Fan vibration data missing intelligent reconstruction method based on deep network structure

Abstract

The invention discloses a fan vibration data missing intelligent reconstruction method based on a deep network structure, and particularly relates to the field of data missing reconstruction, and the method comprises the steps: collecting and normalizing fan vibration and working condition data, designing three missing modes of random points, continuous blocks and channel total loss to construct a data set, and building a normal working condition vibration feature memory library; establishing space-time Transform time sequence prediction, physical information neural network physical simulation and deep memory network contrast learning three types of heterogeneous expert networks, and completing pre-training; constructing a dynamic arbitration fusion network, inputting expert high-level features and meta-task features, and realizing dynamic weight distribution and bifurcation detection; based on the energy flow and a coherence constraint rule base, performing physical verification correction on the bifurcation result; performing weighted fusion according to expert result consistency or divergence scenes, and evaluating and outputting through reconstruction precision, physical rationality and real-time indexes; the method improves the accuracy and reliability of reconstruction in different missing scenes, and meets the requirements of fan vibration monitoring.

Description

Technical Field

[0001] This invention relates to the field of data missing reconstruction technology, and more specifically, to a method for intelligent reconstruction of missing wind turbine vibration data based on deep network structures. Background Technology

[0002] As a core power generation device in the new energy field, the monitoring and fault diagnosis of wind turbines are crucial to ensuring power generation efficiency and equipment safety. Vibration data, as the core carrier reflecting the operating status of key components such as the wind turbine main shaft and gearbox, directly affects the accuracy of monitoring and diagnosis due to its completeness. Currently, wind turbine vibration data acquisition has formed a mature system. Vibration acceleration signals and operating parameters such as speed, torque, and oil temperature are collected synchronously by multi-point vibration sensors and SCADA systems, and the data quality is improved by preprocessing techniques such as wavelet denoising and normalization.

[0003] In actual operation, data loss is a common objective phenomenon, mainly stemming from occasional packet loss in data transmission, short-term sensor failures, or complete disconnections. To address this issue, data reconstruction technology continues to develop. Deep learning, with its powerful feature extraction and fitting capabilities, has become the mainstream approach. Time-series prediction models and physically constrained intelligent algorithms are gradually being applied to vibration data completion. Meanwhile, reference-based reconstruction approaches based on vibration characteristics under normal operating conditions have also been extensively studied, laying the foundation for subsequent optimization of related technologies.

[0004] However, it still has some drawbacks in practical use, such as:

[0005] 1. Insufficient adaptability of single models: Existing technologies mostly rely on a single network model to process vibration data missing reconstruction, which is difficult to adapt to multiple scenarios such as random point missing, continuous block missing, and complete channel loss. The ability to capture complex spatiotemporal dependencies or physical constraints is limited, resulting in large fluctuations in reconstruction accuracy under different missing modes.

[0006] 2. Lack of physical rationality constraints: Most reconstruction methods only focus on the accuracy of data fitting and do not incorporate the dynamic characteristics of the wind turbine transmission chain. This easily leads to unreasonable results that violate the laws of energy transfer and the consistency of vibration sources, and cannot meet the core requirements of industrial monitoring for the physical reliability of data.

[0007] 3. Rigid weight allocation mechanism: Traditional fusion methods often adopt fixed weights or simple weighting strategies, without considering differences in working conditions, missing patterns and the historical performance of expert models. This makes it difficult to dynamically adapt to different reconstruction tasks and fully leverage the advantages of each model.

[0008] 4. Incomplete coverage of missing scenarios: Existing technologies are mostly limited to the simulation of missing patterns at random points, and do not adequately consider scenarios such as continuous block missing caused by short-term sensor failures or channel loss caused by complete disconnection. The training dataset lacks specificity, and the model's generalization ability is limited. Summary of the Invention

[0009] To overcome the aforementioned deficiencies of the prior art, this invention provides an intelligent reconstruction method for missing wind turbine vibration data based on a deep network structure, which addresses the problems mentioned in the background art through the following approach.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent reconstruction of missing wind turbine vibration data based on deep network structures, comprising: S1: Data preprocessing and knowledge base construction: Collect wind turbine vibration data and operating condition data and perform normalization processing. Design three types of missing patterns to construct training and validation sets. Construct a normal operating condition vibration feature memory base based on wind turbine normal operation data. S2: Construction and pre-training of heterogeneous deep network expert groups: Construct three types of heterogeneous deep network expert groups: time series prediction expert, physical simulation expert, and contrastive learning expert. Each expert network outputs its own reconstruction results and corresponding high-level features. Pre-training is completed based on the training set and validation set. S3: Construction and Training of Dynamic Arbitration Fusion Network: Construct a dynamic arbitration fusion network, input expert high-level features and meta-task features, quantify the differences in expert results through a disagreement detection mechanism and output disagreement indicators, and complete the training using a meta-learning paradigm to obtain dynamic weight allocation capabilities. S4: Physical verification implementation: Construct a physical rule base containing energy flow constraints and coherence constraints, trigger verification based on divergence flags, verify and correct the physical rationality of expert reconstruction results, and screen effective reconstruction results; S5: Final Fusion Output and Performance Evaluation: Based on dynamic weights and effective reconstruction results, weighted fusion is performed according to scenarios where expert results are consistent or divergent. The final reconstruction results that meet the requirements of wind turbine vibration monitoring are evaluated through three types of quantitative indicators: reconstruction accuracy, physical rationality, and real-time performance.

[0011] The technical effects and advantages of this invention are as follows: 1. Heterogeneous expert groups achieve multi-dimensional complementarity: Three types of heterogeneous network experts are constructed: time series prediction, physical simulation and comparative learning. They exert their efforts from three dimensions: spatiotemporal dependency capture, physical law constraint and historical normal state matching, respectively, to accurately adapt to different missing modes and greatly improve the reconstruction adaptability in complex scenarios.

[0012] 2. Physical verification ensures the credibility of results: A physical rule base containing energy flow constraints and coherence constraints is built. Through quantitative verification using the dynamic equations of the wind turbine drive chain and coherence functions, divergent results are adaptively corrected to ensure that the reconstructed data meets both the data accuracy requirements and the physical laws of industrial operation.

