synchronously collects offshore wind turbine operating status data, structural dynamic response data, external environmental meteorological data, visual morphology of key components, and full life-cycle operation and maintenance records, including:

[0019] The offshore wind turbine SCADA system acquires low-frequency time-series data reflecting the operating status of the offshore wind turbine. The low-frequency time-series data includes blade speed, power generation, power output, voltage, current, pitch angle, yaw angle, and nacelle position.

[0020] Vibration sensors deployed on the top of the tower, nacelle, and key parts of the foundation through the CMS vibration monitoring system acquire high-frequency acceleration waveform signals, tower sway signals, and blade flapping and oscillation signals of the generator bearings and offshore wind turbine towers.

[0021] The external environmental meteorological data acting on the offshore wind turbine structure are obtained through the wind measurement system in the engine room and the meteorological buoy on the sea surface. The external environmental meteorological data includes wind speed, wind direction, tide level, significant wave height, wave direction, ambient temperature and humidity, and atmospheric pressure.

[0022] By deploying industrial-grade cameras inside the nacelle, at the base of the tower, and on key structural surfaces of offshore wind turbines, continuous visual images of blade surface icing, tower swaying, and blade rotation are acquired, forming a real-time monitoring video stream database for key parts of offshore wind turbines.

[0023] The inspection images were obtained by using drones to conduct close-up inspections and photography of the surfaces of offshore wind turbine blades, nacelles and towers, and by using underwater robots to inspect and photograph the anti-corrosion coatings of offshore wind turbines and submarine cables below the sea surface.

[0024] Based on electronic archives and scanned paper logs in the operation and maintenance management system, we collect periodic maintenance records, fault repair orders, spare parts replacement history and written descriptions filled out by operation and maintenance personnel to obtain maintenance history information reflecting the entire service life cycle of the unit.

[0025] Furthermore, the data preprocessing of the multi-source heterogeneous database of offshore wind turbines includes:

[0026] For the operational status data, signal drift is identified and corrected through horizontal comparison of multiple wind turbines at the wind farm, outlier data in the wind turbine operational status data is removed, and missing values ​​in the operational status data are filled in using the Lagrange interpolation method. The data is then gradually aligned on the timestamps to serve as the benchmark time axis for wind turbine status assessment.

[0027] For the structural dynamic response data, wavelet packet denoising and variational mode decomposition are used to filter out low-frequency background noise caused by wave force and random bandwidth fluctuations of the mechanical transmission system. The cross-sectional features of the high-frequency acceleration waveform signal are extracted using energy operators. The tower sway signal and blade flapping vibration signal are reduced to a sampling frequency consistent with the operating status data. Based on the signal acquisition trigger timestamp, the data is resampled and mapped to the reference time axis to achieve synchronous alignment between dynamic response and operating status.

[0028] For external environmental meteorological data, the anemometer shading bias is corrected, and the tide level and wave height are smoothed to eliminate the instantaneous interference caused by breaking waves. The external environmental meteorological data is mapped to the specific machine position coordinate system through spatial coordinate transformation. At the same time, the linear interpolation algorithm is used to compensate for the communication delay between sensors and the meteorological observation sequence is aligned to the offshore wind turbine status reference time axis.

[0029] For the video monitoring the nacelle, blades, and tower, image enhancement algorithms are run, and optical flow is used to remove image jitter noise caused by wind turbine vibration; key frame extraction technology is adopted to trigger high-frequency frame extraction based on the alarm time point in the offshore wind turbine SCADA system, and low-frequency monitoring is performed at other times to achieve time alignment between visual features and alarm information.

[0030] For inspection images collected by drones and underwater robots, invalid images that are overexposed at sea, reflective of the sea surface, too dark at the seabed, or blurry in focus are removed. For images collected by underwater robots, blue-green tones are restored, and the images are aligned to the reference time axis of offshore wind turbine status according to the image acquisition time.

[0031] For the entire life cycle operation and maintenance records, optical character recognition technology is used to convert paper log scan files into electronic archive files, correct non-standard abbreviations and ambiguous professional terms, and extract key elements; the maintenance time points in the text records are matched with the inflection points of the status feature evolution in the original operating data of the offshore wind turbine SCADA system and CMS monitoring system in a time sequence correlation.

[0032] Furthermore, the step of mapping the preprocessed multi-source heterogeneous data to a unified embedding representation space using physical modeling and deep learning techniques includes:

[0033] The power generation, voltage, current, and blade speed of the offshore wind turbine are mapped to the operating intensity characteristics of the offshore wind turbine; the yaw angle, pitch angle, and nacelle position are mapped to the attitude control characteristics of the offshore wind turbine.

[0034] The high-frequency acceleration waveform signals of generator bearings and offshore wind turbine towers, tower sway signals and blade flapping vibration signals are mapped into offshore wind turbine structural health indicators.

[0035] Semantize external environmental meteorological data into characteristics of external excitation sources;

[0036] The pixel information in the monitoring video and inspection images is converted into blade speed, tower sway amplitude, crack width, coating peeling area and exposed length of submarine cable as visual structured indicators.

[0037] Semantize the maintenance and replacement actions in the full life cycle operation and maintenance records into an operation and maintenance knowledge weight vector of the time point of component wear degree restoration and maintenance experience;

[0038] The characteristics of offshore wind turbine working intensity, attitude control, structural health indicators and external excitation source are processed by min-max normalization. A fully connected embedding layer is used to map physical features of different dimensions to a unified high-dimensional linear vector space to construct a physical perception feature tensor with temporal continuity.

[0039] By fusing visual structured indicators with operation and maintenance knowledge weight vectors, a visual and textual semantic feature tensor that can reflect discrete events and morphological features is formed.

[0040] The physical perception feature tensor and the visual and textual semantic feature tensor are mapped to a unified embedding representation space, and sine / cosine position encoding is introduced for all feature vectors in the unified embedding representation space.

[0041] Furthermore, the construction of a feature interaction operator based on a multi-head cross-attention mechanism to perform feature decoupling and extract a fused feature vector reflecting the panoramic operating status of offshore wind turbines includes:

[0042] A feature interaction operator based on a multi-head cross-attention mechanism is constructed. The physical perception features, visual and textual semantic features embedded in the unified representation space are mapped into query tensors, key tensors and value tensors respectively through linear transformation matrices, forming a multi-source heterogeneous feature tensor group with alignment dimension.

