Fan gear box state monitoring method and system and electronic equipment
By screening the mapping relationship between wind speed and active power and the correlation changes of temperature monitoring parameters, and combining them with an adaptive recurrent neural network model, the problem of insufficient feature selection in the condition monitoring of wind turbine gearboxes was solved, enabling sensitive identification and accurate prediction of early faults, and improving the accuracy and stability of monitoring.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUANENG CHONGQING FENGJIE WIND POWER CO LTD
- Filing Date
- 2026-01-26
- Publication Date
- 2026-04-24
AI Technical Summary
Existing wind turbine gearbox condition monitoring methods do not fully consider the correlation changes between monitoring indicators under different operating conditions, resulting in a lack of state sensitivity in feature selection. Furthermore, the generalization ability of the model decreases under different operating conditions, making it difficult to effectively eliminate atypical operating samples. The prediction model is not sensitive to weak signal variations caused by early faults, affecting the accuracy and stability of condition monitoring.
By acquiring operational data of wind speed, active power, and temperature monitoring parameters, target data segments are selected based on the mapping relationship between wind speed and active power. The monitoring variables are processed using a target time series prediction model. Combined with the correlation changes between temperature monitoring parameters and non-temperature parameters, an adaptive recurrent neural network model is constructed for prediction to identify early faults.
It improves the accuracy and robustness of wind turbine gearbox condition monitoring, enhances sensitivity to fault conditions, can promptly reflect early performance degradation, and improves the model's adaptability and prediction accuracy under complex operating conditions.
Smart Images
Figure CN121916129A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of condition monitoring technology for wind power generation equipment, and in particular to a method, system and electronic equipment for condition monitoring of wind turbine gearboxes. Background Technology
[0002] In wind turbine gearbox condition monitoring, existing methods mostly rely on operational data collected by SCADA systems to build predictive models. However, they fail to fully consider the dynamic changes in the correlation between monitoring indicators under different operating conditions, resulting in a lack of state sensitivity in feature selection. Furthermore, the time-varying nature of data distribution caused by wind speed fluctuations and power variations reduces the model's generalization ability under different operating conditions. Traditional data screening relies on fixed thresholds or simple statistical rules, making it difficult to effectively eliminate atypical operating samples. Moreover, predictive models are insensitive to weak signal variations caused by early faults and are easily affected by operating conditions, leading to reduced fault identification accuracy and impacting the reliability and stability of condition monitoring. Summary of the Invention
[0003] The purpose of this application is to provide a method, system, and electronic device for monitoring the condition of a wind turbine gearbox, so as to alleviate the above-mentioned technical problems existing in the prior art.
[0004] In a first aspect, the present invention provides a method for monitoring the condition of a wind turbine gearbox, comprising: Acquire the operating data of the gearbox of the wind turbine to be monitored, including wind speed, active power and multiple temperature monitoring parameters; Based on the mapping relationship between wind speed and active power, the operating data is filtered to retain the target data segment that meets the preset correlation mode. Based on the correlation changes between each temperature monitoring parameter and non-temperature monitoring parameters under different operating conditions, the monitoring indicators in the target data segment that meet the preset correlation conditions are determined as target monitoring variables. The time series corresponding to the target monitoring variables are processed by the target time series prediction model to determine the predicted values of each monitoring indicator; The operating status of the wind turbine gearbox is monitored based on the deviation between the predicted value and the actual measured value of the corresponding target monitoring variable.
[0005] In an optional implementation, the operating data of the wind turbine gearbox to be monitored is acquired, including: Collect historical operating data of the gearbox of the wind turbine to be monitored. The historical operating data includes wind speed, active power, and measured values of high-speed shaft front temperature, high-speed shaft rear temperature, oil temperature and inlet lubricating oil temperature at multiple time points. Data records in the shutdown state are removed, and abnormal values caused by null values or sensor malfunctions are eliminated to generate preliminary cleaned operating data.
[0006] In an optional implementation, the operating data is filtered based on the mapping relationship between wind speed and active power, retaining target data segments that conform to a preset correlation pattern, including: An ideal mapping curve is constructed based on a preset correlation pattern between wind speed and active power under normal operating conditions. The correspondence between wind speed and active power in the operational data is compared with the ideal mapping curve; Operational data whose actual active power deviates significantly from the expected power corresponding to the ideal mapping curve under wind speed are identified as abnormal data and removed, while target data segments that conform to the correlation pattern are retained.