[0013] 3. Dynamic Arbitration Enhances Integration and Intelligence: The dynamic arbitration network is trained based on the meta-learning paradigm, integrating the characteristics of experts at the senior level with meta-task characteristics such as working conditions and missing patterns. Through dynamic time warping, expert disagreements are detected, and weights are dynamically allocated to fully leverage the advantages of each expert and adapt to diverse reconstruction scenarios.

[0014] 4. Full-process optimization ensures the practicality of the project: From the data preprocessing stage to build three types of missing pattern datasets and normal operating condition feature memory, to scenario-based fusion output and multi-dimensional evaluation, a complete technical chain is formed, taking into account reconstruction accuracy, physical rationality and real-time performance, to meet the needs of wind turbine vibration monitoring projects. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0016] Figure 2 This is a schematic diagram of the S1 process of the present invention.

[0017] Figure 3 This is a schematic diagram of the S2 process of the present invention.

[0018] Figure 4 This is a schematic diagram of the S3 process of the present invention.

[0019] Figure 5 This is a schematic diagram of the S4 process of the present invention.

[0020] Figure 6 This is a schematic diagram of the S5 process of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] refer to Figures 1-6 The intelligent reconstruction method for missing wind turbine vibration data based on deep network structure shown includes: S1: Data Preprocessing and Knowledge Base Construction By collecting vibration and operating condition data and performing normalization processing, a spatiotemporal dataset that can be directly used for model training is formed; three types of missing data patterns are designed for scenarios with missing vibration data, and a supervised training set is constructed; at the same time, features are extracted from long-term normal operation data to establish a normal operating condition memory bank; the details are as follows: S101: Wind Turbine Vibration and Operating Condition Data Acquisition and Construction of Standardized Spatiotemporal Dataset: Taking the wind turbine gearbox and main shaft as key components for monitoring, the placement of vibration sensors, sampling frequency, and specific operating condition parameters to be collected by the SCADA system were determined. Data covering four typical operating conditions—rated speed, variable speed, start-up, and shutdown—was collected over six months. Through time alignment, wavelet threshold denoising, and Z-score normalization, invalid data was removed and the data format was standardized, ultimately forming a standardized spatiotemporal dataset containing vibration sequences and operating condition label vectors, as detailed below: Data acquisition: Vibration data from the fan gearbox and main shaft are used as the source of vibration data. Two types of data are collected: the collection period is set to 6 months, covering four typical operating conditions: rated speed, variable speed, start-up, and shutdown. Vibration monitoring system data: Three-dimensional acceleration sensors are installed at key locations on the gearbox input shaft, intermediate shaft, output shaft bearing housing, and gearbox housing, with a sampling frequency set to [value missing]. Collect vibration acceleration signals; SCADA system data: Simultaneously collects three operating parameters: fan speed, torque, and oil temperature, with the sampling frequency set to... ; Data alignment: Based on the timestamp of the SCADA system, the vibration data is aligned according to time granularity. The data is divided into segments, and each segment of vibration data is time-aligned with the SCADA operating condition parameters within the corresponding time window. Invalid data with mismatched timestamps or signal amplitudes exceeding the physical range are removed. Data denoising: The vibration sequence is processed using wavelet thresholding. The db4 wavelet basis is selected and the number of decomposition layers is set to 5. The threshold of each decomposition layer is calculated by the Birgé-Massart adaptive threshold function to remove high-frequency useless signals such as environmental noise and electromagnetic interference, while retaining the equipment state characteristics in the vibration signal. Data normalization: for the denoised vibration sequence Z-score normalization is used, and the formula is: Among them, vibration sequence Dimensions , The total number of data segments. For the number of measurement points, The number of sampling points for each segment of vibration data. , The mean of the vibration sequence. The standard deviation of the vibration sequence is calculated statistically based on complete historical vibration data to eliminate differences in vibration amplitude magnitudes at different measuring points and under different operating conditions. Label mapping or one-hot encoding is performed on SCADA operating condition parameters to form operating condition label vectors. Constructing a standardized spatiotemporal dataset ; where the working condition label vector Dimensions , For the number of operating condition categories, here This corresponds to four typical working conditions; S102: Design of Missing Patterns in Wind Turbine Vibration Data and Construction of Supervised Training Set: Addressing the actual missing data scenarios of wind turbine vibration data, three missing patterns were designed: random points, continuous blocks, and complete channel loss. Missing data was artificially synthesized on complete vibration data segments of the normalized dataset, forming data pairs containing missing data labels and original complete data. The training and validation sets were divided in an 8:2 ratio to provide a supervised learning foundation for the subsequent training of a heterogeneous deep network expert group, ensuring that the model can specifically learn the reconstruction rules of different missing patterns. The specific process is as follows: Missing Pattern Definition: For actual missing data scenarios in wind turbine vibration data, three typical missing patterns are artificially synthesized within the complete vibration data segment of dataset D: Random point missing: for each vibration sequence According to the proportion of missing items Randomly select data points and set them as missing values ​​to simulate occasional packet loss during data transmission. The proportion of missing random points covers the common probability range of occasional packet loss in data transmission; Continuous block missing: for each vibration sequence , Generate length Each consecutive data block from a sampling point is set as a missing value, and the number of consecutive blocks is [number missing]. indivual, The number of consecutive missing blocks is used to simulate multiple intermittent short-term sensor failures. The length of the continuous missing block corresponds to the typical duration of a short-term sensor failure; the time corresponding to 10 sampling points. 100 sampling points correspond to 3.9ms; All channels lost: Random selection Each vibration sensor channel is set to a missing value for all its corresponding data segments to simulate a complete sensor disconnection. This is the total number of lost channels, covering scenarios where one or more sensors are simultaneously disconnected. Training set generation: For each synthetic missing mode vibration data segment, retain the vibration data containing the missing markers. Compared with the original complete vibration data , To reconstruct the true values, complete vibration data segments without missing data or fault alarms, which have only undergone basic screening and have not been processed, are collected and used to calculate the model reconstruction error; according to the division ratio Divide all data pairs into a supervised training set With the validation set This is used for training and performance validation of subsequent deep network models. The ratio of training set to validation set is 8:2, which is the ratio of training set to validation set in machine learning, ensuring sufficient training data and reliable validation results. S103: Construction of the Vibration Feature Memory Database for Normal Operating Conditions of Wind Turbines: Long-term normal vibration data that meet specific stable operating conditions are selected from the normalized dataset. These data are then segmented into fixed time windows, and time-domain and frequency-domain features are extracted and stored in a structured format to construct the normal operating condition memory database. The specifics are as follows: Normal data filtering: Filter vibration data segments from dataset D that represent long-term normal operation of the wind turbine. The filtering criteria are: no fault alarms in the SCADA system and no speed fluctuations. Oil temperature The screening period is 3 months. This is the rated speed of the fan.

