Dynamic integrated fault detection method and system for wind turbine gearbox based on feature fusion

Abstract

The present application provides a method and system for dynamic integrated fault detection of wind turbine gearboxes based on feature fusion, which relates to the field of data processing technology. The method includes: obtaining each preliminary feature set; quantifying the feature-feature and feature-fault nonlinear relationships of each preliminary feature set to obtain each preliminary screening feature set; analyzing the contribution of each feature in each preliminary screening feature set, performing feature fine screening, and generating each fine screening feature set; calling a multi-agent integrated fault detection system, and performing bottom-up multi-layer agent information sharing and collaborative decision-making based on each fine screening feature set to generate wind turbine gearbox fault detection information. This application solves the technical problem in the prior art that the fault detection accuracy is not high due to the lack of deep mining of features and the reliance on a single-machine diagnostic system. By constructing a multi-agent integrated fault detection system, the false alarm rate and missed alarm rate are significantly reduced, achieving the technical effect of improving the accuracy of fault detection.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method and system for dynamic integrated fault detection of wind turbine gearboxes based on feature fusion. Background Art

[0002] In wind power generation equipment, the wind turbine gearbox is the core transmission component, and its operating status directly affects the reliability and power generation efficiency of the entire machine.

[0003] However, gearbox fault diagnosis faces multiple technical challenges: First, existing feature engineering mostly relies on manual experience to design time-frequency domain indicators (such as kurtosis, energy entropy, etc.), which has limited ability to characterize early weak faults such as gear pitting and bearing micropitting, and is prone to feature redundancy or omission; second, the single-machine diagnosis system adopts a static model architecture. Under dynamic working conditions such as sudden changes in wind speed and load fluctuations, the model parameters are fixed, resulting in a decrease in diagnostic accuracy, and it cannot adapt to complex fault modes such as the gradual change of wear state during long-term operation of the gearbox.

[0004] In summary, the existing technology has technical problems such as low fault detection accuracy due to the lack of in-depth feature mining and reliance on a single-machine diagnosis system. Summary of the Invention

[0005] The purpose of this application is to provide a dynamic integrated fault detection method and system for wind turbine gearboxes based on feature fusion, so as to solve the technical problems in the prior art that the fault detection accuracy is low due to the lack of in-depth mining of features and reliance on a single-machine diagnostic system.

[0006] In view of the above problems, the present application provides a dynamic integrated fault detection method and system for wind turbine gearboxes based on feature fusion.

[0007] In the first aspect, the present application provides a dynamic integrated fault detection method for a wind turbine gearbox based on feature fusion, and the dynamic integrated fault detection method for a wind turbine gearbox based on feature fusion is implemented by a dynamic integrated fault detection system for a wind turbine gearbox based on feature fusion, wherein the dynamic integrated fault detection method for a wind turbine gearbox based on feature fusion includes: collecting vibration data sets of various parts of the wind turbine gearbox and performing noise separation processing, performing multi-dimensional deep feature extraction on the processed data through a multi-dimensional deep feature extraction engine to obtain various preliminary feature sets, wherein the multi-dimensional deep feature extraction engine includes a time-frequency domain feature enhancement layer, a deep feature mining layer and a dynamic optimization mechanism; calling the maximum information coefficient preliminary screening mechanism to quantify the feature-feature and feature-fault nonlinear relationships of the various preliminary feature sets, and screening to obtain various preliminary screening feature sets; analyzing the contribution of each feature in each preliminary screening feature set based on the SHAP value, constructing a feature importance map, performing feature fine screening, and generating various fine screening feature sets; calling a multi-agent integrated fault detection system, and performing bottom-up multi-layer agent information sharing and collaborative decision-making based on the various fine screening feature sets to generate wind turbine gearbox fault detection information.

[0008] In the second aspect, the present application also provides a wind turbine gearbox dynamic integrated fault detection system based on feature fusion, which is used to execute the wind turbine gearbox dynamic integrated fault detection method based on feature fusion as described in the first aspect, wherein the wind turbine gearbox dynamic integrated fault detection system based on feature fusion includes: a feature extraction module for collecting vibration data sets of various parts of the wind turbine gearbox and performing noise separation processing, performing multi-dimensional deep feature extraction on the processed data through a multi-dimensional deep feature extraction engine to obtain various preliminary feature sets, wherein the multi-dimensional deep feature extraction engine includes a time-frequency domain feature enhancement layer, a depth feature Mining layer and dynamic optimization mechanism; a feature initial screening module, used to call the maximum information coefficient initial screening mechanism to quantify the nonlinear relationship between features and features, and features and faults of each preliminary feature set, and screen out each preliminary screening feature set; a feature fine screening module, used to analyze the contribution of each feature in each preliminary screening feature set based on the SHAP value, construct a feature importance map, perform feature fine screening, and generate each fine screening feature set; a fault detection module, used to call the multi-agent integrated fault detection system, perform bottom-up multi-layer agent information sharing and collaborative decision-making based on each fine screening feature set, and generate wind turbine gearbox fault detection information.

[0009] One or more technical solutions provided in this application have at least the following beneficial effects:

[0010] Vibration data sets of various parts of the wind turbine gearbox are collected and noise separation processing is performed. Multi-dimensional deep feature extraction is performed on the processed data through a multi-dimensional deep feature extraction engine to obtain various preliminary feature sets, wherein the multi-dimensional deep feature extraction engine includes a time-frequency domain feature enhancement layer, a deep feature mining layer and a dynamic optimization mechanism; the maximum information coefficient preliminary screening mechanism is called to quantify the feature-feature and feature-fault nonlinear relationships of the various preliminary feature sets, and the various preliminary screening feature sets are screened; the contribution of each feature in each preliminary screening feature set is analyzed based on the SHAP value, a feature importance map is constructed, and feature fine screening is performed to generate various fine screening feature sets; the multi-agent integrated fault detection system is called to perform bottom-up multi-layer agent information sharing and collaborative decision-making based on the various fine screening feature sets to generate wind turbine gearbox fault detection information. By building a multi-agent integrated fault detection system, the bottom-level sensor agent collects vibration signals in real time and triggers abnormal events through dynamic threshold detection. The middle-level component agent associates vibration characteristics with component status based on physical models. The top-level system-level agent integrates multi-source information to generate global diagnostic conclusions. Complex fault analysis tasks are decomposed into independently run sub-modules, realizing distributed collaborative processing of fault diagnosis tasks, significantly reducing false alarm rates and missed alarm rates, and achieving the technical effect of improving fault detection accuracy.

[0011] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, which can be implemented in accordance with the contents of the description, and to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are specifically listed below. It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present application, nor is it intended to limit the scope of the present application. Other features of the present application will become easy to understand through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in this application or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely exemplary, and a person of ordinary skill in the art can obtain other drawings based on the provided drawings without creative work.

[0013] Figure 1 This is a flow chart of the dynamic integrated fault detection method for wind turbine gearboxes based on feature fusion in this application.

[0014] Figure 2 This is a structural diagram of the wind turbine gearbox dynamic integrated fault detection system based on feature fusion in this application.

