Wind turbine generator tower health state assessment method and device based on multi-source data fusion and dynamic baseline modeling

Abstract

The invention discloses a wind turbine generator tower health state assessment method and device based on multi-source data fusion and dynamic baseline modeling. The method comprises the following steps: firstly, constructing a tower drum theoretical dynamics model to form a physical prior knowledge base, and then designing a multi-stage cascade lightweight model; adapting the low-frequency vibration characteristics of the tower drum through a physically constrained depth separable convolutional network and extracting interpretable characteristics; semantic fusion of multi-source heterogeneous data is realized by using a multi-head cross attention mechanism initialized by physical parameters; and a cascade network is adopted to map the fusion features into physical damage sensitive indexes, and four-level health states are output through a fuzzy inference device. And further compressing the model through a knowledge distillation technology, and finally deploying the model on edge equipment to realize real-time online evaluation. According to the method, a physical mechanism and data driving are deeply fused, the problems that an existing method is poor in physical interpretability, weak in small sample generalization and difficult in edge deployment are solved, and high-precision, traceable and low-delay intelligent evaluation of the state of the tower drum is achieved.

Description

Technical Field

[0001] This invention belongs to the field of wind power equipment condition monitoring and health management technology, specifically involving a method and device for assessing the health status of wind turbine towers based on multi-source data fusion and dynamic baseline modeling. Background Technology

[0002] As a critical load-bearing structure supporting the weight of the entire nacelle and blades, the wind turbine tower is subjected to complex alternating loads and harsh environments over long periods, making it prone to structural health issues such as fatigue damage, bolt loosening, and localized buckling. Accurate, real-time, and interpretable assessment of the tower's health status is a core technology for ensuring the safe operation of wind farms, enabling predictive maintenance, and reducing operation and maintenance costs.

[0003] Currently, the health status assessment of wind turbine towers mainly relies on two technical approaches: physical model-based methods and data-driven methods.

[0004] The physical model-based method mainly establishes a finite element model or a simplified dynamic model of the tower, and combines measured data (such as vibration and strain) to correct the model and identify parameters, thereby evaluating the changes in structural stiffness, damping and other characteristics. This type of method has clear physical meaning, but has significant limitations: (1) The establishment and correction of high-fidelity models are highly dependent on expert knowledge, the calculation is complex, and it is difficult to achieve real-time online evaluation; (2) The damage modes of actual structures are diverse, and simplified models often cannot accurately represent complex damage mechanisms and evolution processes, resulting in insufficient evaluation accuracy; (3) The model is heavily dependent on precise structural parameters and boundary conditions, is sensitive to initial errors, and has poor robustness.

[0005] Data-driven approaches, particularly deep learning methods, have gained widespread attention in recent years for automatically learning characteristic patterns of health and injury states from massive amounts of monitoring data. Examples include using convolutional neural networks to process vibration signals or leveraging long short-term memory networks to learn temporal dependencies. These methods have shown potential in damage pattern recognition, but they also face serious challenges: (1) Black box characteristics: The model decision-making process lacks physical interpretability, and it is difficult for maintenance personnel to understand the physical causes of the evaluation results, which limits its application in critical safety decisions; (2) Small sample and imbalanced sample problems: There are extremely few samples of major damage to wind turbine structures, while normal samples account for the vast majority, which makes data-driven models prone to overfitting and have weak generalization ability; (3) Insufficient domain adaptability: Models trained for specific units are difficult to be directly transferred to other units with different design parameters and operating environments, and the cost of repeated modeling is high; (4) Insufficient adaptability to tower structure characteristics: General network architectures (such as standard convolutional neural networks) have not been optimized for the unique structural and response characteristics of towers, such as large aspect ratio, low-frequency vibration dominance, and multi-source signal coupling, and the efficiency and specificity of feature extraction need to be improved.

[0006] Furthermore, existing technical solutions mostly focus on centralized analysis and evaluation in the cloud or on servers, resulting in high data transmission bandwidth pressure and high evaluation latency, making it difficult to meet the needs of real-time monitoring and rapid early warning of tower health status. Although edge computing has been introduced to reduce latency, complex deep learning models are difficult to run efficiently on resource-constrained edge devices.

[0007] Therefore, the current field of wind turbine tower health status assessment urgently needs a new assessment method and device that can integrate the advantages of physical mechanisms and data-driven approaches, possess physical interpretability, small sample generalization ability, lightweight edge deployment characteristics, and can deeply adapt to the dynamic characteristics of tower structures, so as to achieve highly reliable, efficient, and practical online intelligent monitoring and early warning. Summary of the Invention

[0008] The technical problem to be solved by this invention is to overcome the technical defects of existing wind turbine tower health assessment methods, such as poor physical interpretability, weak generalization ability with small samples, and difficulty in edge deployment. The invention provides a complete technical solution that can deeply integrate physical mechanisms and data-driven approaches to achieve interpretable, lightweight, and adaptive assessment.

[0009] In a first aspect, embodiments of this application provide a method for assessing the health status of wind turbine towers based on multi-source data fusion and dynamic baseline modeling, the method comprising: S1. Collect vibration acceleration signals, structural strain data and environmental parameters of the wind turbine tower, and perform standardized processing. S2. Based on the material properties, geometric dimensions and boundary conditions of the tower, establish a theoretical dynamic model, calculate the theoretical natural frequencies, mode shapes and strain distribution at key locations of the tower under healthy reference conditions, and construct a physical prior knowledge base. S3. Construct multiple parallel lightweight feature extraction sub-models to process vibration acceleration signals, structural strain data, and environmental parameter data streams, respectively. Among them, the sub-model for processing vibration acceleration signals adopts a depthwise separable convolutional structure and adaptively configures the temporal receptive field of the convolutional kernel according to the theoretical natural frequency range to match the low-frequency vibration characteristics of the tower. Physical priors are injected during the training process of this sub-model. S4. The features output by each sub-model are concatenated and input into the lightweight adaptive fusion module. This module adopts a multi-head cross-attention mechanism, in which the query vector of a set of attention heads is initialized by the feature parameters in the physical prior knowledge base, so that the fusion process is guided by physical semantics, calculates the correlation weight between features from different data sources and completes the fusion. S5. Input the output fused features into a cascaded quantization network model, which includes: a first lightweight fully connected layer for generating damage-sensitive indicators related to the physical state of the tower; and a second lightweight fuzzy inferencer for mapping the damage-sensitive indicators to multiple predefined health status levels based on a hybrid membership function composed of a Gaussian function and a trapezoidal function. S6. The cascaded quantization network model is trained in stages using source domain data, and the model is compressed using knowledge distillation technology. S7. Deploy the compressed model on the edge computing device at the tower site, connect to the real-time monitoring data stream, and output the health status level of the tower.

[0010] Secondly, embodiments of this application provide a wind turbine tower health status assessment device based on multi-source data fusion and dynamic baseline modeling, applied to the method described in the first aspect, the device comprising: The data acquisition and preprocessing module is used to acquire vibration acceleration signals, structural strain data and environmental parameters of the wind turbine tower, and perform standardized processing. The physical prior knowledge base construction module is used to establish a theoretical dynamic model based on the material properties, geometric dimensions and boundary conditions of the tower, calculate the theoretical natural frequencies, mode shapes and strain distributions at key locations of the tower under healthy reference conditions, and construct the physical prior knowledge base. The parallel lightweight feature extraction module is used to build and run multiple parallel lightweight feature extraction sub-models to process vibration acceleration signals, structural strain data, and environmental parameter data streams, respectively. Among them, the sub-model for processing vibration acceleration signals adopts a depthwise separable convolutional structure and adaptively configures the temporal receptive field of the convolutional kernel according to the theoretical natural frequency range to match the low-frequency vibration characteristics of the tower. This module injects information from the physical prior knowledge base during the training process. The physics-guided multi-source fusion module is used to concatenate the features output by each sub-model and process them using a multi-head cross-attention mechanism. The query vector of a set of attention heads is initialized by the feature parameters in the physics prior knowledge base, so that the fusion process is guided by physical semantics, calculates the correlation weights between features from different data sources and completes the fusion. A cascaded health status quantization module is used to input the fusion features output by the physical guidance multi-source fusion module into a cascaded quantization network model for processing. The cascaded quantization network model includes, in sequence: a first lightweight fully connected layer, used to generate damage-sensitive indicators related to the physical state of the tower; and a second lightweight fuzzy inferencer, used to map the damage-sensitive indicators to multiple predefined health status levels based on a hybrid membership function composed of a Gaussian function and a trapezoidal function. The model optimization and compression module is used to perform phased prediction training on the cascaded health status quantification module using source domain data, and to compress the model using knowledge distillation technology. The edge deployment and online assessment module is used to deploy the compressed model on the edge computing device at the tower site, access the real-time monitoring data stream, and control the device to operate online to output the health status level of the tower.

