Wind turbine generator gearbox state monitoring and fault early warning method and system

Abstract

The invention discloses a wind turbine generator gearbox state monitoring and fault early warning method and system. The method comprises the following steps: defining a key part of a gearbox of a wind turbine generator as a flexible body, and constructing a multi-body dynamic model based on the flexible body; the multi-body dynamics model receives input multi-source time sequence data, then performs feature extraction and multi-index fault early warning algorithm processing, and outputs signal abnormal information for the current multi-source time sequence data; the signal abnormal information is gearbox fault data; constructing a mixed data set guided by a physical mechanism and performing data enhancement; and inputting a spatial-temporal feature coding fault diagnosis model based on the enhanced mixed data set, and then outputting a health index. According to the intelligent early warning system for the wind turbine generator gearbox, through the technical innovation of making up for data missing through modeling simulation, improving diagnosis precision through an intelligent algorithm and guaranteeing real-time early warning through edge calculation, a set of intelligent early warning system for the wind turbine generator gearbox which is high in precision, low in time delay and strong in generalization is successfully constructed.

Description

Technical Field

[0001] This invention relates to the field of wind power equipment condition monitoring technology, and in particular to a method and system for condition monitoring and fault early warning of wind turbine gearboxes. Background Technology

[0002] The green and low-carbon transformation of the energy structure is a global consensus. As a core pillar of clean energy, the large-scale and efficient development of wind power is crucial to achieving the "dual carbon" goal. However, with the continuous growth of wind power installed capacity, the long-term, reliable, and economical operation of wind turbine units faces severe challenges. Wind farms are mostly located in remote and harsh environments (such as offshore, plateaus, and Gobi deserts), resulting in poor accessibility of turbine units and high operation and maintenance costs. Statistics show that the operation and maintenance cost of wind farms accounts for approximately 20%-30% of their total life-cycle cost, with unplanned downtime and replacement / repair due to sudden failures of key components being the main expenditures. Therefore, improving equipment reliability and realizing a smart operation and maintenance model that shifts from "post-failure repair" to "pre-failure early warning" has become an urgent need for the industry.

[0003] Among the many components of a wind turbine, the gearbox, as the core of the transmission system and the power hub, directly determines the overall power generation efficiency and stability of the turbine. The gearbox operates under complex, variable, and highly impactful unsteady conditions: the randomness of wind speed and turbulent characteristics cause drastic fluctuations in its input torque; frequent start-stop, speed changes, and emergency braking subject it to alternating loads and instantaneous impacts far exceeding design specifications. This harsh service environment makes the gearbox one of the components with the highest failure rate in the entire turbine, accounting for more than 30% of all failures. Furthermore, damage to the gearbox not only results in direct economic losses of up to millions of yuan per incident (including power generation losses, high hoisting costs, and spare parts costs), but also poses a significant risk to overall production safety.

[0004] Currently, condition monitoring and fault diagnosis of wind turbine gearboxes mainly rely on offline inspection and online monitoring systems based on vibration analysis. Traditional monitoring methods suffer from technical bottlenecks such as sparse fault samples, poor adaptability to multiple operating conditions, and limitations in cross-model migration. There is an urgent need to overcome technical bottlenecks in simulation-measurement data fusion, dynamic threshold optimization, and lightweight edge deployment to build a high-precision, low-latency intelligent early warning system. Therefore, developing high-precision real-time early warning technology is of great significance for avoiding millions of dollars in downtime losses caused by a single fault and improving the economic benefits of wind farms. At the same time, these traditional methods have revealed several insurmountable technical bottlenecks in engineering practice: serious gearbox faults (such as broken teeth or bearing breakage) are low-probability events, resulting in extremely scarce fault samples that can be used to train intelligent diagnostic models. Data-driven models heavily rely on massive amounts of labeled data; in "zero-sample" or "few-sample" fault scenarios, their diagnostic accuracy and reliability drop sharply, making it difficult to achieve effective early warning. Summary of the Invention

[0005] This invention provides a method for monitoring the condition and providing early warning of faults in the gearbox of a wind turbine, which solves the problems existing in the prior art.

[0006] This invention provides a method for monitoring the condition and providing early warning of faults in a wind turbine gearbox, comprising the following steps:

[0007] Based on the definition of key components of wind turbine gearbox as flexible bodies, a multibody dynamics model is constructed based on the flexible bodies. The multibody dynamics model receives input multi-source time-series data, then performs feature extraction and multi-index fault early warning algorithm processing, and outputs signal anomaly information for the current multi-source time-series data. The signal anomaly information is gearbox fault data.

[0008] A hybrid dataset guided by physical mechanisms is constructed and augmented: a sensor network is deployed at key locations in the gearbox to collect data, constructing a hybrid dataset. This dataset includes measured data from healthy operation, measured data from typical faults, and augmented data generated by a physically constrained deep convolutional generative adversarial network (DGAN). The DGAN is used to augment the data; its generator employs a neural network structure containing m fully connected layers and n deconvolutional layers, while the discriminator consists of a convolutional layer and a fully connected layer. A gradient penalty term is introduced into the adversarial loss function. The augmented data undergoes standardization preprocessing, including sliding window segmentation with a fixed time step and the use of an elliptic filter to eliminate high-frequency interference.

[0009] The enhanced hybrid dataset is then input into the spatiotemporal feature encoding fault diagnosis model, and then the health index is output.

[0010] Preferably, as one feasible implementation; the key components of the wind turbine gearbox are defined as flexible bodies, and a multibody dynamics model is constructed based on the flexible bodies; the multibody dynamics model receives input multi-source time-series data, then performs feature extraction and multi-index fault early warning algorithm processing, and outputs signal anomaly information for the current multi-source time-series data; the signal anomaly information is gearbox fault data, specifically including:

[0011] Based on the structural parameters of the multi-stage planetary gear in the gearbox, its geometric model and parametric assembly are constructed, and key dimension associations are driven by equations.

[0012] The planetary carrier and the housing bearing seat are transformed into flexible bodies, high-order hexahedral meshes are generated, and a flexible body generation method oriented towards multibody dynamics is used for mode reduction, retaining the first few modes to cover the frequency band from 0 to several kilohertz.

[0013] A five-DOF Hertz bearing contact model and a time-varying meshing stiffness gear pair are integrated. The bearing periodic impact function is injected into the dynamic simulation platform to simulate spalling defects, and a stiffness mutation model is used to generate tooth breakage fault data.

[0014] A three-dimensional operating space of wind speed, load, and temperature was established. Multiple sets of operating points were generated using space filling design technology. High-precision simulations were performed on various typical scenarios, including start-up conditions, rated operation conditions, over-power conditions, and emergency shutdown conditions. Multi-channel data was collected simultaneously to form a full life cycle dataset that includes multi-level expansion of tooth surface pitting and multi-stage evolution of tooth breakage.