[0015] Description of the accompanying drawings: feature extraction module 11, feature primary screening module 12, feature fine screening module 13, fault detection module 14. DETAILED DESCRIPTION

[0016] This application solves the technical problem of low fault detection accuracy in the prior art due to the lack of in-depth feature mining and reliance on a single-machine diagnostic system by providing a dynamic integrated fault detection method and system for wind turbine gearboxes based on feature fusion. By constructing a multi-agent integrated fault detection system, the bottom-level sensor agent collects vibration signals in real time and triggers abnormal events through dynamic threshold detection. The middle-level component agent associates vibration characteristics with component status based on physical models. The top-level system-level agent fuses multi-source information to generate global diagnostic conclusions, decomposing complex fault analysis tasks into independently run sub-modules. This realizes distributed collaborative processing of fault diagnosis tasks, significantly reduces false alarm rates and missed alarm rates, and achieves the technical effect of improving fault detection accuracy.

[0017] Below, the technical solutions in this application will be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of this application, rather than all the embodiments of this application. It should be understood that this application is not limited to the example embodiments described herein. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. It should also be noted that, for the convenience of description, only the parts related to this application, rather than all of them, are shown in the accompanying drawings.

[0018] For example, see the attached Figure 1 The present application provides a method for dynamic integrated fault detection of a wind turbine gearbox based on feature fusion, wherein the method is executed by a dynamic integrated fault detection system for a wind turbine gearbox based on feature fusion, and the method specifically comprises the following steps:

[0019] Vibration datasets from various parts of the wind turbine gearbox are collected and subjected to noise separation processing. Multidimensional deep feature extraction is then performed on the processed data using a multidimensional deep feature extraction engine to obtain various preliminary feature sets. The multidimensional deep feature extraction engine includes a time-frequency domain feature enhancement layer, a deep feature mining layer, and a dynamic optimization mechanism.

[0020] Specifically, during fault detection of wind turbine gearboxes, high-precision sensors are first required to collect vibration signals from key equipment components (such as the gear meshing area, bearing seat, and housing) to form a multi-dimensional time-series dataset, i.e., a vibration dataset for each component. Due to the complex operating environment of wind turbines, vibration signals are often accompanied by strong background noise (such as motor electromagnetic noise and aerodynamic noise). Therefore, noise separation processing is required on the original signal. This process utilizes adaptive variational mode decomposition (VMD) technology, dynamically adjusting the decomposition strategy based on the real-time signal-to-noise ratio. When the signal-to-noise ratio is above 10dB, an 80-2000Hz bandpass filter is used for rapid noise reduction. When the signal-to-noise ratio is below 10dB, a combined VMD and wavelet transform noise reduction mechanism is activated. The fast Kurtogram algorithm automatically locates the optimal resonant frequency band, increasing the time-domain energy concentration of the fault signal by 42%, providing a high-quality data foundation for subsequent feature extraction.

[0021] After noise suppression is completed, the multi-dimensional deep feature extraction stage is entered, and each preliminary feature set is obtained based on the extraction results. The time-frequency domain feature enhancement layer fuses HHT marginal spectral entropy and adaptive wavelet packet energy entropy: the former obtains the intrinsic mode function (IMF) through empirical mode decomposition (EMD), and calculates the marginal spectral entropy of each order IMF to quantify the impact energy distribution characteristics. It is particularly suitable for capturing transient impact faults such as gear pitting and bearing spalling; the latter suppresses redundant frequency band energy while retaining the fault characteristic frequency band through adaptive wavelet packet decomposition. For example, in a wind farm measurement, this method successfully separated the 1.2kHz narrowband impact component generated by early micro-pitting of gears. In order to further improve the feature characterization capability, a fast Kurtogram algorithm is introduced to automatically locate the optimal resonant frequency band, and the resolution of the fault-sensitive frequency band is increased to 0.5Hz, which significantly enhances the ability to identify early weak faults.

[0022] The deep feature mining layer utilizes a two-channel, one-dimensional convolutional neural network (CNN) architecture. The primary channel processes the raw vibration waveform, extracting time-domain transient features (such as impulse pulse width and the ratio of adjacent cycle amplitudes) using multi-scale convolution kernels. The secondary channel focuses on differential signal gradient features to capture frequency modulation phenomena (such as gear mesh frequency sidebands). Features from both channels are weighted and integrated using a feature fusion attention mechanism. A dynamic optimization mechanism consists of two parts: first, real-time switching of wavelet basis functions based on the signal kurtosis factor. For impulse-type faults, the Db4 wavelet is used to highlight transient features, while for broadband vibrations, the Sym8 wavelet is used to balance time-frequency resolution. Second, feature dimensionality is compressed using the mutual information criterion, eliminating redundant features while maintaining the core fault feature dimensionality ≤100. This improves model computational efficiency by over 50%, laying a solid foundation for subsequent fault classification.

[0023] The maximum information coefficient preliminary screening mechanism is called to quantify the nonlinear relationships between features and features, and between features and faults of the preliminary feature sets, and the preliminary screening feature sets are obtained by screening.

[0024] Specifically, the maximum information coefficient (MIC) initial screening mechanism achieves efficient feature screening by quantifying the strength of nonlinear correlations between features. It first calculates the maximum information coefficient (MIC) between any two features, capturing both linear and nonlinear dependencies between variables through the principle of maximizing mutual information. This coefficient ranges from 0 to 1, with higher values ​​indicating greater redundancy between features. Feature groups with redundancy exceeding a preset threshold are then removed, effectively retaining only one of the features. For example, in measured data from a particular wind farm, the original feature set included "HHT marginal spectral entropy" and "wavelet packet energy entropy (1-3kHz frequency band)," both of which had a maximum MI of 0.72. The energy entropy feature, with its more specific physical meaning, was automatically retained. The MIC of each feature with the fault label was then screened, eliminating features with discriminability below a threshold (typically set at 0.25). For example, a low-frequency vibration amplitude feature was removed because its MI was only 0.18, making it ineffective in distinguishing fault types. This resulted in the creation of the initial screening feature sets.

[0025] Based on the SHAP value, the contribution of each feature in each preliminary screening feature set is analyzed, a feature importance map is constructed, and feature fine screening is performed to generate each fine screening feature set.

[0026] Specifically, the feature fine-screening mechanism based on SHAP values ​​quantifies the contribution of each feature to model detection, constructs a global feature importance map, and further selects the core feature set that is highly relevant to the fault. This process is divided into three key steps: First, the marginal contribution of each feature to the detection result of a single sample is calculated using the existing SHAP algorithm. The feature weights are dynamically adjusted based on the fault type. For sudden faults (such as gear tooth breakage), the weights of high-frequency impact features (such as kurtosis and pulse factor) are increased; for gradual faults (such as bearing wear), the weights of frequency domain features (such as envelope harmonic ratio and energy entropy) are enhanced. For example, in the scenario of slight wear on the inner ring of the bearing, the weight of the "envelope harmonic ratio" is automatically increased from 0.3 to 0.5, which increases the SHAP contribution of this feature by 32%. Finally, a mapping relationship between physical mechanisms and data features is established. If the SHAP value of a certain feature is high but lacks clear physical meaning (such as a high-frequency noise component), the multimodal verification mechanism is activated to cross-verify its fault correlation through vibration signal envelope demodulation and oil spectral analysis. For example, in a certain experiment, it was found that the SHAP value of a certain high-frequency feature was abnormally high, but there was no clear correlation with gear failure. It was later verified by acoustic emission signals and confirmed to be sensor noise, and finally it was eliminated from the fine screening feature set.

[0027] Feature importance maps constructed using SHAP values ​​can also enable interpretable analysis of fault modes. For example, a SHAP heat map of a bearing failure revealed that "high-frequency impact energy" and "1st harmonic amplitude attenuation rate" were the primary causal features, fully consistent with the modulation effect of impact excitation caused by pitting corrosion in the gearbox and bearing raceway damage, providing operators with a clear basis for fault tracing.