[0007] In an optional implementation, based on the changes in the correlation between each temperature monitoring parameter and non-temperature monitoring parameters under different operating conditions, monitoring indicators in the target data segment that meet preset correlation conditions are determined as target monitoring variables, including: Calculate the first correlation coefficient between each temperature monitoring parameter and non-temperature monitoring parameters under normal operating conditions; Calculate the second correlation coefficient between the same temperature monitoring parameter and non-temperature monitoring parameters before and after the fault occurs; If the absolute difference between the first correlation coefficient and the second correlation coefficient corresponding to a certain temperature monitoring parameter exceeds a preset difference threshold, and both the first correlation coefficient and the second correlation coefficient are in the medium intensity range, then the temperature monitoring parameter is determined as the target monitoring variable.
[0008] In an optional implementation, the method further includes: Construct a target time series prediction model to process the time series of target monitoring variables; The construction of the target time series prediction model includes: The time series of the target monitoring variable is divided into several subsequences to maximize the statistical distribution difference between adjacent subsequences. The distribution distance is determined by the maximum mean difference, and the optimal segmentation strategy is solved by dynamic programming to obtain multiple subsequences with significantly different distribution characteristics. Model pre-training is performed on each subsequence to minimize prediction error loss and initialize the parameters of the target time series prediction model.
[0009] In an optional implementation, constructing the target time series prediction model further includes: On the hidden layer state sequence of a recurrent neural network, the probability distribution difference between different subsequences is calculated, and the influence of each subsequence on the distribution alignment process is determined based on the probability distribution difference. The importance weights of the probability distribution of the corresponding subsequences are dynamically adjusted based on the changes in the distribution differences in consecutive training rounds. The importance weights of the probability distribution are applied to the distribution difference loss term, and the distribution difference loss term is weighted and fused with the prediction error loss term to generate a comprehensive loss function; Based on the comprehensive loss function, the gradient of the parameters is calculated layer by layer and the model weights are updated through the gradient backpropagation algorithm to complete the training of the target time series prediction model.
[0010] In an optional implementation, the operating status of the wind turbine gearbox is monitored based on the deviation between the predicted value and the actual measured value of the corresponding target monitoring variable, including: Calculate the residual sequence between the predicted value and the actual measured value of the target monitoring variable within a continuous time window; Perform time-series trend analysis on the residual series of the same monitored variable to identify any abnormal patterns of continuous increase or periodic fluctuation. Perform intervariate correlation analysis on the residual sequences of multiple target monitoring variables that exceed the preset threshold range, and calculate the consistency index of the changes of each residual sequence in the time dimension; If the consistency index of the change exceeds the set criteria, it is determined that there is a risk of early failure or compound failure in the wind turbine gearbox. Output the corresponding operating status identifier and alarm level information.
[0011] Secondly, the present invention provides a wind turbine gearbox condition monitoring system, comprising: The data acquisition module is used to acquire the operating data of the gearbox of the wind turbine to be monitored. The operating data includes wind speed, active power and multiple temperature monitoring parameters. The data filtering and variable determination module is used to filter the operating data according to the mapping relationship between wind speed and active power, retain the target data segment that meets the preset correlation mode, and determine the monitoring indicators in the target data segment that meet the preset correlation conditions as target monitoring variables based on the correlation changes of each temperature monitoring parameter and non-temperature monitoring parameters under different operating conditions. The time series forecasting module is used to process the time series corresponding to the target monitoring variables through the target time series forecasting model to determine the predicted values of each monitoring indicator; The status monitoring module is used to monitor the operating status of the wind turbine gearbox based on the deviation between the predicted value and the actual measured value of the corresponding target monitoring variable.
[0012] Thirdly, the present invention provides an electronic device including a processor and a memory, wherein the memory stores computer-executable instructions that can be executed by the processor, and the processor executes the computer-executable instructions to implement the wind turbine gearbox status monitoring method of any of the foregoing embodiments.
[0013] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions. When the computer-executable instructions are invoked and executed by a processor, the computer-executable instructions cause the processor to implement the wind turbine gearbox status monitoring method of any of the foregoing embodiments.