[0015] Preferably, as one possible implementation, it also includes deploying a multi-source sensor array to collect first and second monitoring data of the wind turbine gearbox. The first monitoring data includes vibration signals, stress data, and displacement signals; the second monitoring data includes wind speed, power, and temperature parameters.

[0016] Preferably, as one feasible implementation, the enhanced hybrid dataset is then input into a spatiotemporal feature encoding fault diagnosis model, and a health index is output, specifically including:

[0017] A spatiotemporal feature encoder architecture is constructed, with the input layer receiving multidimensional time-series data and the output layer generating fault scores. The feature extraction includes a time-domain analysis branch and a frequency-domain analysis branch. The time-domain analysis branch uses multi-layer separable convolutional groups to capture time-domain features. The frequency-domain analysis branch uses a fast Fourier transform layer to convert the input to the frequency domain and uses adaptive wavelet decomposition to dynamically adjust the scale parameters to match rotational speed fluctuations.

[0018] A working condition feature fusion module is set up to map wind speed, power, and temperature working condition parameters to a high-dimensional vector, and an attention weighting mechanism is adopted to dynamically adjust the weights of different features according to the current working condition.

[0019] A two-stage strategy was adopted to train the model: the first stage used simulation data to pre-train the model, and the second stage introduced reinforcement domain adversarial training, and a constraint loss function was constructed based on the physical indices of envelope entropy and harmonic distortion, and the model was fine-tuned using measured data.

[0020] After fine-tuning, the final spatiotemporal feature-encoded fault diagnosis model is obtained. Then, the enhanced hybrid dataset is input into the spatiotemporal feature-encoded fault diagnosis model, and the health index is output.

[0021] Preferably, as one possible implementation, the health index is an index value within the range of pq.

[0022] Preferably, as one possible implementation, the signal anomaly information includes amplitude abrupt changes and waveform distortion.

[0023] Preferably, as one feasible implementation, it also includes: deploying an edge online early warning system and realizing dynamic fault early warning:

[0024] A lightweight fault scoring model is deployed on the edge computing terminal, and a three-level signal processing pipeline is established, including: dynamic noise reduction using an adaptive FIR filter, time-domain index calculation using a sliding window, and simultaneous execution of time-domain impulse detection and frequency-domain sideband analysis.

[0025] By combining real-time collected wind speed, power, and oil temperature data, a dynamic weight matrix is ​​generated for feature fusion, and a health index is calculated and output.

[0026] Based on the statistical characteristics of the sliding time window, an adaptive dynamic alarm threshold is set, and a hierarchical early warning and response mechanism is implemented. When the health index exceeds different thresholds, local data storage is triggered, an alarm is pushed to the central monitoring system, or the interlocking safety chain is shut down.

[0027] Preferably, as one feasible implementation, the high-precision simulation for various typical scenarios specifically includes:

[0028] Start the operating condition simulation operation;

[0029] Simulated operation under rated operating conditions;

[0030] Overpower condition simulation operation;

[0031] Emergency shutdown scenario simulation operation.

[0032] Preferably, as one possible implementation, the mathematical expression for the generator of the deep convolutional generative adversarial network is: ;

[0033] In the formula: For the input noise vector, The LeakyReLU activation function is used. The output layer is a sigmoid layer; where G(z) represents the generator output, W1, W2, W3, W4, and W5 represent the weight matrices of different layers of the generator, and b1, b2, b3, and b4 represent the bias vectors; the generator adopts a neural network structure consisting of m fully connected layers and n deconvolutional layers.

[0034] The expression for the adversarial loss function L is:

[0035] ;

[0036] In the formula: D is the discriminator, x is the real sample, G(z) is the generated sample, and λ is the penalty coefficient. represents the linear interpolation between real and generated samples; where Ladv is the generative adversarial loss function used to train the GAN, D(x) is the discriminator output, and D(z) is the discriminator's judgment result for the generated sample G(z). This represents the gradient penalty term of the discriminator.

[0037] Preferably, as one possible implementation, the setting of the adaptive dynamic alarm threshold is specifically achieved through the following formula: ;

[0038] In the formula, S(t) is the fault score at time t. Let Q_a be the length of the sliding time window, Q_{a} represent the percentage quantile, μ be the adjustment obtained by fitting high-score events through a generalized Pareto distribution, and Twarn(t) represent the warning value at time T. To compensate for the score.

[0039] Preferably, as one possible implementation, the graded early warning and response mechanism specifically comprises: when the health index is in the range of a1 to a2, triggering local high-frequency data storage; when the health index is in the range of a2 to a3, pushing alarm information to the central monitoring system; when the health index is greater than a3, the interlocking safety chain executes a shutdown operation (where a1...). <a2<a3)。

[0040] This invention provides a wind turbine gearbox condition monitoring and fault early warning system, comprising: a first model processing module, a data processing module, and a second model processing module;

[0041] The first model processing module is used to define the key components of the wind turbine gearbox as flexible bodies, and to construct a multibody dynamics model based on the flexible bodies. The multibody dynamics model receives input multi-source time-series data, then performs feature extraction and multi-index fault early warning algorithm processing, and outputs signal anomaly information for the current multi-source time-series data. The signal anomaly information is gearbox fault data. The signal anomaly information includes amplitude abrupt changes and waveform distortion.

[0042] The data processing module is used to construct a hybrid dataset guided by physical mechanisms and perform data augmentation: A sensor network is deployed at key locations in the gearbox to collect data, constructing a hybrid dataset. This dataset includes measured data from healthy operation, measured data from typical faults, and augmented data generated by a physically constrained deep convolutional generative adversarial network (DBGAN). DBGAN is used to augment the data; its generator employs a neural network structure containing m fully connected layers and n deconvolutional layers, and the discriminator consists of a convolutional layer and a fully connected layer. The adversarial loss function introduces a gradient penalty term. The augmented data undergoes standardization preprocessing, including sliding window segmentation with a fixed time step and the use of an elliptic filter to eliminate high-frequency interference.

[0043] The second model processing module is used to input the spatiotemporal feature encoding fault diagnosis model based on the enhanced hybrid dataset, and then output a health index.

[0044] Accordingly, the present invention provides a computer-readable storage medium storing a computer program, the computer program being executed by a processor of the wind turbine gearbox status monitoring and fault early warning method.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] This invention provides a method for wind turbine gearbox condition monitoring and fault early warning, addressing technical shortcomings in wind turbine gearbox condition monitoring such as model failure due to insufficient fault data, high false alarm rates due to static threshold mechanisms, and poor cross-model generalization ability. This invention uses time-series vibration characteristic values ​​of the low-speed and high-speed shafts of the wind turbine gearbox as input and fault scoring indicators at corresponding times as output. It aims to overcome the limitations of monitoring a single physical quantity, establish a dynamic adaptive early warning model, significantly reduce false alarm and missed alarm rates in complex operating environments, achieve accurate identification and early warning of early faults in wind turbine gearboxes, and improve equipment safety and wind farm economic benefits.

