Intelligent diagnosis system and method for wind turbine impeller-gearbox composite fault

Abstract

The invention provides an intelligent diagnosis system and method for a wind turbine impeller-gearbox composite fault, and relates to the technical field of fault diagnosis, and the system comprises a multi-modal data collection module which is used for carrying out the deployment of a multi-source sensor based on coupling path analysis, and collecting the multi-modal data of the multi-source sensor; the tensor matrix construction module is used for constructing a tensor matrix according to the multi-modal data; the fault analysis module is used for constructing a feature-damage mapping model to perform fault analysis on the tensor matrix to obtain a fault diagnosis result; and the fault maintenance module is used for pushing the fault diagnosis result and carrying out fault maintenance according to the fault diagnosis result. According to the invention, the technical problem that the fault diagnosis accuracy and the operation safety and stability of the wind power equipment are affected due to the fact that tiny early faults of key parts of the wind turbine cannot be effectively identified in the prior art can be solved, and the fault diagnosis accuracy and the system response efficiency are improved; the service life of equipment is prolonged; and the maintenance cost is reduced.

Description

Technical Field

[0001] This application relates to the field of fault diagnosis technology, and in particular to an intelligent diagnostic system and method for combined faults of wind turbine impeller-gearbox. Background Technology

[0002] As the scale of wind power equipment continues to expand and its service life continues to increase, health status monitoring and fault diagnosis of key components of wind turbines have become one of the core technologies to ensure system safety and operational stability.

[0003] Currently, components of wind turbines, such as impellers, main shafts, and gearboxes, are highly susceptible to fatigue cracks, wear, and pitting due to long-term exposure to uneven wind loads and complex mechanical stresses. However, existing technologies primarily rely on single-type sensors for local state sensing or traditional empirical rule methods for signal analysis. These approaches have several shortcomings and are insufficient to meet the needs of identifying multi-field coupling effects and early-stage minor faults in actual operation.

[0004] In summary, existing technologies suffer from technical problems such as the inability to effectively identify minute early faults in key wind turbine components under multi-field coupling due to the limited variety of sensor types, lack of systematic deployment, and reliance on empirical rules in signal processing methods. This further affects the accuracy of fault diagnosis and the safe and stable operation of wind power equipment. Summary of the Invention

[0005] The purpose of this application is to provide an intelligent diagnostic system and method for combined faults of wind turbine rotor-gearbox, in order to solve the technical problems in the prior art, which are due to the single type of sensor, lack of systematic deployment and reliance on empirical rules for signal processing methods, which makes it impossible to effectively identify small early faults of key components of wind turbines under multi-field coupling, and further affects the accuracy of fault diagnosis and the safe and stable operation of wind power equipment.

[0006] In view of the above problems, this application provides an intelligent diagnostic system and method for combined faults of wind turbine impeller and gearbox.

[0007] In a first aspect, this application provides an intelligent diagnostic system for composite faults of wind turbine rotor-gearbox, comprising: a multimodal data acquisition module for deploying multi-source sensors based on coupled path analysis and acquiring multimodal data from the multi-source sensors; a tensor matrix construction module for constructing a tensor matrix based on the multimodal data; a fault analysis module for constructing a feature-damage mapping model to perform fault analysis on the tensor matrix and obtain fault diagnosis results; and a fault maintenance module for pushing the fault diagnosis results and performing fault maintenance based on the fault diagnosis results.

[0008] Preferably, the intelligent diagnostic system for combined wind turbine impeller-gearbox faults further includes: an aerodynamic load simulation unit for performing aerodynamic load simulation on the wind turbine and identifying the dynamic deformation region of the impeller; a wind power transmission simulation unit for performing wind power transmission simulation based on the dynamic deformation region of the impeller and analyzing the stress transmission path of the impeller-main shaft-gearbox; and a multi-source sensor deployment unit for deploying multi-source sensors on the impeller and gearbox respectively according to the stress transmission path.

[0009] Preferably, the intelligent diagnostic system for combined wind turbine rotor-gearbox faults further includes: an indicator system hierarchical structure construction unit, used to construct an indicator system hierarchical structure for deployment optimization, wherein the indicator system hierarchical structure includes a target layer, a criterion layer, and a scheme layer, and the criterion layer includes convenience, signal-to-noise ratio, redundancy, and monitorability; a simulation result filtering unit, used to construct a fuzzy judgment matrix, with the sensor deployment optimization of the target layer as the objective, to filter the simulation results of the scheme layer generated based on the simulation of the criterion layer, and output the optimization results; and a multi-source sensor update and deployment unit, used to update the deployment of multi-source sensors based on the optimization results.

[0010] Preferably, the intelligent diagnostic system for wind turbine impeller-gearbox composite faults further includes: a deployment visualization unit, used to perform deployment visualization of the optimization results based on a three-dimensional model, filter overlapping results based on the visualization results, and obtain updated optimization results.

[0011] Preferably, the intelligent diagnostic system for combined wind turbine rotor-gearbox faults further includes: a pre-aligned data acquisition unit, used to align the time window of the multimodal data, and then perform dynamic resampling and interpolation alignment to obtain pre-aligned data; an aligned data acquisition unit, used to decompose the non-stationary signals in the pre-aligned data to obtain aligned data; and a tensor matrix establishment unit, used to extract the time-domain features, frequency-domain features, and collaborative features of the aligned data, and, in combination with the spatial location of the multi-source sensors, establish a tensor matrix of time-frequency-space three-dimensional features.

[0012] Preferably, the intelligent diagnostic system for combined wind turbine impeller-gearbox faults further includes: a historical fault record identification unit, used to identify crack propagation morphology, abrasive particle distribution, and pitting structure changes in historical fault records; a macroscopic stiffness change modeling unit, used to introduce molecular dynamics simulation, construct the crack tip evolution path based on the crack identification results, and combine finite element simulation to model macroscopic stiffness changes; a degradation sensitivity parameter acquisition unit, used to obtain degradation sensitivity parameters based on the macroscopic stiffness change mapping; and a feature-damage mapping model construction unit, used to construct the feature-damage mapping model based on the degradation sensitivity parameters as training target annotations for the feature-damage mapping model.