[0043] By utilizing a multi-source heterogeneous feature tensor set, scaling dot product operations are performed to capture the implicit correlation strength between physical response and unstructured modalities, and a normalized cross-modal attention score matrix is ​​calculated.

[0044] We use a cross-modal attention scoring matrix to perform weighted mapping on the value tensor and introduce residual links to construct an enhanced physical feature vector that integrates environmental contextual information.

[0045] The enhanced physical feature vectors are mapped to the decoupled subspace by a linear projection matrix, separating independent feature dimensions, including structural service stiffness, external load environment, and short-term operational fluctuations.

[0046] The structural service stiffness is subjected to layer normalization processing to generate a normalized panoramic service state vector.

[0047] Furthermore, the variational Bayesian calibration framework, which incorporates the panoramic service status vector as a global observation term into an integrated physical information neural network, enables dynamic calibration of the physical parameters of the digital twin, generating an offshore wind turbine service status assessment report with uncertainty quantification support, including:

[0048] Based on the design parameters and historical service data of offshore wind turbines, the initial prior distribution of the physical parameters to be calibrated is obtained, and hierarchical hyperparameters are introduced to characterize the common laws and evolution uncertainties of structural properties under different service stages and variable operating conditions, and output the physical parameter space model to be optimized.

[0049] A proxy model integrating physical information neural networks is constructed, and the structural dynamic equilibrium equations are embedded as penalty terms in the loss function, outputting a likelihood function mapping operator from physical parameters to multimodal responses;

[0050] We introduce a parameterized variational distribution to approximate the true posterior distribution, and use the likelihood function mapping operator and the panoramic service state vector to construct an evidence lower bound loss function containing physical constraint terms, outputting the minimum divergence objective function to be solved.

[0051] The variational distribution is sampled using the reparameterization technique, the objective function is optimized using the stochastic gradient descent algorithm, the variational parameters are updated, and the converged posterior probability density function of the physical parameters is output.

[0052] From the posterior probability density function, the mean is extracted as the optimal calibration value of the physical parameters, and the variance is extracted as an uncertainty index to quantify the confidence of the digital twin's service status assessment. The parameter calibration set with confidence intervals is output.

[0053] The parameter calibration values ​​in the parameter calibration set are fed back to the finite element model for modal analysis. The measured residuals are compared and the dynamic monotonicity criterion is executed to eliminate solutions that violate physical logic. The final calibration parameters that pass the physical criteria are output.

[0054] The final calibration parameters are injected into the digital twin model to calculate the basic stiffness degradation rate and fatigue cumulative damage probability, generating an offshore wind turbine service status assessment report with uncertainty quantification support.

[0055] The second objective of this invention can be achieved by adopting the following technical solution:

[0056] A device for multi-source heterogeneous data fusion and condition assessment of offshore wind turbines, the device comprising:

[0057] The acquisition module is used to acquire offshore wind turbine operating status data, structural dynamic response data, external environmental meteorological data, visual morphology of key parts, and full life cycle operation and maintenance records to build a multi-source heterogeneous database for offshore wind turbines.

[0058] The preprocessing module is used to preprocess the multi-source heterogeneous database of offshore wind turbines. The data preprocessing includes data cleaning and spatiotemporal alignment.

[0059] The mapping module is used to map multi-source heterogeneous data after data preprocessing to a unified embedded representation space using physical modeling and deep learning techniques.

[0060] The fusion module is used to construct feature interaction operators based on multi-head cross-attention mechanism in a unified embedded representation space, perform feature decoupling, and extract fusion feature vectors that reflect the panoramic operation status of offshore wind turbines as panoramic service status vectors.

[0061] The evaluation module is used to introduce the panoramic service status vector as a global observation term into the variational Bayesian calibration framework of the integrated physical information neural network, so as to realize the dynamic calibration of the physical parameters of the digital twin and generate an offshore wind turbine service status evaluation report with uncertainty quantification support.

[0062] The third objective of this invention can be achieved by adopting the following technical solution:

[0063] A computer device includes a processor and a memory for storing processor-executable programs, wherein when the processor executes the program stored in the memory, it implements the above-described method for multi-source heterogeneous data fusion and status assessment of offshore wind turbines.

[0064] The fourth objective of this invention can be achieved by adopting the following technical solution:

[0065] A storage medium storing a program, which, when executed by a processor, implements the above-described method for multi-source heterogeneous data fusion and status assessment of offshore wind turbines.

[0066] The present invention has the following advantages over the prior art:

[0067] This invention achieves precise semantic alignment between physically perceived data and unstructured visual and textual information by constructing a unified embedding representation space based on contrastive learning. The method utilizes a cross-attention mechanism to capture implicit cross-modal relationships and perform feature decoupling, extracting a fused feature vector reflecting the panoramic operating status of the wind turbine. Based on this, a variational Bayesian calibration framework integrating physical information neural networks is introduced to achieve dynamic calibration and uncertainty quantification of the physical parameters of the digital twin. This effectively solves problems such as data silos, conservative assessments, and delayed decision-making, generating a service status assessment report with physical interpretability and uncertainty support. Attached Figure Description

[0068] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. 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 the structures shown in these drawings without creative effort.

[0069] Figure 1 This is a flowchart of the method for multi-source heterogeneous data fusion and status assessment of offshore wind turbines according to Embodiment 1 of the present invention.

[0070] Figure 2 This is a schematic diagram of the data acquisition principle of the offshore wind turbine data acquisition system according to Embodiment 1 of the present invention.

[0071] Figure 3 This is a schematic diagram illustrating the data preprocessing principle of the multi-source heterogeneous database for offshore wind turbines in Embodiment 1 of the present invention.

[0072] Figure 4 This is a schematic diagram illustrating the principle of mapping multi-source heterogeneous data to a unified embedded representation space in Embodiment 1 of the present invention.

[0073] Figure 5 This is a flowchart of the panoramic service status vector generation process in Embodiment 1 of the present invention.

[0074] Figure 6 This is a flowchart of the process for generating an offshore wind turbine service status assessment report according to Embodiment 1 of the present invention.

[0075] Figure 7 This is a structural block diagram of the offshore wind turbine multi-source heterogeneous data fusion and status assessment device according to Embodiment 2 of the present invention.