[0047] This invention overcomes the shortcomings of traditional data-driven methods, which struggle to construct effective diagnostic models due to the scarcity of severe gearbox fault samples. This invention innovatively adopts a technical approach that deeply integrates physical mechanisms with data-driven methods: by constructing a high-fidelity multibody dynamics model, it accurately simulates the entire process from a healthy state to fault evolution, generating a simulation dataset covering the entire lifecycle. A physically constrained deep convolutional generative adversarial network is introduced for data augmentation, significantly expanding the fault samples, especially data on rare fault types, while adhering to physical laws. Ultimately, a hybrid dataset consisting of simulation data, measured data, and augmented data is formed, effectively addressing the core pain points of insufficient model training and weak generalization ability under small sample conditions.

[0048] Meanwhile, this invention significantly improves diagnostic accuracy under complex operating conditions. It achieves adaptive diagnosis through multiple technical means; a spatiotemporal feature encoder is constructed to capture both time and frequency domain features of the signal, and the feature weights are dynamically adjusted through an operating condition feature fusion module, enabling the model to intelligently adapt to changes in operating conditions such as wind speed, power, and temperature. A two-stage training strategy and reinforcement domain adversarial training are employed to enhance the model's transfer capability from the simulation domain to the measured domain, improving generalization performance across different machine types and operating environments. This invention proposes an adaptive dynamic alarm threshold algorithm, which automatically adjusts the threshold based on the statistical characteristics of the equipment's recent operating status, overcoming the poor adaptability of fixed thresholds. A tiered early warning and response mechanism is established, triggering differentiated response measures based on different intervals of the health index, achieving precise and efficient closed-loop management from early warning to shutdown, effectively preventing risks while avoiding over-response.

[0049] In summary, this invention, through a three-pronged technological innovation—compensating for data gaps through modeling and simulation, improving diagnostic accuracy through intelligent algorithms, and ensuring real-time early warning through edge computing—successfully constructs a high-precision, low-latency, and highly generalizable intelligent early warning system for wind turbine gearboxes. Attached Figure Description

[0050] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. The accompanying drawings below illustrate some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0051] Figure 1 This is a flowchart of a method for monitoring the condition and providing early warning of faults in a wind turbine gearbox, provided in Embodiment 1 of the present invention.

[0052] Figure 2 This is a flowchart of a wind turbine gearbox condition monitoring and fault early warning system provided in Embodiment 2 of the present invention;

[0053] Labels: First model processing module 10; Data processing module 20; Second model processing module 30. Detailed Implementation

[0054] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings.

[0056] Example 1

[0057] like Figure 1 As shown, this invention proposes a method for monitoring the condition and providing early warning of faults in a wind turbine gearbox, comprising the following steps:

[0058] S10. Based on the definition of key components of the wind turbine gearbox as flexible bodies, a multibody dynamics model is constructed based on the flexible bodies. The multibody dynamics model receives input multi-source time-series data, then performs feature extraction and multi-index fault early warning algorithm processing, and outputs signal anomaly information for the current multi-source time-series data. The signal anomaly information is gearbox fault data (e.g., gearbox fault data may include broken tooth fault data and bearing fault data, etc.).

[0059] S101. Construct the geometric model and parametric assembly of the multi-stage planetary gear structure based on the structural parameters of the gearbox, and drive the key dimension association through equations; wherein the multi-stage planetary gear is, for example, a three-stage planetary gear; at this time, construct the geometric model and parametric assembly of the three-stage planetary gear structure parameters, and drive the key dimension association through equations; the three-stage planetary gear structure parameters include: the first stage sun gear has a diameter of Φ632mm and 42 teeth, and is matched with 4 planetary gears with a diameter of Φ785mm; the second stage sun gear has a diameter of Φ485mm and 38 teeth, and is matched with 5 planetary gears with a diameter of Φ595mm; the third stage parallel shaft pinion has a diameter of Φ550mm, and drives the large gear with a diameter of Φ1460mm;

[0060] S102. The planetary carrier and the housing bearing seat are transformed into flexible bodies, high-order hexahedral meshes are generated, and the Craig-Bampton method is used for mode reduction, retaining the first 35 modes to cover the 0-5500Hz frequency band.

[0061] S103 integrates a five-degree-of-freedom Hertz bearing contact model with a time-varying meshing stiffness gear pair, injects a bearing periodic impact function into the dynamic simulation platform to simulate spalling defects, and uses a stiffness mutation model to generate tooth breakage fault data.

[0062] S104. Establish a three-dimensional working condition space of wind speed-load-temperature, use Latin hypercube sampling technology to generate multiple sets of working condition points, and perform high-precision simulations for various typical scenarios including start-up working condition, rated operation working condition, over-power working condition and emergency shutdown working condition. Simultaneously collect multi-channel data to form a full life cycle dataset containing multi-level expansion of tooth surface pitting and multi-stage evolution of tooth breakage.

[0063] S20. Construct a hybrid dataset guided by physical mechanisms and perform data augmentation:

[0064] S201. Deploy a sensor network at key locations in the gearbox to collect data and construct a hybrid dataset. The hybrid dataset includes measured data of healthy operation, measured data of typical faults, and enhanced data generated by a physically constrained deep convolutional generative adversarial network.

[0065] S202. A deep convolutional generative adversarial network is used to augment the data. The generator adopts a U-Net structure containing five fully connected layers and three deconvolutional layers. The discriminator consists of three convolutional layers and two fully connected layers. The adversarial loss function introduces a gradient penalty term.

[0066] S203. Perform standardization preprocessing on the enhanced data, including sliding window segmentation with a fixed time step and using an elliptic filter to eliminate high-frequency interference.

[0067] It should be noted that, to address the discrepancy between the distribution of simulation and measured data in wind turbine gearbox fault early warning, a three-stage transfer learning framework guided by physical mechanisms was constructed. First, a multi-source sensor network was deployed at the wind farm: lidar monitored the radial displacement of the shaft system at a 10kHz sampling rate, fiber optic strain gauges acquired stress at the root of the planetary carrier at 20kHz, triaxial accelerometers captured vibration of the gearbox bearing housing at 50kHz, acoustic emission sensors monitored crack propagation waves in the gear meshing zone at 100kHz, and oil sensors detected abrasive particle concentration in the return oil pipeline in real time. Health data from eight identical wind turbines operating continuously for three months was collected, covering the entire operating range of wind speeds from 3 to 25 m / s, recording 5 minutes of complete waveform every 10 minutes. Simultaneously, two pre-configured faulty turbines were configured for bearing spalling / gear pitting. When the comprehensive health index fell below 0.8, a high-density sampling mode was automatically triggered, increasing the vibration signal sampling rate to 100kHz and recording 2 minutes of complete data every 5 minutes.

