Wind power gear box online fault diagnosis method based on multi-source data fusion

Abstract

The invention discloses a wind power gear box online fault diagnosis method based on multi-source data fusion, and particularly relates to the field of mechanical fault detection, and the method comprises the steps: S1, collecting high-frequency dynamic, medium-frequency working condition and low-frequency thermal state data, and outputting a standardized data set through time alignment, quality verification and physical constraint verification; s2, performing multi-scale decomposition to retain a fault sensitive frequency band, inverting physical parameters such as gear contact stress and the like, constructing a physical cause and effect graph, and determining fault sensitive characteristics and threshold values; s3, constructing a multi-modal feature tensor, and obtaining a low-dimensional health representation vector through CP decomposition fusion, graph neural network reasoning and variational auto-encoder dimension reduction; s4, calculating a weight by using an entropy weight method, calculating a dynamic health degree in combination with a health benchmark, predicting a trend by using LSTM, and establishing a five-level health system; s5, judging a fault mode through double-layer identification, analyzing a root cause and formulating a hierarchical operation and maintenance suggestion; the method is based on multi-source fusion and data mechanism dual drive, and precise diagnosis and operation and maintenance guidance are achieved.

Description

Technical Field

[0001] This invention relates to the field of mechanical fault detection technology, and more specifically, to a method for online fault diagnosis of wind turbine gearboxes based on multi-source data fusion. Background Technology

[0002] As the core transmission component of a wind turbine generator set, the operating status of the gearbox directly affects the reliability and power generation efficiency of the entire unit. Due to long-term operation under complex conditions such as variable load, high speed, and strong impact, key components of the gearbox are prone to fatigue, wear, pitting, and other failures, which may lead to unplanned shutdowns of the unit and cause significant economic losses.

[0003] Currently, fault diagnosis of wind turbine gearboxes mainly relies on single or a few data sources such as vibration signal analysis, oil monitoring, and temperature detection, which can achieve fault identification to a certain extent. As the single unit capacity of wind turbines continues to increase and the operating environment becomes increasingly complex, the development of more accurate and intelligent online fault diagnosis technology has become an urgent need for the industry.

[0004] However, it still has some drawbacks in practical use, such as: 1. Poor coordination of multi-source data: Existing technologies often collect data such as vibration and temperature of wind turbine gearboxes separately, lacking time alignment and quality verification. Data from different sources have large time deviations, and invalid data is not eliminated through energy conservation and transmission chain kinematic verification, resulting in low reliability of subsequent diagnostic data, easy introduction of interference, and affecting diagnostic accuracy. 2. Lack of mechanistic support for feature extraction: Existing technologies mostly rely on data-driven feature extraction, without combining key parameters such as contact stress and dynamic load with the mechanical transmission mechanism of gearboxes, and without constructing physical causal relationships. The threshold calibration of fault-sensitive features is subjective, making it difficult to accurately capture early fault signals and easily leading to missed or misjudged cases. 3. Insufficient dynamic adaptability of health assessment: The existing technology calculates health with fixed weights, without adjusting according to operating conditions such as input shaft speed, and lacks time series models to predict health trends. It relies solely on static threshold assessment, which cannot provide timely warnings of health deterioration and is difficult to adapt to the complex and ever-changing operating conditions of wind turbine gearboxes. 4. Lack of specificity in maintenance recommendations: Existing technologies often provide general maintenance recommendations after fault diagnosis, without considering the fault mode, root cause, and health level classification. This leads to untimely maintenance response or over-maintenance, increasing equipment downtime and maintenance costs. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, this invention provides an online fault diagnosis method for wind turbine gearboxes based on multi-source data fusion, which addresses the problems mentioned in the background art through the following solutions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an online fault diagnosis method for wind turbine gearboxes based on multi-source data fusion, comprising: S1: Multiphysics Data Synchronous Acquisition and Preprocessing: Acquire three types of data: high-frequency dynamic, medium-frequency operating condition, and low-frequency thermal state data. Through time alignment, quality verification, and physical constraint verification, output a standardized dataset. S2: Causal Feature Extraction and Physical Modeling: For the standardized dataset, fault-sensitive frequency band signals are preserved through multi-scale signal decomposition; physical parameters are inverted based on the standardized dataset and fault-sensitive frequency band signals, and physical causal graph modeling is performed to determine health and fault thresholds; S3: Multimodal Feature Fusion and Health Status Inference: A multimodal feature tensor is constructed based on fault-sensitive features, fault-sensitive frequency band signals, and standardized datasets. The multimodal feature tensor is decomposed, and the factor vectors of each dimension are separated and the weights are optimized to obtain fused features. The graph neural network structure is determined based on the physical causal graph, and the fused features are input to obtain high-dimensional health features. The high-dimensional health features are then dimensionality reduced and health pattern modeled through a variational autoencoder to obtain low-dimensional health representation vectors. S4: Dynamic Health Calculation and Health Status Assessment: The entropy weight method is used to calculate the objective weights of each dimension of the low-dimensional health representation vector and fine-tunes it in combination with working conditions; the health benchmark vector is determined by health operation data, the weighted distance is calculated and converted into dynamic health; the health trend is predicted by time series model and early warning signals are output; and a health status classification system is established by combining dynamic health and trend. S5: Fault Diagnosis and Operation and Maintenance Decision Recommendations: Based on the fault sensitivity feature threshold for coarse classification and machine learning model for fine classification to identify fault modes; combine physical parameters and fault cause-effect graphs to determine the root cause of the fault; formulate hierarchical operation and maintenance recommendations based on health status level, fault mode and fault root cause.