[0013] Preferably, the intelligent diagnostic system for combined wind turbine rotor-gearbox faults further includes: a time-series signal database construction unit, used to extract curves of rotor dynamic load, gear meshing frequency response, and vibration over time to construct a time-series signal database; a path chain construction unit, used to use an information entropy flow algorithm to identify time-lag causal paths between multiple variables in the time-series signal database and construct path chains; and a fault source map acquisition unit, used to dynamically update time confidence levels, obtain a fault source map of state evolution based on the path chains, and add the fault source map to the feature-damage mapping model for auxiliary analysis.

[0014] Preferably, the intelligent diagnostic system for combined wind turbine rotor-gearbox faults further includes: a multi-channel neural network construction unit for constructing a multi-channel neural network, wherein the multi-channel neural network includes a time-frequency feature extraction channel, a temporal evolution identification channel, and a global correlation extraction channel; an unstructured feature acquisition unit for pre-training text vector encoding of unstructured data to obtain unstructured features; a multi-modal state classification construction unit for introducing a multi-head attention mechanism, combining the unstructured features, and weighted selection of the multi-modal features output by the multi-channel neural network model to construct a multi-modal state classification; and a joint output architecture construction unit for generating new samples based on the multi-modal state classification, performing remaining lifetime prediction, constructing a joint output architecture, and adding the joint output architecture to the feature-damage mapping model for auxiliary analysis.

[0015] Preferably, the intelligent diagnostic system for combined wind turbine rotor-gearbox faults further includes: a cloud-edge collaboration unit, used to deploy the feature-damage mapping model on an edge industrial computing platform for cloud-edge collaboration; a push unit, used to push risk levels, anomaly trend charts, potential faulty components, and suggested maintenance measures through cloud-edge collaboration; and a personalized maintenance strategy recommendation unit, used to combine historical case libraries and maintenance records to recommend personalized maintenance strategies based on the push results.

[0016] Secondly, this application also provides an intelligent diagnostic method for composite faults of wind turbine impeller and gearbox, including: deploying multi-source sensors based on coupled path analysis to collect multi-modal data from the multi-source sensors; constructing a tensor matrix based on the multi-modal data; constructing a feature-damage mapping model to perform fault analysis on the tensor matrix to obtain fault diagnosis results; pushing the fault diagnosis results and performing fault maintenance based on the fault diagnosis results.

[0017] The technical solution provided in this application has at least the following technical effects or advantages: by achieving the technical goal of accurate identification and visual tracing of complex faults in key components of wind turbines that integrate multi-source sensor information, it achieves the technical effects of improving fault diagnosis sensitivity, accuracy and system response efficiency, extending equipment service life and reducing maintenance costs.

[0018] The above description is merely an overview of the technical solution of this application. To better understand the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the intelligent diagnostic system for combined faults of wind turbine impeller and gearbox in this application.

[0021] Figure 2 This is a flowchart illustrating the intelligent diagnostic method for combined wind turbine impeller-gearbox faults presented in this application.

[0022] Figure labeling: Multimodal data acquisition module 1, tensor matrix construction module 2, fault analysis module 3, fault maintenance module 4. Detailed Implementation

[0023] This application provides an intelligent diagnostic system and method for combined faults in wind turbine rotors and gearboxes. It addresses the technical problems in existing technologies where the limited variety of sensor types, lack of systematic deployment, and reliance on empirical rules in signal processing prevent the effective identification of minute early-stage faults in critical wind turbine components under multi-field coupling, thus affecting the accuracy of fault diagnosis and the safe and stable operation of wind power equipment. The system achieves the technical goal of accurate identification and visualized tracing of combined faults in critical wind turbine components by integrating multi-source sensor information, thereby improving fault diagnosis sensitivity, accuracy, and system response efficiency, extending equipment lifespan, and reducing maintenance costs.

[0024] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. It should be understood that this application is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. It should also be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not all of them.

[0025] Example 1, please refer to the appendix. Figure 1 This application provides an intelligent diagnostic system for combined faults in wind turbine rotor-gearbox systems, specifically including:

[0026] Multimodal data acquisition module 1 is used for deploying multi-source sensors based on coupled path analysis and acquiring multimodal data from the multi-source sensors.

[0027] Specifically, the deployment of multi-source sensors based on coupling path analysis involves arranging sensors according to the force and vibration transmission paths between components, based on an understanding of the internal mechanical relationships of the wind turbine. The coupling path refers to the mutual influence between the impeller, main shaft, and gearbox. For example, the impeller undergoes slight deformation after being impacted by wind; this deformation is transmitted to the gearbox through the main shaft, causing changes in meshing stiffness. Therefore, there is significant dynamic coupling between the components. Guided by coupling analysis, sensors can be precisely deployed at the most information-sensitive locations, such as the impeller connection section, main bearing area, and gearbox side cover, thereby improving monitoring efficiency and effectiveness. Multi-source sensors refer to sensor devices using different physical principles to jointly collect data, such as fiber optic accelerometers, MEMS vibration sensors, strain gauges, and temperature sensors, reflecting the mechanical state from different perspectives. Multimodal data refers to data types from different sensors having different physical meanings and temporal resolutions. For example, fiber optic accelerometers provide high-frequency responses up to kilohertz, used to identify high-frequency vibrations caused by micro-cracks, while strain gauges primarily record low-frequency load variation trends. By fusing multi-source heterogeneous information, the operating status of equipment can be comprehensively perceived within the same time period, improving the accuracy of fault identification. For example, during a gearbox wear process, the temperature sensor recorded a continuous temperature rise, the vibration sensor detected an abnormal meshing frequency, and the strain gauge showed an imbalance of force on the spindle. The information from these three sources jointly confirms the fault, thereby greatly improving the reliability of the early warning.

[0028] Tensor matrix construction module 2 is used to construct a tensor matrix based on the multimodal data.