[0028] A multi-agent integrated fault detection system is called to perform bottom-up multi-layer agent information sharing and collaborative decision-making based on the respective fine-screened feature sets to generate wind turbine gearbox fault detection information.

[0029] Specifically, the multi-agent integrated fault detection system simulates the collaborative mechanisms of human society, breaking down complex fault diagnosis tasks into independently executable subtask units. The underlying system consists of sensor agents distributed across key locations in the wind turbine gearbox. For example, a vibration sensor agent continuously monitors the vibration acceleration signals in the gear meshing area, while a temperature sensor agent simultaneously collects bearing seat temperature data. When a vibration sensor detects an abnormal impact signal, it triggers adjacent auxiliary sensors (such as acoustic emission sensors) to perform multimodal data cross-validation. This triggering mechanism is implemented using pre-set fault propagation rules, such as automatically activating high-frequency acoustic signal acquisition when the vibration amplitude exceeds a threshold.

[0030] The component agent in the middle layer is responsible for logical reasoning about faults in specific mechanical components. Each component agent establishes a bidirectional data channel with its corresponding sensor agent. For example, the gear agent not only receives vibration signals but also integrates lubricant particle count data. Using a built-in gear wear dynamics model, it dynamically correlates the energy distribution in the vibration spectrum with changes in tooth contact stress. When it detects a 30% increase in the amplitude of the gear mesh frequency harmonics, the agent automatically calculates the probability of tooth damage and transmits this warning information to the upper-level system-level agent. The system-level agent, serving as the decision-making hub, utilizes a federated learning framework to integrate local detection results from component agents such as gears, bearings, and drive shafts, assigning weights to each component's contribution to the system-level fault, ultimately generating a global fault diagnosis conclusion.

[0031] Furthermore, the present application also includes the following steps: the multi-agent integrated fault detection system includes a first-layer autonomous agent, a second-layer autonomous agent and a third-layer autonomous agent, wherein the first-layer autonomous agent is a sensor agent, the second-layer autonomous agent is a component agent, and the third-layer autonomous agent is a system-level agent.

[0032] Specifically, the multi-agent integrated fault detection system uses a layered architecture to decompose complex fault diagnosis tasks into independently operable and collaborative subtask units. The first layer of autonomous agents (sensor agents) are deployed in key areas of the gearbox (such as the gear meshing area, bearing seat, and housing), responsible for collecting and initially analyzing raw data. For example, the vibration sensor agent continuously monitors the vibration acceleration signal of the gear meshing frequency, while the temperature sensor agent synchronously collects bearing seat temperature data. When the vibration sensor detects an abnormal impact signal (such as an acceleration amplitude exceeding a threshold), it triggers adjacent auxiliary sensors (such as acoustic emission sensors) to perform multimodal data cross-validation. This triggering mechanism is implemented through preset fault propagation rules. For example, when the vibration amplitude exceeds a threshold, high-frequency acoustic signal collection is automatically activated to ensure full-dimensional perception of critical events.

[0033] The second-tier autonomous agent (component agent) is built around the mechanical components of the wind turbine gearbox (gears, bearings, lubrication system, etc.). Each component agent establishes a bidirectional data channel with its corresponding sensor agent. For example, the gear agent not only receives vibration signals but also integrates lubricant particle count data. Using a built-in gear wear dynamics model, it dynamically correlates the energy distribution in the vibration spectrum with changes in tooth contact stress. When the amplitude of the gear mesh frequency harmonics increases by 30%, the agent automatically calculates the probability of tooth damage and transmits a warning to the upper-level system-level agent. The fault propagation mechanism embedded within the component agent also considers the physical coupling between components. For example, gear wear can lead to abnormal bearing load distribution, which in turn can cause a 30-50Hz shift in bearing vibration frequency. This causal logic is encoded as executable inference rules.

[0034] The third-tier autonomous agent (system-level agent) serves as the decision-making hub. It uses a federated learning framework to integrate local detection results from component agents, such as gears, bearings, and drive shafts. By weighting the contribution of each component failure to the system-level failure, it ultimately generates a global fault diagnosis conclusion. For example, if the gear agent reports a 75% probability of tooth pitting and the bearing agent reports a 68% probability of inner race damage, the system-level agent eliminates uncertainty through a conflicting evidence synthesis algorithm. By combining historical operating data (such as recent maintenance records and environmental parameters), it ultimately outputs a composite fault diagnosis with an 89% probability of failure.

[0035] Furthermore, the present application also includes the following steps: performing fault sensing analysis on each component of the wind turbine gearbox based on vibration anomalies, determining the sensor agent including vibration sensing and auxiliary sensing, and generating the first-layer autonomous agent; constructing each component agent with the each component, wherein the each component agent is connected to the vibration sensing agent and auxiliary sensing agent of the corresponding component in the first-layer autonomous agent, and configuring a component-level fault transmission mechanism to generate the second-layer autonomous agent; connecting the each component agent to the system-level agent, and constructing a global fault analysis model to generate the third-layer autonomous agent; generating the multi-agent integrated fault detection system with the first-layer autonomous agent, the second-layer autonomous agent, and the third-layer autonomous agent.

[0036] Specifically, the construction of three layers of autonomous agents follows a step-by-step abstraction logic, from bottom-level data collection to top-level decision-making. The first-layer autonomous agent (sensor agent) focuses on vibration anomalies and forms a perception network by deploying vibration sensors and auxiliary sensors (such as temperature and pressure sensors) at key locations in the gearbox. Each sensor agent possesses independent signal preprocessing capabilities. For example, the vibration sensor agent collects real-time vibration acceleration signals from the gear meshing area and filters out low-frequency noise using wavelet transform. When a sudden change in impact energy is detected, it triggers a nearby acoustic emission sensor to synchronously collect high-frequency signals. The two types of data are timestamped to form a multimodal feature set. This vibration anomaly-focused agent design can detect weak signals of faults such as gear pitting and bearing spalling at an early stage. Fault sensing analysis, based on historical fault data and starting with vibration anomalies, analyzes abnormal sensing trends. For example, if the vibration anomaly reaches a preset temperature threshold or if there is an abnormal sound, temperature and sound sensors are used as auxiliary sensors to generate the first-layer autonomous agent.

[0037] The second-tier autonomous agent (component agent) uses the mechanical components of the wind turbine gearbox as a modeling unit. Each component agent establishes a bidirectional data channel with its corresponding sensor agent. For example, the gear agent integrates vibration signals, lubricant particle counts, and oil temperature data. A built-in gear wear dynamics model dynamically correlates the energy distribution in the vibration spectrum with changes in tooth contact stress. When a 30% increase in the amplitude of the gear mesh frequency harmonics is detected, the agent uses a fault propagation rule library to infer possible associated faults. For example, gear wear may lead to abnormal bearing load distribution, triggering vibration frequency offset analysis in the bearing agent. The fault propagation mechanism between component agents utilizes Bayesian network coding, translating physical mechanisms (such as the correlation coefficient between gear wear rate and bearing load) into computable confidence propagation rules that can simulate real-world fault propagation paths.