[0023] Vibration feature extraction: For the filtered normal vibration data, feature extraction is performed within a time window consistent with the data segmentation granularity. Segmentation, extracting two types of features within each window: Time-domain characteristics: mean It reflects the average amplitude level and variance of the vibration signal. This reflects the dispersion and peak value of the vibration signal. Reflects the maximum amplitude and peak-to-peak value of the vibration signal. It reflects the amplitude fluctuation range and kurtosis of the vibration signal. Reflects the pulse characteristics and skewness of the vibration signal This reflects the symmetry of the amplitude distribution of the vibration signal; Frequency domain characteristics: through Fast Fourier Transform Calculate power spectral density Extracting peak frequency harmonic amplitude ratio Frequency band energy ; Where f is the frequency, in Hz. Reflects the energy distribution of the vibration signal at different frequencies; the peak frequency is The frequency corresponding to the maximum amplitude corresponds to the main vibration source frequency. The fundamental frequency amplitude, The first harmonic amplitude is the value of the frequency. The harmonic amplitude ratio reflects the harmonic characteristics of the vibration source. , These are the upper and lower limits of the frequency band, reflecting the vibrational energy within a specific frequency range; Memory storage: the extracted timestamp Operating condition labels eigenvectors A vibration characteristic memory M under normal operating conditions is constructed by storing data in a structured format; efficient retrieval is achieved using a Redis database, and the memory size is [size missing]. Feature fragments, The total number of feature fragments in the memory bank is 100,000, which can cover vibration features under different normal operating conditions, ensuring that similar normal states can be matched during retrieval, and providing historical normal state reference for experts in subsequent comparative learning.

[0024] S2: Construction and Pre-training of a Heterogeneous Deep Network Expert Group: Three types of functionally complementary heterogeneous deep networks are constructed as experts to reconstruct missing vibration data from three dimensions: temporal dependency capture, physical law constraint, and historical normal state comparison. For each expert, a network structure, loss function, and training parameters adapted to its functional positioning are designed. Supervised or semi-supervised training enables each expert to independently process missing data reconstruction, providing high-quality basic reconstruction results for subsequent dynamic arbitration fusion. The specific steps are as follows: S201: Construction and Training of the Spatiotemporal Transformer as a Temporal Prediction Expert: A spatiotemporal Transformer network is constructed as a temporal prediction expert. Through the design of an embedding layer, a multi-layer encoder (including a multi-head attention mechanism), and a decoder, it can simultaneously capture the temporal dependencies across time steps and the spatial correlations across sensors in vibration data. A mean squared error loss function is used, and 100 rounds of training are completed on a supervised training set, with the loss stabilization on the validation set as the termination condition. Ultimately, this expert is equipped with the ability to accurately reconstruct complex global fluctuation vibration data. The details are as follows: Network structure design: To capture the spatiotemporal dependencies of vibration data, a spatiotemporal Transformer network is constructed, with the input being vibration data containing missing information. Dimension B is the batch size, B=32, balancing training efficiency and memory consumption; For the time step, here =64, divide the vibration data of each segment of 25600 sampling points into time steps of 25600 / 64=400 sampling points; For the number of measurement points, =6, corresponding to 6 measuring points in the gearbox, with the specific structure as follows: Embedding layer: Maps the input vibration data into a high-dimensional feature vector, transforming the dimension into... ,in =64, For the embedding layer, 64 dimensions can strike a balance between capturing feature details and controlling computational cost; Encoder: Set the number of encoder layers =3 layers. Three layers effectively capture multi-scale spatiotemporal dependencies; too many layers can easily lead to overfitting. Each layer includes a multi-head attention mechanism and a feedforward network: the multi-head attention mechanism simultaneously calculates attention across time steps and attention across sensors; the number of neurons in the hidden layer of the feedforward network is set to... The activation function is Among them, the number of attention heads in a multi-head attention mechanism With a setup of 8 heads, different types of dependencies can be captured in parallel. Cross-timestep attention captures temporal dependencies at different times for the same measurement point, while cross-sensor attention captures spatial correlations at the same time for different measurement points. 256 neurons represent the number of neurons in the hidden layer of the feedforward network, which can adequately fit the feature mapping relationship. x represents the input value of a neuron in a certain layer of the neural network. Decoder: Determine the number of decoder layers. =2 layers. Two layers can accurately reconstruct the temporal structure of vibration data. Too many layers will increase the calculation delay. The first layer receives the encoder output features, and the second layer is the output of the last hidden layer, which is the high-level feature. Dimension It receives the spatiotemporal features output by the encoder and generates a composite image through the output layer. Vibration data reconstruction results with consistent dimensions ; Loss function definition: The mean squared error (MSE) loss is used to measure the difference between the reconstruction result and the true value. Its specific mathematical function is as follows: ,in, for The value at the t-th time step and the s-th measurement point. To correspond to the true value, MSE loss has an excellent effect on quantifying the reconstruction error of continuous vibration data; Pre-training process: For training data, To validate the data, the optimizer used was the Adam optimizer, with an initial learning rate set to... This allows the model to converge quickly and avoids initial oscillations. The learning rate decay strategy is cosine annealing, which reduces the learning rate by 10%. One cycle of decay. Using the cosine annealing cycle, 10 rounds can balance convergence speed and accuracy; let the training rounds be... =100 rounds, 100 rounds ensures the model fully learns the data patterns, when the validation set... continuous Stop training when the number of cycles does not decrease, and save the model parameters. The extraction interface ensures that the network has a stable ability to reconstruct time-series vibration data. To stop the cycle early, 10 cycles can effectively prevent overfitting.