[0011] Thirdly, embodiments of this application provide an electronic device, including: processor; Memory used to store processor-executable instructions; The processor is used to implement the wind turbine tower health status assessment method based on multi-source data fusion and dynamic baseline modeling as described in the first aspect when executing the instructions.

[0012] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program that instructs a device to perform the wind turbine tower health status assessment method based on multi-source data fusion and dynamic baseline modeling as described in the first aspect.

[0013] By constructing a physical prior knowledge base and injecting it into the deep learning model in various ways (spatial attention, frequency domain regularization, query vector initialization), the feature learning, fusion, and decision-making processes are all guided by physical mechanisms. The final output can be directly correlated with specific physical damage parameters (such as frequency shift and strain anomalies). The introduction of physical priors provides the model with strong domain knowledge, reducing its dependence on large amounts of labeled data. Combining domain adaptive training and knowledge distillation techniques enables the model to quickly transfer and adapt to different units, improving the evaluation accuracy when target unit samples are scarce. A depthwise separable convolutional structure specifically designed based on the low-frequency vibration characteristics of the tower effectively captures long-period vibration modes. A physical-guided fusion mechanism for multi-source heterogeneous data is adopted to fully explore the deep correlation between vibration, strain, and environmental parameters. Through lightweight model design (depthwise separable convolution, lightweight attention head) and knowledge distillation compression, the model complexity and computational overhead are significantly reduced, enabling the complete intelligent evaluation process to run in real time on resource-constrained edge devices, meeting the requirements for low-latency and high-efficiency online monitoring. This invention achieves the complementary advantages of physical models and data-driven models, providing a high-precision, highly interpretable, and predictive maintenance-enabled intelligent health status assessment system for wind turbine towers. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of a method for assessing the health status of wind turbine towers based on multi-source data fusion and dynamic baseline modeling, provided as an embodiment of this application.

[0015] Figure 2 The architecture diagram of the wind turbine tower health status assessment device based on multi-source data fusion and dynamic baseline modeling provided in this application.

[0016] Figure 3 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0018] It should be noted that in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0019] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] Example 1

[0021] Figure 1 This is a schematic flowchart illustrating a method for assessing the health status of wind turbine towers based on multi-source data fusion and dynamic baseline modeling, provided as an embodiment of this application. Figure 1 As shown, a method for assessing the health status of wind turbine towers based on multi-source data fusion and dynamic baseline modeling includes: S1. Collect vibration acceleration signals, structural strain data, and environmental parameters of the wind turbine tower and perform standardized processing. The health status assessment of the wind turbine tower begins with real-time monitoring and effective preprocessing of key physical quantities, which forms the data foundation for building a reliable assessment model. To comprehensively capture the structural dynamic response of the tower and its service environment, this invention uses the following three types of sensor networks for synchronous data acquisition: Vibration acceleration signal acquisition system: High-precision piezoelectric or MEMS accelerometer arrays are arranged along the horizontal X and Y directions (mutually orthogonal) at key locations on the tower structure, typically including the tower top (near the bottom of the nacelle), the middle of the tower body, and near the doorway area. This system continuously acquires vibration acceleration time-history signals at a sampling frequency of no less than 100Hz to effectively cover the tower's main low-order natural frequencies (usually concentrated in the range of 0.1Hz to 2Hz) and high-frequency vibration components.

[0022] Structural strain data acquisition system: Fiber optic strain sensors or resistance strain gauge networks are installed in anticipated high-stress concentration areas, such as flange connections, locations of abrupt changes in tower wall thickness, and near the foundation ring. These sensors are primarily used to monitor the local dynamic strain of the tower under wind loads to reveal potential fatigue damage or loosening of connections.

[0023] Environmental parameter acquisition system: This system collects external environmental data that directly affects the tower load, including but not limited to instantaneous wind speed and direction measured by the anemometer on top of the nacelle, pitch angle information provided by the pitch control system, and ambient temperature and humidity provided by atmospheric sensors. These parameters provide a basis for subsequent operating condition identification and load standardization.

[0024] To ensure the fusion and analysis of monitoring data from different physical mechanisms and dimensions, a rigorous data synchronization and preprocessing process must be implemented. First, data synchronization and alignment are crucial. All sensor timestamps are uniformly calibrated using the same GPS timing module or high-precision network time protocol to ensure that data streams from vibration, strain, and environmental samples at different sampling rates are aligned on the same time base. Next, signal cleaning and filtering are performed. The raw vibration and strain signals undergo necessary preprocessing, including: trend term removal to eliminate long-term trends caused by temperature drift or sensor zero-point drift; outlier removal using statistical methods (such as based on the 3σ principle) or engineering experience thresholds to remove outliers caused by sensor malfunctions or electromagnetic interference; bandpass filtering to apply a bandpass filter (e.g., 0.1-10Hz) matching the tower's primary frequency band of interest to the vibration acceleration signal, preserving the effective structural response and suppressing high-frequency noise and low-frequency drift; and standardization to eliminate the impact of differences in sensor dimensions and magnitudes on model training. All cleaned monitoring data sequences are then standardized. For vibration acceleration and strain signals, a standardization method with zero mean and unit variance is usually used.

[0025] Through the above steps, multi-source standardized monitoring data with time alignment, high signal-to-noise ratio, and uniform scale were obtained, providing high-quality input for subsequent construction of physical prior knowledge base and training of deep learning models.

[0026] S2. Based on the tower's material properties, geometric dimensions, and boundary conditions, a theoretical dynamic model is established to calculate the theoretical natural frequencies, mode shapes, and strain distributions at key locations of the tower under healthy baseline conditions, thus constructing a physical prior knowledge base. To effectively incorporate the tower's structural mechanism knowledge into the subsequent data-driven intelligent evaluation model, this invention constructs a physical prior knowledge base based on a parametric dynamic model, providing quantifiable theoretical physical characteristics under healthy baseline conditions for the neural network. The specific steps for constructing the physical prior knowledge base and effectively injecting it into the lightweight feature extraction process are as follows: S2.1. Based on the tower's material density, elastic modulus, geometric dimensions, and bottom fixed boundary conditions, establish a parametric continuum dynamics model conforming to either the Euler-Bernoulli beam theory or the Timoshenko beam theory. The Euler-Bernoulli beam theory is a classical beam theory that assumes the beam's cross-section remains planar and perpendicular to the neutral axis during bending deformation, neglecting the effects of shear deformation and rotational inertia. It is suitable for structures with a large slenderness ratio. The Timoshenko beam theory is a more accurate beam theory that considers shear deformation and rotational inertia. It is suitable for short, thick beams or high-frequency vibration analysis. A parametric continuum dynamics model refers to a mechanical model whose shape, material, and boundaries are defined and controlled by a set of adjustable parameters (such as dimensions, density, and elastic modulus).

[0027] In this embodiment of the invention, based on the typical structural characteristics of the tower (large height-to-diameter ratio, primarily characterized by bending deformation), the Euler-Bernoulli Beam theory, which offers higher computational efficiency and meets engineering accuracy requirements, is preferentially selected to establish a simplified model. For specific tower designs with large wall thickness or requiring consideration of shear deformation, the Timoshenko Beam theory model can also be used. Input parameters: The tower is made of Q355 steel with a material density ρ = 7850 kg / m³ and an elastic modulus E = 210 GPa. The total height of the tower is 100 meters. To simplify calculations, it is discretized into 10 segments of uniformly shaped conical beams, each 10 meters high, with a bottom diameter of 4.2 meters and a top diameter of 3.0 meters. The wall thickness decreases linearly along the height. Based on the above parameters and the bottom fixed boundary conditions (i.e., zero displacement and rotation at the tower bottom), a parametric finite element model of the tower is established using commercial finite element software (such as ANSYS or ABAQUS) or by writing a program. This model can essentially be categorized as Timoshenko beam theory, balancing accuracy and computational efficiency.