[0038] The third-tier autonomous agent (system-level agent) serves as the decision-making hub, integrating the local detection results of each component agent through a federated learning framework. This agent constructs a global fault analysis model allocation mechanism that quantifies the contribution of component faults, such as gears and bearings, to system-level faults. The federated learning framework allows multiple autonomous agents to perform local computation and learning, then aggregates these local results at the decision-making level. This eliminates the need to centralize all data for processing, effectively reducing the burden of data transmission and processing and improving system efficiency. The federated learning framework coordinates distributed data and model updates, enabling each component agent to operate independently and conduct local learning. Each component agent performs local fault detection based on sensor data from its component (e.g., gears, bearings, etc.). For example, if a gear agent detects abnormal vibration in the gear meshing area, it leverages established physical models (e.g., gear wear dynamics models) to infer whether tooth surface damage or fatigue is present. A bearing agent, on the other hand, combines vibration signals, temperature data, and oil analysis data to analyze the bearing's condition using a bearing wear model. These local detection results are not directly used as the final diagnostic conclusion; instead, they are integrated with fault detection information from other components through the federated learning framework.

[0039] To build a global fault analysis model, the system-level agent first learns from historical fault data and constructs a physics-based fault propagation model through a fault propagation mechanism. Specifically, the system-level agent reads historical fault records for each device. This data contains the fault type, characteristic data when it occurs, and its impact on system performance under different operating conditions. The system-level agent analyzes this historical data to learn how different component failures affect system operation and generates multiple sets of test data. Through multi-level agent collaboration, this test data establishes the fault propagation relationships between different components, thereby building a global fault analysis model. Once the model is established, the system-level agent uses a weighting mechanism to calculate each component's contribution to the system failure based on the fault detection results of each component. Weights are not distributed equally but are based on the importance of each component within the entire system. For example, gears, as critical transmission components, have a direct impact on system stability. Therefore, gear failures are typically weighted higher in system-level fault analysis. In contrast, failures of auxiliary components (such as sensors and lubrication systems) may have a smaller impact on the overall system and therefore receive lower weights.

[0040] When integrating fault detection information from various components, the system-level agent performs cross-validation and conflicting evidence synthesis. If conflicting detection results are found from certain component agents (e.g., a gear agent reports a high probability of failure, while a bearing agent reports a low probability), the system-level agent fuses this information based on historical operating data, operating modes, and fault propagation models to ultimately arrive at a more accurate system-level fault detection result. This result is typically a comprehensive failure probability that accurately reflects the health of the entire wind turbine gearbox.

[0041] To validate the effectiveness of this approach, we illustrate it with typical experimental data. For example, in a fault diagnosis experiment involving a wind turbine gearbox, vibration signals from various fault sources (such as gears, bearings, and drive shafts) were collected. A system-level agent, using a federated learning framework, integrated the local detection results of each component. In this experiment, the gear agent and the bearing agent each provided a probabilistic detection of their respective faults. The system-level agent comprehensively considered the local results of these two components, combined with historical operational data such as load and wind speed fluctuations, and calculated the overall system failure probability using a global fault analysis model. For example, when the gear agent reported a 75% probability of gear damage and the bearing agent reported a 60% probability of bearing damage, the system-level agent, based on the interrelationships between the components and their impact on the system, used a global analysis model to derive a comprehensive system-level failure probability. This result effectively guided maintenance personnel in determining whether equipment downtime and maintenance were necessary.

[0042] Through this integrated fault detection system, the third-layer autonomous agent not only acts as an integrator of fault information, but also realizes the collaborative processing of multi-source information of complex equipment failures through the federated learning framework and the weighted mechanism of the global fault analysis model, providing a more accurate and comprehensive solution for equipment fault diagnosis.

[0043] Furthermore, the present application also includes the following steps: constructing each vibration sensing agent with the vibration sensors of each component of the wind turbine gearbox; performing fault sensing trend analysis on each component with vibration abnormality as the starting point, determining the auxiliary sensors of each component, and constructing each auxiliary sensing agent; connecting each vibration sensing agent and each auxiliary sensing agent with the fault sensing trend, and configuring the first-layer collaborative detection feedback mechanism to generate the first-layer autonomous agent.

[0044] Specifically, the construction of the first-tier autonomous agent (sensor agent) forms the foundation of the entire diagnostic process. This layer deploys high-precision vibration sensors and auxiliary sensors (such as temperature, pressure, and acoustic emission sensors) to form a perception network. Each sensor agent has independent signal preprocessing capabilities. For example, the vibration sensor agent collects real-time vibration acceleration signals from the gear meshing area and filters out low-frequency noise using wavelet transform. When a sudden change in impact energy is detected, it triggers a nearby acoustic emission sensor to synchronously collect high-frequency signals. These two types of data are timestamped and aligned to form a multimodal feature set. This vibration-focused agent design can detect subtle fault signals such as gear pitting and bearing spalling at an early stage. The configuration of auxiliary sensing agents is based on fault sensing trend analysis. By performing time-series cluster analysis on historical fault data, the agent identifies patterns of association between vibration anomalies and multi-physics field parameters. For example, gear pitting failures are often accompanied by a sudden increase in the energy of the high-frequency vibration component and an increase in bearing temperature. Bearing wear is manifested as an increase in the energy proportion of the low-frequency vibration envelope spectrum and a decrease in lubricant viscosity. Based on these association patterns, auxiliary sensors are deployed, such as temperature sensors to monitor bearing seat oil temperature, acoustic emission sensors to capture high-frequency impact signals, and oil sensors to analyze metal particle concentration. Sensor selection must balance sensitivity and durability. For example, in high-speed gearbox operation scenarios, IEPE piezoelectric accelerometers are used, with a frequency response range exceeding 20kHz.

[0045] The first-tier collaborative detection and feedback mechanism utilizes a fault propagation rule base to achieve multimodal data fusion. This mechanism establishes a fault propagation model based on physical mechanisms. For example, gear wear reduces mesh stiffness, leading to dynamic bearing load fluctuations, while bearing pitting excites the rotor system's natural frequency, causing gear mesh frequency modulation. The fusion strategy utilizes a weighted fusion algorithm, with weights set to 0.6 for the vibration signal, 0.3 for the temperature signal, and 0.1 for the acoustic emission signal. Data fusion is achieved through Kalman filtering. A dynamic weight adjustment module optimizes parameters in real time based on operating conditions, for example, increasing the weight of the vibration signal in sudden wind speed changes. The collaborative feedback loop incorporates triggered data acquisition and an online confidence assessment mechanism. When the vibration sensor detects an impact energy exceeding a threshold, high-frequency sampling of the acoustic emission sensor is initiated. If the vibration and temperature sensor data conflict, a data quality verification process is triggered. This maintains high sensitivity even under complex operating conditions. For example, in sudden wind speed changes (±20 m / s), the collaborative mechanism between the vibration and auxiliary sensors keeps the fault miss rate below 0.8%.

[0046] Furthermore, the present application also includes the following steps: inputting the various fine screening feature sets into the various vibration sensing agents of the first-layer autonomous agent, triggering the abnormality verification of the auxiliary sensor through vibration abnormality discrimination, and generating first-layer agent fault detection information; sharing the first-layer agent fault detection information to the second-layer autonomous agent, using the component-level fault transmission mechanism to perform component-level fault detection, and generating second-layer agent fault detection information; sharing the first-layer agent fault detection information and the second-layer agent fault detection information to the third-layer autonomous agent, using the global fault analysis model to integrate the first-layer agent fault detection information and the second-layer agent fault detection information to obtain the wind turbine gearbox fault detection information.