[0068] The model training adopts a phased optimization strategy: In the first phase, the generated simulation samples are used for basic model training. The cross-entropy loss of fault classification is minimized by an adaptive moment estimation optimizer. The initial learning rate is set to 0.001 and dynamically adjusted using a cosine annealing strategy. In the second phase, real-world samples are introduced for transfer fine-tuning. Label smoothing technology is used to alleviate the distribution difference between simulation data and real-world data.

[0069] The feature distributions of simulation and measured data are aligned through domain adversarial training: a domain classifier with gradient inversion layers is constructed, enabling the feature extractor to generate domain-invariant features, and its loss function includes a measure of the domain difference between simulation and measured data. A physical constraint loss function is constructed based on the fault dynamics equation, forcing the neural network output to conform to the fault modulation mechanism during the fine-tuning stage. Correlation constraints between simulation and measurement are established through three physical indices: envelope entropy, kurtosis, and harmonic distortion. An active learning strategy is adopted, automatically triggering high-density sampling when the comprehensive health index of the faulty unit is abnormal, forming high-quality labeled samples. To address the sample imbalance problem, a physically constrained DCGAN enhancement model is developed: the generator takes a 32-dimensional operating condition vector and fault code as input, and generates multi-channel vibration data conforming to the envelope spectrum characteristics through a three-layer transposed convolutional network; the discriminator innovatively introduces spectral correlation constraints, requiring the spectral feature error between the generated data and the measured data to be less than 8%, ensuring the physical rationality of the generated samples. Finally, a hybrid dataset containing more than 200,000 samples was constructed, of which 25% was rigid-flexible coupling simulation data, 60% was health test data, 7.5% was fault test data, and 7.5% was GAN-generated data, covering all working conditions of 8 types of fault modes.

[0070] S30. Based on the enhanced hybrid dataset, the spatiotemporal feature encoding fault diagnosis model is input again, and then the health index is output. It also includes deploying a multi-source sensor array to collect the first monitoring data and the second monitoring data of the wind turbine gearbox. The first monitoring data includes vibration signals, stress data and displacement signals; the second monitoring data includes wind speed, power and temperature operating parameters.

[0071] S301. Construct a spatiotemporal feature encoder architecture. The input layer receives multi-dimensional time-series data, and the output layer generates a fault score. The feature extraction includes a time-domain analysis branch and a frequency-domain analysis branch. The time-domain analysis branch uses multi-layer separable convolutional groups to capture time-domain features. The frequency-domain analysis branch uses a fast Fourier transform layer to convert the input to the frequency domain and uses adaptive wavelet decomposition to dynamically adjust the scale parameter to match the rotational speed fluctuation.

[0072] S302. Set up a working condition feature fusion module to map wind speed, power, and temperature working condition parameters to a high-dimensional vector, and adopt an attention weighting mechanism to dynamically adjust the weights of different features according to the current working condition.

[0073] S303. A two-stage strategy is adopted to train the model: the first stage uses simulation data to pre-train the model, and the second stage introduces enhanced domain adversarial training, and constructs a constraint loss function based on the physical indices of envelope entropy and harmonic distortion, and fine-tunes the model using measured data.

[0074] After fine-tuning, the final spatiotemporal feature-encoded fault diagnosis model is obtained. Then, the enhanced hybrid dataset is input into the spatiotemporal feature-encoded fault diagnosis model, and the health index is output.

[0075] In the above steps, during the spatiotemporal feature encoding operation of multi-source data, the temporal analysis employs a deep separable convolutional network for temporal feature extraction, capturing transient impacts and waveform distortions of the vibration signal through convolution operations. Frequency domain analysis utilizes a hybrid architecture of adaptive wavelet packets and FFT to perform frequency domain analysis on the signal, extracting harmonic information of gear meshing frequency and bearing fault characteristics. Operating condition features combine operating parameters such as wind speed, power, and temperature with spatiotemporal features through an attention mechanism, dynamically adjusting the feature weights. The spatiotemporal feature encoding fault diagnosis model inputs the spatiotemporal features and operating condition features into a fault scoring network, employing a three-layer fully connected neural network to quantify the health status. It outputs a health index (fault score) between 0 and 1, representing the operational health status of the wind turbine gearbox. Detailed subsequent operation procedures will be provided later.

[0076] Following step S30, the following step is also included: S40, deploying an edge online early warning system and implementing dynamic fault early warning:

[0077] S401. Deploy a lightweight fault scoring model on the edge computing terminal and establish a three-level signal processing pipeline, including: dynamic noise reduction using an adaptive FIR filter, time-domain index calculation using a sliding window, and simultaneous execution of time-domain impulse detection and frequency-domain sideband analysis.

[0078] S402. Combine real-time collected wind speed, power, and oil temperature data to generate a dynamic weight matrix for feature fusion, calculate and output the health index;

[0079] S403. Based on the statistical characteristics of the sliding time window, set an adaptive dynamic alarm threshold and implement a graded early warning and response mechanism; when the health index exceeds different thresholds, trigger local data storage, push alarms to the central monitoring system or shut down the interlocked safety chain.

[0080] It should be noted that an intelligent monitoring terminal industrial computer is deployed at the edge of the wind turbine to collect raw data from the sensor array via a serial bus and save it to the industrial computer. The system collects 50kHz high-frequency vibration waveforms from the triaxial accelerometer of the gearbox bearing housing, 20kHz stress data from the planetary carrier fiber optic strain gauge, and 10kHz displacement signals from the shaft laser alignment instrument in real time. Preprocessing is completed through a three-stage hardware acceleration pipeline: first, an adaptive FIR filter is used for real-time noise reduction, with its cutoff frequency dynamically adjusted according to the spindle speed; second, time-domain statistical characteristics, including RMS values, kurtosis factors, and peak values, are calculated using a 100ms sliding window. The lightweight inference engine is equipped with a fault warning model optimized by knowledge distillation, and the model parameter size is compressed from the initial 85MB to 8.3MB, meeting the edge resource constraints of memory usage not exceeding 50MB and computational load less than one billion floating-point operations per second.

[0081] The real-time processing flow processes three sensor signals simultaneously within a 200ms analysis window. The feature extraction layer outputs 256-dimensional time-domain features and 192-dimensional frequency-domain features. The operating condition fusion module combines real-time wind speed, output power, and lubricating oil temperature to generate dynamic weighting coefficients. The fault scoring network outputs a health index ranging from 0 to 1. When the index value exceeds a dynamic threshold, a tiered warning is triggered—this threshold is calculated based on the 98th percentile of 1000 health index samples within the last 100 seconds and is automatically updated every 10 seconds. The warning result is transmitted to the site monitoring center via 4G / 5G narrowband communication, implementing a three-level response mechanism: Level 1 warning (index 0.8-0.9) initiates local high-resolution data storage; Level 2 warning (0.9-0.95) pushes alarm information to the operation and maintenance platform; Level 3 warning (>0.95) triggers a safety chain to execute shutdown protection.