[0076] Figure 8 This is a structural block diagram of the computer device according to Embodiment 3 of the present invention. Detailed Implementation

[0077] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0078] Example 1:

[0079] like Figure 1 As shown in the figure, this embodiment provides a method for multi-source heterogeneous data fusion and status assessment of offshore wind turbines. The method includes the following steps:

[0080] S101. Acquire offshore wind turbine operating status data, structural dynamic response data, external environmental meteorological data, visual morphology of key parts, and full life cycle operation and maintenance records to construct a multi-source heterogeneous database for offshore wind turbines.

[0081] like Figure 2 As shown, this embodiment uses an offshore wind turbine data acquisition system to simultaneously collect offshore wind turbine operating status data, structural dynamic response data, external environmental meteorological data, visual morphology of key parts, and full life cycle operation and maintenance records, and constructs a multi-source heterogeneous database of offshore wind turbines that integrates low-frequency time series, high-frequency waveforms, continuous video streams, discrete high-definition images, and semi-structured text.

[0082] The offshore wind turbine data acquisition system in this embodiment includes an offshore wind turbine SCADA (Supervisory Control and Data Acquisition) system, a CMS (Condition Monitoring System) monitoring system, a nacelle wind measurement system, a surface meteorological buoy, industrial-grade visual monitoring equipment, UAV and underwater robot intelligent inspection equipment, and an operation and maintenance management system. The offshore wind turbine SCADA system is used to collect offshore wind turbine operating status data; the CMS monitoring system is used to collect structural dynamic response data; the nacelle wind measurement system and surface meteorological buoy are used to collect external environmental meteorological data; the industrial-grade visual monitoring equipment uses industrial-grade cameras and, together with the UAV and underwater robot intelligent inspection equipment, is used to collect visual morphology of key components; and the operation and maintenance management system is used to collect operation and maintenance records throughout the entire lifecycle.

[0083] The data acquisition process of the offshore wind turbine data acquisition system in this embodiment includes:

[0084] S1011. The offshore wind turbine SCADA system acquires low-frequency time-series data reflecting the operating status of the offshore wind turbine. The low-frequency time-series data includes blade speed, power generation, power output, voltage, current, pitch angle, yaw angle, and nacelle position. These data can reflect the macroscopic operating status of the unit.

[0085] S1012. Vibration sensors deployed on the top of the tower, nacelle, and key parts of the foundation through the CMS vibration monitoring system acquire high-frequency acceleration waveform signals of the generator bearings and offshore wind turbine towers, tower sway signals, and blade flapping and oscillation signals.

[0086] S1013. Obtain external environmental meteorological data acting on the offshore wind turbine structure through the engine room wind measurement system and sea surface meteorological buoys. The external environmental meteorological data includes parameters such as wind speed, wind direction, tide level, significant wave height, wave direction, ambient temperature and humidity, and atmospheric pressure.

[0087] S1014. By deploying industrial-grade cameras inside the nacelle, at the base of the tower, and on key structural surfaces of offshore wind turbines, continuous visual images are acquired to monitor information such as icing on the blade surface, tower swaying, and blade rotation, forming a real-time monitoring video stream database for key parts of offshore wind turbines.

[0088] S1015. Use drones to conduct close-range inspections and photography of the surfaces of offshore wind turbine blades, nacelles, and towers, and use underwater robots to conduct inspections and photography of the anti-corrosion coatings of offshore wind turbines and submarine cables below the sea surface, and obtain inspection images.

[0089] S1016. Based on the electronic archive files and paper log scan files in the operation and maintenance management system, collect the periodic maintenance records, fault repair orders, spare parts replacement history and text descriptions filled in by operation and maintenance personnel to obtain maintenance history information reflecting the entire service life cycle of the unit. This maintenance history information is a semi-structured text of the operation and maintenance process.

[0090] S102. Perform data preprocessing on the multi-source heterogeneous database of offshore wind turbines. The data preprocessing includes data cleaning and spatiotemporal alignment.

[0091] This embodiment addresses the signal interference and multi-source data sampling mismatch issues caused by complex marine environments. It performs signal repair and outlier removal, multi-scale denoising and feature dimensionality reduction, environmental bias correction and spatial mapping, visual image enhancement and keyframe alignment, and text semantic normalization and performance anchor matching. This eliminates the noise effects caused by salt spray corrosion, wave loads, and sensor heterogeneity. Finally, using UNIX timestamps as a common benchmark, it achieves precise alignment of physical time sequence, visual appearance, and operation and maintenance events on the spatiotemporal axis, constructing a standardized and consistent data sequence.

[0092] like Figure 3 As shown, the data preprocessing process of the multi-source heterogeneous database for offshore wind turbines in this embodiment includes:

[0093] S1021. For the operating status data, signal drift is identified and corrected through horizontal comparison of multiple wind turbines at the wind farm. Outlier data in the wind turbine operating status data is removed based on the 3-sigma criterion. Lagrange interpolation is used to complete the missing values ​​in the operating status data. The sampling frequency is 1Hz, and the data is gradually aligned on the UNIX timestamp as the reference time axis for wind turbine status assessment.

[0094] S1022. For the structural dynamic response data, wavelet packet denoising and variational mode decomposition are used to filter out low-frequency background noise caused by wave force and random bandwidth fluctuations of the mechanical transmission system. The energy operator is used to extract the cross-sectional features of the high-frequency acceleration waveform signal. The tower sway signal and blade flapping vibration signal are reduced to the same sampling frequency as the operating status data. Based on the signal acquisition trigger timestamp, the data is resampled and mapped to the reference time axis to achieve synchronous alignment between the dynamic response and the operating status.

[0095] S1023. Based on the external environmental meteorological data, correct the anemometer shading deviation and smooth the tide level and wave height to eliminate the instantaneous interference caused by breaking waves; map the external environmental meteorological data to the specific machine position coordinate system through spatial coordinate transformation, and at the same time use linear interpolation algorithm to compensate for the communication delay between sensors, and align the meteorological observation sequence to the offshore wind turbine status reference time axis.

[0096] S1024. For the video monitoring the nacelle, blades, and tower, run the image enhancement algorithm to reduce the impact of dense fog and rain / snow at sea, and use optical flow to remove the image jitter noise caused by wind turbine vibration; adopt key frame extraction technology, trigger high-frequency frame extraction according to the alarm time point in the offshore wind turbine SCADA system, and perform low-frequency monitoring at other times to achieve time alignment between visual features and alarm information.