[0025] S202: Construction and Training of Physics Simulation Expert: A Physical Information Neural Network (PINN) is constructed as the physics simulation expert. Vibration data features are extracted through a multi-layer fully connected hidden layer design. The core innovation lies in incorporating the simplified dynamic equations of the wind turbine drivetrain as a regularization term into the loss function. A composite loss function of mean squared error and physical regularization is adopted. After 80 rounds of training, the loss on the validation set is stabilized as the termination condition to ensure that the expert's reconstruction results meet basic dynamic constraints and avoid physically unreasonable outputs. The specific process is as follows: Network structure design: To ensure that the reconstruction results meet the physical constraints of the wind turbine drivetrain, a physical information neural network PINN is constructed. The input is the same as S201, and the output is the vibration data reconstruction results. The specific structure is as follows: Input layer: receiving Dimension ; Hidden layers: Set the number of hidden layers. Four layers can fully extract the physical features of vibration data, with the number of neurons being respectively... , , , , The number of neurons in each hidden layer is [number], decreasing layer by layer to achieve feature dimension compression and focus on key information. The activation function is [function name]. The fourth layer, which is the last hidden layer, outputs high-level features. Dimension ; Output layer: Receiver , generation and Dimensional consistency ; Physical regularization term design: Based on the dynamic characteristics of the wind turbine drive chain (main shaft, gearbox, and coupling), simplified physical equations are established: ,in, This is a mass matrix, in kg, reflecting the mass distribution of each component in the transmission chain; The damping matrix, in N·s / m, reflects the energy dissipation characteristics of the transmission chain. is the stiffness matrix, in N / m, reflecting the deformation resistance of the transmission chain; x is the vibration displacement vector, in m. This is the external excitation force vector, in N, reflecting the load excitation experienced by the transmission chain, where: , , The wind turbine drivetrain was modeled using ANSYS finite element analysis software, taking into account material properties such as the elastic modulus of steel. =206 GPa, Poisson's ratio =0.3, density =7850kg / m^3 was calculated; Rotation speed of SCADA system Torque The specific mathematical function for data calculation is: , Main spindle diameter, unit: m; The residuals of the physical equations are used as regularization terms, and their specific mathematical function is as follows: , for The displacement vector at time step t is obtained by integrating the vibration acceleration, and the regularization term ensures that the reconstruction result conforms to the dynamic law; Loss function and training: The total loss function is: , for and MSE, =0.1 is the regularization weight, determined through cross-validation. 0.1 balances the weights of data fitting and physical constraints; the training data and optimizer are the same as S201, and the training epochs are set to... =80 rounds. 80 rounds allow the model to fully fit the data while satisfying physical constraints, when the validation set... continuous Stop after 8 rounds without any decrease, save model parameters and... Extraction interface; S203: Construction and Training of a Deep Memory Network for Contrastive Learning Experts: A deep memory network is constructed as a contrastive learning expert. Through a CNN feature extractor, a cosine similarity-based memory addressing module, and an LSTM decoder, it is designed to retrieve relevant normal feature fragments from a normal operating condition memory bank based on the context of missing data and then fuse and reconstruct them. A composite loss function of mean squared error and contrastive loss is used. 60 rounds of training are completed, with validation set loss stability as the termination condition, ensuring that the expert's reconstruction results closely approximate historical normal conditions and providing stability guarantees. The specific details are as follows: Network Structure Design: A deep memory network is constructed with the goal of reconstructing missing data based on historical normal states. The input is... Contextual features, before and after the missing segment Vibration data at each time step, The context time step is 32 steps, which provides sufficient contextual information for retrieval. The output is the reconstruction result. The specific structure is as follows: Feature extractor: Assume the number of CNN layers Two layers can effectively extract local correlations of contextual features, and the kernel size is... A 3×3 approach can balance the feature extraction range and computational cost, with a step size of [missing information]. ,filling To ensure that the input and output dimensions are consistent, the activation function is... Transform the context data into a low-dimensional feature vector Q, with dimension... , For querying feature dimensions, 128 dimensions can fully represent contextual information; Memory addressing module: calculates the feature vector Q and all feature vectors in the normal operating condition memory M. The cosine similarity, specifically its mathematical function, is: , Let M be the i-th eigenvector. Cosine similarity can effectively measure the similarity between vectors. Select the top-K similarity vectors (top-K=10) to form a relevant memory subset. K represents the number of search features; 10 features can balance search accuracy and computational efficiency. Decoder: Set the number of LSTM layers Two layers can effectively handle the fusion of temporal features, and the number of hidden layer neurons... 128 neurons can adequately fit temporal fusion relationships; the second layer, i.e., the last hidden layer, outputs high-level features. Dimension ; with Q and Fusion, generated through the output layer and Dimensional consistency ; Loss function and training: Contrastive loss is used, and its specific mathematical function is as follows: , To compare the loss weight, 0.05 can ensure that the reconstruction result is close to the historical normal state; for The feature vectors in the data; the optimizer used is the AdamW optimizer, with a weight decay coefficient. A learning rate of 0.01 can effectively prevent overfitting. The training rounds are set to Rounds, 60 rounds, allow the model to fully learn the associations of historical normal features, when the validation set... continuous Stop when the wheel stops descending, save model parameters and Extraction interface.

[0026] S3: Construction and Training of the Dynamic Arbitration Fusion Network: A lightweight dynamic arbitration fusion network is constructed. By designing a composite input that includes expert high-level features and meta-task features, accurate adaptation to different missing scenarios is achieved. A disagreement detection mechanism based on dynamic time warping distance is introduced, combined with a meta-learning training paradigm, enabling the network to learn dynamic weight allocation from experience. The output is a fusion reconstruction result that takes into account both expert advantages and scenario adaptability, while identifying expert disagreements to trigger subsequent physical verification. The specific steps are as follows: S301: Arbitration Network Structure and Input Design: A lightweight feedforward network is designed as the arbitration network. A composite input is defined, containing high-level features from three experts and meta-task features. Through the design of hidden and output layers, the network can simultaneously output dynamic weights and pre-featured divergence markers, providing a foundation for subsequent divergence detection and weight allocation. This ensures that the input information fully reflects the reconstructed scene and expert state. Its specific construction is as follows: Network Structure: With the goal of dynamically integrating expert reconstruction results, a lightweight feedforward network is constructed as the arbitration network G, with the following specific structure: Input layer: Receives arbitration input ,in, These are the outputs of the last hidden layer of the three expert networks S201-S203, which are then unified through feature concatenation. ; The feature vector for the meta-task has the following dimensions: , Includes: operating condition label (one-hot encoding), missing mode (random point = 1, continuous block = 2, channel total loss = 3), and missing ratio. Recent confidence level of experts (past =The reciprocal of the MSE of expert output and truth value in 100 tasks To ensure the reliability of the confidence level, the number of tasks is counted and normalized to 0-1. Hidden Layers: Determine the number of hidden layers in the arbitration network. The number of neurons are respectively , The activation function is ; Output layer: Outputs two types of results: dynamic weights ; dimension ,pass Normalization, satisfy ; Pre-disagreement characteristics Dimension This provides characteristic support for the judgment of discrepancies.