[0082] The aforementioned S40, deploying an edge-based online early warning system and implementing dynamic fault early warning, specifically achieves tiered early warning and response. That is, when the health index reaches a specific threshold, the system triggers an early warning and responds according to the following levels: Level 1 warning (health index 0.8-0.9): Local data storage is initiated for further analysis. Level 2 warning (health index 0.9-0.95): Alarm information is pushed to the operation and maintenance platform. Level 3 warning (health index > 0.95): The safety chain is activated to execute shutdown protection. The above warning information is transmitted to the wind farm's monitoring center via 4G / 5G narrowband communication to ensure timely response.

[0083] Preferably, as one possible implementation, step S104, the high-precision simulation for multiple typical scenarios specifically includes:

[0084] S1041. Start-up condition simulation: Simulates an acceleration process lasting 200 seconds within a wind speed range of 5-8 m / s, and dynamically adjusts the yaw angle using a formula;

[0085] S1042, Rated operating condition simulation: Reproduce the stable operating state for 700 seconds at a speed of 10.5±0.4 rpm;

[0086] S1043, Overpower Condition Simulation: Construct a turbulent wind speed of 26±4m / s and an overspeed condition greater than 15rpm to simulate extreme load operation for 150 seconds.

[0087] S1044, Emergency Stop Scenario Simulation: Reproduce the step deceleration process from 25m / s wind speed to 9m / s wind speed, and collect dynamic response data during the braking process in conjunction with a 42MPa hydraulic brake.

[0088] Preferably, as one possible implementation; in step S202, the mathematical expression of the generator of the deep convolutional generative adversarial network is: ;

[0089] In the formula: For the input noise vector, The LeakyReLU activation function is used. The output layer is a sigmoid layer; where G(z) represents the generator output, W1, W2, W3, W4, and W5 represent the weight matrices of different layers of the generator, and b1, b2, b3, and b4 represent the bias vectors; the generator adopts a U-Net structure consisting of five fully connected layers and three deconvolutional layers.

[0090] The expression for the adversarial loss function L is:

[0091] ;

[0092] In the formula: D is the discriminator, x is the real sample, G(z) is the generated sample, and λ is the penalty coefficient. =10 is the penalty coefficient. represents the linear interpolation between real and generated samples; where Ladv is the generative adversarial loss function used to train the GAN, D(x) is the discriminator output, and D(z) is the discriminator's judgment result for the generated sample G(z). This represents the gradient penalty term of the discriminator.

[0093] The generated sample is λ, where λ is the penalty coefficient and x̂ is the linear interpolation between the real sample and the generated sample.

[0094] Preferably, as one possible implementation; in step S403, the setting of the adaptive dynamic alarm threshold is specifically achieved through the following formula: ;

[0095] In the formula, S(t) is the fault score at time t. Let Q_{0.95} be the length of the sliding time window, Q_{0.95} represent the 95th percentile, μ be the adjustment amount obtained by fitting high-score events through a generalized Pareto distribution, and Twarn(t) represent the warning value at time T. To compensate for the score.

[0096] The tiered early warning and response mechanism is as follows: when the health index is in the range of 0.8 to 0.9, local high-frequency data storage is triggered; when the health index is in the range of 0.9 to 0.95, alarm information is pushed to the central monitoring system; when the health index is greater than 0.95, the interlocking safety chain executes a shutdown operation.

[0097] The following is a specific example to illustrate the distance in this embodiment: Taking a 10MW wind turbine in a certain region as an example, this invention mainly includes the following steps: First, a rigid-flexible coupled multibody dynamics model of the gearbox is constructed, with three-stage planetary gear structure parameters (first-stage sun gear Φ632mm / 42 teeth meshing with 4 Φ785mm planetary gears, second-stage sun gear Φ485mm / 38 teeth meshing with 5 Φ595mm planetary gears, third-stage parallel shaft Φ550mm pinion driving Φ1460mm large gear). A geometric model is constructed based on 3D modeling software, a parametric assembly is constructed, and key dimensions are associated through equations:

[0098] ;

[0099] Indicates the diameter of the ring gear, Indicates the diameter of the sun gear, Indicates the diameter of the planetary gear, This refers to the backlash, which is the free clearance caused by wear and tear in the gear system.

[0100] The planetary carrier and housing bearing housing are transformed into flexible bodies. Simultaneously, the mesh shape is determined, and a high-order hexahedral mesh is generated. Modal reduction is performed using the Craig-Bampton method, retaining the first 35 modes covering the 0-5500Hz frequency band.

[0101] ;

[0102] In the formula: b is the boundary node, i is the internal node, K is the stiffness, M is the mass, and u is the displacement. The eigenvalue is the square of the angular frequency. A five-DOF Hertz bearing contact model is integrated with a time-varying meshing stiffness gear pair. Periodic impact functions of the bearing are injected into the Simpack platform to simulate spalling defects, and a stiffness mutation model is used to generate tooth breakage fault data.

[0103] Subsequently, multi-condition fault data was generated, establishing a three-dimensional operating condition space of wind speed, load, and temperature. Latin hypercube sampling technology was used to generate 1200 sets of operating condition points. High-precision simulations with a 0.1ms step size were performed for four typical scenarios: the simulation started with a 200-second acceleration process in the 5-8 m / s wind speed range. The formula enables dynamic adjustment of the yaw angle; the rated operation scenario is used to cover the fault causes under stable power generation conditions, reproducing 700 seconds of stable operation at 10.5±0.4 rpm; the overpower operation scenario focuses on the fault evolution simulation under extreme loads, constructing a turbulent wind speed of 26±4 m / s and an overspeed state of >15 rpm for 150 seconds; the emergency shutdown scenario reproduces the emergency response process of the fault, setting a step deceleration of 25 m / s to 9 m / s combined with a 42 MPa hydraulic brake (peak torque 3.8 MN·m), and simultaneously collecting 60 kHz triaxial vibration, 25 kHz planetary carrier stress, and 200 kHz acoustic emission data to form a full life cycle dataset including 12 levels of tooth surface pitting expansion and 6 stages of tooth breakage evolution.

[0104] In the transfer learning and sample augmentation phase, an enhanced sensor network was deployed at key locations in the 10MW gearbox: a 60kHz triaxial accelerometer monitored the vibration of the large bearing housing, a 25kHz fiber optic strain gauge acquired stress to strengthen the planetary carrier, and a 200kHz acoustic emission sensor monitored the crack propagation of the large-module gear. The transfer learning employed a combination of simulated and measured data for model training. Simultaneously, during data augmentation, a generative adversarial network with physical constraints was used to generate defect samples, compensating for the scarcity of fault data.