[0014] The wind turbine gearbox condition monitoring method, system, and electronic equipment provided in this application effectively eliminate distribution offset samples caused by wind shear, turbulence, or abnormal operating conditions by filtering operating data based on the mapping relationship between wind speed and active power, thereby improving the quality and representativeness of input data and enhancing the model's adaptability to real operating environments. By utilizing the changing characteristics of the correlation between temperature monitoring parameters and non-temperature parameters under different operating conditions to filter target monitoring variables, the sensitivity of selected features to fault states is improved, solving the problems of feature redundancy and insufficient discrimination ability in traditional methods. Combining target time-series prediction models to perform time-series modeling of key variables can capture the dynamic evolution patterns under complex operating conditions and improve prediction accuracy; condition monitoring based on prediction deviations can promptly reflect early performance degradation. Overall, this improves the accuracy and robustness of wind turbine gearbox condition monitoring. Attached Figure Description
[0015] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0016] Figure 1 A flowchart of a wind turbine gearbox condition monitoring method provided in this application embodiment; Figure 2 A structural diagram of a wind turbine gearbox condition monitoring system provided in this application embodiment; Figure 3 A structural diagram of another wind turbine gearbox condition monitoring system provided in this application embodiment; Figure 4 This is a structural diagram of an electronic device provided in an embodiment of this application. Detailed Implementation
[0017] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0018] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.
[0019] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.
[0020] This application provides a method for monitoring the condition of a wind turbine gearbox. (See also...) Figure 1 As shown, the method mainly includes the following steps: S110 acquires the operating data of the gearbox of the wind turbine to be monitored, including wind speed, active power, and multiple temperature monitoring parameters.
[0021] The acquired operational data consists of multi-dimensional time-series parameters recorded by the Supervisory Control and Data Acquisition (SCADA) system during continuous wind turbine operation. Wind speed characterizes external excitation conditions, active power characterizes energy conversion efficiency, and temperature monitoring parameters mainly include high-speed shaft inlet temperature, high-speed shaft outlet temperature, gearbox oil temperature, and inlet lubricating oil temperature; their trends are closely related to mechanical friction conditions and lubrication performance. This step involves accessing the wind farm's historical database to collect complete cycle data, including data from normal operation, before a fault occurs, and after fault handling, ensuring the sample covers typical operating conditions and providing fundamental data support for subsequent analysis.
[0022] S120 filters the operating data based on the mapping relationship between wind speed and active power, retains the target data segment that meets the preset correlation mode, and determines the monitoring indicators in the target data segment that meet the preset correlation conditions as target monitoring variables based on the correlation changes of each temperature monitoring parameter and non-temperature monitoring parameters under different operating conditions.
[0023] The above data filtering based on the mapping relationship between wind speed and active power involves constructing an ideal response curve of wind speed and active power under normal conditions, eliminating data points that deviate from the curve due to sensor errors, extreme turbulence, or shutdown disturbances, and retaining target data segments that conform to typical energy conversion laws.
[0024] In one implementation, based on the feature selection of dynamic correlation changes, the Spearman correlation coefficient can be used to quantify the correlation strength between each temperature monitoring parameter and non-temperature monitoring parameters. The correlation coefficient is calculated separately under normal and abnormal conditions. If the absolute difference of the correlation of a certain parameter exceeds the set threshold and is in the medium correlation range in both states, it is considered to have the ability to distinguish the operating state and is selected as the target monitoring variable.
[0025] S130: The time series corresponding to the target monitoring variables is processed by the target time series prediction model to determine the predicted values of each monitoring indicator.
[0026] The target time series prediction model described in this embodiment is an Adaptive Recurrent Neural Network (AdaRNN) that integrates time distribution representation and matching mechanisms. First, the time series of the target monitoring variable is divided into several subsequences according to their statistical distribution characteristics, maximizing the distribution difference between adjacent subsequences. The maximum mean difference is used to measure the similarity between sequences, and a dynamic programming algorithm is used to solve for the optimal segmentation strategy. Then, pre-training is performed on each subsequence to minimize the prediction error loss and initialize the model parameters. Next, a time distribution matching module is introduced to align the probability distributions of different subsequences on the hidden layer state sequence, dynamically adjusting the importance weights of each distribution, and weightedly fusing the prediction loss and distribution alignment loss to complete model training. After training, the model can perform rolling predictions on the input time series and output predicted values for future or current times.