[0047] Specifically, feature information is progressively transferred and integrated from bottom-level perception to top-level decision-making through a layered agent architecture. First, refined feature sets are fed into the vibration sensing agents of the first-tier autonomous agent layer. These agents are deployed in key locations within the gearbox (such as the gear mesh area and bearing housing). They receive vibration signals in real time and perform preliminary analysis. Upon detecting unusual patterns in the vibration signature (such as sudden changes in high-frequency impact energy or an abnormal increase in the energy proportion of a specific frequency band), they trigger auxiliary sensors (such as acoustic emission sensors and temperature sensors) at corresponding locations to perform multimodal data verification. For example, if the vibration signal indicates an abnormally high amplitude of the gear mesh frequency harmonics, an acoustic emission sensor is activated to capture the high-frequency impact signal. The two data types are aligned using timestamps to generate first-tier agent fault detection information containing multi-dimensional features such as vibration amplitude, impact energy density, and temperature change rate. This detection information primarily reflects abnormal local sensor responses. For example, if the impact intensity in a particular gear mesh area exceeds a threshold, this may indicate an early pitting failure.

[0048] The fault detection information generated by the first-tier agent is then shared with the second-tier autonomous agent (component agent) via the edge computing gateway. Each component agent corresponds to a specific mechanical component of the gearbox (such as gears, bearings, and drive shafts) and embeds a physics-based fault transmission model. For example, after receiving an abnormal vibration signal, the gear agent combines lubricant particle count data and the tooth contact stress model to analyze whether gear wear has caused abnormal bearing load distribution. If the bearing agent simultaneously monitors vibration frequency deviation and temperature rise, the component-level fault transmission mechanism calculates the correlation probability between the two types of faults using a Bayesian network, generating second-tier agent fault detection information that includes component health status scores. For example, it might infer a 65% probability of gear wear and a 58% probability of bearing fatigue, and generate a comprehensive score using confidence weighting.

[0049] Finally, the detection information from the first and second layers is uploaded to the system-level agent at the third layer. This agent uses a federated learning framework to integrate multi-source heterogeneous data. Through cross-validation and fusion of multi-level information, it ultimately generates wind gearbox fault detection information including fault type, occurrence probability, and impact range.

[0050] The wind turbine gearbox fault detection information output by the third-tier autonomous agent (system-level agent) is fed back to the second-tier autonomous agent (component-level agent) for further component-level fault analysis and optimization. This process is essentially an iterative optimization process, aiming to guide and adjust the fault detection accuracy of each component agent using the global diagnostic results generated by the system-level agent, thereby improving the accuracy and flexibility of overall fault detection. The fault detection information output by the system-level agent not only reflects the status of individual components but also reveals interactions between components and possible fault propagation paths. In this regard, the feedback provided by the system-level agent is not only the final decision output but also contains optimization suggestions and adjustment directions. After receiving the feedback from the system-level agent, the component-level agent first analyzes the fault detection information generated by the system-level agent. If the system-level agent's diagnostic conclusion is that a certain component (such as a gear) has a high probability of failure, the component-level agent will use this information to optimize the fault detection of the current component. For example, if the system-level agent believes that the probability of gear failure is 75%, while the probability of bearing failure is lower, this information will prompt the gear agent to re-examine the analysis results of its vibration signal, recalculate the risk of tooth surface wear, and adjust the gear fault detection based on historical operating data and physical models, thereby improving the detection ability of each component agent for its own component faults.

[0051] Furthermore, the present application also includes the following steps: reading the model and specification information of the wind turbine gearbox, retrieving pre-processed historical fault record data; executing the test of the first-layer autonomous agent and the second-layer autonomous agent with the historical fault record data to generate multiple sets of test data; executing the cross- and co-occurrence features of the first-layer autonomous agent and the second-layer autonomous agent under the same fault source with the multiple sets of test data, and training the global fault analysis model with the cross- and co-occurrence features.

[0052] Specifically, the training process of the global fault analysis model deeply integrates historical fault data with real-time sensor information, building a highly generalizable fault diagnosis model through collaborative verification using multi-level agents. First, the model and specifications of the wind turbine gearbox are read, and matching historical fault records are screened from a preprocessed database. These records cover typical fault scenarios such as gear pitting, bearing wear, and poor lubrication. After standardization and time alignment, this historical data is transformed into a multidimensional fault sample set, providing a benchmark for subsequent testing.

[0053] Next, test data is generated for the first and second-tier agents. The vibration sensing agent and auxiliary sensing agent simulate the signal acquisition process under real-world operating conditions. For example, for a "gear pitting" fault sample for a certain wind turbine model, the vibration sensing agent outputs a high-frequency impact energy spectrum, the acoustic emission agent outputs an impact pulse count, and the temperature agent outputs a temperature rise curve. By comparing historical fault labels, the first-tier test data, including the sensor's raw characteristics and preliminary failure probabilities, is generated. The component agents utilize built-in physical models to analyze fault propagation. For example, the gear agent calculates the gear pitting area percentage based on vibration spectrum distortion and lubricant particle concentration, while the bearing agent detects the degree of bearing rolling element damage by combining vibration frequency offset and temperature gradient. This generates the second-tier test data, which correlates component-level failure probabilities with physical parameters. When mining cross- and co-occurrence features for the same fault source, the team extracted the temporal correlation between vibration impact energy and acoustic emission pulse count as cross-features for gear pitting faults, analyzed the probability of concurrent gear mesh frequency modulation and bearing outer race faults as co-occurrence features, and screened feature combinations with high contributions to fault classification through SHAP value analysis. For example, the combined feature of "vibration impact energy × oil temperature gradient" in a certain fan model improved the discrimination of early-stage micropitting faults by 45%. These features were used to train a global fault analysis model, which uses a federated learning framework to fuse multi-source heterogeneous data and multi-level agent detection results. Comparative learning enhances the ability to distinguish similar fault modes, and an online learning mechanism is introduced to regularly update model parameters with new data, thereby improving fault detection accuracy.

[0054] Furthermore, the present application also includes the following steps: calculating the maximum information coefficient for every two features in each preliminary feature set to generate several nonlinear relationships between features; calculating the maximum information coefficient for each feature and fault label in each preliminary feature set to generate various feature-fault nonlinear relationships; eliminating redundant features based on the several nonlinear relationships between features, and then performing feature screening based on the ability of each feature in each feature-fault nonlinear relationship to distinguish fault types to generate the various preliminary screening feature sets.

[0055] Specifically, the maximum information coefficient (MIC) initial screening mechanism achieves efficient feature selection by quantifying the strength of nonlinear correlations between features. It first calculates the maximum information coefficient (MIC) between any two features, capturing linear and nonlinear dependencies between variables using the principle of maximizing mutual information. The value ranges from 0 to 1, with higher values ​​indicating greater redundancy between features. For example, in measured data from a wind farm, the original feature set included "HHT marginal spectral entropy" and "wavelet packet energy entropy (1-3kHz frequency band)." Both had a maximum MI of 0.72, so only one feature was automatically retained. Each feature was then screened based on its maximum MIC against the fault label, eliminating features with discrimination below a threshold (typically set at 0.25). For example, a low-frequency vibration amplitude feature with a maximum MI of only 0.18 was rejected because it could not effectively distinguish fault types. To enhance the ability to identify subtle faults, an improved ReliefF algorithm was introduced, assigning higher weights to minority class samples (for example, doubling the weight when the proportion of faulty samples is less than 5%). This enhances feature selection for subtle faults such as micropitting on bearing inner rings. To adapt to dynamic changes in working conditions, an adaptive threshold update mechanism is also set up to periodically adjust the maximum information coefficient threshold based on the distribution of historical data, and set a 7-day half-life for time domain features (such as kurtosis and pulse factor). After the expiration, the correlation strength needs to be recalculated to ensure that the feature screening results always conform to the current equipment status.