[0029] Specifically, a tensor matrix is ​​constructed based on multimodal data to integrate sensor data of different types, time scales, and spatial sources into a unified three-dimensional array structure for subsequent input into a deep learning model. Multimodal data refers to data from multiple sensors, which may include accelerometers recording high-frequency vibrations, encoders recording rotational speeds, thermal sensors recording temperature changes, and fiber optic sensors detecting structural strain, each representing different aspects of the device's state in the time, frequency, or spatial dimensions. The tensor matrix is ​​constructed by preprocessing the signals acquired by each sensor over a period of time and organizing them along a three-axis structure of "time-frequency-space." For example, one dimension represents the time step, the second represents frequency components or statistical characteristics (such as mean, variance, and spectral peak value), and the third represents the sensor's location or category. Through neural networks, temporal trends, spectral characteristics, and spatial distributions can be learned simultaneously under a unified input format, thereby achieving more accurate state identification or fault prediction.

[0030] Fault analysis module 3 is used to construct a feature-damage mapping model to perform fault analysis on the tensor matrix and obtain fault diagnosis results.

[0031] Specifically, a feature-damage mapping model is constructed to perform fault analysis on tensor matrices. Machine learning or deep learning techniques are used to build a model that learns the correspondence between multi-dimensional features extracted from the tensor matrix and known damage states, thereby achieving automatic diagnosis of new data. Features refer to the statistical regularities or physical characteristics reflecting the operating state of equipment from multi-source sensor data, such as the peak value, frequency energy density, and mean thermal distribution of vibration signals, which are extracted and used as input. Damage represents the real problems occurring in the mechanical structure, such as cracks, fatigue, wear, or pitting. The mapping model is essentially a functional relationship that learns how to predict the corresponding damage state from the input features. For example, a convolutional neural network can be used to identify the fatigue crack level of a wind turbine gearbox from a time-frequency-space tensor. The fault analysis process involves inputting a new tensor matrix into the model, allowing the model to determine whether a fault exists in the equipment, what type of fault it is, and its severity.

[0032] The fault maintenance module 4 is used to push the fault diagnosis results and perform fault maintenance based on the fault diagnosis results.

[0033] Specifically, the analyzed fault information is transmitted in real time to equipment managers, remote operation and maintenance systems, or cloud platforms via communication networks to facilitate timely maintenance decisions. Push notifications are a proactive transmission mechanism that uses edge computing nodes or IoT platforms to output diagnostic results in structured data format. Fault diagnosis results are used to determine the extent of damage by analyzing sensor-collected data, identifying the specific problem type, location, and severity. Fault maintenance refers to targeted operations based on the diagnostic results, including component replacement, lubrication, structural adjustments, or downtime for repairs. This enables efficient transmission and accurate response to fault information, reducing equipment failure rates and improving operational efficiency and resource utilization.

[0034] Furthermore, this application also includes: an aerodynamic load simulation unit for performing aerodynamic load simulation on the wind turbine and identifying the dynamic deformation region of the impeller; a wind power transmission simulation unit for performing wind power transmission simulation based on the dynamic deformation region of the impeller and analyzing the stress transmission path of the impeller-main shaft-gearbox; and a multi-source sensor deployment unit for deploying multi-source sensors on the impeller and gearbox respectively according to the stress transmission path.

[0035] Specifically, aerodynamic load simulation of the wind turbine is performed. Computational fluid dynamics software (such as Fluent) is used to simulate the aerodynamic force distribution on the wind turbine rotor under different wind speeds, directions, and turbulence conditions. This obtains a spatial distribution map of the wind force acting on the rotor under actual operating conditions, thereby identifying areas prone to significant dynamic deformation. The dynamic deformation region of the rotor refers to the areas that experience periodic bending, torsion, or vibration due to repeated wind forces during rotation, thus serving as high-risk points for fatigue failure, such as the blade tip and root. Identifying the dynamic deformation region of the rotor provides boundary conditions for subsequent structural stress analysis.

[0036] Next, aerodynamic loads are applied to the three-dimensional model of the impeller using a multibody dynamics simulation platform (such as RecurDyn or Adams). By tracing how wind force is transmitted from the impeller to the main shaft and then to components such as the gearbox, the complete stress path from the impeller to the gearbox output can be calculated. The wind force acting on the impeller will cause torsional or bending stresses on the main shaft, which are then transmitted to the inside of the gearbox through the coupling, thus affecting the gear meshing state and bearing stress. Analyzing the stress transmission path helps to identify key nodes in the load transmission chain, thus revealing where mechanical wear, crack propagation, or pitting damage is most likely to occur.

[0037] Next, based on the stress transmission path, multi-source sensors are deployed at key locations in the impeller and gearbox. Multi-source sensor deployment refers to the combined use of sensors with different functions and dimensions to achieve comprehensive perception of the structural state. Distributed fiber optic sensors are used on the impeller to monitor minute strain changes along the blade length in real time, offering high spatial resolution and resistance to electromagnetic interference. Inside the gearbox, microelectromechanical systems (MEMS) accelerometers are more suitable due to their small size and fast response, making them ideal for installation in confined spaces to detect vibration signals.

[0038] Furthermore, this application also includes: an indicator system hierarchical structure construction unit, used to construct a hierarchical structure of the indicator system for deployment optimization, wherein the hierarchical structure of the indicator system includes a target layer, a criterion layer, and a scheme layer, wherein the criterion layer includes convenience, signal-to-noise ratio, redundancy, and monitoring capability; a simulation result filtering unit, used to construct a fuzzy judgment matrix, with the sensor deployment optimization of the target layer as the objective, to filter the simulation results of the scheme layer generated based on the simulation of the criterion layer, and output the optimization results; and a multi-source sensor update and deployment unit, used to update the deployment of multi-source sensors based on the optimization results.