[0027] S140 monitors the operating status of the wind turbine gearbox based on the deviation between the predicted value and the actual measured value of the corresponding target monitoring variable.
[0028] The bias, or prediction residual, is used to characterize the degree of fit between the aforementioned target time-series prediction model and the actual operating state. By performing trend analysis on the residual sequence within a continuous time window, it is possible to identify abnormal patterns such as continuous increases or periodic abrupt changes. When multiple target monitoring variables simultaneously show a significant increase in residuals, cross-variable correlation analysis is further performed to calculate the consistency index of their changes over time. If this index exceeds the set criteria, it indicates that the system as a whole has deviated from the normal evolution path, and it is determined that there is a risk of early failure or compound failure. The corresponding operating status identifier and alarm level information are then output.
[0029] For ease of understanding, the following provides a detailed description of the wind turbine gearbox condition monitoring method provided in the embodiments of this application. This method aims to improve the accuracy and robustness of existing prediction models under complex operating conditions. Based on multi-source operating data collected by the SCADA system, and combined with data quality control, sensitive feature identification, and adaptive time-series modeling techniques, this method effectively identifies early-stage and complex gearbox faults.
[0030] First, obtain the operating data of the gearbox of the wind turbine to be monitored. In specific implementation, this may include the following steps 1.1 and 1.2: Step 1.1: Collect historical operating data of the gearbox of the wind turbine to be monitored. The historical operating data includes the wind speed, active power, and measured values of high-speed shaft front temperature, high-speed shaft rear temperature, oil temperature and inlet lubricating oil temperature at multiple time points.
[0031] Historical operating data refers to multi-dimensional time-series data that is continuously recorded over a long period of time through the wind farm monitoring and data acquisition system (SCADA), including complete operating condition information of the unit in multiple stages such as normal operation, potential failure occurrence and maintenance and repair.
[0032] Wind speed is used to characterize external excitation conditions, while active power reflects energy conversion efficiency and is a key indicator for evaluating the output performance of the unit. The high-speed shaft inlet temperature and high-speed shaft outlet temperature mainly characterize the working state of the high-speed stage bearings of the gearbox, and abnormal increases are related to poor lubrication or increased wear. Oil temperature is used to characterize the overall thermal balance of the gearbox, and the inlet lubricating oil temperature is related to the cooling system efficiency and the initial thermal state of the lubricating medium.
[0033] The data acquisition process can synchronously acquire time series data of each variable at a fixed sampling frequency (such as 10-minute average) to ensure data time sequence alignment and sufficient coverage period, providing high-quality original samples for subsequent modeling.
[0034] Step 1.2: Remove data records from the shutdown state and remove abnormal values caused by null values or sensor malfunctions to generate preliminary cleaned operating data.
[0035] The shutdown state refers to the operating condition where the fan is not connected to the grid, the main shaft is stationary, or the speed is close to zero. In this state, parameters such as temperature are not affected by the load and cannot reflect the true operating characteristics; retaining such data would interfere with model training. Therefore, it is necessary to identify and exclude data records from this time period based on signals such as power and speed. Null values refer to data loss caused by communication interruptions, storage failures, etc.; abnormal values caused by sensor malfunctions include extreme values exceeding physically reasonable ranges (such as oil temperature below environmental limits or above the material's tolerance limit), abrupt changes, or prolonged periods of constant change.
[0036] In practical applications, low-quality data points can be identified and removed or marked for rejection by setting filtering rules or using interpolation detection algorithms. The pre-cleaned operational data generated after this process has higher integrity and reliability, avoids noise interference caused by factors other than the equipment itself, and ensures the reliability of the data foundation for subsequent analysis.
[0037] The above method comprehensively covers key operating parameters and temperature monitoring indicators by systematically collecting historical operating data, ensuring the integrity and timeliness of input information; and effectively improves data quality and reduces the interference of irrelevant variables on model learning by removing downtime data and various outliers.
[0038] Furthermore, the above-mentioned filtering of operational data based on the mapping relationship between wind speed and active power, retaining target data segments that conform to the preset correlation pattern, may include the following steps 2.1 to 2.3 in specific implementation: Step 2.1: Construct an ideal mapping curve based on the preset correlation mode between wind speed and active power under normal operating conditions.