[0056] During feature selection, a feature dimension threshold can be introduced. This provides an effective constraint mechanism, enabling feature selection to be optimized not only based on information relevance but also based on the model's computational complexity and performance requirements. Setting a feature dimension threshold ensures that the number of features ultimately retained remains within a reasonable range, thus avoiding overfitting and excessive model complexity. For example, relying solely on the information coefficient for feature selection may result in the retention of some high-dimensional, low-information features, which can affect model performance. Therefore, setting a feature dimension threshold helps ensure a concise and effective feature set. For example, limiting the number of selected features to a preset range (e.g., 100 features) and dynamically adjusting this threshold based on the actual model requirements can control feature selection results and prevent excessive redundant features from burdening model performance. For example, when the number of features exceeds the threshold, the feature set can be ranked based on the information coefficient, automatically eliminating features that are highly redundant with other features. The remaining features with high information gain are ultimately selected to ensure accurate fault diagnosis. The feature dimension threshold can be obtained by professional and technical personnel in this field based on experience or through actual testing. For example, the minimum dimension when the fault detection accuracy meets the preset threshold is used as the feature dimension threshold.

[0057] Furthermore, the present application also includes the following steps: the time-frequency domain feature enhancement layer performs time-frequency domain feature enhancement by introducing a composite information entropy mechanism of HHT marginal spectral entropy and adaptive wavelet packet energy entropy; the deep feature mining layer is a parallel architecture based on a dual-channel one-dimensional convolutional neural network, and the dual-channel one-dimensional convolutional neural network includes an original waveform convolution channel, a differential signal convolution channel and a feature fusion attention mechanism; the dynamic optimization mechanism includes an adaptive wavelet basis function switching strategy based on the kurtosis factor and a feature dimension compression strategy based on the mutual information criterion.

[0058] Specifically, the time-frequency domain feature enhancement layer constructs a multi-dimensional time-frequency feature enhancement mechanism by fusing the Hilbert-Huang transform (HHT) marginal spectral entropy with the adaptive wavelet packet energy entropy. The HHT marginal spectral entropy decomposes the vibration signal into multiple intrinsic mode functions (IMFs) through empirical mode decomposition (EMD). The marginal spectral entropy of each order of IMFs is calculated to quantify the distribution characteristics of the impact energy in the time domain. This is particularly suitable for capturing the energy mutation characteristics of transient impact faults such as gear pitting and bearing spalling. For example, when early micropitting occurs in gears, the marginal spectral entropy value of the high-frequency impact component will increase significantly. The optimal resonant frequency band is automatically located through a fast Kurtogram algorithm, and the resolution of the fault-sensitive frequency band is increased to 0.5Hz. Even in a strong noise environment with a signal-to-noise ratio below 5dB, a 42% improvement in the fault feature signal-to-noise ratio can be maintained. At the same time, the adaptive wavelet packet energy entropy dynamically adjusts the number of decomposition layers (impact signals are decomposed into 4 layers, and broadband vibrations are decomposed into 6 layers) and adaptively selects the wavelet basis function (Db4 wavelet is used for impact signals, and Sym8 wavelet is used for broadband signals) based on the signal kurtosis value, thereby suppressing the energy of redundant frequency bands while highlighting the fault-sensitive frequency bands.

[0059] The deep feature mining layer utilizes a dual-channel, one-dimensional convolutional neural network (CNN) architecture to process the raw vibration waveform and differential signal gradient features, respectively. The raw waveform convolution channel uses multi-scale convolution kernels (3×1, 5×1, and 7×1 windows) to extract time-domain transient features, such as impulse pulse width, amplitude ratio between adjacent cycles, and waveform symmetry. This is particularly suitable for capturing periodic impact signals caused by bearing inner race damage. The differential signal convolution channel focuses on the first-order difference of the signal and uses gradient-enhanced convolution kernels (such as the Sobel operator) to extract frequency modulation features, such as the sideband modulation phenomenon of gear mesh frequency. The features from both channels are weighted and integrated using a feature fusion attention mechanism, with traditional time-frequency features receiving 30% weight and deep learning features receiving 70%. This ratio is dynamically optimized using SHAP value analysis.

[0060] The dynamic optimization mechanism consists of two parts: first, real-time switching of wavelet basis functions based on the signal kurtosis factor. When the kurtosis value exceeds 3.5, the system automatically activates the Db4 wavelet, which is sensitive to transient shocks, and switches to the Sym8 wavelet in broadband vibration scenarios to balance time-frequency resolution. Second, feature dimensions are compressed using the mutual information criterion, eliminating redundant features while keeping the core fault feature dimension ≤100.

[0061] Furthermore, the present application also includes the following steps: for each of the preliminary feature sets, evaluating the overall contribution of the interaction between features to the ability to distinguish fault types based on the combined mutual information index; performing a joint redundant feature elimination analysis based on the overall contribution, generating a joint redundant feature elimination strategy, and optimizing feature screening.

[0062] Specifically, feature screening should not only focus on the discriminative ability of a single feature, but also evaluate the overall contribution of the interaction between features to fault classification. To this end, the combined mutual information indicator is introduced to optimize the feature screening strategy by quantifying the dependency between the joint distribution of multiple features and the fault label. Specifically, the combined mutual information measures the amount of shared information between multiple features and the fault label through the joint probability distribution. It can not only capture the synergistic effect between features (such as the superimposed effect of gear wear and bearing fatigue), but also identify redundant effects (such as repeated information of vibration amplitude and acceleration). For each set of feature combinations (such as "vibration kurtosis + temperature gradient"), its combined mutual information value is calculated and compared with the mutual information value of a single feature. If the combined mutual information is significantly higher than the sum of the mutual information of the individual features, it indicates that there is a synergistic enhancement effect between the features; if it is close to or lower than the sum of the mutual information of the individual features, there may be redundancy. Redundancy determination rules based on combined mutual information are divided into two categories: Synergistically enhanced feature combinations (for example, in gear pitting faults, the combined mutual information value of "vibration high-frequency energy + oil metal particle concentration" increases by 40% compared to individual features) are retained, while redundantly suppressed features (for example, a low-frequency vibration feature combination with a contribution of only 2%) are eliminated. By combining combined mutual information and a joint redundancy elimination strategy, efficient compression and dynamic optimization of feature sets are achieved while maintaining fault differentiation capabilities, providing key technical support for intelligent diagnosis of wind turbine gearboxes.

[0063] In summary, the wind turbine gearbox dynamic integrated fault detection method based on feature fusion provided by this application has the following beneficial effects:

[0064] Vibration data sets of various parts of the wind turbine gearbox are collected and noise separation processing is performed. Multi-dimensional deep feature extraction is performed on the processed data through a multi-dimensional deep feature extraction engine to obtain various preliminary feature sets, wherein the multi-dimensional deep feature extraction engine includes a time-frequency domain feature enhancement layer, a deep feature mining layer and a dynamic optimization mechanism; the maximum information coefficient preliminary screening mechanism is called to quantify the feature-feature and feature-fault nonlinear relationships of the various preliminary feature sets, and the various preliminary screening feature sets are screened; the contribution of each feature in each preliminary screening feature set is analyzed based on the SHAP value, a feature importance map is constructed, and feature fine screening is performed to generate various fine screening feature sets; the multi-agent integrated fault detection system is called to perform bottom-up multi-layer agent information sharing and collaborative decision-making based on the various fine screening feature sets to generate wind turbine gearbox fault detection information. By building a multi-agent integrated fault detection system, the bottom-level sensor agent collects vibration signals in real time and triggers abnormal events through dynamic threshold detection. The middle-level component agent associates vibration characteristics with component status based on physical models. The top-level system-level agent integrates multi-source information to generate global diagnostic conclusions. Complex fault analysis tasks are decomposed into independently run sub-modules, realizing distributed collaborative processing of fault diagnosis tasks, significantly reducing false alarm rates and missed alarm rates, and achieving the technical effect of improving fault detection accuracy.