[0027] Input preprocessing: Normalize each part: The mean value is calculated using the S201 training set. with standard deviation , The mean value is calculated using the S202 training set. with standard deviation , The mean value is calculated using the S203 training set. with standard deviation ; Min-Max normalization To reach [0, 1], avoid the influence of feature magnitude differences on decision-making; S302: Design of the divergence detection mechanism: Based on the dynamic time warping (DTW) distance, the divergence detection mechanism quantifies the difference by calculating the DTW distance of the reconstruction results of the three experts in S2, and judges the consistency of the expert results by combining the adaptive threshold corresponding to the meta-task features, and outputs a divergence flag to determine whether to trigger physical verification; the specific process is as follows: Dynamic Time Warping (DTW) Distance Calculation: To quantify the differences in reconstruction results from three experts, the DTW distance between any two experts' outputs is calculated. Data preprocessing: Reconstructing the results output from S201-S203 , , Unfold into a one-dimensional sequence at time steps , , After expansion, the dimensions are [B*384]; Distance Matrix Construction: For any two experts' expansion sequences, construct the Euclidean distance matrix: ; Minimum cumulative distance calculation: The minimum cumulative distance of the distance matrix is ​​calculated using a dynamic programming algorithm, i.e. Finally, three sets of distances were obtained: , , .

[0028] Adaptive threshold Determined: Based on meta-learning support set data, from For the 12 task samples extracted, the 95th percentile of the corresponding DTW distance was calculated for each combination of meta-task features. This percentile was used as the divergence judgment threshold for that meta-task. Construct a mapping table containing 12 types of thresholds; Divergence indicator: For the current reconstruction task, calculate the maximum value of the three DTW distances. ; Query the mapping table to obtain the corresponding metata task. ,like Then output the divergence flag. Conversely, output a divergence flag. .

[0029] S303: Meta-learning training method: The arbitration network is trained using a meta-learning paradigm. Task units are constructed by combining missing modes and working conditions. The network learns task patterns through support sets and verifies the effect through query sets. The network parameters are optimized using query set loss to ensure that the arbitration network has the ability to dynamically allocate weights under different missing scenarios. The specific process is as follows: Task Construction: Twelve training tasks were constructed using a combination dimension of 3 missing patterns × 4 working conditions; each task started from... 150 samples were drawn from the dataset and divided into a support set S (100 samples, used to learn task features) and a query set Q (50 samples, used to verify the learning effect) in a 100:50 ratio. Training process: Support set learning phase: Input the support set S into the S201-S203 expert network to obtain the reconstruction results. , , Characteristics of high-level personnel , , ; Calculate the meta-task features of the support set Includes working condition labels, missing patterns, missing percentages, and recent expert confidence levels; splicing. , , and To form a support set for arbitration input ;Will Input the arbitration network G to obtain the dynamic weights of the support set. ; Computation supports integrated output ,by With support set truth value MSE is loss Perform a one-step gradient descent on G to update the network parameters in order to learn the weight distribution pattern of the task. Query set validation phase: Input the query set Q into the S201-S203 expert network to obtain the reconstruction results. , , Characteristics of high-level personnel , , ;Calculate the meta-task features of the query set , forming a query set arbitration input ;Will Input the updated G to obtain the dynamic weight of the query set. With fusion output ; Calculate query set loss , To query the truth value of the set.

[0030] Global optimization phase: Traverse all 12 training tasks and accumulate the results of all tasks. Total loss The Adam optimizer is used, with a learning rate of [missing information]. Perform a global update on the parameters of G to complete one task iteration; Training Termination: Repeat the above training process for a total of 500 task iterations; calculate the total loss after each iteration. If 50 consecutive iterations No decrease (decline < 10⁻) 6 If the condition is met, training is stopped, and the parameter file of the arbitral network G, including weights, biases, and normalized parameters, is saved.

[0031] S4: Physical Verification Implementation: Based on the energy transfer characteristics of the wind turbine drivetrain and the principle of vibration source consistency, a quantifiable physical rule base is constructed. When S3 outputs a divergence flag, the physical rationality of the expert reconstruction results is verified through energy flow constraints and coherence constraints. Results that do not conform to the rules are adaptively corrected to ensure that the final output meets the physical laws of wind turbine operation. The specific process is as follows: S401: Construction of the Physical Rule Base for Wind Turbine Vibration Data: Based on the engineering physical characteristics of the wind turbine drivetrain, two types of core physical rules are constructed: energy flow constraint rules clarify the frequency band energy transfer relationship between upstream and downstream measuring points, and coherence constraint rules define the degree of correlation of signals from the same vibration source; all rule parameters are determined based on statistical data from normal operating conditions to ensure the engineering applicability and computational operability of the rules, as detailed below: Rule 1, Energy Flow Constraint Rule: Based on the characteristic that energy is transferred unidirectionally from input to output in the wind turbine drive chain and that there are losses, it is stipulated that the energy in a specific frequency band at the upstream measuring point is greater than or equal to the energy in the same frequency band at the downstream measuring point. The specific definition is as follows: Upstream and downstream relationship of measuring points: Based on the wind turbine transmission chain structure: main shaft → gearbox input shaft → gearbox output shaft → coupling, the energy transmission sequence of measuring points is determined. The measuring point in the previous link is upstream, and the measuring point in the next link is downstream. Meshing frequency bandwidth determination: Based on the gear parameters of the gearbox, the number of gear teeth z is known, combined with the real-time speed of the SCADA system. (Unit: r / min), calculate the meshing frequency (Unit: Hz); The bandwidth is defined as 50 Hz before and after the meshing frequency. Hz, covering the meshing frequency and its 1st and 2nd harmonics; Energy calculation and constraint relationship: The power spectral density (PSD) of the vibration data at the measurement point within the above bandwidth is calculated using Fast Fourier Transform (FFT), and the frequency band energy is obtained by integrating the PSD. The specific mathematical function is as follows: , , Set the energy transfer efficiency coefficient. Based on three months of normal operating data, this reflects energy losses due to friction and vibration during transmission, with the following constraint relationships: , For the upstream measuring point frequency band energy, Energy in the same frequency band as downstream measuring points; Rule 2, Coherence Constraint Rule: Based on the physical principle that vibration signals generated from the same source have high correlation at characteristic frequencies, the specific definition is as follows: Coherence function calculation: For two measurement points from the same vibration source, at the characteristic frequency... Calculate the coherence function and take the meshing frequency. Or bearing failure characteristic frequency, here taken as For example, its specific mathematical function is: ;in For the two measuring points at The cross-power spectral density at a given point reflects the degree of signal correlation. , The two measuring points are respectively at The self-power spectral density at a given location reflects the energy distribution of a single signal; Constraint threshold setting: Based on statistical data from normal operating conditions, the coherence function calculation results of 3 months of normal data are used, and the 80th percentile of the coherence function is taken. As a constraint threshold; when If the signals from the two measuring points originate from the same vibration source, the reconstruction result will be consistent with the vibration source consistency; otherwise, it will not be consistent.