[0039] The aforementioned preset correlation mode refers to the energy conversion law of wind turbines gradually increasing their output power as the wind speed increases within the rated wind speed range (e.g., 3 m / s to 16 m / s) under healthy operating conditions. This preset correlation mode is used to characterize the physical response characteristics of the unit under ideal conditions such as no faults, normal control, and clean blades. The ideal mapping curve is the mathematical representation of this energy conversion law. This curve is established based on the statistical average wind speed-power response relationship of a large number of wind turbines in the wind farm under normal operating conditions. That is, it is constructed based on a set of sample points in a confirmed normal state from a large amount of historical operating data, and a continuous curve representing the average response trend is generated through statistical fitting methods (such as polynomial regression or piecewise smoothing interpolation).
[0040] Step 2.2: Compare the correspondence between wind speed and active power in the operating data with the ideal mapping curve.
[0041] The correspondence between wind speed and active power in the operational data refers to the data points consisting of each set of synchronously recorded wind speeds and corresponding active power values extracted from the dataset after initial cleaning.
[0042] During the comparison, these measured data points can be projected onto a two-dimensional space with wind speed as the abscissa and active power as the ordinate, and compared point by point with the expected power value at the same wind speed on the ideal mapping curve constructed in step 2.1. This comparison not only focuses on the magnitude of the absolute deviation, but also considers the consistency of local slope changes, thereby determining whether the actual output falls within the theoretical response band.
[0043] Step 2.3: Operational data whose actual active power deviates significantly from the expected power corresponding to the ideal mapping curve under wind speed are identified as abnormal data and removed, while target data segments that conform to the correlation mode are retained.
[0044] Significant deviation refers to a difference exceeding a set threshold between the measured active power and the expected power at the corresponding wind speed on the ideal mapping curve. This deviation is not instantaneous or random, but rather exhibits a persistent or systematic underestimation or overestimation. Such deviations may be caused by various unhealthy factors, such as decreased gearbox mechanical efficiency leading to increased energy loss, sensor calibration errors, control system malfunctions, or extreme environmental influences. These data points are identified as outliers and removed from the training set, retaining only data segments that are substantially consistent with the ideal response trend to generate target data segments that conform to the correlation pattern.
[0045] The above method constructs an ideal mapping curve as a benchmark for healthy operation, and combines the measured data with a systematic comparison of its response trend. This effectively identifies and eliminates abnormal samples caused by performance degradation or abnormal operating conditions, while retaining target data segments that conform to typical energy conversion laws. This improves the adaptability and accuracy of the entire condition monitoring method in complex field environments.
[0046] Furthermore, based on the changes in the correlation between various temperature monitoring parameters and non-temperature monitoring parameters under different operating conditions, the monitoring indicators in the target data segment that meet the preset correlation conditions are determined as target monitoring variables. In specific implementation, this may include the following steps 3.1 to 3.3: Step 3.1: Calculate the first correlation coefficient between each temperature monitoring parameter and non-temperature monitoring parameters under normal operating conditions.
[0047] The aforementioned temperature monitoring parameters include key variables reflecting the thermodynamic state of the gearbox, such as high-speed shaft inlet temperature, high-speed shaft outlet temperature, oil temperature, and inlet lubricating oil temperature. Non-temperature monitoring parameters include parameters such as wind speed, active power, rotational speed, and vibration acceleration, used to characterize the overall operating conditions of the unit. The first correlation coefficient refers to the statistical dependence of a certain temperature monitoring parameter on these operating parameters when the fan is in a confirmed normal operating state, quantified using the Spearman correlation coefficient. The Spearman correlation coefficient is defined as the Pearson correlation coefficient between ordinal variables, and the calculation formula is as follows:
[0048] in, These are the temperature monitoring parameters. and The average value was calculated. Analysis of 100,000 monitoring data samples (50,000 in normal conditions and 50,000 in abnormal conditions) revealed significant differences in the correlation between key monitoring indicators such as high-speed shaft temperature (before and after) and oil temperature and other features under normal and abnormal conditions. Based on this, monitoring indicators meeting the following conditions were selected as predictive features: ① Under both normal and fault conditions, the correlation with the target indicator was within the range of (0.2, 0.8); ② The absolute difference in correlation between normal and fault conditions exceeded 0.2, ensuring the feature's ability to distinguish between different operating states.