[0028] S2.2. Using the parameterized continuum dynamics model, the first M natural frequencies of the tower structure's lateral vibration under a damage-free baseline state, the corresponding mode shape functions, and the theoretical normal strain distribution at the strain sensor placement locations are analytically solved. Based on the model established in the previous step, the fingerprint characteristics of the tower in a healthy, intact state are obtained through mathematical solutions, including its inherent vibration frequencies, mode shapes, and strain response modes at key locations. For simple models, exact solutions (analytical solutions) can be obtained directly through differential equations; for complex models, approximate solutions are usually obtained using numerical methods such as the finite element method. The natural frequencies are specific frequencies of the structural system during free vibration, determined by the system's mass and stiffness distribution, and are inherent properties of the structure. Mode shapes are the structural vibration displacement shapes corresponding to each natural frequency. The theoretical normal strain distribution refers to the spatial distribution of the strain (deformation per unit length) along the axial (normal) direction on the structural surface under a specific mode shape.

[0029] Specifically, modal analysis is performed on the finite element model established in the previous step. The output includes: the first M transverse vibration natural frequencies; for example, the first four transverse bending frequencies are extracted: f1=0.30Hz, f2=1.08Hz, f3=2.45Hz, f4=4.80Hz. The corresponding mode shape functions are the displacement vectors (data) of each node along the height direction of the tower under each mode shape, output by the software. For example, the first mode shape is an overall S-shaped bending. The theoretical normal strain distribution is calculated or directly derived from the mode shape results, based on the axial strain values ​​at the 10 preset strain sensor locations under each mode shape.

[0030] S2.3. The natural frequency values, modal function, and theoretical normal strain distribution are normalized to construct a structured physical prior knowledge base containing frequency domain feature vectors, spatial mode shape matrices, and strain distribution vectors. The various physical features (frequency, displacement, strain) calculated in S2.2 are standardized and structured, transforming them into a unified and standardized digital knowledge base that is easily processed by computers and directly usable by subsequent neural networks. A structured physical prior knowledge base refers to a systematic digital set formed by organizing the normalized physical features into standard data structures such as matrices and vectors according to their logical relationships (frequency domain, spatial domain) and establishing indexes.

[0031] S3. Construct multiple parallel lightweight feature extraction sub-models to process vibration acceleration signals, structural strain data, and environmental parameter data streams, respectively. The sub-model processing vibration acceleration signals employs a depthwise separable convolutional structure, and adaptively configures the temporal receptive field of the convolutional kernels according to the theoretical natural frequency range to match the low-frequency vibration characteristics of the tower. Physical priors are injected during the training of this sub-model. Specifically, the parallel lightweight feature extraction sub-model refers to multiple (usually three) independent, structurally streamlined deep learning models running simultaneously, learning and extracting high-level abstract features from different input data streams. It simultaneously processes three types of monitoring data—vibration, strain, and environment—with different physical meanings and characteristics, avoiding interference between different signal types during extraction by a single model and preserving the independent features of each modality. The lightweight design uses a parametrically efficient network architecture (such as depthwise separable convolutions) to lay the foundation for subsequent deployment on resource-constrained edge devices and real-time evaluation. Physical characteristic adaptation adaptively adjusts the receptive field of the network structure to effectively capture long-period vibration patterns, improving the targeting of feature extraction, specifically targeting the low-frequency dominant characteristics of tower vibration. Depthwise separable convolution is an efficient convolution operation that decomposes standard convolution into two steps: depthwise convolution (spatial convolution for each input channel individually) and pointwise convolution (1x1 convolution, mixing channel information). This significantly reduces model parameters and computational cost, making it a core component of lightweight design. The temporal receptive field (NRF) is the length of time a convolution operation can see or perceive in a time-series processing sequence. The NRF size is determined by the kernel size, dilation rate, and number of network layers. Physical prior injection refers to incorporating domain knowledge (in this case, structural dynamics knowledge) into the model during the training or initialization phase of a neural network in a specific form (e.g., modifying the loss function, setting initial weights, or as attention guidance) to constrain or guide its learning direction. By injecting the physical prior knowledge base established in step S2 as constraints or initial conditions into the neural network's learning process, it ensures that the extracted features have clear physical meaning (e.g., related to specific mode shapes) rather than purely mathematical statistical patterns, fundamentally enhancing the model's interpretability and generalization ability with small samples.

[0032] Step S3 further includes: S31: Three parallel, lightweight feature extraction sub-models are constructed to process vibration acceleration signals, structural strain data, and environmental parameter data streams, respectively. Each sub-model adopts an encoder architecture based on depthwise separable convolution. Among them, the vibration acceleration signal refers to the time history data collected by multiple acceleration sensors, reflecting the overall dynamic response of the tower.

[0033] Structural strain data refers to time-history data collected from multiple strain gauges, reflecting the deformation and stress state of local parts of the tower (such as welds and flanges). Environmental parameter data streams refer to continuously monitored external condition data such as wind speed, wind direction, and temperature, which are the main load sources causing structural response. Three independent feature extraction sub-models are constructed for the three types of data streams with different physical characteristics and frequency ranges: vibration, strain, and environmental parameters. This dedicated network design avoids the problems of mutual interference and feature confusion between different types of signals in an integrated model, ensuring that the unique patterns of each type of data can be extracted more accurately. All sub-models use depthwise separable convolution as the core building block. This design significantly reduces model parameters and computational complexity while maintaining the ability to learn complex spatiotemporal features from data, which is key to achieving real-time evaluation for edge deployment. Multiple neural networks run simultaneously, each model processing only one type of input data and outputting a set of feature vectors, which are then fused. For example, the input data includes: vibration signals: shape [batch size, sequence length = 256, number of channels = 10] (10 accelerometers). Strain signal: Shape [batch size, sequence length = 256, number of channels = 6] (6 strain gauges). Environmental parameters: Shape [batch size, sequence length = 256, number of channels = 3] (wind speed, wind direction, temperature). Each sub-model is a 4-layer encoder. Each layer structure includes a depthwise separable convolutional layer, including: Depthwise convolution: using one-dimensional convolution, the convolution kernel slides in the time dimension, processing each input channel independently. Pointwise convolution: using 1x1 convolution, mixing feature information from different channels. Batch normalization layer: accelerating training and stabilizing convergence. ReLU activation function: introducing nonlinearity. Average pooling layer (used in the first 3 layers): reducing the sequence length and expanding the receptive field of subsequent layers. The three sub-models output feature vectors of shape [batch size, feature dimension], which are prepared for fusion in step S4.

[0034] S32: For the vibration acceleration signal processing sub-model, based on the fundamental frequency value f1 in the theoretical natural frequency range obtained in step S2, determine the size K and dilation rate D of the one-dimensional convolution kernel used in the depthwise separable convolution layer, where K×D≥f s / f1,f sThe sampling frequency for the vibration signal is set to ensure that the temporal receptive field of the convolutional layer can completely cover at least one fundamental frequency vibration cycle. Specifically, the fundamental frequency value f1 is the first-order lateral bending natural frequency of the tower calculated in step S2. It is the lowest frequency and usually the highest energy vibration frequency, and is a core indicator of structural stiffness. The sampling frequency f... s This is the number of data points collected per second from the vibration acceleration signal. It determines the highest frequency (Nyquist frequency) that can be represented without distortion. The kernel size K is the length of the one-dimensional convolution kernel in the time dimension, i.e., the number of original data points covered in one convolution calculation. The dilation rate D is the spacing between the elements of the convolution kernel. When D=1, it is a standard convolution; when D>1, it is a dilated convolution, which can exponentially expand the receptive field without increasing the parameters. The temporal receptive field is the time span of the input sequence corresponding to a point in the output of the convolutional layer. For dilated convolution, the receptive field is calculated as: RF=K+(K-1)*(D-1). The inequality K×D≥f s / f1, its physical meaning is the required time span covered by the convolution kernel (K*D / f s The period of the fundamental frequency oscillation (1 / f1 second) is at least greater than or equal to the period of the fundamental frequency oscillation.