[0097] S1025. For inspection images collected by drones and underwater robots, including images of crack detection on offshore wind turbine blades collected by drones, images of anti-corrosion coating peeling detection and images of exposed submarine cables collected by underwater robots, invalid images with excessive exposure, sea surface reflection, excessively dark seabed brightness or blurry focus are removed. Blue-green tone restoration is performed on images collected by underwater robots, and the images are aligned to the offshore wind turbine status reference time axis according to the image acquisition time.

[0098] S1026. For the full life cycle operation and maintenance records, use optical character recognition (OCR) technology to convert paper log scan files into electronic archive files, correct non-standard abbreviations and ambiguous professional terms, and extract key elements; perform time-series correlation matching between the maintenance time points of the text records and the inflection points of the status feature evolution in the original operating data of the offshore wind turbine SCADA system and CMS monitoring system.

[0099] S103. Using physical modeling and deep learning techniques, the multi-source heterogeneous data after data preprocessing is mapped to a unified embedding representation space.

[0100] This embodiment extracts physical sensing feature tensors reflecting unit operating conditions, structural health, and external excitation sources by dimensionally reorganizing operating parameters, vibration characteristics, and environmental loads. Simultaneously, it utilizes graph convolutional networks and natural language processing to transform visual images and operation and maintenance logs into quantitative damage indicators and expert prior knowledge, and constructs unstructured semantic feature tensors. Finally, a linear embedding layer projects physical and semantic features onto a high-dimensional unified embedding representation space, and introduces sine and cosine position encoding to preserve the temporal information of state evolution. This achieves deep fusion of signals with different dimensions in logical semantics, spatial dimension, and temporal position, providing a standardized input tensor set for multimodal state assessment models.

[0101] like Figure 4 As shown, the process of mapping multi-source heterogeneous data to a unified embedded representation space in this embodiment includes:

[0102] S1031. Map the power generation, voltage, current, blade speed and power generation of the offshore wind turbine to the working intensity characteristics of the offshore wind turbine; map the yaw angle, pitch angle and nacelle position to the attitude control characteristics of the offshore wind turbine to characterize the macroscopic operating state of the unit.

[0103] S1032. Map the high-frequency acceleration waveform signals of generator bearings and offshore wind turbine towers, tower sway signals and blade flapping vibration signals into offshore wind turbine structural health indicators to quantify the degree of damage evolution of components.

[0104] S1033. Semantize external environmental meteorological data into external excitation source characteristics to provide standardized boundary constraint inputs for structural response analysis.

[0105] S1034. The pixel information in the monitoring video and inspection images is converted into quantitative indicators such as blade rotation speed, tower sway amplitude, crack width, coating peeling area and exposed length of submarine cable through GCN technology, which are used as visual structured indicators.

[0106] S1035. Semantize the maintenance and replacement actions in the full life cycle operation and maintenance records into operation and maintenance knowledge weight vectors of the recovery time point of part wear degree and maintenance experience, and use them as expert prior correction factors in condition assessment.

[0107] S1036. The operating intensity characteristics, attitude control characteristics, structural health indicators, and external excitation source characteristics of offshore wind turbines are processed by min-max normalization. A fully connected embedding layer is used to map physical features of different dimensions to a unified high-dimensional linear vector space, thus constructing a physical perception feature tensor with temporal continuity. .

[0108] S1037. Integrate visual structured indicators with operation and maintenance knowledge weight vectors to form a visual and textual semantic feature tensor that can reflect discrete events and appearance features. .

[0109] S1038, Transform the physical perception feature tensor and visual and textual semantic feature tensors Mapped to a unified embedding representation space, sine / cosine positional encoding is introduced for all feature vectors within the unified embedding representation space to preserve the temporal evolution information of offshore wind turbine state evolution; thus, the original data of different modalities are transformed into a unified embedding tensor group with the same dimension, the same time axis reference and containing temporal position information.

[0110] S104. Within the unified embedding representation space, construct a feature interaction operator based on a multi-head cross-attention mechanism, perform feature decoupling, and extract a fusion feature vector that reflects the panoramic operation status of offshore wind turbines, which serves as the panoramic service status vector.

[0111] In this embodiment, a feature interaction operator based on the cross-attention mechanism is constructed within a unified embedding representation space. This operator is used to perform co-evolution and association weighting on multimodal features within the unified embedding representation space. By capturing the cross-modal implicit connections between physical signal waveforms and visual and textual modalities, a decoupled fusion feature vector reflecting the panoramic operating status of the wind turbine is generated. .

[0112] like Figure 5 As shown, the panoramic service state vector generation process in this embodiment includes:

[0113] S1041. Construct a feature interaction operator based on a multi-head cross-attention mechanism, and embed the physical perception features in the unified representation space. Visual and textual semantic features Through linear transformation matrix respectively , , Mapped to query tensor Key tensors Sum tensor This forms a multi-source heterogeneous feature tensor set with alignment dimension, as shown below:

[0114] ;

[0115] In this embodiment, , and It is a linear transformation weight matrix with learnable parameters in the multi-head cross-attention mechanism. Its construction logic is based on a preset feature mapping dimension, and through gradient iteration optimization during the model training phase, it forms an adaptive operator with specific semantic extraction capabilities. In terms of computational logic, these three matrices realize the spatial projection from the original heterogeneous data space to a unified aligned feature manifold: among them, Responsible for the physical sensing features of one-dimensional time series This is transformed into a search-oriented query vector Q, which represents the current physical state's retrieval needs for related information; and and Then, high-dimensional, unstructured visual and textual semantic features are simultaneously integrated. These are mapped to a key vector K and a value vector V, respectively aligned with the dimension of the matrix, to identify feature attributes and carry the substantive evaluation content. Through the linear transformation of these three matrices, the originally disparate physical observation domain and semantic visual domain are mapped to the same linear representation space, thereby eliminating the computational barriers between heterogeneous data at the mathematical level.

[0116] S1042. Using a multi-source heterogeneous feature tensor set, perform a scaled dot product operation to capture the implicit correlation strength between the physical response and the unstructured modality, and calculate the normalized cross-modal attention scoring matrix A, as shown in the following formula:

[0117] ;

[0118] in, The feature dimension of the key vector is used as a scaling factor to smooth the distribution of attention weights and avoid gradient saturation in numerical computation, thereby ensuring the stability of cross-modal data fusion.