[0007] The technical effects and advantages of this invention are as follows: Multi-source data quality controllability: High-frequency dynamic, medium-frequency operating condition, and low-frequency thermal data are collected synchronously. Through time alignment, quality factor screening, and energy conservation / transmission chain verification, a standardized dataset is output to ensure data reliability from the source, provide high-quality data support for subsequent diagnosis, and avoid interference data from affecting the results. Dual-drive feature extraction mechanism and data: By preserving fault-sensitive frequency bands through multi-scale wavelet decomposition, inverting gear contact stress and bearing dynamic load, and combining physical causal graph modeling, the modulation depth, nonlinearity and other features are accurately extracted and the threshold is calibrated. This takes into account both mechanical mechanism and data regularity, greatly improving the ability to identify early faults and reducing missed and false judgments. Dynamic and accurate health assessment: The entropy weight method is used to calculate objective weights, combined with fine-tuning of working conditions, and dynamic health is calculated by weighted Euclidean distance. Then, the LSTM model is used to predict future trends and output early warnings, thus constructing a five-level health system to realize real-time status quantification and trend prediction. The assessment accuracy is adapted to complex working conditions. Precise tiered operation and maintenance decision-making: Fault modes are determined through threshold coarse classification and machine learning subdivision, and root causes are analyzed by combining physical parameters and cause-effect graphs. Operation and maintenance recommendations are formulated according to health level, clarifying the repair methods, time windows and post-repair verification indicators, avoiding over-repair, reducing downtime costs and improving operation and maintenance efficiency. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0009] Figure 2 This is a schematic diagram of the S1-S2 process of the present invention.

[0010] Figure 3 This is a schematic diagram of the S2-S3 process of the present invention.

[0011] Figure 4 This is a schematic diagram of the S3-S4 structure of the present invention.

[0012] Figure 5 This is a schematic diagram of the S4-S5 structure of the present invention. Detailed Implementation

[0013] 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.

[0014] refer to Figures 1-5 The online fault diagnosis method for wind turbine gearboxes based on multi-source data fusion shown includes: S1: Synchronous Acquisition and Preprocessing of Multiphysics Data: Acquiring three types of data: high-frequency dynamic, mid-frequency operating condition, and low-frequency thermal state data. Through time alignment, quality verification, and physical constraint validation, a standardized dataset is output. The specific steps are as follows: S101: Multi-source data synchronous acquisition: The acquisition system is deployed according to three dimensions: high-frequency dynamics, medium-frequency operating conditions, and low-frequency thermal conditions. Sampling rate configuration: High frequency... , intermediate frequency low frequency The specifics are as follows: S1011: High-frequency dynamic signal: captures the impact and motion characteristics of gearbox faults, as detailed below: Vibration acceleration signal acquisition uses a PCB352C33 sensor, specifically including: Deployed on the outer walls of the bearing housings on both sides of the planetary carrier of the gearbox, symmetrically arranged along the central axis of the planetary carrier, the vibration acceleration signals of the bearing housings on the left and right sides of the planetary carrier are collected and recorded as follows: , ; Vibration acceleration signals were collected from the bearing housings of the high-speed shaft near the gear meshing end, the intermediate shaft near the middle bearing housing, and the low-speed shaft near the output bearing housing, distributed along the gearbox transmission axis. These signals were recorded as follows: , , ; Vibration acceleration signals were collected from the input shaft end housing (10cm from the input shaft axis), the top surface of the output shaft end housing (10cm from the output shaft axis), and the middle side of the housing at the center of the gearbox, respectively. These signals were denoted as follows: , , t represents the data acquisition time; The sound pressure signal acquisition uses a BK4189 sensor, specifically including: Deployed at 50cm directly in front of the meshing area of ​​the high-speed shaft gear and 50cm to the side of the meshing area of ​​the medium-speed shaft gear, respectively, near-field sound pressure signals of the high-speed shaft meshing area and the medium-speed shaft meshing area were collected, and recorded as follows: , ; Deployed at a distance of more than 1.5m from the outer wall of the gearbox, the background sound pressure signal of the engine room was collected and recorded as: ; The rotational speed signal acquisition uses a Keyence OP-87251 photoelectric sensor, specifically including: The sensor is deployed on the outer side of the addendum circle of the speed measuring gear near the coupling end of the input shaft, at a distance of 5mm from the addendum circle, to collect the pulse signal of the speed measuring gear on the input shaft, denoted as: ; pass pulse frequency Calculate the input shaft speed: ; S1012: Intermediate frequency operating condition signal, reflecting load and energy characteristics: Power signal acquisition: The output power is read from the generator power monitoring module of the wind power main control system and recorded as: Input power is calculated using wind speed data from the anemometer tower. ; in, Where is air density, v is real-time wind speed, and S is the swept area of ​​the fan blades. The wind energy utilization coefficient of the wind turbine; Torque signal acquisition: The generator torque signal is read through the generator torque monitoring module of the wind power main control system and recorded as the generator output shaft torque signal. ; Wind speed signal acquisition: A wind speed sensor is deployed on the anemometer tower at a height of 70m above the ground. The wind speed signal at this height is collected and recorded as the wind speed signal. ; S1013: Low-frequency thermal signal, reflecting lubrication and temperature characteristics: Temperature signals are acquired using a PT100 platinum resistance thermometer, specifically including: Sensors were embedded in the inner walls of the high-speed shaft bearing housing, near the rolling elements, the intermediate-speed shaft bearing housing, the low-speed shaft bearing housing, and the planetary carrier bearing housing, respectively. Temperature signals from the high-speed shaft bearing housing, intermediate-speed shaft bearing housing, low-speed shaft bearing housing, and planetary carrier bearing housing were collected and denoted as follows: , , , ; The sensor was installed on the inner wall of the bottom of the gearbox oil sump, 5cm from the bottom surface of the sump, to collect the lubricating oil temperature signal, which was recorded as: ; Lubricating oil pressure signal: A Hydac HDA4700 pressure sensor is used, deployed near the oil pump outlet in the gearbox lubricating oil inlet line to collect the lubricating oil inlet line pressure signal, which is recorded as: ; S102: Time Alignment and Quality Verification Timestamp alignment: using high-frequency vibration signals Based on the timestamp, The timestamps are generated by the sensor's built-in clock module. Linear interpolation and resampling are performed on other modal data to ensure alignment, satisfying the following: Based on the slowest sampling rate among the multi-source data, the maximum deviation of all data timestamps is controlled within 1 / 10 of the time interval corresponding to the lowest sampling rate, i.e., satisfying the following function: Where i and j represent any two different data sources, , These represent the timestamps of the i-th and j-th data groups, respectively. , These represent the sampling rates of the i-th and j-th data groups, respectively. Data quality screening: Constructing quality factors by fusing missing rate and signal-to-noise ratio (SNR) ,filter The data is considered valid data, and its specific mathematical function is as follows:

[0015] in, This represents the proportion of invalid values ​​in the i-th type of data. Invalid values ​​are defined as data that are outside the range of physical quantities. The lowest signal-to-noise ratio for data in a healthy state; S103: Physical constraint consistency verification: Energy conservation verification: Compare the deviations between input power and output power and power loss, and eliminate those with deviations exceeding the allowable range. The data, and its specific mathematical function are as follows:

[0016] in, This refers to the power loss during gearbox operation, including gear meshing losses. Friction loss with bearings , For gear meshing force, The gear meshing linear velocity is the linear velocity of the contact point on the tooth surface moving along the meshing line when the gears mesh; it is a fixed design parameter. For the dynamic load on the bearing, at time t, the dynamic force borne by the bearing in the gearbox; The bearing friction coefficient, The radius of the bearing rolling element is a fixed design parameter. This refers to the allowable range of energy deviation. Transmission chain kinematic verification: Based on the gearbox design transmission ratio, the matching between the output shaft speed and the input shaft speed is verified by comparing the input shaft speed with the gearbox design transmission ratio. Convert the output shaft speed: ; This refers to the allowable range of rotational speed error. S104: Standardized Data Output: The output is a valid dataset that has been time-aligned, quality-filtered, and physically-constrained, and includes the following categories: High-frequency dynamic signal: Vibration acceleration signal of the left bearing housing of the planetary carrier Vibration acceleration signal of the right bearing housing of the planetary carrier Vibration acceleration signal of high-speed shaft meshing end bearing housing Vibration acceleration signal of intermediate bearing housing of intermediate speed shaft Vibration acceleration signal of the bearing housing at the low-speed shaft output end Input shaft end housing vibration acceleration signal Vibration acceleration signal of the output shaft end housing Vibration acceleration signal in the middle of the box Cabin background sound pressure signal Near-field acoustic pressure signal in the high-speed shaft meshing area Near-field sound pressure signal in the meshing region of the mid-speed shaft Input shaft speed Output shaft speed Gearbox design transmission ratio ; Intermediate frequency operating condition signal: fan input power Generator output power Generator output shaft torque signal Wind speed signal at a height of 70m from the meteorological tower ; Low-frequency thermal signal: High-speed shaft bearing housing temperature signal Temperature signal of intermediate speed shaft bearing housing Low-speed shaft bearing housing temperature signal Planetary carrier bearing housing temperature signal Oil bath lubricating oil temperature signal Lubricating oil inlet pipeline pressure signal ; Quality and Constraint Parameters: Data Quality Factor Energy deviation Energy tolerance permissible speed error ; S2: Causal Feature Extraction and Physical Modeling Based on the standardized data output by S1, fault-sensitive features are extracted through multi-scale signal decomposition, physical parameter inversion, and physical causal graph modeling. A feature system driven by both data and mechanism is established, providing feature input for subsequent multimodal fusion, health inference, and fault diagnosis. The specific steps are as follows: S201: Multi-scale time series decomposition: Decomposing all 8 vibration acceleration signals output from S104. - Wavelet decomposition was performed separately to uniformly separate signal components in different frequency bands, eliminate noise and useless frequency bands, and retain fault-sensitive frequency band signals to provide core vibration data for subsequent physical parameter inversion and causal graph modeling. For each vibration acceleration signal Let i = 1, 2, 3...8, representing the signal acquired by the i-th vibration sensor. The wavelet decomposition formula is: ; For the i-th vibration signal, there are the k-th layer wavelet coefficients, corresponding to a specific frequency band vibration signal; The residual of the i-th vibration signal is the decomposition result, which includes low-frequency interference and trend terms. Signal components and processing rules for each frequency band: The first layer of wavelet coefficients (k=1): frequency range 12.8-25.6kHz, signal component is high-frequency noise signal, denoted as... This indicates that the signal contains irrelevant interference such as electromagnetic interference and airflow noise, and there is no fault information. The second layer of wavelet coefficients (k=2): frequency range 6.4-12.8kHz, signal component is the bearing fault frequency band signal, denoted as... This indicates the impact signal generated by the collision and spalling of the bearing rolling elements with the inner and outer rings, which is the core characteristic frequency band of bearing failure. The third layer wavelet coefficients (k=3): frequency range 3.2-6.4kHz, signal component is gear meshing frequency band signal, denoted as... ; indicates that it contains meshing vibration signals generated by gear tooth surface contact, including fault information such as tooth surface pitting and broken teeth; The fourth layer wavelet coefficients (k=4): frequency range 1.6-3.2kHz, signal components are structural resonance frequency band signals, denoted as... This indicates that the vibration signal includes the resonance of the gearbox housing and shaft system; the resonance amplitude will change abnormally when a fault occurs. The fifth layer wavelet coefficients (k=5): frequency range 0.8-1.6kHz, signal component is the low-frequency trend term signal, denoted as... This indicates that the signal contains slow changes caused by enclosure deformation and installation errors, but no real-time fault information. Decomposition residual Frequency range <0.8kHz, signal components are low-frequency interference residuals; includes environmental low-frequency interference, no fault information; The final output consists of 8 vibration signals, each corresponding to a fault-sensitive frequency band. Specifically, these are: 8 bearing fault frequency band signals. - 8-channel gear meshing frequency band signal - 8-channel structural resonant frequency band signal - ; S202: Physical Parameter Inversion Calculation: Based on the standardized data output from S104 and the fault-sensitive frequency band signal output from S201, combined with gearbox design parameters, key physical parameters such as gear contact stress and bearing dynamic load, which cannot be directly collected, are inverted, providing quantitative parameter support for subsequent physical cause-effect graph modeling; the details are as follows: Gear meshing force and contact stress inversion: Gear meshing force Based on generator output shaft torque signal Input shaft speed Combined with gear design parameters, gear pitch circle radius Obtained from the gearbox manual, and also referencing the gear meshing frequency band vibration signal output from S201. calculate: ;in, For gear contact efficiency, the fixed value under healthy conditions is 0.98. The correction rule for fault conditions is: when... For every 10% increase in amplitude compared to the healthy baseline, Decrease of 5%; Gear contact stress Gear meshing force obtained from inversion Combined with gear design parameters obtained from the gearbox manual, the gear tooth surface contact area calculate: ; bearing dynamic load Inversion: Vibration signals based on bearing fault frequency band Bearing stiffness in combination with bearing design parameters Bearing damping Calculation: For each path After performing a second integration to eliminate the trend term, the relative displacement of the corresponding bearing is obtained: ,right Differentiating yields the displacement velocity of the corresponding