[0035] For example, given the parameters: fundamental frequency f1 = 0.32Hz, sampling frequency fs = 50Hz, the number of sampling points corresponding to the fundamental frequency period is: N. period =f s / f1=50 / 0.32≈156 points. Parameter configuration: Select kernel size K=7, select dilation rate D=23. Verification: K×D=7×23=161, which is greater than 156, meeting the requirements. Actual receptive field calculation: RF=7+(7-1)*(23-1)=7+132=139 points. Although less than 156, this formula is the precise receptive field, while K×D is the approximate effective length. To ensure complete coverage, in practice, K×D will be slightly larger than f1. s / f1, in this example 161>156, is a reasonable configuration. This setting allows the first layer of the network to perceive a time window as long as 161 / 50Hz = 3.22 seconds.

[0036] This step addresses the significant characteristics of low-frequency vibrations in tower structures (fundamental frequencies typically in the range of 0.1-1Hz) by abandoning common network parameter settings. By dynamically adjusting the temporal receptive field of the convolutional layers, it ensures that the network can capture at least one complete fundamental frequency vibration cycle in the first instant (first convolutional layer), thereby effectively capturing the fundamental vibration modes most relevant to the overall stiffness of the tower structure and enhancing the specificity and physical significance of feature extraction.

[0037] S33: During the training process of the vibration acceleration signal processing sub-model, information from the physical prior knowledge base constructed in step S2 is injected through at least one of the following methods: Theoretical physical knowledge representing the tower's health baseline state is injected as a tutor or rule into the data-driven learning process. This forces the neural network to shift from blindly fitting data to learning under the constraints of physical laws, thereby extracting features with clear physical interpretability (such as first-order modal energy and frequency shift) rather than incomprehensible statistical features.

[0038] Method 1: The theoretical modal shape matrix is ​​used as the initial spatial attention weight matrix, and element-wise weighted with the input multi-channel acceleration signal feature map. At the beginning of feature extraction, the importance weights of each sensor signal are predefined according to the theoretical modal shape, guiding the network to focus on the spatial modes most relevant to the overall structural dynamics. The theoretical modal shape matrix is ​​derived from a knowledge base. Its elements Indicates the first The first mode shape Normalized displacement in each sensor direction. The spatial attention weight matrix is ​​derived from... Extract (usually first-order) mode shape information and construct a weight vector [1, 1, number of channels] that matches the channel dimension of the input feature map shape [batch, length, channel]. For example, from... Extract the first-order vibration mode values ​​for all 10 sensors in the X and Y directions to obtain the vector. Calculate the vector magnitude of each sensor. This yields a 10-dimensional weight vector W. W is then normalized (e.g., using softmax) to make its sum equal to 1. After the first convolutional layer of the vibrating submodel, the feature map is element-wise multiplied with W (via a broadcast mechanism).

[0039] Method 2: Introduce a frequency domain consistency regularization term into the model's loss function. This regularization term calculates the matching error between the spectrum obtained after Fourier transform of the intermediate layer features of the sub-model and the theoretical intrinsic frequency vector. During training, the loss function penalizes features that deviate too much from the theoretical intrinsic frequency in the frequency domain, ensuring that the frequency components of the features learned by the network are strongly correlated with the structurally healthy physical essence. The frequency domain consistency regularization term is an additional loss term added to the main task loss (such as classification loss). The matching error is a measure of the difference between the spectrum A(f) of the features extracted by the network and the theoretical frequency target M(f).

[0040] For example, select intermediate layer features, specifically the feature map Z after activation of the second layer of the vibrator model. Calculate the spectrum; for each sample and each channel of Z, perform an FFT along the time dimension to obtain the amplitude spectrum A(f). Construct a target mask based on... =[0.32,1.15,2.68,5.02]Hz, for each frequency ,exist[ *0.95, *1.05] Set M(f) = 1 within the interval, otherwise set it to 0. Calculate the regularization loss, for example, using negative cosine similarity: Summing is performed across all samples and channels. Total loss function: , where λ (e.g., set to 0.1) is a hyperparameter controlling the strength of physical constraints. Through optimization As the network learns to classify, its internal features are forced to peak at specified frequencies.

[0041] This is the cross-entropy loss function. In the context of this invention, it belongs to the main task loss or supervision signal loss. It measures the difference between the probability distribution predicted by the model and the probability distribution of the true labels. The smaller the difference, the lower the loss value, and the more accurate the model prediction. Input sources include: model prediction output. and real labels , The oscillator model (and subsequent fusion models and cascaded quantization networks) will eventually output a probability vector representing four health status levels, such as [0.85, 0.10, 0.04, 0.01], which means that the model considers the probability of being healthy to be 85%. This originates from a batch of labeled data generated during the training phase. For each data sample, its true state is represented by a one-hot encoded vector, for example, a healthy state is [1,0,0,0]. The calculation formula is: The summation applies to all categories. This forces... Towards near. It is driven by ensuring the accuracy of the model in answering questions correctly.

[0042] Furthermore, in step S33, the frequency domain consistency regularization term The specific calculation method is as follows: Global average pooling along the channel dimension is performed on the feature map Z output by a certain intermediate layer of the vibration acceleration signal processing sub-model to obtain the temporal feature vector z(t). Let Z be the feature map output by the Lth layer (e.g., after the 2nd convolutional layer) of the vibration acceleration signal processing sub-model, with dimensions B, T, and C, where B is the batch size (e.g., 32), T is the time step size (e.g., 64), and C is the number of channels (e.g., 32). Global average pooling along the channel dimension is performed on this feature map Z, calculated as follows: t and c are index variables in the formula, used to represent the time step and channel dimensions respectively, resulting in a temporal feature vector z(t) of shape B,T, which integrates the average activation patterns of all channels in the time dimension, representing the comprehensive vibrational features learned by the network at present.

[0043] Perform a Fast Fourier Transform (FFT) on z(t) and obtain its amplitude spectrum A(f). Specifically, perform a Fast Fourier Transform (FFT) on the time vector z(t) of each sample in each batch. Assuming a time step T = 64, the complex spectrum is obtained after the FFT, and its absolute value is taken to obtain the amplitude spectrum A(f): The shape of A(f) is B,F, where F is the number of frequency points (usually T / 2+1=33). Each A(f)[b] is a 33-dimensional vector representing the energy distribution of the b-th sample feature in the frequency domain.

[0044] Calculate the difference between the amplitude spectrum A(f) and the ideal spectral mask M(f) constructed from the theoretical natural frequency vector obtained in step S2, wherein the ideal spectral mask M(f) is calculated at each theoretical natural frequency f. i The value is 1 within the preset bandwidth Δf centered on the frequency, and 0 otherwise. Based on the theoretical natural frequency vector calculated in step S2, for example: Set the preset bandwidth Δf to 5% of the theoretical frequency (Δf = 0.05 * fi). For each theoretical frequency f... i Determine the corresponding index range on the frequency axis: ,in The Nyquist frequency (sampling frequency f) s / 2), assuming f s =50Hz, then =25Hz. Initialize a zero vector M(f) of the same length as A(f) (length F). Set the value of M(f) to 1 within each of the above index ranges, and 0 for the rest. M(f) is the ideal binarized spectrum mask.

[0045] The frequency domain consistency regularization term It is defined as the negative correlation measure or mean square error between the amplitude spectrum A(f) and the ideal spectral mask M(f). The frequency domain consistency regularization loss term belongs to the category of auxiliary constraint loss or physical knowledge loss. It measures the degree of matching between the frequency components of the feature maps in the intermediate layers of the neural network and the prior theoretical frequencies. The higher the matching degree, the lower this loss term. It is not directly aimed at making the classification more accurate, but rather at standardizing the model's learning process to ensure that the extracted features have the desired physical meaning, i.e., are related to the theoretical intrinsic frequencies of the tower. This formula is the mathematical core for implementing physics prior injection, using negative cosine similarity to quantify the deviation between network features and theoretical physics knowledge. This formula measures the degree of mismatch between the spectrum of features extracted by the network and the ideal physical spectrum template. The smaller (more negative) the value, the worse the match; during training, minimizing this value is crucial. In effect, this forces the value to increase (due to the negative sign), thus improving the matching degree. Here, A(f) is the feature amplitude spectrum vector, a frequency representation derived from the feature map of the intermediate layer of the neural network. Its length is equal to the number of frequency points after the FFT (usually half the time step plus one). For example, if the time step is 256, then A(f) is a 129-dimensional vector, where each element represents the vibrational energy (amplitude) of a specific frequency component. M(f) is the ideal spectral mask vector, a binary target template constructed based on the theoretical intrinsic frequencies in the physical prior knowledge base. M(f) visually indicates the positions where A(f) should have significant peaks in the frequency domain, indicating the correct frequency components related to the tower's health baseline state. A(f)•M(f) (dot product) actually only sums the amplitudes of A(f) at those frequency positions in M(f) where the value is 1 (i.e., near the theoretical frequency). This value represents the total energy of the network feature spectrum within the correct frequency band. The larger this value, the more concentrated the network features are on the physical frequency band we are interested in. ||A(f)|| and ||M(f)|| (norm / modulus) represent normalization factors used to eliminate the influence of absolute energy magnitude, focusing only on whether the shape of the energy distribution matches. (A(f)•M(f)) / (||A(f)||*||M(f)||) (cosine similarity) measures the directional consistency between the energy distribution pattern of the network's feature spectrum and the ideal physical spectrum pattern. The closer the value is to 1, the higher the matching degree.