[0049] Step 3.2: Calculate the second correlation coefficient between the same temperature monitoring parameter and non-temperature monitoring parameters before and after the fault occurs.
[0050] The period before and after a fault refers to the time window centered on a known fault event (such as alarm triggering or downtime maintenance records), including data segments from a period before the fault (e.g., 7 days) and during the fault. During this stage, the Spearman correlation coefficient between the temperature parameter and the same operating parameter is recalculated; this is the second correlation coefficient. Because internal friction intensifies and lubrication efficiency decreases during gearbox performance degradation, the coupling relationship between some temperature variables and their driving factors may change. For example, oil temperature, which normally rises gradually with load, may show a steep increase under fault conditions, leading to a significant increase or decrease in its correlation with active power. By comparing the difference in the correlation of the same parameter under the two conditions, characteristic variables sensitive to faults can be identified.
[0051] In one specific implementation, the 10-minute average active power can be used as the independent variable to analyze various temperature monitoring indicators (high-speed shaft inlet temperature) under different operating conditions. High-speed shaft rear temperature Oil temperature Inlet lubricating oil temperature The distribution characteristics were analyzed. 30,000 random data points within an ambient temperature range of -10℃ to 10℃, excluding shutdown conditions, were selected for analysis. The results showed that at low power levels, the temperature distribution under different operating conditions was dispersed, gradually concentrating as power increased; under fault conditions, only a few indicators reached the emergency shutdown or alarm threshold; under normal conditions, the inlet oil temperature was negatively correlated with power, while under fault conditions, it was positively correlated, and the temperature difference became more significant with increasing power. This demonstrates that high power levels can more effectively reflect the gearbox's operating status, providing a basis for subsequent data selection.
[0052] Step 3.3: If the absolute difference between the first correlation coefficient and the second correlation coefficient corresponding to a certain temperature monitoring parameter exceeds the preset difference threshold, and both the first correlation coefficient and the second correlation coefficient are in the medium intensity range, then the temperature monitoring parameter is determined as the target monitoring variable.
[0053] If the absolute difference exceeds a preset difference threshold, it indicates that the correlation characteristics of the temperature parameter have changed significantly under different operating conditions, demonstrating its ability to effectively distinguish between healthy and abnormal modes. This threshold (e.g., 0.2) is set to eliminate misjudgments caused by minor fluctuations. The moderate intensity range can be defined as (0.2, 0.8], which eliminates noise interference that may be caused by weak correlation (|r|≤0.2) and avoids redundant binding problems caused by strong correlation (|r|>0.8). Temperature monitoring parameters that meet these dual conditions are determined to be target monitoring variables with state discrimination capabilities.
[0054] The above method, through dynamic correlation analysis of temperature monitoring parameters under different operating conditions, identifies the changing characteristics of their correlation with non-temperature parameters, effectively selecting monitoring indicators sensitive to fault response as modeling inputs. This process enhances the physical meaning of feature selection, reduces the interference of irrelevant variables on the prediction model, and improves the discriminative ability and robustness of the condition monitoring method.
[0055] To process the time series corresponding to the aforementioned target monitoring variables, the method further includes: constructing a target time series prediction model, wherein constructing the target time series prediction model includes steps 4.1 and 4.2: Step 4.1: Divide the time series of the target monitoring variable into several subsequences to maximize the statistical distribution difference between adjacent subsequences. Determine the distribution distance by the maximum mean difference and solve the optimal segmentation strategy by dynamic programming to obtain multiple subsequences with significantly different distribution characteristics.
[0056] The target time series prediction model is a core algorithmic framework used to learn the future trends of target monitoring variables. A time series refers to the set of target monitoring variable observations collected by the SCADA system and arranged in chronological order. To improve the model's adaptability to complex operating conditions, time distribution representation (TDC) processing is performed before training.
[0057] When training the model, the original time series is first divided into multiple continuous subsequences. The goal of this division is to maximize the statistical distribution difference between any two adjacent subsequences. The maximum mean difference (MMD) is used as a metric to measure the distribution distance; a larger value indicates that the two data segments are statistically dissimilar. Dynamic programming is used to solve this extremum problem, finding a set of optimal split points while satisfying the minimum / maximum subsequence length constraints. This yields a set of subsequences with representative differences in power levels, load variations, etc. This step aims to expose the distribution drift phenomenon in the data, laying the foundation for subsequent learning of invariant features.