bearing: Substitute the values ​​into the general formula for bearing dynamics to calculate the dynamic load: i = (1, 2, 3, 4), corresponding to four bearings; S203: Structural Modeling Based on Physical Cause-Effect Graphs Based on standardized data, fault-sensitive frequency band signals, and inverted physical parameters, a cause-effect graph of gearbox faults is constructed. The node set V corresponds to the parameters output by S1, and the edge set E defines causal relationships based on mechanical transmission mechanisms or thermodynamic laws. The specific analysis is as follows: The node set includes: load-related nodes, speed-related nodes, stress and load-related nodes, temperature-related nodes, and vibration and sound pressure-related nodes; Load-type nodes include: generator output shaft torque signal Fan input power ; Speed-related nodes include: input shaft speed. Output shaft speed ; Stress and load type nodes include: gear contact stress Bearing dynamic load ; Temperature-related nodes include: bearing housing temperature signal - Oil bath lubricating oil temperature signal ; Vibration and sound pressure level nodes include: vibration signals in fault-sensitive frequency bands. - Near-field acoustic pressure signal in the high-speed shaft meshing area ; The set of causal edges includes causal edge identifiers: ,express Increase As the generator output torque increases, the torque transmitted by the gears also increases synchronously, leading to an increase in tooth surface contact force and ultimately an increase in gear contact stress. ,express Increase As the input shaft speed increases, the gear meshing frequency increases proportionally with the speed. Z represents the number of teeth on the gear. ,express abnormal When the amplitude increases and the gear contact stress is abnormal, the vibration amplitude during gear meshing will increase significantly, which will be reflected in the meshing frequency signal. superior; ,express When it increases As the dynamic load on the bearing increases, the friction between the rolling elements and the inner and outer rings intensifies, leading to increased frictional heat generation and consequently higher temperatures in the corresponding bearing housing. Increase; ,express abnormal Nonlinear enhancement occurs when the oil temperature rises abnormally, causing a decrease in lubricating oil viscosity and an increase in frictional noise between gears and bearings, thus increasing the sound pressure signal. The nonlinear harmonic components increase; Edge weights are calculated based on historical data of the influencing factors of each edge, and the Person correlation coefficient is used as the edge weight, i.e., the correlation coefficient of causal edges. S204: Fault Sensitive Feature Extraction: For gear and bearing faults, based on standardized data, fault-sensitive frequency band signals, and inverted physical parameters, quantitative fault sensitive features are extracted to clarify health and fault thresholds; the details are as follows: S2041: Define modulation depth Vibration signals based on gear meshing frequency band Combined with the input shaft speed The specific calculation process for calculating the number of teeth Z, a gear design parameter, is as follows: For each target Perform FFT transforms on each gear to obtain the vibration spectrum of the corresponding gear meshing frequency band. , where i is the target sensor number; Calculate the gear rotation frequency based on the input shaft speed: ; In each path, identify Carrier frequency and sideband frequency , ; Extract key amplitudes from each spectrum: , , ;in, for The amplitude of the spectrum at that point is denoted as the engagement frequency amplitude. The maximum value at the upper edge. This is the maximum amplitude value at the lower edge; Calculate the modulation depth of each target signal: Finally, the average value of the multi-path results is taken as the global modulation depth. ; Define threshold criteria: Fault condition: Gear pitting or broken teeth ; S2042: Define bearing fault nonlinearity Based on Near-field acoustic pressure signal in the high-speed shaft meshing area after noise cancellation Combined with the input shaft speed Calculation of bearing design parameters, including the number of bearing rolling elements. Bearing rolling element diameter Bearing pitch circle diameter Bearing contact angle The specific calculation process is as follows: After noise reduction Perform an FFT transform to obtain the fault sound pressure spectrum. ; Calculate the failure frequency of the bearing outer ring: ; exist Extract from and 2-5 times the frequency The amplitude; Calculate the nonlinearity: ; Define threshold: Health status: Fault condition: When the bearing is worn or spalled: ; S3: Multimodal Feature Fusion and Health Status Inference: Based on fault-sensitive frequency band signals, inverted physical parameters, and fault-sensitive features, a low-dimensional health representation vector is extracted through multimodal feature tensor construction, tensor decomposition and fusion, physical constraint graphical neural network (GNN) inference, and variational autoencoder (VAE) learning. This vector provides feature input for health assessment and fault diagnosis. The specific steps are as follows: S301: Construction of multimodal feature tensor: Using fault-sensitive features, fault-sensitive frequency band signals and standardized data as data sources, a four-dimensional multimodal feature tensor is constructed to integrate the dispersed multi-physical quantity signals into a structured feature matrix; Define the dimension of a four-dimensional tensor: Time dimension: Based on S2 fault sensitivity characteristics , The calculation cycle is divided using a sliding window; the window length is 10s, corresponding to a sampling rate of 50Hz for the S1 intermediate frequency signal, containing 500 sampling points, with a step size of 5s, and each window outputs one feature value; the dimension size is T=20. Frequency and scale dimensions: For fault-sensitive frequency bands corresponding to S201 multi-scale decomposition, only frequency bands containing fault information are retained; specifically classified as follows: 6.4-12.8kHz, bearing fault frequency band, corresponding to , 3.2-6.4kHz, gear meshing frequency band, corresponding to , 1.6-3.2kHz, structural resonance frequency band, corresponding to Dimension size F=3; Spatial dimension: Select the sensor locations in S1 with the highest correlation to faults, covering key areas such as bearings, gears, and housings; specific locations are: The left bearing housing of the planetary carrier, corresponding to , ; High-speed shaft meshing end bearing housing, corresponding , ; Intermediate speed shaft intermediate bearing housing, corresponding , ; Low-speed shaft output end bearing housing, corresponding , ; Near field of high-speed shaft meshing area, corresponding Dimension size: S=5; Modal dimension: Select physical quantities directly related to the health state from S1-S2 to avoid redundancy; specific modes are: Vibration acceleration, corresponding to S201 , , ; Sound pressure level, corresponding to S1 ; Temperature, corresponding to S1 , ; Fault-sensitive characteristics, corresponding to S204 , ; Load, corresponding to S202 , Dimension size: M=5; Define tensor elements: the characteristic tensor is identified as The element values ​​are the feature values ​​corresponding to the time window t, frequency band f, sensor position s, and physical mode m. The specific calculation rules are as follows: Vibration modes Take the corresponding frequency band f and position s sensor signals. , , The root mean square (RMS) value within the time window t; Sound pressure mode :Pick The peak value within the time window t; Temperature modes : Take the temperature signal at the corresponding position s , The average value within the time window t; Fault characteristic modes Take directly , The calculated value is given within a time window t, with one value per window. Load modes :Pick , The maximum value within the time window t.