[0046] S4. The features output by each sub-model are concatenated and input into the lightweight adaptive fusion module. This module employs a multi-head cross-attention mechanism, where the query vectors of one set of attention heads are initialized by feature parameters from the physical prior knowledge base. This allows the fusion process to calculate the correlation weights between features from different data sources and complete the fusion under the guidance of physical semantics. The function of step S4 is to deeply integrate multiple data-driven perception results under a decision framework guided by physical laws (a physical-guided attention mechanism), thereby generating a high-level feature representation that combines data complexity and physical consistency. This lays a crucial foundation for achieving accurate and interpretable health assessments.

[0047] Step S4 further includes: S41: Align and concatenate the feature vectors output by the vibration acceleration signal processing sub-model, structural strain data processing sub-model, and environmental parameter data processing sub-model described in step S3 to form a multi-source fusion feature matrix X. This achieves spatial unification of heterogeneous feature vectors. The features output by the three parallel sub-models of vibration, strain, and environment (which may have different dimensions) are projected onto a unified feature dimension through a fully connected layer or similar method, and then concatenated along the feature dimension to form a multi-source fusion feature matrix X that can be uniformly processed by the subsequent fusion module. This is a necessary prerequisite for deep fusion.

[0048] S42: Construct a lightweight adaptive fusion module, the core of which is a multi-head cross-attention layer containing H attention heads, at least one of which is designated as a physics-guided attention head. Design the core architecture and division of labor strategy of the fusion module. Construct a lightweight module with a multi-head cross-attention layer as its core, and clearly define the division of labor within it: at least one attention head is given a special role, namely the physics-guided attention head, which is specifically responsible for examining and fusing features from a physical perspective; the rest are standard attention heads, responsible for freely discovering potential complex correlations in the data. This hybrid design of specialized and general attention heads is key to balancing physical interpretability and model representation capabilities.

[0049] S43: For the physical guidance attention head, query its vector Q. phy The vector is initialized with features from the physical prior knowledge base obtained in step S2. This vector is mapped to the dimension required by the attention mechanism through a linear transformation layer; simultaneously, its key vector K and value vector V both originate from the multi-source fusion feature matrix X. This achieves targeted injection of physical prior knowledge. An initial search direction with clear physical meaning is set for the physical guidance attention head. Its query vector Q is then used. phy The initialization is performed using vectors extracted and transformed from a priori physical knowledge base (such as theoretical frequency vectors and principal mode vectors). This is equivalent to giving this attention head a physical compass. Its key vector K and value vector V come from the data feature matrix X to be fused. This asymmetric initialization means that the attention head will always use physical knowledge as the question or benchmark to retrieve and extract the most relevant answer from the data features. Essentially, it encodes physics-based assumptions into the network structure, making the fusion process constrained by physical laws from the beginning.

[0050] In step S43, the physical guidance attention head query vector Q is initialized. phyThe characteristic parameters are encoded representations of any one or a combination of the following: theoretical natural frequency vector; principal component vector of the theoretical modal shape matrix after dimensionality reduction by principal component analysis; and theoretical strain distribution vector. The function of this step is to provide the physical guidance attention head with an initial query instruction with clear physical semantics, ensuring that its fusion process is always anchored to the core physical properties of structural health. The three parameter selections provide guidance from different physical dimensions: the theoretical natural frequency vector is used to guide the attention head to focus on frequency domain features. The theoretical frequency representing the healthy reference stiffness of the tower (e.g., [0.32, 1.15, 2.68, 5.02] Hz) is used as the initial query. This is equivalent to instructing the attention head to find the information most relevant to these key frequency components from multi-source features and fuse them, so that the fused features can prominently reflect changes in the overall structural stiffness (frequency) and are highly sensitive to damage affecting the overall dynamic characteristics, such as tower stiffness degradation and crack propagation. The principal component vector of the theoretical modal shape matrix after dimensionality reduction by principal component analysis is used to guide the attention head to focus on spatial vibration morphology features. The theoretical modal shape matrix describes the spatial distribution of vibration energy (along the tower height and direction). Principal component analysis (PCA) extracts its main components (e.g., the first principal component represents the most significant overall bending mode), resulting in a low-dimensional vector representing the most significant vibration mode, which serves as the query. This is equivalent to instructing the user to focus on spatial characteristic modes matching the tower's main vibration mode (e.g., first-order bending) for fusion, making the fused features more sensitive to damage affecting the vibration mode, such as mass imbalance, loose connections, and local stiffness changes. The theoretical strain distribution vector guides the focus to local stress / strain characteristics. This vector represents the theoretical strain response sensitivity at key locations (e.g., welds, flanges) under the theoretical mode shape. It serves as the query. This is equivalent to instructing the user to filter and integrate the features most relevant to strain changes at these high-stress-risk points from the data, allowing the fusion process to prioritize information related to local fatigue, stress concentration, bolt preload loss, and other local damage. Encoding the above two or three parameters together as the query vector provides multi-dimensional comprehensive physical guidance. For example, by combining frequency vectors and mode principal components, attention can be directed to focus on characteristic modes with specific spatial morphology at specific frequencies, thereby enabling more precise identification of complex composite damage that simultaneously affects both frequency and mode.

[0051] Regardless of the parameters used, their essential function is to transform expert knowledge in the field of physics (frequency, mode shape, strain) into search keywords that machine learning models can understand. This transforms the feature fusion of the model from blind correlation to a controllable process with clear objectives and physical meaning, greatly enhancing the interpretability of subsequent health status assessments and their targeting of specific damage patterns.

[0052] S44: The physically guided attention head calculates the attention weight matrix under physical prior guidance, and obtains the fusion feature representation with enhanced physical semantics through weighted summation; the remaining attention heads adopt the standard cross-attention mechanism, and their query vector, key vector, and value vector are all derived from different linear transformations of the multi-source fusion feature matrix X. This step is used to perform attention calculations in two different modes to generate complementary feature representations. Specifically, in the physically guided mode: the physically guided attention head calculates the physical query Q based on its physicalized query Q. phy The attention weights are calculated using K and V, derived from the data features. This weight matrix explicitly reflects the correlation between the data features and prior physical knowledge. Through weighted summation, a fusion feature representation with enhanced physical semantics is obtained. Data-driven mode: For other standard attention heads, Q, K, and V are derived from different linear transformations of the data feature X. These parameters are not physically constrained, freely exploring any possible complex correlation patterns within the data, including unknown or nonlinear ones, generating a data-driven discovery fusion feature representation.

[0053] S45: The feature sub-representations output by each attention head are concatenated and linearly projected to obtain the final multi-source physical fusion feature representation. This step integrates the fusion results from different perspectives to generate the final unified feature. The physics-guided feature sub-representation is concatenated with multiple data-driven feature sub-representations to form a rich feature set that integrates the prior physical perspective and the perspective of free data exploration. The concatenated feature is then reduced in dimensionality and integrated through a linear projection layer to obtain the final multi-source physical fusion feature representation. This representation contains both core information guided by physical laws and retains supplementary information mined from the data. It is a high-quality feature with complete information and clear guidance, providing a solid foundation for the final health status decision.

[0054] The S4 series of steps is designed to build and implement a fusion mechanism that allows physics experts and data algorithms to work together. First, heterogeneous data is aligned. Then, a well-defined attention committee (including physics consultants and data analysts) reviews and fuses the data from multiple perspectives. Finally, the opinions from all parties are integrated to form a high-quality joint report (final fusion feature) for subsequent final decision-making (health status assessment).