[0032] S402: Reconstruction Result Verification and Correction Process: Based on the physical rules constructed in S401, the verification logic after divergence is clarified: First, the energy rationality of the downstream measuring point reconstruction results is verified through energy flow constraints, and unreasonable results are corrected. Then, the consistency of the vibration source is verified through coherence constraints. Finally, expert results that simultaneously satisfy both types of rules are output; as detailed below: Triggering condition: When S302 outputs the branching flag. At that time, the physical verification step will be automatically invoked, and the input data will be the reconstruction results of the three experts in S2. , , And vibration data from upstream measuring points collected in S1 that are not missing. and the speed under current operating conditions Number of gear teeth z; Energy Flow Constraint Verification and Correction: Real Upstream Energy Calculation: Extracting Vibration Data from Upstream Measuring Points (without missing data) from any set of upstream and downstream measuring points. The signal is converted to a frequency domain signal using FFT, and the bandwidth of the signal at the corresponding meshing frequency is calculated. By integrating the PSD within Hz, the true frequency band energy of the upstream measurement point can be obtained. ; Calculation of reasonable upper limit of downstream energy: According to the energy flow constraint rules, the reasonable upper limit of frequency band energy at the corresponding downstream measuring point is calculated using the formula... calculate; Expert Result Energy Validation: Reconstruction results from three experts for the aforementioned downstream measuring points were extracted. , , Using the same method as the upstream measurement points, the frequency band energy of each reconstruction result within the same meshing frequency bandwidth was calculated. , , The criteria for determining significant exceedance of the upper limit are set as follows: Of this, 10% is the allowable error range for engineering, used to avoid misjudgments caused by minute energy fluctuations. If an expert's reconstruction result meets this condition, it is judged to be physically unreasonable. Adaptive correction: For expert reconstruction results deemed physically unreasonable, an amplitude scaling method is used for correction. The correction formula is as follows: By scaling the amplitude of the reconstructed result proportionally, the corrected bandwidth energy is made exactly equal to a reasonable upper limit, ensuring that the energy flow constraint is met; if the bandwidth energy of the expert reconstruction result satisfies... Then the original reconstruction result will be retained. constant; Coherence constraint verification: Vibration source measurement point selection: Select two measurement points corresponding to the same vibration source from the measurement point layout of the wind turbine vibration monitoring system; Coherence function calculation: For the three-expert reconstruction results verified or corrected by energy flow constraints, the reconstructed data corresponding to the two vibration source measurement points are extracted respectively, and the two sets of data are calculated at the characteristic frequency. The coherence function at the corresponding gear meshing frequency. ; Consistency determination: If the coherence function of an expert's reconstruction result satisfies If the result meets the requirements, the vibration source consistency is determined; otherwise, it is determined that the consistency is not met. Results selection: Based on the verification results of both energy flow constraints and coherence constraints, expert reconstruction results that satisfy both types of rules are retained; if there are reconstruction results that only satisfy a single rule, results that satisfy energy flow constraints are retained first.

[0033] S5: Final Fusion Output and Performance Evaluation: Based on the expert reconstruction results after S3 dynamic arbitration weights and S4 physical verification, the final fusion output for each scenario is achieved. Simultaneously, quantitative evaluation indicators are designed from three dimensions: reconstruction accuracy, physical rationality, and real-time performance, ensuring that the output results meet the engineering application requirements of wind turbine vibration monitoring. The specific process is as follows: S501: Scenario-Specific Final Fusion Strategy: Two fusion strategies are designed to address the divergence indicators in the S3 output: when expert results are consistent, fusion is directly performed based on arbitration weights; when expert results diverge, weights are reallocated based on physical verification screening results, ensuring the final output combines accuracy and physical plausibility. Details are as follows: Scenario 1: Divergence Symbol The experts' conclusions were consistent: Input data: Original reconstruction results from three experts in S2 , , and the dynamic weights output by S3 ; Fusion calculation: The final reconstruction result is calculated using a weighted summation method, and the formula is as follows: ; Output: Direct output and the corresponding weight allocation report; Scenario 2: Divergence Symbol Experts disagreed on the results, but the S4 physical verification has been passed. Input data: Valid expert reconstruction results after S4 filtering, set as , A maximum of 3, a minimum of 1, and the original weight W of S3; Weight redistribution: If the number of valid results is K, Then, the weights of ineffective experts in the original weights are proportionally allocated to effective experts, and the specific mathematical function is as follows: ; Fusion calculation: A weighted summation method is used, the formula is as follows: ; Output: Output Physical verification report and weighted report after reallocation.

[0034] Special handling: If only one valid expert result remains after S4 screening, that result will be used directly. The output report should indicate that there is only one valid expert result and no fusion operation.

[0035] S502: Design of Quantifiable Performance Evaluation Metrics: Three types of quantifiable performance evaluation metrics are designed to measure the accuracy, physical rationality, and real-time performance of the reconstruction results, ensuring the objectivity and practicality of the evaluation results; the specifics are as follows: Reconstruction accuracy metrics: Metric 1: Root Mean Square Error (RMSE), measures the overall deviation between the reconstructed results and the actual data. Its specific mathematical function is: ;in This is real vibration data without any missing data; the smaller the RMSE value, the higher the reconstruction accuracy.