[0017] Output Tensor: The final output is a four-dimensional multimodal feature tensor. 20 time windows × 3 frequency bands × 5 sensor locations × 5 physical modes; S302: Multimodal Feature Fusion Based on Tensor Decomposition: The four-dimensional feature tensor constructed in S301 is subjected to CP decomposition to separate the factor vectors of each dimension. Multimodal information is fused through weight optimization, redundant features are eliminated, and low-dimensional key fused features are extracted to provide input for subsequent GNN inference. The specific steps are as follows: S3021: CP decomposition model: for feature tensors The mathematical function for CP decomposition is: , where ○ represents the tensor outer product operation; R is the decomposition rank (fusion feature dimension); The r-th time factor vector The r-th frequency factor vector The r-th spatial factor vector The r-th modal factor vector; S3022: Decomposition of rank and factor vectors: Decomposition rank R is determined by minimizing the tensor reconstruction error (root mean square error RMSE) through 5-fold cross-validation; value: R=12, reconstruction error <5%, ensuring that the fused features retain more than 95% of the original information, while reducing the dimensionality by 60%.

[0018] Factor vector: time factor : Reflects the weight of fault information in different time windows; frequency factor : Reflects the fault contribution rate of different frequency bands; space factor : Reflects the sensitivity of different sensor locations; modal factor : Reflects the importance of different physical modes; S3023: Output of fused feature vectors: The four factor vectors after decomposition are concatenated according to the time, frequency, space, and modality dimensions to form the fused feature vector corresponding to each decomposition rank r. The final output is: Fusion feature matrix: 20 time windows × 33 fused feature dimensions; S303: Physically Constrained Graph Neural Network (GNN) Health Inference: Using the fused feature matrix output from S302 as input, the GNN topology is determined based on the fault cause-effect graph G=(V,E) constructed from S203. Node features are learned through a message passing mechanism, and physical parameter limit constraints are embedded to output high-dimensional health state features. The specific process is as follows: Define GNN topology and fault cause graph Consistency: Node set V: Nodes corresponding to the cause-effect graph of S203, a total of 9: 2 load type, 2 speed type, 2 stress and load type, 2 temperature type, and 1 vibration and sound pressure type; Edge set E: The five causal edges corresponding to the S203 causal graph, with edge weights being the Pearson correlation coefficients calculated based on historical data in S203; Message passing mechanism: for each node In the l-th layer, the features of the node itself are updated by weighting the features of its neighboring nodes. The specific mathematical function is as follows: ;in, Let v be the feature vector of the l-th node. Let v be the set of neighboring nodes. The weights of edges u and v are taken from the causal edge correlation coefficients of S203. For learnable weight matrix, For activation functions; Physical constraint embedding: Limit constraints of the physical parameters S1-S2 are added to the GNN loss function to prevent the inference results from exceeding the reasonable range of engineering. The total loss function is: ;in, Cross-entropy loss is used for pre-training in binary classification of health and faults; The physical constraint loss is calculated by determining the deviation between the inference result and the physical parameter limit value. The formula for calculation is as follows: ; For the node v parameter values ​​of GNN inference, For the limit value of the parameter, The weights are determined through optimization using the validation set; Output: After the GNN training is complete, the output is the feature vector of the nodes in the last layer (layer 3): a high-dimensional health feature matrix. 20 time windows × 128-dimensional node features; S304: Variational Autoencoder (VAE) Low-Dimensional Health Representation Learning: Based on a high-dimensional health feature matrix, a low-dimensional health representation vector is learned through the encoder and decoder structure of a VAE, achieving feature dimensionality reduction and health pattern modeling, and outputting a core vector that can be directly used for health score calculation; the details are as follows: VAE structural design: Encoder: Input high-dimensional health features Each time window corresponds to a 128-dimensional feature vector; the mean of the latent variable z is output. With variance The dimension of the latent variables was set to 4, which was determined by the reconstruction error verification. The network structure is a 2-layer fully connected layer. The input is a 128-dimensional health feature vector, which is compressed into a 64-dimensional hidden feature after the first layer. After the second layer, the output is an 8-dimensional vector, of which 4 are the mean and 4 are the variance. Decoder: Inputting a low-dimensional latent variable z, a two-layer fully connected layer is used. The first layer uses a linear transformation combined with the ReLU activation function to map the 4-dimensional z into a 64-dimensional latent feature, introducing non-linearity to adapt to the complex patterns of the health state. The second layer uses a linear transformation combined with the Sigmoid activation function to map the 64-dimensional latent feature into a 128-dimensional reconstructed feature, constraining the output value to the interval [0, 1]. After physical constraint verification, the reconstructed high-dimensional feature is output. ; VAE Loss Function: The total loss function includes reconstruction loss, KL divergence, and physical constraint loss. Its mathematical function is: ; in, To calculate the reconstruction loss, mean squared error (MSE) is used to calculate H and To minimize deviations and ensure reconstruction accuracy; Let KL divergence be the constraint that the latent variable z follows a standard normal distribution. To avoid overfitting; The physical constraint loss is calculated based on the conservation relationship between input power and output power, and its specific mathematical function is as follows: ;in , The power value reconstructed for VAE. The model reconstruction values ​​for gearbox power loss; weights Determined through validation set optimization; After VAE training is complete, the output encoder calculates the latent variable mean vector, which is the low-dimensional health representation vector: Low-dimensional health representation matrix: 20 time windows × 4-dimensional health representation vector; S4: Dynamic Health Calculation and Health Status Assessment: Based on a low-dimensional health representation vector, a real-time quantitative, trend-based, and status-level health assessment system is constructed through dynamic weight allocation, multi-dimensional health quantification