[0039] Specifically, a hierarchical structure of indicators for deployment optimization is constructed, comprising an objective layer, a criterion layer, and a scheme layer. The objective layer represents the ultimate optimization goal, namely, sensor deployment optimization. The criterion layer comprises evaluation standards for measuring the merits of different schemes, including ease of use, signal-to-noise ratio (SNR), redundancy, and monitorability. Ease of use indicates the ease of sensor installation, especially under conditions of limited space inside wind turbines or high-altitude operations. SNR represents whether the sensor can acquire a clear signal and whether background noise interferes with identification. Redundancy indicates whether deployed sensors have the ability to supplement and back up each other across different locations. Monitorability refers to whether the deployed sensors can effectively cover key fault areas, ensuring the representativeness and integrity of the data.

[0040] Next, each sensor deployment scheme in the scheme layer is compared pairwise with each indicator in the criteria layer, and a fuzzy score of relative importance is given, such as "slightly better" or "significantly better". Then, the performance of each scheme under each indicator is weighted and summed according to the fuzzy comprehensive evaluation method, and finally the degree of superiority or inferiority of each scheme relative to the deployment optimization target is calculated. The output is the comprehensive score of each sensor deployment scheme, thereby selecting the optimal or second-best deployment layout.

[0041] Finally, based on the optimization results, the deployment of multi-source sensors is updated, and the types and locations of sensors are readjusted. For example, if the analysis reveals that a certain MEMS sensor has a high signal-to-noise ratio and is easy to deploy near the gearbox input shaft, then this type of sensor can be prioritized for deployment in that area in the final deployment plan.

[0042] Furthermore, this application also includes: a deployment visualization unit, used to perform deployment visualization of the optimization results based on a three-dimensional model, filter overlapping results based on the visualization results, and obtain updated optimization results.

[0043] Specifically, the optimization results are visualized using a 3D model. The sensor placement scheme obtained from the optimization algorithm is mapped onto a 3D geometric model with a spatial structure, allowing the position, orientation, and quantity of sensors to be visually displayed within the physical structure of the wind turbine. The 3D model consists of a digital twin model or CAD structural drawing of the wind turbine, including the actual dimensions and positional relationships of key components such as the impeller, main shaft, and gearbox. During deployment visualization, each sensor location is embedded in the model as spatial coordinates and distinguished by color, shape, or number, clearly revealing any issues such as overly dense deployments, omissions, or positional conflicts.

[0044] Subsequently, overlapping results were filtered based on the visualization results. In the visualization, areas where multiple sensor deployment locations spatially overlapped or functionally redundant were identified and optimized. Overlapping results typically include multiple sensors located in extremely close spatial positions, or excessive redundancy in the monitored physical quantities, failing to effectively improve the system's monitoring capabilities. For example, if two accelerometers are deployed simultaneously above a gearbox with the same monitoring direction, duplicate data collection and wasted resources may occur.

[0045] After 3D visualization and overlay filtering, the updated optimization results were obtained, eliminating spatial and functional conflicts between deployments and resulting in higher execution feasibility and operational efficiency. For example, the initial optimization results might have suggested deploying two accelerometers on each of the left and right sides of the gearbox. However, 3D visualization revealed that the space on the right side was too narrow, posing an installation interference problem. Therefore, only two accelerometers were ultimately deployed on the left side, and the right-side sensors were moved to a location near the upper support, thus maintaining the same monitoring coverage while improving actual accessibility.

[0046] Furthermore, this application also includes: a pre-aligned data obtaining unit, used to align the time window of the multimodal data, and then perform dynamic resampling and interpolation alignment to obtain pre-aligned data; an aligned data obtaining unit, used to decompose the non-stationary signals in the pre-aligned data to obtain aligned data; and a tensor matrix establishing unit, used to extract the time-domain features, frequency-domain features, and collaborative features of the aligned data, and combine them with the spatial location of the multi-source sensors to establish a tensor matrix of time-frequency-space three-dimensional features.

[0047] Specifically, data from different sensors suffers from time asynchrony due to varying sampling frequencies, inconsistent start-up times, or signal loss. After aligning the time windows of the multimodal data, dynamic resampling and interpolation alignment are performed to obtain pre-aligned data. Multimodal data includes observation information from different types of devices such as fiber optics, accelerometers, and temperature sensors, resulting in time-axis misalignment during acquisition. Time window alignment refers to slicing all data into uniform time segments to ensure subsequent analysis is conducted on the same time scale. After time alignment, dynamic resampling and interpolation methods are needed to fill in missing data points or adjust the data sampling density. Dynamic resampling adjusts the sampling rate based on the actual rate of change of sensor data, while interpolation uses mathematical algorithms to estimate the values ​​of missing data points, such as commonly used linear interpolation and spline interpolation. The final result is a preliminary unified time series data, called pre-aligned data.

[0048] Non-stationary signals in the pre-aligned data are decomposed to obtain aligned data. Non-stationary signals refer to signals whose statistical characteristics, such as mean and variance, change over time. These are frequently found in wind turbines, such as acceleration fluctuations caused by changes in blade load. Non-stationarity affects the accuracy of subsequent signal analysis, thus requiring decomposition methods. Decomposition methods include variational mode decomposition, empirical mode decomposition, and wavelet packet decomposition, which can break down complex non-stationary signals into several stable sub-signals. By decomposing the pre-aligned data, noise components in the signal can be eliminated, key patterns related to equipment failure can be highlighted, and clearer aligned data can be obtained, providing a high-quality foundation for feature extraction.

[0049] Temporal, frequency, and collaborative features of the aligned data are extracted and combined with the spatial locations of multiple sensors to establish a tensor matrix of time-frequency-space three-dimensional features. Temporal features refer to statistics directly extracted from the time series, such as mean, standard deviation, kurtosis, and skewness, reflecting the overall volatility of the signal. Frequency features are frequency distribution information extracted through Fourier transforms and wavelet transforms, such as dominant frequency, band energy, and frequency center, reflecting the periodicity or resonance characteristics of the signal. Collaborative features refer to the interrelationships between multiple sensors, such as coherence, cross-correlation, and phase difference, used to identify the coupling behavior between different components in the system. Since the sensors are deployed at different locations on the wind turbine, each type of feature also needs to be organized in conjunction with the spatial coordinate information of the sensors. Finally, the features are integrated into a tensor matrix across the time, frequency, and spatial dimensions.