[0119] S1043. A weighted mapping of the value tensor is performed using a cross-modal attention scoring matrix, and residual links are introduced to construct an enhanced physical feature vector that incorporates environmental contextual information. As shown in the following formula:

[0120] ;

[0121] S1044. The enhanced physical feature vectors are mapped to the decoupled subspace using a linear projection matrix, separating independent feature dimensions, as shown in the following equation:

[0122] ;

[0123] Linear decoupling is performed on the high-dimensional features after co-evolution to separate the components that reflect the structural service stiffness. External load environment and short-term fluctuations Independent feature dimensions.

[0124] S1045, Regarding structural service stiffness Execution layer normalization processing generates a normalized panoramic service state vector. As shown in the following formula:

[0125] ;

[0126] S105. Introduce the panoramic service status vector as a global observation term into the variational Bayesian calibration framework of the integrated physical information neural network to achieve dynamic calibration of the physical parameters of the digital twin and generate an offshore wind turbine service status assessment report with uncertainty quantification support.

[0127] This embodiment presents a panoramic service status vector. Introduced into the Bayesian update framework as a global observation, physical parameters such as pile-soil stiffness, damping ratio, and elastic modulus are first set based on design parameters. initial prior distribution Furthermore, hierarchical hyperparameters are introduced to characterize the uncertainties of the operating conditions; a surrogate model integrating physical information neural networks (PINN) is constructed, and the dynamic equilibrium equations are used as penalty terms to construct a likelihood function operator from the parameter space to the multimodal response. Introducing variational distribution The true posterior distribution is approximated by minimizing the KL divergence and optimizing the evidence lower bound (ELBO) loss function. The variational distribution is sampled using a reparameterization technique, and stochastic gradient descent is performed to output the converged posterior probability density function, from which the mean is extracted. As a calibration value and using variance The uncertainty is quantified; finally, the calibration value is fed back to the finite element model for physical consistency verification, and the final parameters are injected into the digital twin model to calculate the basic stiffness degradation rate and fatigue damage probability, generating an offshore wind turbine service status assessment report with uncertainty quantification support.

[0128] like Figure 6 As shown, the process for generating the offshore wind turbine service status assessment report in this embodiment includes:

[0129] S1051. Based on the design parameters and historical service data of offshore wind turbines, obtain the physical parameters to be calibrated. initial prior distribution The system introduces hierarchical hyperparameters to characterize the common laws and evolution uncertainties of structural properties under different service stages and variable working conditions, and outputs a physical parameter space model to be optimized. The physical parameter space model is essentially a parameter pool. All specific physical parameters (such as material stiffness, damping ratio, fatigue coefficient, etc.) involved in the calculation, calibration and optimization in subsequent steps are extracted from this physical parameter space model.

[0130] S1052. Construct a surrogate model integrating physical information neural networks, embedding the structural dynamics equilibrium equations as penalty terms into the loss function, and outputting a likelihood function mapping operator from physical parameters to multimodal responses. .

[0131] S1053, Introducing parameterized variational distribution To approximate the true posterior distribution, a likelihood function mapping operator is used to map the data to the panoramic service state vector. Construct an evidence lower bound loss function that includes physical constraint terms, and output the objective function to be solved: minimize the KL divergence. .

[0132] ;

[0133] : Measure the model's ability to represent the original observation data, ensuring a panoramic service state vector It can accurately reproduce input features and avoid information loss.

[0134] : Calculate the KL divergence between the approximate posterior distribution and the prior distribution, and use it as a regularization constraint. The distribution pattern helps prevent overfitting and improves generalization performance.

[0135] The physical residuals are calculated based on the wind turbine mechanism equations. By penalizing outputs that violate physical laws, the model evaluation results are forced to conform to the dynamic logic of the wind turbine.

[0136] S1054. The variational distribution is sampled using the reparameterization technique, the objective function is optimized using the stochastic gradient descent algorithm, the variational parameters are updated, and the converged posterior probability density function of the physical parameters is output.

[0137] ;

[0138] The variational parameters to be updated determine the panoramic service state vector. Distribution pattern.

[0139] The learning rate controls the step size for updating parameters.

[0140] : Objective function.

[0141] The gradient operator represents the rate of change of the objective function with respect to the variational parameters.

[0142] S1055. Extract the mean from the posterior probability density function. As the optimal calibration value for physical parameters, the variance is extracted. As an uncertainty indicator for quantifying confidence in the service status assessment of digital twins, it outputs a set of parameter calibrations with confidence intervals.

[0143] S1056. Calculate the parameter calibration values ​​from the parameter calibration set. The data is fed back to the finite element model for modal analysis. The measured residuals are compared with those of the model, and the dynamic monotonicity criterion is applied to eliminate solutions that violate physical logic. Finally, the calibration parameters that pass the physical criteria are output.

[0144] S1057. Inject the final calibration parameters into the digital twin model, calculate the basic stiffness degradation rate and fatigue cumulative damage probability, and generate an offshore wind turbine service status assessment report with uncertainty quantification support.

[0145] It should be noted that although the method operations of the above embodiments are described in a specific order, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. On the contrary, the described steps may be performed in a different order. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0146] Example 2:

[0147] like Figure 7 As shown in the figure, this embodiment provides a device for multi-source heterogeneous data fusion and status assessment of offshore wind turbines. The device includes an acquisition module 701, a preprocessing module 702, a mapping module 703, a fusion module 704, and an assessment module 705. The specific descriptions of each module are as follows:

[0148] The acquisition module 701 is used to acquire offshore wind turbine operating status data, structural dynamic response data, external environmental meteorological data, visual morphology of key parts and full life cycle operation and maintenance records, and to build a multi-source heterogeneous database of offshore wind turbines.

[0149] Preprocessing module 702 is used to preprocess data from a multi-source heterogeneous database of offshore wind turbines. The data preprocessing includes data cleaning and spatiotemporal alignment.

[0150] The mapping module 703 is used to map multi-source heterogeneous data that has undergone data preprocessing to a unified embedded representation space using physical modeling and deep learning techniques.

[0151] The fusion module 704 is used to construct a feature interaction operator based on a multi-head cross-attention mechanism in a unified embedded representation space, perform feature decoupling, and extract a fusion feature vector that reflects the panoramic operation status of offshore wind turbines as a panoramic service status vector.