[0105] A hybrid dataset guided by physical mechanisms was constructed, comprising 30% simulated samples (65,000+) generated by rigid-flexible coupling, 55% from four months of healthy operation data from six 10MW units, 8% from measured data from pre-set peeling / pitting fault units, and the remaining 7% generated using a physically constrained DCGAN (spectral feature error <7%). Data augmentation was performed using a deep convolutional generative adversarial network (DCGAN). The generator employs a U-Net structure consisting of five fully connected layers and three deconvolutional layers, mathematically expressed as follows:

[0106] ;

[0107] In the formula: For the input noise vector, The LeakyReLU activation function is used. This is the output layer of the Sigmoid algorithm. Here, G(z) represents the generator output, W1, W2, W3, W4, and W5 represent the weight matrices of different layers of the generator, and b1, b2, b3, and b4 represent the bias vectors.

[0108] The discriminator consists of three convolutional layers and two fully connected layers, and a gradient penalty term is introduced into the adversarial loss function:

[0109] ;

[0110] In the formula: =10 is the penalty coefficient. This represents the linear interpolation between the real sample and the generated sample. Here, Ladv is the generative adversarial loss function used to train the GAN, D(x) is the discriminator output, and D(z) is the discriminator's judgment result for the generated sample G(z). This represents the gradient penalty term of the discriminator;

[0111] The enhanced data underwent a standardized preprocessing procedure, with a sliding window segmentation of 128 time steps (corresponding to 1.28 seconds) and a step size of 32 steps. An 8th-order elliptic filter was used to eliminate high-frequency interference. The transfer function is:

[0112] ;

[0113] In the formula: =0.1 is the ripple factor. =20kHz is the cutoff frequency. Where: It is the standard sine wave response function and damping ratio Frequency-related Indicates the system damping ratio. Cutoff frequency, s represents a complex frequency domain variable that describes the frequency response of a signal.

[0114] The model employs a spatiotemporal feature encoder architecture. The input layer receives 128×8-dimensional time-series data (8 channels including vibration / speed / torque, etc.), and the output layer generates fault scores. Feature extraction is divided into a time-domain branch and a frequency-domain branch. The time-domain analysis branch uses 5 separable convolutional layers, each containing a 3*1 convolutional kernel, with a channel expansion rate of 2.5 times, to capture ripple distortion features. The formula is as follows:

[0115] ;

[0116] Where: Ftime represents the output of the time domain analysis, Wdw represents the weights of the frequency domain filter, X is the input signal, bdw is the bias term, Wpw is the frequency domain weight, and bpw is the bias term.

[0117] The frequency domain analysis branch uses a Fast Fourier Transform layer to transform the input to the frequency domain, and simultaneously employs adaptive wavelet decomposition to dynamically adjust the scaling parameter to match rotational speed fluctuations, as shown in the following formula:

[0118] ;

[0119] Where: Wj,k(t) represents the frequency domain weighting function. Let t represent the basis function, where t is time and k is the time step.

[0120] The operating condition feature fusion module embeds operating condition parameters, mapping wind speed v, power P, and temperature T to a 32-dimensional vector.

[0121] ;

[0122] Where: ev is the feature error, ReLU is the activation function, and Wv is the wind speed weight. It represents the change in wind speed, and bv is the wind speed offset term.

[0123] By employing an attention-weighted mechanism, the weight of the rotational speed characteristic is increased to 0.68±0.05 under low wind speed conditions (v<6m / s), and the weight of the electromagnetic torque characteristic reaches 0.72±0.03 under overpower conditions (v>22m / s).

[0124] The model training employs a two-stage training strategy: the first stage uses simulation data to pre-train the model with an Adam optimizer and an initial learning rate of 0.0012; the second stage introduces reinforcement domain adversarial training, using constrained loss functions constructed through physical indices such as envelope entropy and harmonic distortion for fine-tuning with measured data. In the dynamic threshold alarm system, the fault score S(t) is generated through a three-layer fully connected network, and its mathematical expression is:

[0125] ;

[0126] In the formula: F is the 256-dimensional feature vector output by the spatiotemporal feature encoder, and W1, W2, and W3 are weight matrices. The activation function is LeakyReLU. The warning threshold is adaptive, based on the statistical characteristics of the sliding time window.

[0127] ;

[0128] In the formula: =24 is the window length, Q0.95 represents the 95th percentile, and the alarm threshold extreme value modeling uses a generalized Pareto distribution to fit high-score events. In the formula, S(t) is the fault score at time t. Let Q_{0.95} be the length of the sliding time window, Q_{0.95} represent the 95th percentile, μ be the adjustment amount obtained by fitting high-score events through a generalized Pareto distribution, and Twarn(t) represent the warning value at time T. To compensate for the score (fixed preset value).

[0129] The final deployment of the edge online early warning system: This embodiment of the invention uses dual NVIDIA Jetson AGX Orin edge terminals equipped with wideband signal conditioning circuits to process 60kHz vibration and 25kHz stress signals in real time; a three-level processing pipeline is established—adaptive FIR filter dynamically tracks speed for noise reduction (cutoff frequency automatically adjusts with the rated speed of 10.5rpm), a 250ms sliding window calculates the effective value / kurtosis / peak time-domain index, a lightweight fault scoring model (9.1MB of parameters) simultaneously performs time-domain deep separable convolutional impact detection (capturing 0.3-4ms pulses) and frequency-domain adaptive wavelet packet-FFT sideband analysis; combined with real-time wind speed / power / oil temperature to generate a dynamic weight matrix for feature fusion, outputting a 0-1 health index;

[0130] A dynamic threshold is set based on the 99th percentile value of a 120-second sliding window, and a three-level response is implemented: a health index of 0.8-0.9 triggers local 120kHz data storage, 0.9-0.95 pushes an alarm to the central monitoring system, and >0.95 triggers an interlocking safety chain shutdown.

[0131] In actual deployment at the East China Sea wind farm, the system successfully triggered a continuous early warning 68 hours before the planetary gear tooth breakage failure occurred. The health index rose from 0.84 to 0.98 at the time of the failure, and the damage location prediction error was only 3.2% upon unpacking verification. The study found that the system can issue early warnings 68 hours in advance with a fault location error of 3.2%, proving that the system effectively improves the accuracy of fault identification and the timeliness of early warning.

[0132] This invention addresses a method for early warning of gearbox faults in wind turbines. Its technical advantages can be summarized as follows: This invention overcomes the shortcomings of traditional data-driven methods, which struggle to construct effective diagnostic models due to the scarcity of severe gearbox fault samples. This invention innovatively adopts a deep integration of physical mechanisms and data-driven approaches: by constructing a high-fidelity multibody dynamics model, it accurately simulates the entire process from a healthy state to fault evolution, generating a simulation dataset covering the entire lifecycle. A physically constrained deep convolutional generative adversarial network is introduced for data augmentation, significantly expanding the fault samples, especially data on rare fault types, while adhering to physical laws. Ultimately, a hybrid dataset consisting of simulation data, measured data, and augmented data is formed, effectively solving the core pain points of insufficient model training and weak generalization ability under small sample conditions.