[0050] Furthermore, this application also includes: a historical fault record identification unit, used to identify crack propagation morphology, abrasive grain distribution, and pitting structure changes in historical fault records; a macroscopic stiffness change modeling unit, used to introduce molecular dynamics simulation, construct the crack tip evolution path based on the crack identification results, and combine finite element simulation to model macroscopic stiffness changes; a degradation sensitivity parameter acquisition unit, used to obtain degradation sensitivity parameters based on the macroscopic stiffness change mapping; and a feature-damage mapping model construction unit, used to construct the feature-damage mapping model based on the degradation sensitivity parameters as training target annotations for the feature-damage mapping model.

[0051] Specifically, historical fault records are analyzed to identify crack propagation morphology, abrasive particle distribution, and pitting structure changes, extracting typical patterns of microscopic damage evolution from existing equipment malfunctions. Crack propagation morphology refers to the geometric changes in the crack shape during its propagation within the material, such as the crack gradually widening or bifurcating, reflecting the changing trend of stress concentration areas within the material. Abrasive particle distribution refers to the spatial distribution of metal particles or impurities generated by friction on the surface of lubricating oil or materials, used to determine whether excessive wear exists between components. Pitting structure changes indicate variations in the number, shape, and depth of small pits formed on the metal surface due to chemical or electrochemical reactions, representing the occurrence and development process of corrosive damage.

[0052] Molecular dynamics simulation is introduced to construct the crack tip evolution path based on crack identification results. This is combined with finite element method (FEM) simulation to model macroscopic stiffness changes, thus linking microscopic damage mechanisms with macroscopic structural performance changes. Molecular dynamics simulation is a computational method that simulates the motion of atoms and molecules at the nanoscale. It helps to study how materials fracture and recombine at the atomic level when subjected to crack impact, thereby deducing the direction, rate, and path of crack tip propagation. Finite element method (FEM) simulation is an engineering modeling technique that divides the entire structure into discrete elements and evaluates the stress and stiffness changes of the overall structure under load through numerical calculations. When the crack tip path obtained from molecular dynamics is introduced into the FEM model as input, the overall stiffness reduction of the material due to crack propagation can be accurately predicted, for example, from an initial 400 kN / mm to 370 kN / mm, achieving a bridge between microscopic damage and macroscopic performance.

[0053] Degradation sensitivity parameters are obtained by mapping macroscopic stiffness changes, and the trend of structural stiffness changes with time or damage is mathematically modeled to quantify the impact of different damages on structural performance. Macroscopic stiffness is an indicator reflecting a structure's resistance to deformation; as internal cracks propagate, corrosion occurs, or fatigue accumulates, stiffness gradually decreases. Degradation sensitivity parameters are quantitative descriptions of stiffness changes under different damage modes, such as the slope, fluctuation amplitude, or critical point.

[0054] Using degradation sensitivity parameters as training targets for the feature-damage mapping model, a feature-damage mapping model is constructed. This model is then used to train an artificial intelligence model to predict the mapping between input features and actual damage states. The input to the feature-damage mapping model can include tensor-processed time-frequency-space features such as signal amplitude, dominant frequency, band energy, and signal correlation. The output target is the corresponding degradation sensitivity parameter. Using the training target as a label reflects the changing trend of damage severity. The model is trained using supervised learning, enabling it to predict the current damage level of the wind power structure based on newly acquired sensor data, and even estimate the damage development trend over a future period.

[0055] Furthermore, this application also includes: a time-series signal database construction unit, used to extract curves of impeller dynamic load, gear meshing frequency response, and vibration changing over time to construct a time-series signal database; a path chain construction unit, used to use an information entropy flow algorithm to identify time-lag causal paths between multiple variables in the time-series signal database and construct path chains; and a fault source map acquisition unit, used to dynamically update time confidence, obtain a fault source map of state evolution based on the path chains, and add the fault source map to the feature-damage mapping model for auxiliary analysis.

[0056] Specifically, the dynamic load of the impeller, the gear meshing frequency response, and the vibration over time curves are extracted to construct a time-series signal database. This involves collecting various dynamic response information of wind power equipment during operation and forming a structured database with time as the main axis. The impeller dynamic load refers to the periodic load generated by the impeller under wind force, which fluctuates with changes in wind speed and blade angle, causing periodic impacts on the main shaft and gearbox. The gear meshing frequency response reflects the characteristic vibration frequency generated by the contact and disengagement of gear teeth during operation, revealing whether there are gear defects, uneven loads, or abnormal lubrication. The vibration over time curve is an important signal of equipment operational stability; for example, a sudden increase in vibration amplitude at a certain time period may indicate loosening or damage to internal components.

[0057] The information entropy flow algorithm is used to identify time-lag causal paths between multiple variables in a time-series signal database, constructing path chains to reveal the temporal interactions between various types of sensor data. Information entropy is an indicator of signal uncertainty and information content. When a change in the information entropy of one variable can predict a change in another variable, it indicates a causal relationship between the two. The information entropy flow algorithm is a technique used to determine causality between variables. By analyzing the trends in information entropy changes of each variable within different time windows, it identifies the sequential relationships, thereby obtaining time-lag causal paths. A path chain represents a causal sequence in which a change in one variable causes corresponding changes in other variables within a certain time period. For example, a change in impeller load may cause an abnormality in gear meshing frequency one second later, which in turn triggers an increase in vibration signals, forming a complete dynamic transmission path.