[0152] The evaluation module 705 is used to introduce the panoramic service status vector as a global observation term into the variational Bayesian calibration framework of the integrated physical information neural network, so as to realize the dynamic calibration of the physical parameters of the digital twin and generate an offshore wind turbine service status evaluation report with uncertainty quantification support.

[0153] The specific implementation of each module in this embodiment can be found in Embodiment 1 above, and will not be repeated here. It should be noted that the device provided in this embodiment is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure can be divided into different functional modules to complete all or part of the functions described above.

[0154] Example 3:

[0155] This embodiment provides a computer device, such as... Figure 8 As shown, it includes a processor 802, a memory, and a network interface 803 connected via a system bus 801. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium 804 and internal memory 805. The non-volatile storage medium 804 stores an operating system, computer programs, and a database. The internal memory 805 provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. When the processor 802 executes the computer programs stored in the memory, it implements the offshore wind turbine multi-source heterogeneous data fusion and status assessment method of Embodiment 1 above, as follows:

[0156] This process involves acquiring operational status data, structural dynamic response data, external environmental meteorological data, visual morphology of key components, and full life-cycle maintenance records of offshore wind turbines to construct a multi-source heterogeneous database. The database undergoes preprocessing, including data cleaning and spatiotemporal alignment. Physical modeling and deep learning techniques are used to map the preprocessed heterogeneous data to a unified embedding representation space. Within this unified embedding representation space, a feature interaction operator based on a multi-head cross-attention mechanism is constructed to perform feature decoupling and extract a fused feature vector reflecting the panoramic operational status of the offshore wind turbine, serving as the panoramic service status vector. This panoramic service status vector is then introduced as a global observation term into a variational Bayesian calibration framework integrating physical information neural networks to achieve dynamic calibration of the physical parameters of the digital twin, generating an offshore wind turbine service status assessment report with uncertainty quantification support.

[0157] Example 4:

[0158] This embodiment provides a storage medium, which is a computer-readable storage medium, storing a computer program. When the program is executed by a processor, the processor executes the computer program stored in the memory to implement the offshore wind turbine multi-source heterogeneous data fusion and status assessment method of Embodiment 1 above, as follows:

[0159] This process involves acquiring operational status data, structural dynamic response data, external environmental meteorological data, visual morphology of key components, and full life-cycle maintenance records of offshore wind turbines to construct a multi-source heterogeneous database. The database undergoes preprocessing, including data cleaning and spatiotemporal alignment. Physical modeling and deep learning techniques are used to map the preprocessed heterogeneous data to a unified embedding representation space. Within this unified embedding representation space, a feature interaction operator based on a multi-head cross-attention mechanism is constructed to perform feature decoupling and extract a fused feature vector reflecting the panoramic operational status of the offshore wind turbine, serving as the panoramic service status vector. This panoramic service status vector is then introduced as a global observation term into a variational Bayesian calibration framework integrating physical information neural networks to achieve dynamic calibration of the physical parameters of the digital twin, generating an offshore wind turbine service status assessment report with uncertainty quantification support.

[0160] It should be noted that the computer-readable storage medium in this embodiment can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor device, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0161] In this embodiment, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used or combined with an instruction execution device, apparatus, or device. In this embodiment, the computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying a computer-readable program. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable storage medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use or combined with an instruction execution device, apparatus, or device. The computer program contained on the computer-readable storage medium can be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination thereof.

[0162] The computer-readable storage medium described above can be used to write computer programs for executing this embodiment in one or more programming languages ​​or combinations thereof. These programming languages ​​include object-oriented programming languages—such as Java, Python, and C++—and conventional procedural programming languages—such as C or similar programming languages. The program can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0163] In summary, this invention achieves precise semantic alignment between physically perceived data and unstructured visual and textual information by constructing a unified embedding representation space based on contrastive learning. This method utilizes a cross-attention mechanism to capture implicit cross-modal relationships and perform feature decoupling, extracting a fused feature vector reflecting the panoramic operating status of the wind turbine. Based on this, a variational Bayesian calibration framework integrating physical information neural networks is introduced to achieve dynamic calibration and uncertainty quantification of the physical parameters of the digital twin. This effectively solves problems such as data silos, conservative assessments, and delayed decision-making, generating a service status assessment report with physical interpretability and uncertainty support.

[0164] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope disclosed in the present invention, based on the technical solution and inventive concept of the present invention, shall fall within the scope of protection of the present invention.

Claims

1. A method for fusing multi-source heterogeneous data and assessing the condition of offshore wind turbines, characterized in that, The method includes: Acquire offshore wind turbine operating status data, structural dynamic response data, external environmental meteorological data, visual morphology of key components, and full life cycle operation and maintenance records to construct a multi-source heterogeneous database for offshore wind turbines; Data preprocessing is performed on a multi-source heterogeneous database of offshore wind turbines, including data cleaning and spatiotemporal alignment. Using physical modeling and deep learning techniques, multi-source heterogeneous data that has undergone data preprocessing is mapped to a unified embedded representation space; Within a unified embedding representation space, a feature interaction operator based on a multi-head cross-attention mechanism is constructed to perform feature decoupling and extract a fused feature vector reflecting the panoramic operation status of offshore wind turbines, serving as the panoramic service state vector. This process includes: constructing a feature interaction operator based on a multi-head cross-attention mechanism, mapping the physical perception features, visual and textual semantic features in the unified embedding representation space to query tensors, key tensors, and value tensors respectively through linear transformation matrices, forming a multi-source heterogeneous feature tensor group with alignment dimensions; using the multi-source heterogeneous feature tensor group, performing scaling dot product operations to capture the implicit correlation strength between physical responses and unstructured modes, and calculating a normalized cross-modal attention scoring matrix; using the cross-modal attention scoring matrix to perform weighted mapping on the value tensors and introducing residual links to construct an enhanced physical feature vector that integrates environmental context information; mapping the enhanced physical feature vector to a decoupled subspace through a linear projection matrix to separate independent feature dimensions, including structural service stiffness, external load environment, and short-term operational fluctuations; and performing layer normalization on the structural service stiffness to generate a normalized panoramic service state vector. This paper introduces the panoramic service state vector as a global observation term into a variational Bayesian calibration framework integrated with a physical information neural network to achieve dynamic calibration of the physical parameters of a digital twin. This generates a service status assessment report for offshore wind turbines with uncertainty quantification support. The process includes: obtaining the initial prior distribution of the physical parameters to be calibrated based on the offshore wind turbine design parameters and historical service data; introducing hierarchical hyperparameters to characterize the common laws and evolutionary uncertainties of structural properties under different service stages and varying operating conditions; outputting a physical parameter space model to be optimized; constructing a surrogate model integrated with the physical information neural network; embedding the structural dynamic equilibrium equation as a penalty term into the loss function; and outputting a likelihood function mapping operator from physical parameters to multimodal responses. Finally, it introduces a parameterized variational distribution to approximate the true posterior distribution and uses the likelihood function mapping operator and the panoramic service state vector to construct an evidence lower bound loss containing physical constraint terms. The function outputs the objective function to minimize the divergence. It samples the variational distribution using reparameterization techniques, optimizes the objective function using stochastic gradient descent, updates the variational parameters, and outputs the converged posterior probability density function of the physical parameters. From the posterior probability density function, it extracts the mean as the optimal calibration value of the physical parameters and the variance as an uncertainty index to quantify the confidence in the service status assessment of the digital twin, outputting a parameter calibration set with confidence intervals. The parameter calibration values ​​in the parameter calibration set are fed back to the finite element model for modal analysis, compared with the measured residuals, and a dynamic monotonicity criterion is applied to eliminate solutions that violate physical logic, outputting the final calibration parameters verified by physical criteria. The final calibration parameters are then injected into the digital twin model to calculate the foundation stiffness degradation rate and fatigue cumulative damage probability, generating a service status assessment report for the offshore wind turbine with uncertainty quantification support.