calculation, time-series trend prediction, and hierarchical assessment. This system outputs the health indicators, trend characteristics, and status labels required for S5 fault diagnosis; specifically as follows: S401: Dynamic allocation of health weights: For the low-dimensional health representation vector Z, combined with the standardized data output from S1, the entropy weight method is used to assign weights and dynamically adjust the weights of each dimension; the details are as follows: Objective weighting using entropy weighting method: based on S304 The four dimensions of Z are: denoted as: Vibration health status Temperature and health status Load health status Fault characteristic health status; calculate the information entropy of each dimension. The smaller the entropy value, the higher the data dispersion and the greater the information contribution. The calculation process is as follows: Standardized Z is obtained ; Calculate the information entropy of the j-th dimension ; Calculate objective weights: Output objective weight vector ,satisfy ; Adaptive Weighting for Operating Conditions: Input Shaft Speed ​​Based on S104 Divide the working conditions into 3 categories, in Based on this, the weights are finely adjusted to adapt to the differences in fault sensitivity of physical quantities under different operating conditions, as follows: Low-speed operating conditions ( <800r / min): The load fluctuation is small, and the sensitivity of temperature to faults is increased. Therefore, the weight of the temperature dimension is increased by 5%, and the weight of the load dimension is decreased by 5%. High-speed operating conditions ( >1500r / min): Vibration and noise have a more significant impact on the response to faults, therefore the weight of the vibration dimension is increased by 5%, and the weight of the temperature dimension is decreased by 5%. Medium speed operating conditions (800r / min≤ ≤1500r / min): Fault sensitivity across all dimensions is balanced, and the weights remain unchanged; The final output is the fine-tuned weight vector. ; S402: Dynamic Health Measurement Calculation: With Z and As input, the weighted Euclidean distance method is used to calculate the real-time health score, quantify the current health status, and output a smoothed health score sequence, as follows: Determining the health baseline vector: Collect 100 hours of health operation data for the gearbox after it leaves the factory, and obtain the health characterization set according to the S3 process. The mean of each dimension is used as the benchmark to output a health benchmark vector. ; Dynamic health score calculation: Weighted distance to health conversion: health and arrive The weighted distance is negatively correlated, and the formula is: , ,in, Weighted deviation distance; =0.8, a threshold based on historical fault data, during a fault. ≥0.8; 1 = perfectly healthy, 0 = faulty; Smoothing: A 3-window moving average is used to eliminate noise, outputting a smoothed health sequence. ; Result verification: Under healthy conditions: and , Under fault conditions: and , ; S403: Health Trend Prediction: Based on the HD sequence of S402, an LSTM time series model is used to predict future changes in health, outputting the trend slope and early warning signal, as detailed below: LSTM model construction: Input the health status of the previous 10 time windows to form Output the predicted health status for the next 5 time windows. ; Model structure: Network layers: 2 LSTM layers (32 nodes → 16 nodes) and 1 fully connected output layer (5 nodes); Activation function: tanh for hidden layers, sigmoid for output layers; Training parameters: batch size 32, 100 iterations, Adam optimizer, learning rate 0.001; Trend characteristics and early warning: Trend slope calculation: based on Linear fitting trend line, slope The specific mathematical function reflecting the rate of change in health status is: ; Its slope This indicates a rebound. Indicates stability. This indicates a deterioration; Warning threshold: Each window's health level decreases by 0.02; the warning signal is: if and Output 1 indicates a warning, and 0 indicates normal. S404: Health Status Grading Assessment: Combining with S402 With S403 Establish a 5-level health status system, and output status labels and confidence levels, as follows: Five-level status classification standard: when and This indicates complete health and is labeled as ; when and This indicates mild degradation and is marked as ; when and This indicates moderate degradation, and is marked as... ; when and This indicates severe degradation and is identified as... ; when and This indicates a fault status and is marked as such. ; State and confidence calculation: Quantifying graded reliability, its mathematical function is: ;in, This represents the median threshold for the current level. The higher the value, the stronger the confidence level. Output real-time health status level State confidence Health sequence HD, predicted health Warning signals ; S5: Fault Diagnosis and Maintenance Decision Recommendations: Based on fault sensitivity characteristics, low-dimensional health representations, and health status data, through fault mode recognition and root cause analysis, a complete diagnostic result of fault type and root cause is output. Combined with equipment operating conditions and health trends, hierarchical maintenance decision recommendations are generated. The specific steps are as follows: S501: Fault Mode Recognition: Taking the quantitative fault sensitivity features of S204, the health representation differences of S3, and the state level of S4 as inputs, a two-layer recognition logic of feature threshold judgment and machine learning classification is adopted to distinguish the typical fault modes of gears and bearings and output a unique fault mode label. Input feature set and weight assignment: The input features are: gear modulation depth Bearing nonlinearity Low-dimensional representation vector Z, health status level The weights are defined as follows: 0.35, 0.35, 0.2, and 0.1. Dual-layer recognition logic: First layer: Initial judgment of feature thresholds: based on the calibrated fault thresholds: healthy: , ;Fault: or First, perform a rough classification: like and The fault is identified as gear-related, and the system will proceed to gear sub-mode identification. like and The fault is identified as bearing-related, and the system proceeds to bearing sub-mode identification. like and Both exceed limits: This is determined to be a compound fault, and gear and bearing sub-mode identification is initiated simultaneously; Second layer: Machine learning classification: comparing the Z (4-dimensional) of S304 with that of S204. and Construct a 5-dimensional input vector, input it into a pre-trained improved CNN model (containing 1 convolutional layer, 1 pooling layer, and 2 fully connected layers), and output the specific fault mode: Gear Sub-mode: Tooth surface pitting, , Tooth surface wear, , Broken tooth ; Bearing sub-mode: Rolling element wear, , The inner and outer rings peeled off. , Cage failure .