[0058] The system dynamically updates time-based confidence levels, derives a fault source map based on the path chain to represent the state evolution, and adds this map to the feature-damage mapping model for auxiliary analysis. This allows for dynamic management and visual integration of the identified causal paths. Confidence level is a quantitative measure of the validity of a causal relationship. Dynamic updates refer to the real-time reassessment of the effectiveness and stability of the path chain after acquiring new sensor data during equipment operation. The fault source map is a graph-structured representation used to describe the entire process of tracing possible causes upstream from the surface fault phenomenon. Nodes in the graph represent different signal variables or structural states, edges represent causal relationships, and edge weights represent confidence levels. Introducing the fault source map into the feature-damage mapping model enhances the model's understanding and prediction capabilities of complex state change patterns, improving the accuracy and response time of fault identification.

[0059] Furthermore, this application also includes: a multi-channel neural network construction unit for constructing a multi-channel neural network, wherein the multi-channel neural network includes a time-frequency feature extraction channel, a temporal evolution identification channel, and a global correlation extraction channel; an unstructured feature acquisition unit for pre-training text vector encoding of unstructured data to obtain unstructured features; a multimodal state classification construction unit for introducing a multi-head attention mechanism, combining the unstructured features, and weighted selection of the multimodal features output by the multi-channel neural network model to construct a multimodal state classification; and a joint output architecture construction unit for generating new samples based on the multimodal state classification, performing remaining lifetime prediction, constructing a joint output architecture, and adding the joint output architecture to the feature-damage mapping model for auxiliary analysis.

[0060] Specifically, to process the multi-dimensional, multi-modal signal data generated during the operation of complex equipment such as wind turbines, a neural network structure with multiple functional branches is constructed. Multiple channels serve as independent data processing paths, with each channel focusing on processing a specific type of feature information. The time-frequency feature extraction channel is primarily for data like vibration signals that can be converted into a spectrum or time-frequency graph, using convolutional neural networks to extract the patterns between frequency components and time variations. The temporal evolution identification channel is suitable for analyzing the evolution patterns in time series data, such as the changing trends of dynamic loads or temperature over time, often implemented using recurrent neural networks or long short-term memory networks. The global correlation extraction channel focuses on the overall correlation between different signal sources, such as the inherent coupling relationship between multiple signals from the impeller and gearbox, using graph neural networks or attention mechanisms to achieve information fusion.

[0061] Pre-trained text vector encoding of unstructured data is used to process input information that cannot be directly represented as numerical values ​​in a fixed format, such as maintenance logs, operation and maintenance records, or manually annotated reports. Unstructured data contains natural language or symbolic information. Pre-trained text vector encoding refers to using a pre-trained natural language processing model (such as BERT or Word2Vec) to convert text information into vector form so that neural networks can recognize it.

[0062] The introduction of a multi-head attention mechanism aims to more accurately identify the most critical information when processing multimodal features. An attention mechanism is an algorithm that mimics human concentration, automatically focusing on the most relevant parts of the input data based on the needs of the current task. A multi-head attention mechanism involves multiple attention mechanisms operating in parallel, with each "head" learning information from different dimensions or angles, thereby improving the model's ability to understand the relationships between complex features. Combining unstructured features, the encoded text data is used as supplementary information and input into the model along with structured data such as sensor signals for weighted fusion. Weighted selection means that the model automatically assigns weights to each type of information based on the specific task, thereby extracting the most relevant features to construct accurate multimodal state classifications, that is, classifying the current device operating status into different types such as healthy, warning, and fault.

[0063] New samples are generated based on multimodal state classification, and the current model's classification results are used to simulate and expand potential unseen states. The generation of new samples typically employs Generative Adversarial Networks (GANs) or data augmentation strategies to improve the model's robustness in situations with scarce samples. Next, remaining lifetime prediction is performed, estimating how much longer the equipment can continue to operate normally based on the identified state development trends. Combining historical data curves, degradation models, and state evolution patterns, a specific remaining lifetime value is output, such as 300 hours or 500 revolutions of remaining operating time. Constructing a joint output architecture means simultaneously outputting multiple results within a single model, such as state classification results, lifetime prediction values, and feature embedding vectors, unifying the processing flow and improving overall efficiency. Finally, the joint output architecture is added to the feature-damage mapping model as an auxiliary analysis tool to further enhance the ability to identify and predict complex failure mechanisms.

[0064] Furthermore, this application also includes: a cloud-edge collaboration unit, used to deploy the feature-damage mapping model on an edge industrial computing platform for cloud-edge collaboration; a push unit, used to push risk levels, anomaly trend charts, potential faulty components and suggested maintenance measures through cloud-edge collaboration; and a personalized maintenance strategy recommendation unit, used to combine historical case libraries and maintenance records to recommend personalized maintenance strategies based on the push results.

[0065] Specifically, the feature-damage mapping model is deployed on an edge industrial computing platform for cloud-edge collaboration. The trained damage identification model is embedded into edge devices close to the data source, such as wind turbine tower controllers or intelligent data acquisition terminals near the impeller. The edge industrial computing platform refers to a small industrial computer with certain processing capabilities, capable of processing sensor data in real time and making preliminary judgments. Cloud-edge collaboration means that the edge device is responsible for data preprocessing and preliminary analysis, while the cloud is responsible for training, updating, and complex calculations of the deep learning model. The two communicate via network to achieve iterative model updates, algorithm distribution, and data synchronization, improving the response speed of fault warnings and reducing bandwidth dependence. For example, edge devices can identify abnormal trends within one second and update model parameters via the cloud.

[0066] By collaboratively pushing risk levels, anomaly trend charts, potential faulty components, and suggested maintenance measures through cloud-edge processing, the edge device detects potential anomalies and uploads relevant data to the cloud for further processing. The cloud model then analyzes the data to generate clear diagnostic results. Risk levels reflect the degree of danger in the current equipment's operating status, categorized as low, medium, and high risk. Anomaly trend charts display curves showing how key equipment indicators change over time, helping maintenance personnel intuitively determine if the problem is worsening. Potential faulty components identify critical areas that may be damaged, such as impeller connecting sections or gearbox input shafts, based on multimodal signals. Suggested maintenance measures are responses based on existing knowledge and reasoning mechanisms.