[0036] Indicator 2: Mean Absolute Error (MAE), measures the average deviation of the reconstruction results, reducing the impact of extreme values. Its specific mathematical function is: The smaller the MAE value, the better the stability of the reconstruction results. Physical rationality index: Physical rule satisfaction rate (RPR), which measures the proportion of the final reconstruction result that conforms to the S4 physical rules, is calculated using the following formula: The rule requirement is that both energy flow constraints and coherence constraints are met simultaneously; an RPR ≥ 90% is considered physically reasonable. Real-time performance metric: Single-sample processing latency (T), which measures the time from input vibration data containing missing values ​​to output. Total execution time, in milliseconds (ms); test environment configuration: CPU (Intel Core i7-12700K) and GPU (NVIDIA RTX 3090), batch size B=32; requirements This meets the engineering requirements of second-level response for wind turbine vibration monitoring systems.

[0037] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent reconstruction of missing fan vibration data based on a deep network structure, characterized in that, Comprise: S1: data preprocessing and knowledge base construction: collect fan vibration data and working condition data and perform standardized processing, design three types of missing patterns to construct training set and validation set, construct normal working condition vibration feature memory based on fan normal operation data; S2: construction and pre-training of heterogeneous deep network expert group: construct three types of heterogeneous deep network experts including time series prediction expert, physical simulation expert and comparative learning expert, each expert network outputs its own reconstruction result and corresponding high-level feature, and pre-training is completed based on the training set and the validation set; S3: construction and training of dynamic arbitration fusion network: construct a dynamic arbitration fusion network, input expert high-level features and meta-task features, quantify expert result differences through a divergence detection mechanism and output a divergence flag, and complete training to obtain dynamic weight allocation capability using a meta-learning paradigm; S4: physical verification: construct a physical rule base containing energy flow constraints and coherence constraints, trigger verification based on the divergence flag, perform physical reasonableness verification and correction on the expert reconstruction results, and select valid reconstruction results; S5: final fusion output and performance evaluation: based on dynamic weights and valid reconstruction results, weighted fusion is performed according to expert result consistency or divergence scenarios, and the final reconstruction result meeting the fan vibration monitoring requirements is output through three types of quantitative indicators such as reconstruction accuracy, physical reasonableness and real-time performance.

2. The fan vibration data missing intelligent reconstruction method based on a deep network structure according to claim 1, characterized in that: The data preprocessing and knowledge base construction comprises: Data acquisition and standardized processing: collect vibration data of each measuring point of the fan and corresponding SCADA system working condition data, align the vibration data and working condition data according to the time stamp, denoise the vibration data using wavelet threshold algorithm, map the vibration data to the [0, 1] interval through Z-score normalization, and perform one-hot encoding on the working condition data to form a spatio-temporal data set; Data set construction: design three types of missing patterns: random point missing, continuous block missing and channel complete missing, synthesize missing data on the complete vibration data of the spatio-temporal data set according to the preset proportion, wherein the random point missing ratio is 5%-30%, the continuous block missing length is 10-100 sampling points, and the supervised training set and validation set are obtained in an 8:2 ratio; Memory bank construction: select fan continuous 7-day fault-free normal operation data, extract time domain features and frequency domain features of the data, time domain features include peak value, mean value, variance, frequency domain features include meshing frequency amplitude and harmonic amplitude, store the features according to the working condition to construct a normal working condition vibration feature memory bank, and the memory bank is updated based on the newly added normal data every month.

3. The fan vibration data missing intelligent reconstruction method based on a deep network structure according to claim 2, characterized in that: The time series prediction expert comprises: An spatio-temporal Transformer network is used, the input is the normalized fan vibration data with missing data, the network sequentially includes an embedding layer, an encoder and a decoder; the embedding layer maps the input vibration data to a high-dimensional feature vector, the encoder has 3 layers and each layer contains 8 heads of multi-head attention mechanism and feedforward network, the decoder has 2 layers, the last hidden layer of the decoder outputs high-level features, and the decoder output layer generates a reconstruction result consistent with the dimension of the complete vibration data; With supervised training set as training data, validation set as validation data, mean square error as loss function, loss calculation focuses on the deviation of reconstruction results and real data at each time step and each measuring point; select Adam optimizer, set the initial learning rate to 10 -4 , adopt cosine annealing learning rate decay strategy every 10 rounds The maximum training round is set to 100 rounds, and the training is stopped when the validation set loss does not decrease for 10 consecutive rounds, and the network parameters and high-level feature extraction interface are saved.

4. The fan vibration data missing intelligent reconstruction method based on a deep network structure according to claim 3, characterized in that: The physical simulation expert comprises: The physical information neural network is adopted, the input is the normalized vibration data containing missing data, the network sequentially comprises an input layer, four full-connection hidden layers and an output layer, the number of hidden layer neurons is set to 256, 128, 64 and 32 in sequence, the activation function is all ReLU, the fourth hidden layer is the last hidden layer, the output is high-level features integrating the physical constraint information of the fan transmission chain, and the output layer receives the high-level features and generates a reconstruction result matching the dimension of the complete vibration data; Based on the dynamic characteristics of the fan transmission chain, a dynamic equation is established, the equation residual is taken as a physical regularization term, and the regularization weight is set to 0.1 to balance data fitting and physical constraints. With the training set, the validation set as data, select Adam optimizer, initial learning rate 10 -4 , set the maximum training rounds 80 rounds; when the validation set total loss continuous 8 rounds no decline, stop training, save network parameters and high layer feature extraction interface.

5. The fan vibration data missing intelligent reconstruction method based on a deep network structure according to claim 4, characterized in that: The contrast learning expert comprises: The deep memory network is adopted, the normalized vibration data containing missing data is taken as the input, and the network comprises a feature extraction module and a contrast matching module; the feature extraction module is provided with three layers of LSTM networks, the second layer of LSTM hidden layer outputs high-level features containing information associated with normal working conditions, the contrast matching module receives the high-level features and combines the normal working condition vibration feature memory bank to complete feature matching, and outputs a reconstruction result consistent with the dimension of the complete vibration data; Normal features of the same working condition as the current reconstruction task are retrieved from the normal working condition vibration feature memory bank as positive samples, and normal features of different working conditions are randomly selected as negative samples, the cosine similarity of the features of the data to be reconstructed and the positive and negative samples is calculated, a triplet loss function is used to minimize the distance between the features to be reconstructed and the positive samples and maximize the distance between the features to be reconstructed and the negative samples, and the triplet loss weight is set to 0.2, and the mean square error loss of the reconstruction result is combined into a total loss; Take the training set as input, and evaluate the effect of the verification set. Select Adam optimizer, initial learning rate 10 -4 Learning rate decay is performed every 15 rounds, decay coefficient 0.8; set the maximum training rounds to 90 rounds, stop training when the total loss of the verification set decreases by less than 10 -6 times for 10 consecutive rounds, save the network parameters and high-level feature extraction interface.