[0055] S5. Input the output fused features into a cascaded quantization network model, which includes: a first lightweight fully connected layer for generating damage-sensitive indicators related to the physical state of the tower; and a second lightweight fuzzy inferencer for mapping the damage-sensitive indicators to multiple predefined health status levels based on a hybrid membership function composed of a Gaussian function and a trapezoidal function.

[0056] Specifically, step S5 further includes: S51: The fused features output from step S4 are input to the first lightweight fully connected layer. The output dimension of this fully connected layer corresponds to multiple preset physical damage sensitivity indicators, including but not limited to: tower first-order natural frequency offset, damping ratio variation coefficient, strain energy concentration index, and modal correlation coefficient. This step is used to map the abstract high-dimensional fused features into quantifiable physical damage characteristic parameters with clear engineering significance. Feature decoding refers to the first lightweight fully connected layer acting as a decoder, decoding the high-level fused features rich in multi-source information output from step S4 into a series of low-dimensional indicators representing specific physical damage modes. Physical interpretability assurance means that each output neuron directly corresponds to a preset physical damage sensitivity indicator. For example: tower first-order natural frequency offset: reflects the overall stiffness change. Damping ratio variation coefficient: reflects the change in structural energy dissipation characteristics or connection state. Strain energy concentration index: reflects local stress concentration or fatigue damage risk. Modal correlation coefficient: reflects the degree of distortion of vibration mode relative to a healthy baseline. This step realizes the key transformation from a machine-understandable feature space to a human-understandable physical parameter space. The output set of indicators constitutes a multi-faceted and quantitative description of the tower's health status, providing direct and clear input for knowledge-based final decision-making.

[0057] S52: The damage-sensitive indicators are input into a second lightweight fuzzy inference engine, simulating human expert thinking, to synthesize multiple imprecise, continuous physical indicators into a discrete, actionable health status level. This inference engine includes: Fuzzy Interface: For each damage-sensitive indicator, a combination of Gaussian and trapezoidal membership functions is used to define its membership degree across four health status levels: healthy, sub-healthy, mildly damaged, and severely damaged. This transforms each precise physical damage indicator value into the probability or degree to which it belongs to each health level. Membership functions for each indicator across four health status levels (healthy, sub-healthy, mildly damaged, and severely damaged) are defined. A hybrid approach is used, employing Gaussian functions (to describe smooth transitions between good / acceptable intermediate states) and trapezoidal functions (to describe explicit thresholds for extreme states of health / danger). For example, for frequency offset, -2% to 0% might belong to healthy (trapezoidal), and 0% to 5% to sub-healthy (Gaussian).

[0058] Rule base: Contains predefined fuzzy reasoning rules based on domain knowledge, used to logically combine the membership degrees of various indicators, encapsulate domain expert experience, and define how multiple physical indicators collaboratively determine the final state. Includes a set of IF-THEN fuzzy rules. For example: IF small frequency offset (healthy) AND low strain concentration (healthy) THEN state is healthy; IF large frequency offset (dangerous) OR low mode shape correlation coefficient (dangerous) THEN state is dangerous. The fuzzy terms such as small, large, and low in the rules have been quantized by a fuzzification interface.

[0059] Defuzzification Interface: Using either the centroid method or the maximum membership method, the final comprehensive health status level is calculated based on the output of the rule base. The membership information scattered across various status levels from the rule base is aggregated and transformed back into a clear, single health status level conclusion. Specifically, the centroid method calculates a weighted average of all statuses to obtain a continuous health value; the maximum membership method directly selects the status level with the highest membership degree. The latter is more intuitive and provides a clear output. This step addresses the inherent uncertainty and multi-indicator conflicts in health assessment. It does not require precise mathematical thresholds but rather employs a comprehensive approach, weighing and judging factors like an expert, ultimately outputting a robust, reliable, and easily understandable decision.

[0060] S53: The comprehensive health status level output by the second lightweight fuzzy inference engine is used as a quantitative assessment result of the tower's current health status, delivering a final user-readable assessment conclusion. The comprehensive health status level (e.g., minor damage) output by the S52 fuzzy inference engine is used as the system's final, authoritative quantitative assessment result of the tower's current health status. This result is a discrete classification label, intuitive and clear, and can be directly used to trigger corresponding operation and maintenance decisions (e.g., enhanced monitoring or planned maintenance).

[0061] S6. The cascaded quantization network model is trained in stages using source domain data, and the model is compressed using knowledge distillation techniques. Specifically, step S6 further includes: S61: Using historical monitoring data from multiple similar wind turbine units as the source domain training set, the cascaded quantization network model described in step S5 is trained in the first stage to optimize model parameters. Sufficient external data (source domain) provides the model with a general knowledge base and good initial parameter values. Historical monitoring data collected from multiple other similar wind turbine units (source domain units) is used. This data volume is usually large, containing rich data on normal operation and various known fault modes. The complete cascaded quantization network model (teacher model) constructed in step S5 is fully trained on this data. The goal of this stage is to enable the model to learn general characteristics, damage patterns, and discrimination modes regarding tower health status, i.e., to master the general knowledge for the task of wind turbine tower health assessment. The model obtained through prediction training (teacher model) is already a high-performing expert, and its parameters are no longer random. This provides a high starting point for subsequent application to the target unit (data scarce), greatly alleviating the dependence on a large amount of labeled data for the target unit, and is a key step in solving the small sample problem.

[0062] S62: Using the network model trained in step S61 as the teacher model, construct a student model with a more streamlined structure and fewer parameters. Design a model architecture that is functionally similar to the teacher model but with significantly reduced computational and storage overhead. The student model retains the core functional structure of the teacher model (such as cascaded feature extraction, fusion, and quantization networks), but simplifies aspects such as network width, network depth, and more efficient operations. This ensures that the student model has fewer parameters and lower computational complexity, enabling deployment on resource-constrained edge devices, while maintaining sufficient architectural capacity to handle the knowledge transferred from the teacher model. A streamlined model framework with deployment feasibility is thus created.

[0063] S63: In the knowledge distillation process, a soft-target distillation method is adopted. The softened probability distribution output by the teacher model in the damage-sensitive index generation layer is used as a supervisory signal to guide the training of the student model, transferring the implicit knowledge contained in the teacher model that is difficult to learn directly from the data to the lightweight student model. The distillation focuses on the output of the first lightweight fully connected layer (damage-sensitive index generation layer) in the teacher model. The output of this layer is a set of continuous values ​​(or a probability distribution after temperature softening by softmax). Hard labels only tell the student that the image corresponds to mild damage (one category). Soft targets are the probability distribution (soft targets) output by the teacher model, which contains richer information. For example, the image has 70% of the characteristics of mild damage, but some features (20%) resemble sub-health, and there is a very small chance (10%) that it is normal but has interference. This relative relationship between different categories is valuable implicit knowledge. During distillation training, the student model not only learns to match the real health status labels (hard targets), but more importantly, learns to imitate the softened probability distribution (soft targets) output by the teacher model. The core value of knowledge distillation lies in enabling student models to inherit the more refined discriminative abilities and generalization characteristics of teacher models, achieving the effect of "small model, great wisdom."

[0064] S64: The student model is trained in the second stage by combining standard classification loss and distillation loss to obtain a compressed, lightweight cascaded quantization network model. Within a unified training framework, the classification accuracy and knowledge transfer of the student model are simultaneously optimized, ultimately resulting in a high-performance lightweight model. The student model is trained in the second stage using potentially limited data from the target unit (or combined with some source domain data). By minimizing the joint loss, the parameters of the student model are simultaneously pushed in two directions: one is to directly fit the target data, and the other is to approximate the thinking process of the teacher model. After training, the resulting student model has performance close to or sometimes even surpasses the teacher model, but with orders of magnitude advantages in model size and computational speed. It is a final product specifically optimized for the target unit and suitable for edge deployment.

[0065] S7. Deploy the compressed model on the edge computing device at the tower site, connect to the real-time monitoring data stream, and output the tower's health status level. Specifically, S7 includes: The lightweight, cascaded quantized network model, after knowledge distillation and compression, is converted into a format suitable for edge computing device inference engines and deployed in embedded edge computing units located in the nacelle or base of wind turbine towers. This step achieves seamless migration and adaptation of the model from the training framework to the deployment environment. Specifically, model files trained under frameworks such as PyTorch / TensorFlow are converted into efficient, fixed formats (e.g., .engine, .tflite) supported by edge device inference engines (e.g., TensorRT, OpenVINO, TFLite). This process typically includes graph optimization, operator fusion, and precision quantization (e.g., FP32 to INT8) to further reduce model size and improve inference speed. The converted model files, necessary preprocessing libraries, and inference engine are then installed into the embedded edge computing unit located in the nacelle or base of the wind turbine tower. This unit is a rugged, low-power computing device specifically designed for industrial environments.