[0067] By combining historical case databases and maintenance records, personalized maintenance strategies are recommended based on the push notifications. The current analysis results are compared with past fault cases accumulated in the system to identify the most similar historical scenarios. The historical case database includes the damage evolution process of equipment under different operating conditions and corresponding maintenance methods, while maintenance records refer to detailed logs of each maintenance session, including the time, method, and effect. By comparing the pushed risk level and fault location, the system can recommend the most similar handling experience, thereby generating more targeted and economical maintenance plans. For example, if a wind turbine previously experienced bolt loosening after a sudden change in gearbox meshing frequency and successfully restored normal operation through tightening, then when another wind turbine experiences a similar frequency anomaly, tightening and inspection will be prioritized instead of directly replacing parts, thus saving maintenance costs. Table 1 shows a partial record of the most recent wind turbine equipment fault diagnosis and maintenance strategy.

[0068] Table 1: Partial Record of the Most Recent Wind Turbine Equipment Fault Diagnosis and Maintenance Strategy

[0069]

[0070] In summary, the intelligent diagnostic system for combined wind turbine rotor-gearbox faults provided in this application has the following technical effects: by achieving the technical goal of accurate identification and visual tracing of combined faults in key wind turbine components by integrating multi-source sensor information, it achieves the technical effects of improving fault diagnosis sensitivity, accuracy and system response efficiency, extending equipment service life and reducing maintenance costs.

[0071] Example 2: Based on the same inventive concept as the intelligent diagnostic system for combined wind turbine rotor-gearbox faults in the foregoing examples, this application also provides an intelligent diagnostic method for combined wind turbine rotor-gearbox faults. Please refer to the appendix. Figure 2 ,include:

[0072] Furthermore, the intelligent diagnosis method for combined wind turbine impeller-gearbox faults is also used for: performing aerodynamic load simulation on the wind turbine to identify the dynamic deformation area of ​​the impeller; performing wind power transmission simulation based on the dynamic deformation area of ​​the impeller to analyze the stress transmission path of the impeller-main shaft-gearbox; and deploying multi-source sensors on the impeller and gearbox respectively according to the stress transmission path.

[0073] Furthermore, the intelligent diagnosis method for wind turbine rotor-gearbox composite faults is also used for: constructing a hierarchical structure of an indicator system for deployment optimization, wherein the hierarchical structure of the indicator system includes a target layer, a criterion layer, and a scheme layer, wherein the criterion layer includes convenience, signal-to-noise ratio, redundancy, and monitorability; constructing a fuzzy judgment matrix, taking the sensor deployment optimization of the target layer as the objective, filtering the simulation results of the scheme layer generated based on the simulation of the criterion layer, and outputting the optimization results; and updating the deployment of multi-source sensors based on the optimization results.

[0074] Furthermore, the intelligent diagnosis method for wind turbine impeller-gearbox composite faults is also used to: visualize the optimization results based on a three-dimensional model, filter overlapping results based on the visualization results, and obtain updated optimization results.

[0075] Furthermore, the intelligent diagnosis method for wind turbine impeller-gearbox composite faults is also used for: aligning the time window of the multimodal data, performing dynamic resampling and interpolation alignment to obtain pre-aligned data; decomposing the non-stationary signals in the pre-aligned data to obtain aligned data; extracting the time-domain features, frequency-domain features, and collaborative features of the aligned data, and combining them with the spatial location of the multi-source sensors to establish a tensor matrix of time-frequency-space three-dimensional features.

[0076] Furthermore, the intelligent diagnostic method for wind turbine impeller-gearbox composite faults is also used for: identifying crack propagation morphology, abrasive particle distribution, and pitting structure changes in historical fault records; introducing molecular dynamics simulation to construct the crack tip evolution path based on crack identification results, and combining finite element simulation to model macroscopic stiffness changes; mapping degradation sensitivity parameters based on the macroscopic stiffness changes; and constructing the feature-damage mapping model by using the degradation sensitivity parameters as training target labels for the feature-damage mapping model.

[0077] Furthermore, the intelligent diagnosis method for combined wind turbine rotor-gearbox faults is also used to: extract the curves of rotor dynamic load, gear meshing frequency response, and vibration over time to construct a time-series signal database; use the information entropy flow algorithm to identify the time-lag causal paths between multiple variables in the time-series signal database and construct path chains; dynamically update the time confidence level; obtain a fault source map of state evolution based on the path chains; and add the fault source map to the feature-damage mapping model for auxiliary analysis.

[0078] Furthermore, the intelligent diagnosis method for wind turbine rotor-gearbox composite faults is also used for: constructing a multi-channel neural network, wherein the multi-channel neural network includes a channel for extracting time-frequency features, a channel for identifying temporal evolution, and a channel for extracting global correlation; pre-training text vector encoding for unstructured data to obtain unstructured features; introducing a multi-head attention mechanism, combining the unstructured features, and weighting the selection of multimodal features output by the multi-channel neural network model to construct a multimodal state classification; generating new samples based on the multimodal state classification, performing remaining lifetime prediction, constructing a joint output architecture, and adding the joint output architecture to the feature-damage mapping model for auxiliary analysis.

[0079] Furthermore, the intelligent diagnosis method for wind turbine impeller-gearbox composite faults is also used to: deploy the feature-damage mapping model on an edge industrial computing platform for cloud-edge collaboration; push risk levels, anomaly trend charts, potential faulty components, and suggested maintenance measures through cloud-edge collaboration; and recommend personalized maintenance strategies based on the push results, combined with historical case libraries and maintenance records.

[0080] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The intelligent diagnosis system and specific examples of wind turbine rotor-gearbox composite faults in the aforementioned embodiment one are also applicable to the intelligent diagnosis method of wind turbine rotor-gearbox composite faults in this embodiment. Through the foregoing detailed description of the intelligent diagnosis system of wind turbine rotor-gearbox composite faults, those skilled in the art can clearly understand the intelligent diagnosis method of wind turbine rotor-gearbox composite faults in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0081] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0082] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application also intends to include such modifications and variations.