[0133] Meanwhile, this invention significantly improves diagnostic accuracy under complex operating conditions. It achieves adaptive diagnosis through multiple technical means; a spatiotemporal feature encoder is constructed to capture both time and frequency domain features of the signal, and the feature weights are dynamically adjusted through an operating condition feature fusion module, enabling the model to intelligently adapt to changes in operating conditions such as wind speed, power, and temperature. A two-stage training strategy and reinforcement domain adversarial training are employed to enhance the model's transfer capability from the simulation domain to the measured domain, improving generalization performance across different machine types and operating environments. This invention proposes an adaptive dynamic alarm threshold algorithm, which automatically adjusts the threshold based on the statistical characteristics of the equipment's recent operating status, overcoming the poor adaptability of fixed thresholds. A tiered early warning and response mechanism is established, triggering differentiated response measures based on different intervals of the health index, achieving precise and efficient closed-loop management from early warning to shutdown, effectively preventing risks while avoiding over-response.

[0134] In summary, this invention, through a three-pronged technological innovation—compensating for data gaps through modeling and simulation, improving diagnostic accuracy through intelligent algorithms, and ensuring real-time early warning through edge computing—successfully constructs a high-precision, low-latency, and highly generalizable intelligent early warning system for wind turbine gearboxes.

[0135] Example 2

[0136] like Figure 2 As shown, the present invention also proposes a wind turbine gearbox condition monitoring and fault early warning system, comprising: a first model processing module 10, a data processing module 20, and a second model processing module 30;

[0137] The first model processing module 10 is used to define the key components of the wind turbine gearbox as flexible bodies, and construct a multibody dynamics model based on the flexible bodies. The multibody dynamics model receives input multi-source time-series data, then performs feature extraction and multi-index fault early warning algorithm processing, and outputs signal anomaly information for the current multi-source time-series data. The signal anomaly information is gearbox fault data. The signal anomaly information includes amplitude abrupt changes and waveform distortion.

[0138] The data processing module 20 is used to construct a hybrid dataset guided by physical mechanisms and perform data augmentation: A sensor network is deployed at key locations in the gearbox to collect data, constructing a hybrid dataset. This hybrid dataset includes measured data from healthy operation, measured data from typical faults, and augmented data generated by a physically constrained deep convolutional generative adversarial network (DGAN). The DGAN is used to augment the data; its generator employs a U-Net structure containing five fully connected layers and three deconvolutional layers, and the discriminator consists of three convolutional layers and two fully connected layers. The adversarial loss function introduces a gradient penalty term. The augmented data undergoes standardized preprocessing, including sliding window segmentation with a fixed time step and the use of an elliptic filter to eliminate high-frequency interference.

[0139] The second model processing module 30 is used to input the spatiotemporal feature encoding fault diagnosis model based on the enhanced hybrid dataset, and then output a health index.

[0140] Example 3

[0141] Accordingly, the present invention provides a computer-readable storage medium storing a computer program, the computer program being executed by a processor of the wind turbine gearbox status monitoring and fault early warning method.

[0142] In summary, the wind turbine gearbox condition monitoring and fault early warning method and system proposed in this invention, in practical applications, firstly, defines the key components of the wind turbine gearbox as flexible bodies, and then constructs a multibody dynamics model based on these flexible bodies. The multibody dynamics model receives input multi-source time-series data, then performs feature extraction and multi-index fault early warning algorithm processing, and outputs signal anomaly information for the current multi-source time-series data. The signal anomaly information is gearbox fault data.

[0143] Then, a hybrid dataset guided by physical mechanisms was constructed and data augmentation was performed: a sensor network was deployed at key locations in the gearbox to collect data, constructing a hybrid dataset. This hybrid dataset included measured data of healthy operation, measured data of typical faults, and augmented data generated by a physically constrained deep convolutional generative adversarial network (DGAN). The DGAN was used to augment the data, with its generator employing a U-Net structure containing five fully connected layers and three deconvolutional layers, and the discriminator consisting of three convolutional layers and two fully connected layers. The adversarial loss function introduced a gradient penalty term. The augmented data underwent standardization preprocessing, including sliding window segmentation with a fixed time step and the use of an elliptic filter to eliminate high-frequency interference. Finally, the augmented hybrid dataset was input into a spatiotemporal feature encoding fault diagnosis model, and a health index was output.

[0144] This invention provides a method for condition monitoring and fault early warning of wind turbine gearboxes, addressing technical shortcomings in gearbox condition monitoring such as model failure due to insufficient fault data, high false alarm rates in static threshold mechanisms, and poor cross-model generalization ability. This invention uses time-series vibration characteristic values ​​of the low-speed and high-speed shafts of the wind turbine gearbox as input and fault scoring indicators at corresponding times as output. It aims to overcome the limitations of monitoring a single physical quantity, establishing a dynamic adaptive early warning model to significantly reduce false alarm and missed alarm rates in complex operating environments. This enables accurate identification and early warning of early faults in wind turbine gearboxes, improving equipment safety and the economic benefits of wind farms.

[0145] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; those skilled in the art can modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for monitoring the condition and providing early warning of faults in a wind turbine gearbox, characterized in that, The following steps are included: Based on the definition of key components of wind turbine gearbox as flexible bodies, a multibody dynamics model is constructed based on the flexible bodies. The multibody dynamics model receives input multi-source time-series data, then performs feature extraction and multi-index fault early warning algorithm processing, and outputs signal anomaly information for the current multi-source time-series data; the signal anomaly information is gearbox fault data. A hybrid dataset guided by physical mechanisms is constructed and augmented: a sensor network is deployed at key locations in the gearbox to collect data, and a hybrid dataset is constructed. The hybrid dataset includes measured data of healthy operation, measured data of typical faults, and augmented data generated by a physically constrained deep convolutional generative adversarial network. The data is augmented using a deep convolutional generative adversarial network. The generator adopts a neural network structure containing m fully connected layers and n deconvolutional layers, and the discriminator consists of a convolutional layer and b fully connected layers. The adversarial loss function introduces a gradient penalty term. The enhanced data is preprocessed by standardization, including sliding window segmentation with a fixed time step and elliptic filter to eliminate high-frequency interference. The enhanced hybrid dataset is then input into the spatiotemporal feature encoding fault diagnosis model, and then the health index is output.