[0065] In the second embodiment, based on the same inventive concept as the wind turbine gearbox dynamic integrated fault detection method based on feature fusion in the aforementioned first embodiment, the present application also provides a wind turbine gearbox dynamic integrated fault detection system based on feature fusion, please refer to the attached Figure 2 , the wind turbine gearbox dynamic integrated fault detection system based on feature fusion includes:

[0066] The feature extraction module 11 is used to collect vibration data sets of various parts of the wind turbine gearbox and perform noise separation processing, and perform multi-dimensional deep feature extraction on the processed data through a multi-dimensional deep feature extraction engine to obtain various preliminary feature sets, wherein the multi-dimensional deep feature extraction engine includes a time-frequency domain feature enhancement layer, a deep feature mining layer and a dynamic optimization mechanism.

[0067] The feature initial screening module 12 is used to call the maximum information coefficient initial screening mechanism to quantify the nonlinear relationship between features and features, and between features and faults of the various preliminary feature sets, and screen to obtain various preliminary screening feature sets.

[0068] The feature fine screening module 13 is used to analyze the contribution of each feature in each preliminary screening feature set based on the SHAP value, construct a feature importance map, perform feature fine screening, and generate each fine screening feature set.

[0069] The fault detection module 14 is used to call the multi-agent integrated fault detection system, perform bottom-up multi-layer agent information sharing and collaborative decision-making based on the various fine-screening feature sets, and generate wind turbine gearbox fault detection information.

[0070] Furthermore, the fault detection module 14 in the dynamic integrated fault detection system for wind turbine gearboxes based on feature fusion is also used for: the multi-agent integrated fault detection system includes a first-layer autonomous agent, a second-layer autonomous agent and a third-layer autonomous agent, wherein the first-layer autonomous agent is a sensor agent, the second-layer autonomous agent is a component agent, and the third-layer autonomous agent is a system-level agent.

[0071] Furthermore, the fault detection module 14 in the dynamic integrated fault detection system for wind gearboxes based on feature fusion is also used to: perform fault sensing analysis on each component of the wind gearbox starting from vibration abnormality, determine the sensor agent including vibration sensing and auxiliary sensing, and generate the first-layer autonomous agent; construct each component agent with the each component, wherein the each component agent is connected to the vibration sensing agent and auxiliary sensing agent of the corresponding component in the first-layer autonomous agent, and configure the component-level fault transmission mechanism to generate the second-layer autonomous agent; connect the each component agent to the system-level agent, and construct a global fault analysis model to generate the third-layer autonomous agent; generate the multi-agent integrated fault detection system with the first-layer autonomous agent, the second-layer autonomous agent and the third-layer autonomous agent.

[0072] Furthermore, the fault detection module 14 in the dynamic integrated fault detection system for wind gearboxes based on feature fusion is also used to: construct various vibration sensing agents using the vibration sensors of various components of the wind gearbox; perform fault sensing trend analysis on the various components based on vibration anomalies, determine the auxiliary sensors of the various components, and construct various auxiliary sensing agents; connect the various vibration sensing agents and the various auxiliary sensing agents using fault sensing trends, and configure a first-layer collaborative detection feedback mechanism to generate the first-layer autonomous agent.

[0073] Furthermore, the fault detection module 14 in the wind turbine gearbox dynamic integrated fault detection system based on feature fusion is further configured to:

[0074] The various fine-screening feature sets are input into the various vibration sensing agents of the first-layer autonomous agent, and the abnormality verification of the auxiliary sensor is triggered by vibration abnormality discrimination to generate first-layer agent fault detection information; the first-layer agent fault detection information is shared with the second-layer autonomous agent, and component-level fault detection is performed using the component-level fault transmission mechanism to generate second-layer agent fault detection information; the first-layer agent fault detection information and the second-layer agent fault detection information are shared with the third-layer autonomous agent, and the first-layer agent fault detection information and the second-layer agent fault detection information are integrated using a global fault analysis model to obtain the wind turbine gearbox fault detection information.

[0075] Furthermore, the fault detection module 14 in the dynamic integrated fault detection system for wind turbine gearboxes based on feature fusion is also used to: read the model and specification information of the wind turbine gearbox, and retrieve the pre-processed historical fault record data; perform tests on the first-layer autonomous agent and the second-layer autonomous agent using the historical fault record data to generate multiple sets of test data; and perform cross- and co-occurrence features of the first-layer autonomous agent and the second-layer autonomous agent under the same fault source using the multiple sets of test data, and train the global fault analysis model using cross- and co-occurrence features.

[0076] Furthermore, the feature preliminary screening module 12 in the dynamic integrated fault detection system for wind turbine gearboxes based on feature fusion is also used to: calculate the maximum information coefficient for every two features in each preliminary feature set to generate several nonlinear relationships between features; calculate the maximum information coefficient for each feature and fault label in each preliminary feature set to generate various feature-fault nonlinear relationships; eliminate redundant features based on the several nonlinear relationships between features, and then perform feature screening based on the ability of each feature in each feature-fault nonlinear relationship to distinguish fault types to generate the various preliminary screening feature sets.

[0077] Furthermore, the feature extraction module 11 in the dynamic integrated fault detection system for wind turbine gearboxes based on feature fusion is also used for: the time-frequency domain feature enhancement layer performs time-frequency domain feature enhancement by introducing a composite information entropy mechanism of HHT marginal spectral entropy and adaptive wavelet packet energy entropy; the deep feature mining layer is a parallel architecture based on a dual-channel one-dimensional convolutional neural network, and the dual-channel one-dimensional convolutional neural network includes an original waveform convolution channel, a differential signal convolution channel and a feature fusion attention mechanism; the dynamic optimization mechanism includes an adaptive wavelet basis function switching strategy based on the kurtosis factor and a feature dimension compression strategy based on the mutual information criterion.

[0078] Furthermore, the dynamic integrated fault detection system for wind turbine gearboxes based on feature fusion also includes a joint redundant feature elimination module, which is used to: evaluate the overall contribution of the interaction between features to the ability to distinguish fault types based on the combined mutual information index for each preliminary feature set; perform a joint redundant feature elimination analysis based on the overall contribution, generate a joint redundant feature elimination strategy, and optimize feature screening.

[0079] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. Figure 1 The dynamic integrated fault detection method for wind gearbox based on feature fusion and the specific examples in Example 1 are also applicable to the dynamic integrated fault detection system for wind gearbox based on feature fusion in this embodiment. Through the above detailed description of the dynamic integrated fault detection method for wind gearbox based on feature fusion, those skilled in the art can clearly understand the dynamic integrated fault detection system for wind gearbox based on feature fusion in this embodiment, so for the sake of brevity of the specification, it will not be described in detail here.

[0080] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

[0081] Obviously, for those skilled in the art, several improvements and modifications can be made to the present application without departing from the principles of the present application, and these improvements and modifications also fall within the scope of protection of the present application.