[0066] The embedded edge computing unit acquires the tower's vibration acceleration signals, structural strain data, and environmental parameters in real time, and performs the same standardized preprocessing as in S1. Then, the processed data is input into the deployed lightweight model, sequentially executing feature extraction, physical-guided fusion, and health status quantification inference. This step establishes an end-to-end automated data processing and intelligent analysis pipeline at the edge. Specifically, the edge computing unit acquires raw data streams from various sensors on the tower in real time and synchronously through its I / O interface. Then, it immediately executes a standardized preprocessing process (such as filtering, detrending, and normalization) completely identical to S1 locally, ensuring that the data quality input to the model is consistent with that during training. The preprocessed data serves as input, driving the deployed lightweight model to execute the complete intelligent analysis chain from S3 to S5 online, automatically, and cyclically. All calculations are completed locally on the tower, eliminating the need to upload massive amounts of raw data to the cloud, greatly reducing communication bandwidth requirements, eliminating network latency, and ensuring data security and real-time processing (typically with sub-second or even sub-second response times).

[0067] The embedded edge computing unit periodically outputs real-time health status assessment results for the tower. When the health status deteriorates to a preset threshold, it simultaneously generates and sends an early warning signal to the remote monitoring center. This transforms the model's results into decision support information directly usable by maintenance personnel, enabling proactive early warning. The edge computing unit uploads the assessed health status level (e.g., healthy, sub-healthy) and optional key damage indicators (e.g., frequency offset) to the remote monitoring center via low-bandwidth communication (e.g., 4G / 5G, fiber optic) at a set interval (e.g., once per minute), updating the digital twin system or monitoring screen and providing continuous situational awareness for maintenance personnel. The edge computing unit has preset early warning logic (e.g., health status deteriorating to mild damage or worse). When the real-time inference results meet the early warning conditions, the unit immediately and proactively generates a high-priority early warning signal (including time, location, specific level, possible causes, etc.) and sends it to the monitoring center and relevant personnel (e.g., via SMS, App push). This achieves a leap from status monitoring to predictive early warning, saving valuable time for scheduling preventative maintenance.

[0068] Example 2

[0069] like Figure 2 As shown in the figure, this application provides an architecture diagram of a wind turbine tower health status assessment device based on multi-source data fusion and dynamic baseline modeling, which is applied to the wind turbine tower health status assessment device based on multi-source data fusion and dynamic baseline modeling as described in Embodiment 1. It includes: a data acquisition and preprocessing module 210, a physical prior knowledge base construction module 220, a parallel lightweight feature extraction module 230, a physical-guided multi-source fusion module 240, a cascaded health status quantification module 250, a model optimization and compression module 260, and an edge deployment and online assessment module 270.

[0070] The data acquisition and preprocessing module 210 is used to acquire vibration acceleration signals, structural strain data and environmental parameters of the wind turbine tower and perform standardized processing.

[0071] The physical prior knowledge base construction module 220 is used to establish a theoretical dynamic model based on the material properties, geometric dimensions and boundary conditions of the tower, calculate the theoretical natural frequencies, mode shapes and strain distributions at key locations of the tower under healthy reference conditions, and construct a physical prior knowledge base.

[0072] The parallel lightweight feature extraction module 230 is used to construct and run multiple parallel lightweight feature extraction sub-models to process vibration acceleration signals, structural strain data, and environmental parameter data streams, respectively. Among them, the sub-model for processing vibration acceleration signals adopts a depthwise separable convolutional structure and adaptively configures the temporal receptive field of the convolutional kernel according to the theoretical natural frequency range to match the low-frequency vibration characteristics of the tower. This module injects information from the physical prior knowledge base during the training process.

[0073] The physics-guided multi-source fusion module 240 is used to concatenate the features output by each sub-model and process them using a multi-head cross-attention mechanism. The query vector of a set of attention heads is initialized by the feature parameters in the physics prior knowledge base, so that the fusion process calculates the correlation weights between features from different data sources and completes the fusion under the guidance of physical semantics.

[0074] The cascaded health status quantization module 250 is used to input the fusion features output by the physical guidance multi-source fusion module into the cascaded quantization network model for processing. The cascaded quantization network model includes, in sequence: a first lightweight fully connected layer, used to generate damage-sensitive indicators related to the physical state of the tower; and a second lightweight fuzzy inferencer, used to map the damage-sensitive indicators to multiple predefined health status levels based on a hybrid membership function composed of a Gaussian function and a trapezoidal function.

[0075] The model optimization and compression module 260 is used to perform phased prediction training on the cascaded health status quantification module using source domain data, and to compress the model using knowledge distillation technology.

[0076] The edge deployment and online assessment module 270 is used to deploy the compressed model on the edge computing device at the tower site, access the real-time monitoring data stream, and control the device to operate online to output the health status level of the tower.

[0077] Figure 3 This is an electronic device provided in one embodiment of this application. For example... Figure 3 As shown, the electronic device includes at least the following components: processor 301 and memory 300, communication interface 303, and bus 302.

[0078] In this embodiment of the application, the memory 300 is used to store executable instructions of the processor 301, which is used to implement the method as described in the first aspect when executing the instructions.

[0079] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0080] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Information processed by these devices is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (FlashROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0081] It should be noted that a portion of the electronic device described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0082] It should be noted that the computer mentioned here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, computer-readable recording media refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage devices such as hard drives built into the computer.

[0083] Furthermore, computer-readable recording media can include: media that dynamically stores programs for short periods of time, such as communication lines used when transmitting programs via networks like the Internet or communication lines like telephone lines; and media that store programs for fixed periods of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining them with programs already recorded in the computer.

[0084] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (device group) composed of multiple devices. Each device constituting the device group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a device group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0085] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A method for assessing the health status of wind turbine towers based on multi-source data fusion and dynamic baseline modeling, characterized in that, Includes the following steps: S1. Collect vibration acceleration signals, structural strain data and environmental parameters of the wind turbine tower, and perform standardized processing. S2. Based on the material properties, geometric dimensions and boundary conditions of the tower, establish a theoretical dynamic model, calculate the theoretical natural frequencies, mode shapes and strain distribution at key locations of the tower under healthy reference conditions, and construct a physical prior knowledge base. S3. Construct multiple parallel lightweight feature extraction sub-models to process vibration acceleration signals, structural strain data, and environmental parameter data streams, respectively. Among them, the sub-model for processing vibration acceleration signals adopts a depthwise separable convolutional structure and adaptively configures the temporal receptive field of the convolutional kernel according to the theoretical natural frequency range to match the low-frequency vibration characteristics of the tower. Physical priors are injected during the training process of this sub-model. S4. The features output by each sub-model are concatenated and input into the lightweight adaptive fusion module. This module adopts a multi-head cross-attention mechanism, in which the query vector of a set of attention heads is initialized by the feature parameters in the physical prior knowledge base, so that the fusion process is guided by physical semantics, calculates the correlation weight between features from different data sources and completes the fusion. S5. Input the output fused features into a cascaded quantization network model, which includes: a first lightweight fully connected layer for generating damage-sensitive indicators related to the physical state of the tower; and a second lightweight fuzzy inferencer for mapping the damage-sensitive indicators to multiple predefined health status levels based on a hybrid membership function composed of a Gaussian function and a trapezoidal function. S6. The cascaded quantization network model is trained in stages using source domain data, and the model is compressed using knowledge distillation technology. S7. Deploy the compressed model on the edge computing device at the tower site, connect to the real-time monitoring data stream, and output the health status level of the tower.

2. The method according to claim 1, characterized in that, S2 specifically includes: Based on the material density, elastic modulus, geometric profile dimensions, and bottom fixed boundary conditions of the tower, a parametric continuum dynamics model conforming to Euler-Bernoulli beam theory or Timoshenko beam theory is established. The parameterized continuum dynamics model is used to analytically solve for the first M natural frequencies of the tower structure in the undamaged reference state, the corresponding mode shape functions, and the theoretical normal strain distribution at the location of the strain sensor. The natural frequency values, mode shape functions, and theoretical normal strain distributions are normalized to construct a structured physical prior knowledge base containing frequency domain eigenvectors, spatial mode shape matrices, and strain distribution vectors.