Claims

1. An intelligent diagnostic system for combined faults in wind turbine impeller and gearbox, characterized in that, include: The multimodal data acquisition module is used for deploying multi-source sensors based on coupled path analysis and acquiring multimodal data from the multi-source sensors. The tensor matrix construction module is used to construct a tensor matrix based on the multimodal data; The fault analysis module is used to construct a feature-damage mapping model to perform fault analysis on the tensor matrix and obtain fault diagnosis results. The fault maintenance module is used to push the fault diagnosis results and perform fault maintenance based on the fault diagnosis results.

2. The intelligent diagnostic system for combined wind turbine impeller-gearbox faults as described in claim 1, characterized in that, The multimodal data acquisition module includes: The aerodynamic load simulation unit is used to perform aerodynamic load simulation on wind turbines and identify the dynamic deformation area of ​​the impeller. The wind power transmission simulation unit is used to simulate wind power transmission based on the dynamic deformation region of the impeller and analyze the stress transmission path from the impeller to the main shaft to the gearbox. The multi-source sensor deployment unit is used to deploy multi-source sensors on the impeller and gearbox according to the stress transmission path.

3. The intelligent diagnostic system for combined wind turbine impeller-gearbox faults as described in claim 2, characterized in that, The multi-source sensor deployment unit includes: The indicator system hierarchical structure construction unit is used to construct and optimize the indicator system hierarchical structure. The indicator system hierarchical structure includes a target layer, a criterion layer, and a scheme layer. The criterion layer includes convenience, signal-to-noise ratio, redundancy, and monitorability. The simulation result filtering unit is used to construct a fuzzy judgment matrix, with the sensor deployment optimization of the target layer as the objective, to filter the simulation results of the scheme layer generated based on the simulation of the criterion layer, and output the optimization results; A multi-source sensor update and deployment unit is used to update the deployment of multi-source sensors based on the optimization results.

4. The intelligent diagnostic system for combined wind turbine impeller-gearbox faults as described in claim 3, characterized in that, The multi-source sensor update deployment unit includes: A deployment visualization unit is used to visualize the optimization results based on a 3D model, filter overlapping results based on the visualization results, and obtain updated optimization results.

5. The intelligent diagnostic system for combined wind turbine impeller-gearbox faults as described in claim 1, characterized in that, The tensor matrix construction module includes: The pre-aligned data acquisition unit is used to align the time window of the multimodal data, perform dynamic resampling and interpolation alignment, and obtain pre-aligned data. The alignment data obtaining unit is used to decompose the non-stationary signals in the pre-aligned data to obtain aligned data; The tensor matrix establishment unit is used to extract the temporal, frequency, and collaborative features of the aligned data, and, in combination with the spatial location of the multi-source sensors, establish a tensor matrix of time-frequency-space three-dimensional features.

6. The intelligent diagnostic system for combined wind turbine impeller-gearbox faults as described in claim 1, characterized in that, The fault analysis module includes: The historical fault record identification unit is used to identify crack propagation morphology, abrasive grain distribution, and pitting structure changes in historical fault records. The macroscopic stiffness variation modeling unit is used to introduce molecular dynamics simulation, construct the crack tip evolution path based on crack identification results, and combine finite element simulation to model macroscopic stiffness variation. A degradation sensitivity parameter acquisition unit is used to obtain the degradation sensitivity parameter based on the macroscopic stiffness change mapping. The feature-damage mapping model construction unit is used to construct the feature-damage mapping model by using the degradation sensitivity parameter as the training target label of the feature-damage mapping model.

7. The intelligent diagnostic system for combined wind turbine impeller-gearbox faults as described in claim 6, characterized in that, The feature-damage mapping model construction unit includes: The time-series signal database construction unit is used to extract the curves of impeller dynamic load, gear meshing frequency response and vibration over time to construct a time-series signal database. The path chain construction unit is used to identify time-lag causal paths between multiple variables in the time-series signal database using the information entropy flow algorithm, and to construct path chains. The fault source map acquisition unit is used to dynamically update the time confidence level. Based on the path chain, a fault source map of state evolution is obtained, and the fault source map is added to the feature-damage mapping model for auxiliary analysis.

8. The intelligent diagnostic system for combined wind turbine impeller-gearbox faults as described in claim 6, characterized in that, The feature-damage mapping model construction unit also includes: A multi-channel neural network building unit is used to build a multi-channel neural network, wherein the multi-channel neural network includes a time-frequency feature extraction channel, a time-series evolution recognition channel, and a global correlation extraction channel; Unstructured feature generation units are used to pre-train text vector encoding of unstructured data to obtain unstructured features; A multimodal state classification construction unit is used to introduce a multi-head attention mechanism, combine the unstructured features, and weight the selection of multimodal features output by the multi-channel neural network model to construct a multimodal state classification. The joint output architecture construction unit is used to generate new samples based on the multimodal state classification, perform remaining lifetime prediction, construct the joint output architecture, and add the joint output architecture to the feature-damage mapping model for auxiliary analysis.

9. The intelligent diagnostic system for combined wind turbine impeller-gearbox faults as described in claim 1, characterized in that, The fault maintenance module includes: The cloud-edge collaboration unit is used to deploy the feature-damage mapping model on an edge industrial computing platform for cloud-edge collaboration. The push unit is used to push risk levels, anomaly trend charts, potential faulty components, and recommended maintenance measures through cloud-edge collaboration. The personalized maintenance strategy recommendation unit is used to combine historical case databases and maintenance records to recommend personalized maintenance strategies based on the push results.

10. An intelligent diagnostic method for combined faults in wind turbine impellers and gearboxes, characterized in that, The intelligent diagnostic system for combined wind turbine rotor-gearbox faults as described in any one of claims 1 to 9 is executed, including: Deployment of multi-source sensors based on coupling path analysis to collect multimodal data from multiple sources; Construct a tensor matrix based on the multimodal data; A feature-damage mapping model is constructed to perform fault analysis on the tensor matrix, and fault diagnosis results are obtained. The fault diagnosis results are pushed out, and fault maintenance is performed based on the fault diagnosis results.

Images