Claims

1. A dynamic integrated fault detection method for wind turbine gearbox based on feature fusion, characterized in that: include: Vibration data sets of various parts of the wind turbine gearbox are collected and noise separation processing is performed. Multi-dimensional deep feature extraction is performed on the processed data through a multi-dimensional deep feature extraction engine to obtain various preliminary feature sets, wherein the multi-dimensional deep feature extraction engine includes a time-frequency domain feature enhancement layer, a deep feature mining layer and a dynamic optimization mechanism. The time-frequency domain feature enhancement layer enhances the time-frequency domain features by introducing a composite information entropy mechanism of HHT marginal spectrum entropy and adaptive wavelet packet energy entropy; the deep feature mining layer is a parallel architecture based on a dual-channel one-dimensional convolutional neural network, which includes an original waveform convolution channel, a differential signal convolution channel and a feature fusion attention mechanism; the dynamic optimization mechanism includes an adaptive wavelet basis function switching strategy based on the kurtosis factor and a feature dimension compression strategy based on the mutual information criterion; Calling the maximum information coefficient preliminary screening mechanism to quantify the nonlinear relationships between features and features, and between features and faults of each preliminary feature set, and screening to obtain each preliminary screening feature set; Analyze the contribution of each feature in each preliminary screening feature set based on the SHAP value, construct a feature importance map, perform feature fine screening, and generate each fine screening feature set; A multi-agent integrated fault detection system is called to perform bottom-up multi-layer agent information sharing and collaborative decision-making based on the respective fine-screened feature sets to generate wind turbine gearbox fault detection information.

2. The wind turbine gearbox dynamic integrated fault detection method based on feature fusion according to claim 1 is characterized in that: The multi-agent integrated fault detection system includes a first-layer autonomous agent, a second-layer autonomous agent, and a third-layer autonomous agent, wherein the first-layer autonomous agent is a sensor agent, the second-layer autonomous agent is a component agent, and the third-layer autonomous agent is a system-level agent.

3. The wind turbine gearbox dynamic integrated fault detection method based on feature fusion according to claim 2, characterized in that: The steps of constructing the multi-agent integrated fault detection system include: Perform fault sensing analysis on each component of the wind turbine gearbox, starting with vibration anomalies, to determine sensor agents that include vibration sensing and auxiliary sensing, and generate the first-layer autonomous agent; Constructing each component agent with each component, wherein each component agent is connected to the vibration sensing agent and auxiliary sensing agent of the corresponding component in the first-layer autonomous agent, and configuring a component-level fault transmission mechanism to generate the second-layer autonomous agent; Connecting each component agent to the system-level agent and building a global fault analysis model to generate the third-layer autonomous agent; The multi-agent integrated fault detection system is generated using the first-layer autonomous agent, the second-layer autonomous agent, and the third-layer autonomous agent.

4. The method for dynamic integrated fault detection of wind turbine gearboxes based on feature fusion according to claim 3, characterized in that: Perform fault sensing analysis on each component of the wind turbine gearbox, starting with vibration anomalies, to determine sensor agents that include vibration sensing and auxiliary sensing, and generate the first-layer autonomous agent, including: Build individual vibration sensing agents using vibration sensors from various components of the wind turbine gearbox; Performing fault sensing trend analysis on each component starting from vibration anomaly, determining auxiliary sensors for each component, and constructing each auxiliary sensing agent; The vibration sensing agents and the auxiliary sensing agents are connected with each other according to the fault sensing trend, and a first-layer collaborative detection feedback mechanism is configured to generate the first-layer autonomous agent.

5. The method for dynamic integrated fault detection of wind turbine gearboxes based on feature fusion according to claim 4, characterized in that: The multi-agent integrated fault detection system is called to perform bottom-up multi-layer agent information sharing and collaborative decision-making based on the respective fine-screened feature sets to generate wind turbine gearbox fault detection information, including: Inputting each fine screening feature set into each vibration sensing agent of the first layer autonomous agent, triggering abnormality verification of the auxiliary sensor through vibration abnormality discrimination, and generating first layer agent fault detection information; Sharing the first-layer agent fault detection information with the second-layer autonomous agent, performing component-level fault detection using a component-level fault transmission mechanism, and generating second-layer agent fault detection information; The first-layer agent fault detection information and the second-layer agent fault detection information are shared with the third-layer autonomous agent, and the first-layer agent fault detection information and the second-layer agent fault detection information are integrated using a global fault analysis model to obtain the wind turbine gearbox fault detection information.

6. The method for dynamic integrated fault detection of wind turbine gearboxes based on feature fusion according to claim 3, characterized in that: Build a global fault analysis model, including: Read the model and specification information of the wind turbine gearbox and retrieve pre-processed historical fault record data; Execute tests on the first-layer autonomous agent and the second-layer autonomous agent using the historical fault record data to generate multiple sets of test data; The cross and co-occurrence features of the first-layer autonomous agent and the second-layer autonomous agent under the same fault source are executed using the multiple groups of test data, and the global fault analysis model is trained using the cross and co-occurrence features.

7. The method for dynamic integrated fault detection of wind turbine gearboxes based on feature fusion according to claim 1, characterized in that: The maximum information coefficient preliminary screening mechanism is called to quantify the nonlinear relationship between features and features, and features and faults of each preliminary feature set, and each preliminary screening feature set is obtained by screening, including: Calculating the maximum information coefficient for every two features in each preliminary feature set to generate a number of nonlinear relationships between the features; Calculating the maximum information coefficient for each feature and fault label in each preliminary feature set to generate each feature-fault nonlinear relationship; After redundant features are eliminated based on the nonlinear relationships among the plurality of features, feature screening is performed based on the fault type distinguishing capability of each feature according to each feature-fault nonlinear relationship to generate the respective primary screening feature sets.

8. The method for dynamic integrated fault detection of wind turbine gearboxes based on feature fusion according to claim 7, characterized in that: Generating each of the primary screening feature sets further includes: For each of the preliminary feature sets, the overall contribution of the interaction between features to the ability to distinguish fault types is evaluated based on the combined mutual information index; A joint redundant feature elimination analysis is performed based on the overall contribution, a joint redundant feature elimination strategy is generated, and feature screening is optimized.

9. A dynamic integrated fault detection system for wind turbine gearboxes based on feature fusion, characterized in that: The method for implementing the dynamic integrated fault detection method for a wind turbine gearbox based on feature fusion according to any one of claims 1 to 8 is provided, wherein the dynamic integrated fault detection system for a wind turbine gearbox based on feature fusion comprises: A feature extraction module is used to collect vibration data sets from various parts of the wind turbine gearbox and perform noise separation processing. The processed data is subjected to multidimensional deep feature extraction by a multidimensional deep feature extraction engine to obtain various preliminary feature sets, wherein the multidimensional deep feature extraction engine includes a time-frequency domain feature enhancement layer, a deep feature mining layer, and a dynamic optimization mechanism; A feature preliminary screening module is used to call the maximum information coefficient preliminary screening mechanism to quantify the nonlinear relationship between features and features, and between features and faults of each preliminary feature set, and screen each preliminary screening feature set; The feature fine-screening module is used to analyze the contribution of each feature in each preliminary screening feature set based on the SHAP value, construct a feature importance map, perform feature fine-screening, and generate each fine-screening feature set; The fault detection module is used to call the multi-agent integrated fault detection system, perform bottom-up multi-layer agent information sharing and collaborative decision-making based on the various fine-screening feature sets, and generate wind turbine gearbox fault detection information.

Images