3. The method according to claim 1, characterized in that, Step S3 further includes: S31: Construct three parallel lightweight feature extraction sub-models for processing vibration acceleration signals, structural strain data and environmental parameter data streams, respectively. Each sub-model adopts an encoder architecture based on depthwise separable convolution. S32: For the vibration acceleration signal processing sub-model, based on the fundamental frequency value f1 in the theoretical natural frequency range obtained in step S2, determine the size K and dilation rate D of the one-dimensional convolution kernel used in the depthwise separable convolution layer, where K×D≥f s / f1,f s The sampling frequency of the vibration signal is set to ensure that the temporal receptive field of the convolutional layer can fully cover at least one fundamental frequency vibration cycle. S33: During the training process of the vibration acceleration signal processing sub-model, information from the physical prior knowledge base constructed in step S2 is injected through at least one of the following methods: The theoretical mode shape matrix is ​​used as the initial spatial attention weight matrix, and then weighted element-wise with the feature map of the input multi-channel acceleration signal. A frequency domain consistency regularization term is introduced into the model's loss function. This regularization term calculates the matching error between the spectrum obtained after Fourier transform of the intermediate layer features of the sub-model and the theoretical intrinsic frequency vector.

4. The method according to claim 3, characterized in that, In step S33, the frequency domain consistency regularization term is calculated as follows: Global average pooling along the channel dimension is performed on the feature map Z output by a certain intermediate layer of the vibration acceleration signal processing sub-model to obtain the temporal feature vector z(t); Perform a fast Fourier transform on z(t) and take its amplitude spectrum A(f); Calculate the difference between the amplitude spectrum A(f) and the ideal spectral mask M(f) constructed from the theoretical natural frequency vector obtained in step S2, wherein the ideal spectral mask M(f) is calculated at each theoretical natural frequency f. i The value is 1 within the preset bandwidth Δf centered on the center, and 0 otherwise; The frequency domain consistency regularization term is defined as the negative correlation measure or mean square error between the amplitude spectrum A(f) and the ideal spectral mask M(f).

5. The method according to claim 4, characterized in that, Step S4 further includes: S41: Align and stitch together the feature vectors output by the vibration acceleration signal processing sub-model, structural strain data processing sub-model and environmental parameter data processing sub-model described in step S3 to form a multi-source fusion feature matrix X. S42: Construct a lightweight adaptive fusion module, the core of which is a multi-head cross-attention layer containing H attention heads, where at least one attention head is designated as a physically guided attention head; S43: For the physical guidance attention head, query its vector Q. phy The vector is initialized as a vector composed of feature parameters from the physical prior knowledge base obtained in step S2. This vector is mapped to the dimension required by the attention mechanism through a linear transformation layer. At the same time, its key vector K and value vector V are both derived from the multi-source fusion feature matrix X. S44: The attention weight matrix under physical prior guidance is calculated through the physical guidance attention head, and the fusion feature sub-representation with physical semantic enhancement is obtained by weighted summation; the other attention heads adopt the standard cross-attention mechanism, and their query vector, key vector and value vector are all derived from different linear transformations of the multi-source fusion feature matrix X; S45: The feature representations output by each attention head are concatenated and linearly projected to obtain the final multi-source physical fusion feature representation.

6. The method according to claim 5, characterized in that, In step S43, the physical guidance attention head query vector Q is initialized. phy The characteristic parameters are encoded by any one or a combination of the following: the theoretical natural frequency vector, the principal component vector of the theoretical mode shape matrix after dimensionality reduction by principal component analysis, and the theoretical strain distribution vector.

7. The method according to claim 1, characterized in that, Step S5 further includes: S51: Input the fusion feature output from step S4 into the first lightweight fully connected layer. The output dimension of the fully connected layer corresponds to multiple preset physical damage sensitive indicators, including but not limited to: tower first natural frequency offset, damping ratio variation coefficient, strain energy concentration index and mode shape correlation coefficient. S52: Input the damage sensitivity index into a second lightweight fuzzy inference engine, which includes: The fuzzy interface is used to define the membership degree of each damage-sensitive indicator in four health status levels: healthy, sub-healthy, mildly damaged, and severely damaged, by using a combination of Gaussian membership function and trapezoidal membership function. The rule base contains predefined fuzzy reasoning rules based on domain knowledge, which are used to logically combine the membership degrees of various indicators; The defuzzification interface is used to calculate the final comprehensive health status level based on the output of the rule base using the centroid method or the maximum membership method. S53: The comprehensive health status level output by the second lightweight fuzzy inference engine is used as a quantitative assessment result of the tower's current health status.

8. The method according to claim 1, characterized in that, Step S6 further includes: S61: Using historical monitoring data from multiple similar wind turbine units as the source domain training set, perform the first stage training of the cascaded quantization network model described in step S5 to optimize the model parameters. S62: Use the network model obtained after training in step S61 as the teacher model to build a student model with a more streamlined structure and fewer parameters. S63: In the knowledge distillation process, a soft target distillation method is adopted, and the softening probability distribution output by the teacher model in the damage-sensitive index generation layer is used as a supervision signal to guide the training of the student model; S64: Combine the standard classification loss and distillation loss to perform a second-stage training on the student model to obtain a compressed lightweight cascaded quantization network model.

9. The method according to claim 1, characterized in that, S7 includes: The lightweight cascaded quantization network model, after knowledge distillation and compression, is converted into a format suitable for the inference engine of edge computing devices and deployed in the embedded edge computing unit of the wind turbine tower nacelle or tower base. The embedded edge computing unit collects the tower's vibration acceleration signal, structural strain data, and environmental parameters in real time, and performs the same standardized preprocessing as S1. Then, the processed data is input into the deployed lightweight model, and feature extraction, physical-guided fusion, and health status quantitative inference are performed in sequence. The embedded edge computing unit periodically outputs the real-time health status level assessment results of the tower, and when it detects that the health status level has deteriorated to a preset threshold, it synchronously generates and sends an early warning signal to the remote monitoring center.

10. A wind turbine tower health status assessment device based on multi-source data fusion and dynamic baseline modeling, applied to the method described in any one of claims 1 to 9, characterized in that, The device includes: The data acquisition and preprocessing module is used to acquire vibration acceleration signals, structural strain data and environmental parameters of the wind turbine tower, and perform standardized processing. The physical prior knowledge base construction module is used to establish a theoretical dynamic model based on the material properties, geometric dimensions and boundary conditions of the tower, calculate the theoretical natural frequencies, mode shapes and strain distributions at key locations of the tower under healthy reference conditions, and construct the physical prior knowledge base. The parallel lightweight feature extraction module is used to build and run multiple parallel lightweight feature extraction sub-models to process vibration acceleration signals, structural strain data, and environmental parameter data streams, respectively. Among them, the sub-model for processing vibration acceleration signals adopts a depthwise separable convolutional structure and adaptively configures the temporal receptive field of the convolutional kernel according to the theoretical natural frequency range to match the low-frequency vibration characteristics of the tower. This module injects information from the physical prior knowledge base during the training process. The physics-guided multi-source fusion module is used to concatenate the features output by each sub-model and process them using a multi-head cross-attention mechanism. The query vector of a set of attention heads is initialized by the feature parameters in the physics prior knowledge base, so that the fusion process is guided by physical semantics, calculates the correlation weights between features from different data sources and completes the fusion. A cascaded health status quantization module is used to input the fusion features output by the physical guidance multi-source fusion module into a cascaded quantization network model for processing. The cascaded quantization network model includes, in sequence: a first lightweight fully connected layer, used to generate damage-sensitive indicators related to the physical state of the tower; and a second lightweight fuzzy inferencer, used to map the damage-sensitive indicators to multiple predefined health status levels based on a hybrid membership function composed of a Gaussian function and a trapezoidal function. The model optimization and compression module is used to perform phased prediction training on the cascaded health status quantification module using source domain data, and to compress the model using knowledge distillation technology. The edge deployment and online assessment module is used to deploy the compressed model on the edge computing device at the tower site, access the real-time monitoring data stream, and control the device to operate online to output the health status level of the tower.

Images