6. The fan vibration data missing intelligent reconstruction method based on a deep network structure according to claim 5, characterized in that: The dynamic arbitration fusion network is constructed, comprising: The light full-connection network is adopted, the input comprises high-level features output by the three types of expert networks and meta-task features, the meta-task features comprise working condition labels converted from working condition data, preset missing mode labels, missing proportion quantitative values and expert confidence calculated from the historical reconstruction accuracy of the expert network; the network is provided with three hidden layers, the number of neurons is 128, 64 and 32 in sequence, and the output layer outputs dynamic weights and a divergence flag corresponding to the three types of experts; The reconstruction results of the three types of experts are extracted, the dynamic time warping algorithm is used to calculate the distance value of any two expert results, and the maximum distance value is taken as a difference quantitative index; based on the normal working condition data, a difference threshold is statistically obtained, and when the quantitative index exceeds the threshold, the divergence flag is output as 1, otherwise as 0; The training set is used to construct a missing mode and working condition combination task unit, and each unit is divided into a support set and a query set; an Adam optimizer is used, and an initial learning rate is 5×10 -5 The mean square error of the fusion result after weight allocation and the true data is used as the loss; the maximum training round is set to 60 rounds, the training is stopped when the query set loss does not decrease for 8 consecutive rounds, and the network parameters are saved.

7. The method according to claim 6, wherein the method is characterized in that: The divergence detection mechanism comprises: The complete reconstruction results output by the time series prediction expert, the physical simulation expert and the contrast learning expert are extracted, each reconstruction result is preprocessed by time series length alignment and amplitude normalization, and the missing-free normal vibration data of the same working condition as the current reconstruction task in S1 are synchronously retrieved as a reference benchmark; The dynamic time warping algorithm is used to calculate the temporal distance between any two types of expert reconstruction results. The local distance metric of the DTW algorithm is Euclidean distance, and the window constraint is set to 10% of the temporal length of the reconstruction result. The arithmetic mean of the three sets of temporal distances is taken, and then compared with the mean distance between the reference benchmark and the expert results to obtain a normalized difference quantification index. Based on the vibration characteristic memory of normal working conditions, 1000 sets of normal data without missing data were randomly selected. The reconstruction results were generated through three types of expert networks and the difference quantification index was calculated. The 95th percentile of the statistical index was used as the divergence judgment threshold. The threshold was stored according to the working condition type to adapt to different operating scenarios. The difference quantification index of the current reconstruction task is compared with the judgment threshold of the corresponding working condition. If the index is greater than the threshold, the divergence flag 1 is output, indicating that there is a divergence in the expert results; if the index is less than or equal to the threshold, the divergence flag 0 is output, indicating that the expert results are consistent. The flag is output to S4. If the flag is 1, the physical verification process is triggered; otherwise, the physical verification process is skipped.

8. The method according to claim 6, wherein the method is characterized in that: The meta-learning paradigm includes: Based on the supervised training set, a meta-task unit is constructed by combining three types of missing patterns and four types of wind turbine operating conditions. Each unit is divided into a support set for scene feature learning and a query set for weight verification. The missing patterns and operating condition types of the support set and the query set are consistent. A dual-loop mechanism is adopted, with the inner loop supporting the updating of local network parameters and the outer loop querying the set to calculate the fusion loss and update the global parameters. The inner loop loss is the mean squared error of the support set fusion result, and the outer loop loss is the weighted sum of the mean squared error of the query set and the entropy value of the weight allocation. Test with validation set after every 10 outer loop iterations, when the RMSE of the validation set fusion result is continuously decreasing and the decrease is less than 10 -5 times, the adaptability is determined to be up to standard, until the training termination condition is met. 9.The method of claim 1, wherein the method further comprises: The physical verification implementation includes: Construct a physical rule base containing energy flow constraints and coherence constraints: energy flow constraints are based on the energy transfer law of the wind turbine drive chain, limiting the energy relationship of the characteristic frequency bands of upstream and downstream measuring points; coherence constraints take the vibration source correlation as the core, and clarify the standard of coherence coefficient of characteristic frequency of the measuring point group with the same vibration source. Triggering and Verification: When the divergence flag is 1, verification is triggered. The reconstruction results of each expert are substituted into the rule base to complete the energy flow and coherence compliance verification in sequence. Correction and screening: Adaptive correction is performed on non-compliant results with single constraints using amplitude adjustment or phase correction, and results that are non-compliant with both constraints are marked as invalid; results that satisfy both constraints after screening and correction are taken as valid reconstruction results, and a verification report is generated and transmitted to S5 simultaneously.

10. The method of claim 1, wherein the method is characterized by: The final fusion output and performance evaluation include: If the disagreement flag is 0, the weights output by the dynamic arbitration fusion network are directly used to sum the reconstruction results of the three types of experts to obtain a preliminary result; if the flag is 1, based on the effective reconstruction results screened by S4, the weights of invalid experts are allocated according to the proportion of historical reconstruction accuracy of effective experts, and then weighted fusion is performed to generate a preliminary result. Reconstruction accuracy is measured by root mean square error and mean absolute error, and the deviation between the preliminary results and the real data is calculated; physical rationality is indicated by the rate of physical rule satisfaction, and the percentage of measurement points that meet the constraints is statistically analyzed; real-time performance is tested by the processing delay of a single sample from input to output. When the accuracy index is lower than the preset threshold, the reasonableness is ≥95%, and the delay is <50ms, the preliminary result is deemed qualified and the final reconstruction result is output; if it is unqualified, it is fed back to S3 for fine-tuning of weights and then re-fused until the wind turbine vibration monitoring requirements are met.

Images