2. The method according to claim 1, characterized in that, Based on the definition of key components of wind turbine gearbox as flexible bodies, a multibody dynamics model is constructed based on the flexible bodies. The multibody dynamics model receives input multi-source time-series data, then performs feature extraction and multi-index fault early warning algorithm processing, and outputs signal anomaly information for the current multi-source time-series data. The abnormal signal information is gearbox fault data, specifically including: Based on the structural parameters of the multi-stage planetary gear in the gearbox, its geometric model and parametric assembly are constructed, and key dimension associations are driven by equations. The planetary carrier and the housing bearing seat are transformed into flexible bodies, high-order hexahedral meshes are generated, and a flexible body generation method oriented towards multibody dynamics is used for mode reduction, retaining the first few modes to cover the frequency band from 0 to several kilohertz. A five-DOF Hertz bearing contact model and a time-varying meshing stiffness gear pair are integrated. The bearing periodic impact function is injected into the dynamic simulation platform to simulate spalling defects, and a stiffness mutation model is used to generate tooth breakage fault data. A three-dimensional operating space of wind speed, load, and temperature was established. Multiple sets of operating points were generated by space filling design. High-precision simulations were performed on various typical scenarios, including start-up, rated operation, overpower, and emergency shutdown. Multi-channel data were collected simultaneously to form a full life cycle dataset that includes multi-level expansion of tooth surface pitting and multi-stage evolution of tooth breakage.

3. The method according to claim 2, characterized in that, It also includes deploying a multi-source sensor array to collect first and second monitoring data of the wind turbine gearbox. The first monitoring data includes vibration signals, stress data, and displacement signals; the second monitoring data includes wind speed, power, and temperature parameters.

4. The method according to claim 3, characterized in that, The enhanced hybrid dataset is then input into the spatiotemporal feature encoding fault diagnosis model, which outputs a health index, specifically including: A spatiotemporal feature encoder architecture is constructed, with the input layer receiving multidimensional time-series data and the output layer generating fault scores. The feature extraction includes a time-domain analysis branch and a frequency-domain analysis branch. The time-domain analysis branch uses multi-layer separable convolutional groups to capture time-domain features. The frequency-domain analysis branch uses a fast Fourier transform layer to convert the input to the frequency domain and uses adaptive wavelet decomposition to dynamically adjust the scale parameters to match rotational speed fluctuations. A working condition feature fusion module is set up to map wind speed, power, and temperature working condition parameters to a high-dimensional vector, and an attention weighting mechanism is adopted to dynamically adjust the weights of different features according to the current working condition. A two-stage strategy was adopted to train the model: the first stage used simulation data to pre-train the model, and the second stage introduced reinforcement domain adversarial training, and a constraint loss function was constructed based on the physical indices of envelope entropy and harmonic distortion, and the model was fine-tuned using measured data. After fine-tuning, the final spatiotemporal feature-encoded fault diagnosis model is obtained. Then, the enhanced hybrid dataset is input into the spatiotemporal feature-encoded fault diagnosis model, and the health index is output.

5. The method according to claim 4, characterized in that, The health index is an index value within the range of pq; the signal anomaly information includes amplitude abrupt changes and waveform distortion.

6. The method according to claim 1, characterized in that, The steps, after inputting the enhanced hybrid dataset into the spatiotemporal feature encoding fault diagnosis model and outputting the health index, also include: deploying an edge online early warning system and implementing dynamic fault early warning. A lightweight fault scoring model is deployed on the edge computing terminal, and a three-level signal processing pipeline is established, including: dynamic noise reduction using an adaptive FIR filter, time-domain index calculation using a sliding window, and simultaneous execution of time-domain impulse detection and frequency-domain sideband analysis. By combining real-time collected wind speed, power, and oil temperature data, a dynamic weight matrix is ​​generated for feature fusion, and a health index is calculated and output. Based on the statistical characteristics of the sliding time window, an adaptive dynamic alarm threshold is set, and a hierarchical early warning and response mechanism is implemented. When the health index exceeds different thresholds, local data storage is triggered, an alarm is pushed to the central monitoring system, or the interlocking safety chain is shut down.

7. The method according to claim 6, characterized in that, The high-precision simulation for various typical scenarios specifically includes: Start the operating condition simulation operation; Simulated operation under rated operating conditions; Overpower condition simulation operation; Emergency shutdown scenario simulation operation.

8. The method according to claim 7, characterized in that, The mathematical expression for the generator in the deep convolutional generative adversarial network is: ; In the formula: For the input noise vector, The LeakyReLU activation function is used. The output layer is the Sigmoid layer; where G(z) represents the generator output, W1, W2, W3, W4, and W5 represent the weight matrices of different layers of the generator, and b1, b2, b3, and b4 represent bias vectors; the generator adopts a neural network structure consisting of m fully connected layers and n deconvolutional layers, and the expression for the adversarial loss function L is: ； In the formula: D is the discriminator, x is the real sample, G(z) is the generated sample, and λ is the penalty coefficient. is a linear interpolation between real and generated samples; where Ladv is the generative adversarial loss function used to train the GAN, D(x) is the discriminator output, and D(z) is the discriminator's judgment result for the generated sample G(z). This represents the gradient penalty term of the discriminator.

9. The method according to claim 8, characterized in that, The adaptive dynamic alarm threshold is set using the following formula: ; In the formula, S(t) is the fault score at time t. Let Q_{a} be the length of the sliding time window, Q_{a} represent the percentage quantile, μ be the adjustment obtained by fitting high-score events through a generalized Pareto distribution, and Twarn(t) represent the warning value at time T. To compensate for the score.

10. The method according to claim 9, characterized in that, The tiered early warning and response mechanism is as follows: when the health index is in the range of a1 to a2, local high-frequency data storage is triggered; when the health index is in the range of a2 to a3, alarm information is pushed to the central monitoring system; when the health index is greater than a3, the interlocked safety chain executes a shutdown operation (where a1...). <a2<a3)。 11. A wind turbine gearbox condition monitoring and fault early warning system, characterized in that, The wind turbine gearbox condition monitoring and fault early warning method as described in any one of claims 1-10 includes: a first model processing module, a data processing module, and a second model processing module; The first model processing module is used to define the key components of the wind turbine gearbox as flexible bodies, and to construct a multibody dynamics model based on the flexible bodies. The multibody dynamics model receives input multi-source time-series data, then performs feature extraction and multi-index fault early warning algorithm processing, and outputs signal anomaly information for the current multi-source time-series data. The signal anomaly information is gearbox fault data. The signal anomaly information includes amplitude abrupt changes and waveform distortion. The data processing module is used to construct a hybrid dataset guided by physical mechanisms and perform data augmentation: A sensor network is deployed at key locations in the gearbox to collect data, constructing a hybrid dataset. This dataset includes measured data from healthy operation, measured data from typical faults, and augmented data generated by a physically constrained deep convolutional generative adversarial network (DBGAN). DBGAN is used to augment the data; its generator employs a neural network structure containing m fully connected layers and n deconvolutional layers, and the discriminator consists of a convolutional layer and a fully connected layer. The adversarial loss function introduces a gradient penalty term. The augmented data undergoes standardization preprocessing, including sliding window segmentation with a fixed time step and the use of an elliptic filter to eliminate high-frequency interference. The second model processing module is used to input the spatiotemporal feature encoding fault diagnosis model based on the enhanced hybrid dataset, and then output a health index.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the wind turbine gearbox condition monitoring and fault early warning method as described in any one of claims 1 to 10.

Images