2. The method for multi-source heterogeneous data fusion and condition assessment of offshore wind turbines according to claim 1, characterized in that, The acquisition of offshore wind turbine operating status data, structural dynamic response data, external environmental meteorological data, visual morphology of key components, and full life-cycle operation and maintenance records includes: The offshore wind turbine data acquisition system synchronously collects offshore wind turbine operating status data, structural dynamic response data, external environmental meteorological data, visual morphology of key components, and full life cycle operation and maintenance records. It constructs a multi-source heterogeneous database of offshore wind turbines that integrates low-frequency time series, high-frequency waveforms, continuous video streams, discrete high-definition images, and semi-structured text. The offshore wind turbine data acquisition system includes an offshore wind turbine SCADA system, a CMS monitoring system, a nacelle wind measurement system, sea surface meteorological buoys, industrial-grade visual monitoring equipment, intelligent inspection equipment for drones and underwater robots, and an operation and maintenance management system.

3. The method for multi-source heterogeneous data fusion and condition assessment of offshore wind turbines according to claim 2, characterized in that, The offshore wind turbine data acquisition system synchronously collects offshore wind turbine operating status data, structural dynamic response data, external environmental meteorological data, visual morphology of key components, and full life-cycle operation and maintenance records, including: The offshore wind turbine SCADA system acquires low-frequency time-series data reflecting the operating status of the offshore wind turbine. The low-frequency time-series data includes blade speed, power generation, power output, voltage, current, pitch angle, yaw angle, and nacelle position. Vibration sensors deployed on the top of the tower, nacelle, and key parts of the foundation through the CMS vibration monitoring system acquire high-frequency acceleration waveform signals, tower sway signals, and blade flapping and oscillation signals of the generator bearings and offshore wind turbine towers. The external environmental meteorological data acting on the offshore wind turbine structure are obtained through the wind measurement system in the engine room and the meteorological buoy on the sea surface. The external environmental meteorological data includes wind speed, wind direction, tide level, significant wave height, wave direction, ambient temperature and humidity, and atmospheric pressure. By deploying industrial-grade cameras inside the nacelle, at the base of the tower, and on key structural surfaces of offshore wind turbines, continuous visual images of blade surface icing, tower swaying, and blade rotation are acquired, forming a real-time monitoring video stream database for key parts of offshore wind turbines. The inspection images were obtained by using drones to conduct close-up inspections and photography of the surfaces of offshore wind turbine blades, nacelles and towers, and by using underwater robots to inspect and photograph the anti-corrosion coatings of offshore wind turbines and submarine cables below the sea surface. Based on electronic archives and scanned paper logs in the operation and maintenance management system, we collect periodic maintenance records, fault repair orders, spare parts replacement history and written descriptions filled out by operation and maintenance personnel to obtain maintenance history information reflecting the entire service life cycle of the unit.

4. The method for multi-source heterogeneous data fusion and condition assessment of offshore wind turbines according to claim 3, characterized in that, The data preprocessing of the multi-source heterogeneous database of offshore wind turbines includes: For the operational status data, signal drift is identified and corrected through horizontal comparison of multiple wind turbines at the wind farm, outlier data in the wind turbine operational status data is removed, and missing values ​​in the operational status data are filled in using the Lagrange interpolation method. The data is then gradually aligned on the timestamps to serve as the benchmark time axis for wind turbine status assessment. For the structural dynamic response data, wavelet packet denoising and variational mode decomposition are used to filter out low-frequency background noise caused by wave force and random bandwidth fluctuations of the mechanical transmission system. The cross-sectional features of the high-frequency acceleration waveform signal are extracted using energy operators. The tower sway signal and blade flapping vibration signal are reduced to a sampling frequency consistent with the operating status data. Based on the signal acquisition trigger timestamp, the data is resampled and mapped to the reference time axis to achieve synchronous alignment between dynamic response and operating status. For external environmental meteorological data, the anemometer shading bias is corrected, and the tide level and wave height are smoothed to eliminate the instantaneous interference caused by breaking waves. The external environmental meteorological data is mapped to the specific machine position coordinate system through spatial coordinate transformation. At the same time, the linear interpolation algorithm is used to compensate for the communication delay between sensors and the meteorological observation sequence is aligned to the offshore wind turbine status reference time axis. For the video monitoring the nacelle, blades, and tower, image enhancement algorithms are run, and optical flow is used to remove image jitter noise caused by wind turbine vibration; key frame extraction technology is adopted to trigger high-frequency frame extraction based on the alarm time point in the offshore wind turbine SCADA system, and low-frequency monitoring is performed at other times to achieve time alignment between visual features and alarm information. For inspection images collected by drones and underwater robots, invalid images that are overexposed at sea, reflective of the sea surface, too dark at the seabed, or blurry in focus are removed. For images collected by underwater robots, blue-green tones are restored, and the images are aligned to the reference time axis of offshore wind turbine status according to the image acquisition time. For the entire life cycle operation and maintenance records, optical character recognition technology is used to convert paper log scan files into electronic archive files, correct non-standard abbreviations and ambiguous professional terms, and extract key elements; the maintenance time points in the text records are matched with the inflection points of the status feature evolution in the original operating data of the offshore wind turbine SCADA system and CMS monitoring system in a time sequence correlation.