[0019] Output: Fault mode labels This includes gear sub-mode and bearing sub-mode.

[0020] S502: Root cause analysis of the fault: using the inverse physical parameters of S202 , Fault cause-effect graph of S203 G=(V,E), health trend of S403 Based on this, the root cause of the failure is traced from four dimensions: load, temperature, lubrication, and operating conditions, and the root cause code and mechanism explanation are output; as follows: The dimensions and judgment rules for root cause analysis are as follows: Load anomaly dimensions: , The corresponding root cause code is The cause was determined to be overload leading to excessive contact stress, which accelerated wear or spalling. Temperature anomaly: , The corresponding root cause code is The cause was determined to be excessively high temperature, which led to a decrease in lubricant viscosity and increased friction. Lubrication failure: Normal but Nonlinear enhancement, The corresponding root cause code is This indicates that the lubricating oil has deteriorated or is insufficient, resulting in direct metal-to-metal contact. Operating condition fluctuations: Fluctuation range greater than 15%, The corresponding root cause code is The damage was determined to be caused by impact loads and fatigue due to frequent fluctuations in rotational speed. Root cause priority ranking: When there are multi-dimensional anomalies, the root cause is ranked based on the influence strength of the causal graph. The higher the weight of the causal edge, the higher the root cause priority. Output the root cause code and explanation of the root cause mechanism; S503: Tiered Operation and Maintenance Decision Recommendations: Combining S404 Health Status Levels S501 Fault Modes For the root cause R of S502, tiered maintenance recommendations are formulated based on urgency, repair cost, and downtime impact, specifying repair methods, spare parts requirements, and time windows, as detailed below: Operation and maintenance grading standards and recommendations: When the fault status level is 5, it is determined to be an emergency. It is recommended to stop the machine immediately, replace the faulty parts, and run it under no-load for 2 hours after the repair. When the fault status level is 4, it is judged as high priority. It is recommended to stop the machine within 72 hours, disassemble and check the wear of the faulty parts, replace or repair them, and add lubricating oil. When the fault status level is 3, it is judged as medium priority. It is recommended to develop a maintenance plan within one month, increase the monitoring frequency to once a day, and optimize the operating conditions in a targeted manner. When the fault status level is 2, it is judged as low priority. It is recommended to monitor the health status once every two weeks, improve lubrication conditions, and avoid overload operation. Define the monitoring indicators: Post-repair verification metrics: , , ; Long-term monitoring indicator: Health trend slope The root cause-related parameters are restored to their rated range.

[0021] 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 wind turbine gearbox online fault diagnosis method based on multi-source data fusion, characterized in that, Comprise: S1: multi-physical field data synchronous acquisition and preprocessing: collect dynamic, working condition, and thermal state three types of data by frequency, output standardized data set through time alignment, quality verification, and physical constraint verification; S2: causal feature extraction and physical modeling: based on the standardized data set, the fault sensitive frequency band signal is reserved through multi-scale signal decomposition; the physical parameters are inverted based on the standardized data set and the fault sensitive frequency band signal, and the physical causal diagram modeling is performed to determine the health and fault threshold; S3: multi-modal feature fusion and health state reasoning: based on the fault sensitive feature, the fault sensitive frequency band signal and the standardized data set, a multi-modal feature tensor is constructed, the multi-modal feature tensor is decomposed, the factor vectors of each dimension are separated and the weight is optimized to obtain the fusion feature, the graph neural network structure is determined based on the physical causal diagram, and the high-dimensional health feature is obtained by inputting the fusion feature; The high-dimensional health feature is reduced and health pattern modeling is performed through the variational autoencoder to obtain a low-dimensional health representation vector; S4: dynamic health degree calculation and health state evaluation: the objective weight of each dimension of the low-dimensional health representation vector is calculated by entropy weight method and combined with working condition fine tuning; The health benchmark vector is determined based on the health running data, the weighted distance is calculated and converted into dynamic health degree; the health trend is predicted through the time series model and the early warning signal is output, and the health state grading system is established based on the dynamic health degree and the trend; S5: fault diagnosis and operation and maintenance decision suggestion: based on the fault sensitive feature threshold coarse classification and the machine learning model fine classification, the fault mode is identified; the fault root cause is determined based on the physical parameters and the fault causal diagram; the hierarchical operation and maintenance suggestion is made according to the health state level, the fault mode and the fault root cause.

2. The wind turbine gearbox online fault diagnosis method based on multi-source data fusion according to claim 1, characterized in that: The standardized data set comprises: The high-frequency dynamic data, medium-frequency working condition data and low-frequency thermal state data collected synchronously are processed to form; wherein the high-frequency dynamic data includes gear box vibration and sound pressure signal, the medium-frequency working condition data includes input and output speed, torque and power signal, and the low-frequency thermal state data includes bearing seat and lubricating oil temperature signal; the standardized data set is output after time stamp alignment, data quality screening, energy conservation verification and transmission chain kinematics verification, the high frequency is 25.6 kHz, the medium frequency is 50 Hz, and the low frequency is 20 Hz.

3. The wind turbine gearbox online fault diagnosis method based on multi-source data fusion according to claim 2, characterized in that: The multi-scale signal decomposition comprises: Based on the standardized data set, different frequency band signals in the data are separated by frequency scale layering; by matching the characteristic frequency range corresponding to the typical fault of the wind turbine gearbox, the fault sensitive frequency band signal is screened and reserved, and noise and redundant frequency band signals irrelevant to the fault are simultaneously removed.

4. The wind turbine gearbox online fault diagnosis method based on multi-source data fusion according to claim 3, characterized in that: The inverted physical parameters comprise: Bearing dynamic load, gear contact stress, gear meshing force; based on the standardized data set and the fault sensitive frequency band signal, the mechanical transmission mechanism, thermodynamic law and equipment design parameters of the wind turbine gearbox are combined to calculate the mapping relationship between data and mechanism.

5. The wind turbine gearbox online fault diagnosis method based on multi-source data fusion according to claim 4, characterized in that: The physical causal diagram modeling comprises: The standardized data set, the fault sensitive frequency band signal and the inverted physical parameters are taken as nodes, the causal relationship and the influence strength between the nodes are defined according to the mechanical transmission mechanism and the thermodynamic law of the wind turbine gearbox, and in the physical causal diagram, the nodes correspond to the equipment operation state parameters or characteristics, and the connection between the nodes represents the causal relationship between the parameters.

6. The wind turbine gearbox online fault diagnosis method based on multi-source data fusion according to claim 5, characterized in that: The multi-modal feature tensor is constructed, including: Based on the fault sensitive features, the fault sensitive frequency band signal and the standardized data set, three types of multi-source information are divided and associated according to signal attributes and dimensions, forming a multi-modal feature tensor containing dynamic signals, quantitative features and basic parameters.

7. The wind turbine gearbox online fault diagnosis method based on multi-source data fusion according to claim 6, characterized in that: The high-dimensional health feature includes: The multi-modal feature tensor is dimensionally decomposed to separate the dimensional factor vectors corresponding to dynamic signals, quantitative features and basic parameters; the weights of the factor vectors are optimized according to the requirements of wind turbine gearbox fault diagnosis, the weight of fault related information is strengthened and the weight of redundant noise is suppressed, and the fusion features are integrated; the topology of the graph neural network is determined based on the physical causal diagram, so that the nodes of the graph neural network correspond to the nodes of the physical causal diagram one by one, and the edges correspond to the causal relationship between the nodes; through the message passing mechanism of the graph neural network, the dimensional information and the associated relationship in the fusion features are learned, and the health state reasoning is performed simultaneously by embedding the physical parameter limit constraints of the wind turbine gearbox, and finally the high-dimensional health feature is output.

8. The wind turbine gearbox online fault diagnosis method based on multi-source data fusion according to claim 7, characterized in that: The low-dimensional health representation vector includes: Based on the high-dimensional health feature, a variational autoencoder is used for dimension reduction processing, which retains the health state associated information while compressing the data dimension; a health mode model is established based on the differences between the health and fault modes of the wind turbine gearbox, and the redundant noise in the high-dimensional data is removed, and finally the low-dimensional health representation vector is output.

9. The wind turbine gearbox online fault diagnosis method based on multi-source data fusion according to claim 8, characterized in that: The dynamic health degree includes: Based on the low-dimensional health representation vector, the objective weights of each health dimension are calculated by using the entropy weight method, and the objective weights are adaptively fine-tuned according to the real-time working condition of the wind turbine gearbox to obtain the working condition adaptive weight; the gearbox health running data is collected, and the low-dimensional representation vector under the health state is obtained by repeating steps S2 and S3, and the mean value of each dimension is taken as the health reference vector; based on the working condition adaptive weight, the weighted distance between the low-dimensional health representation vector and the health reference vector is calculated, the weighted distance is converted into a health degree value according to a preset mapping rule, and the health degree value is smoothed to eliminate transient fluctuations, and finally the dynamic health degree is output.

10. The wind turbine gearbox online fault diagnosis method based on multi-source data fusion according to claim 9, characterized in that: The health trend is predicted by the time series model, including: Based on the dynamic health degree, a time series data set of the dynamic health degree is constructed, which contains the historical and real-time health degree change information of the gearbox; a time series prediction model is used to learn the time series data set to mine the evolution law of the dynamic health degree over time; based on the learned law, the health degree change in a future preset time period is predicted, and the health trend slope and the future health degree interval are output; the health trend slope warning threshold and the health degree threshold are set, and if the health trend slope and the future health degree are less than or equal to the corresponding threshold, an early warning signal is triggered and output.

Images