5. The method for multi-source heterogeneous data fusion and condition assessment of offshore wind turbines according to claim 3, characterized in that, The process of mapping multi-source heterogeneous data to a unified embedding representation space using physical modeling and deep learning techniques includes: The power generation, voltage, current, blade speed, and power generation of offshore wind turbines are mapped to the operating intensity characteristics of offshore wind turbines; the yaw angle, pitch angle, and nacelle position are mapped to the attitude control characteristics of offshore wind turbines. The high-frequency acceleration waveform signals of generator bearings and offshore wind turbine towers, tower sway signals and blade flapping vibration signals are mapped into offshore wind turbine structural health indicators. Semantize external environmental meteorological data into characteristics of external excitation sources; The pixel information in the monitoring video and inspection images is converted into blade speed, tower sway amplitude, crack width, coating peeling area and exposed length of submarine cable as visual structured indicators. Semantize the maintenance and replacement actions in the full life cycle operation and maintenance records into an operation and maintenance knowledge weight vector of the time point of component wear degree restoration and maintenance experience; The characteristics of offshore wind turbine working intensity, attitude control, structural health indicators and external excitation source are processed by min-max normalization. A fully connected embedding layer is used to map physical features of different dimensions to a unified high-dimensional linear vector space to construct a physical perception feature tensor with temporal continuity. By fusing visual structured indicators with operation and maintenance knowledge weight vectors, a visual and textual semantic feature tensor that can reflect discrete events and morphological features is formed. The physical perception feature tensor and the visual and textual semantic feature tensor are mapped to a unified embedding representation space, and sine / cosine position encoding is introduced for all feature vectors in the unified embedding representation space.

6. A device for multi-source heterogeneous data fusion and condition assessment of offshore wind turbines, characterized in that, The device includes: The acquisition module is used to acquire offshore wind turbine operating status data, structural dynamic response data, external environmental meteorological data, visual morphology of key parts, and full life cycle operation and maintenance records to build a multi-source heterogeneous database for offshore wind turbines. The preprocessing module is used to preprocess the multi-source heterogeneous database of offshore wind turbines. The data preprocessing includes data cleaning and spatiotemporal alignment. The mapping module is used to map multi-source heterogeneous data after data preprocessing to a unified embedded representation space using physical modeling and deep learning techniques. The fusion module is used to construct a feature interaction operator based on a multi-head cross-attention mechanism within a unified embedding representation space, perform feature decoupling, and extract a fused feature vector reflecting the panoramic operation status of offshore wind turbines as a panoramic service status vector. This includes: constructing a feature interaction operator based on a multi-head cross-attention mechanism, mapping the physical perception features, visual and textual semantic features in the unified embedding representation space to query tensors, key tensors, and value tensors respectively through linear transformation matrices, forming a multi-source heterogeneous feature tensor set with alignment dimensions; and using the multi-source heterogeneous feature tensor set to perform scaling point... The product operation captures the implicit correlation strength between the physical response and the unstructured modes, and calculates the normalized cross-modal attention scoring matrix. The cross-modal attention scoring matrix is ​​used to perform a weighted mapping on the value tensor, and residual links are introduced to construct an enhanced physical feature vector that integrates environmental context information. The enhanced physical feature vector is mapped to the decoupled subspace through a linear projection matrix to separate independent feature dimensions, which include structural service stiffness, external load environment, and short-term operational fluctuations. Layer normalization is performed on the structural service stiffness to generate a normalized panoramic service state vector. The evaluation module incorporates the panoramic service state vector as a global observation term into a variational Bayesian calibration framework integrated with a physical information neural network. This enables dynamic calibration of the physical parameters of the digital twin, generating a service state evaluation report for offshore wind turbines with uncertainty quantification support. The process includes: obtaining the initial prior distribution of the physical parameters to be calibrated based on the offshore wind turbine design parameters and historical service data; introducing hierarchical hyperparameters to characterize the commonalities and evolutionary uncertainties of structural properties under different service stages and varying operating conditions; outputting a physical parameter space model to be optimized; constructing a surrogate model integrated with a physical information neural network, embedding the structural dynamic equilibrium equation as a penalty term into the loss function, and outputting a likelihood function mapping operator from physical parameters to multimodal responses; introducing a parameterized variational distribution to approximate the true posterior distribution; and using the likelihood function mapping operator and the panoramic service state vector to construct an evidence-based model containing physical constraint terms. The bounded loss function is used to output the objective function that minimizes the divergence to be solved. The variational distribution is sampled using a reparameterization technique, and the objective function is optimized using a stochastic gradient descent algorithm to update the variational parameters, outputting the converged posterior probability density function of the physical parameters. From the posterior probability density function, the mean is extracted as the optimal calibration value of the physical parameters, and the variance is extracted as an uncertainty index to quantify the confidence of the digital twin's service status assessment, outputting a parameter calibration set with confidence intervals. The parameter calibration values ​​in the parameter calibration set are fed back to the finite element model for modal analysis, compared with the measured residuals, and the dynamic monotonicity criterion is executed to eliminate solutions that violate physical logic, outputting the final calibration parameters that pass the physical criteria verification. The final calibration parameters are injected into the digital twin model to calculate the foundation stiffness degradation rate and fatigue cumulative damage probability, generating an offshore wind turbine service status assessment report with uncertainty quantification support.

7. A computer device comprising a processor and a memory for storing a processor-executable program, characterized in that, When the processor executes the program stored in the memory, it implements the offshore wind turbine multi-source heterogeneous data fusion and status assessment method according to any one of claims 1-5.

8. A storage medium storing a program, characterized in that, When the program is executed by the processor, it implements the method for multi-source heterogeneous data fusion and status assessment of offshore wind turbines as described in any one of claims 1-5.

Images