[0058] Step 4.2: Perform model pre-training on each subsequence, minimize prediction error loss, and initialize the parameters of the target time series prediction model.
[0059] The pre-training phase involves independent initial learning for each segmented subsequence. The goal is to minimize the prediction error loss between the model output and the actual measurement, using mean squared error as the loss function. By optimizing the network weights on data segments with different distributions, the model initially grasps the local prediction patterns under various conditions. This process initializes the parameters of the target time-series prediction model, avoiding convergence difficulties caused by unreasonable initial weights during direct end-to-end training.
[0060] The above approach achieves explicit modeling of multi-condition data by optimally segmenting the time series according to distributional differences and pre-training on each subsequence. This strategy helps reveal the distributional shift characteristics in the data, improves the model's initial adaptability to complex operating conditions, and provides a reliable starting point for building state prediction models with high generalization performance.
[0061] Furthermore, the above-mentioned construction of the target time series prediction model also includes the following steps 5.1 to 5.4: Step 5.1: Calculate the probability distribution differences between different subsequences on the hidden layer state sequence of the recurrent neural network, and determine the degree of influence of each subsequence on the distribution alignment process based on the probability distribution differences.
[0062] Recurrent Neural Networks (RNNs) serve as the foundational architecture for target time series prediction models, capturing long-term dependencies in time series. Deep feature distributions of each subsequence can be extracted from the state sequences output by their hidden layers. By comparing the differences between these distributions (still using the maximum mean difference metric), the degree of deviation of different subsequences in the feature space is assessed. The degree of influence is determined by the magnitude of this distribution difference: the greater the difference, the more difficult it is for the subsequence to be uniformly represented by existing models, and the greater the challenge it poses to the overall distribution alignment task.
[0063] Step 5.2: Dynamically adjust the probability distribution importance weights of the corresponding subsequences based on the changes in distribution differences during consecutive training rounds.
[0064] The importance weights of the probability distribution are used to adjust the contribution ratio of each subsequence in the distribution alignment process. If the distribution distance of a certain subsequence in the (n+1)th round of training does not decrease significantly or even increases compared with the nth round, it is considered that it has not been fully learned, and its weight should be increased. The weight update mechanism draws on the idea of boosting learning, automatically adjusting according to the changes in the distribution gap between previous and subsequent rounds, ensuring that the model gradually focuses on the most difficult data segments to align, thereby achieving an adaptive learning process.
[0065] Step 5.3: Apply the importance weights of the probability distribution to the distribution difference loss term, and then perform a weighted fusion of the distribution difference loss term and the prediction error loss term to generate a comprehensive loss function.
[0066] The distribution difference loss term represents the average distance between the feature distributions of each subsequence, multiplied by the importance weight of its corresponding probability distribution to form the weighted distribution loss. The prediction error loss term reflects the model's fitting accuracy on each subsequence. The two are weighted and summed according to a certain balance coefficient to form the comprehensive loss function.
[0067] Step 5.4: Based on the comprehensive loss function, the gradient of the parameters is calculated layer by layer and the model weights are updated through the gradient backpropagation algorithm to complete the training of the target time series prediction model.
[0068] Error signals are propagated along the network layers using the backpropagation algorithm. The gradient directions of each connection weight from the hidden layer to the input layer are calculated sequentially, and the parameters are updated using an optimizer (such as Adam). After multiple iterations, the model gradually converges to an optimal solution that balances prediction accuracy and robustness, thus completing the end-to-end training of the target time series prediction model.
[0069] To ensure the quality of training data, historical SCADA monitoring data was cleaned using the mapping relationship between wind speed and power and statistical distribution methods: ① Data from outage states, null values, and abnormal data caused by sensor malfunctions were removed; ② Data with a 10-minute average active power less than zero was removed; ③ Based on the normal correlation between 10-minute average wind speed and average active power, abnormal data deviating from this correlation due to wind shear, turbulence, etc., were removed. Taking 3000 SCADA monitoring data points from WTG 13 of a wind farm as an example, the cleaned data points better conform to the normal distribution of "wind speed-power," effectively avoiding the impact of abnormal data on model accuracy and generalization ability.
[0070] The above approach introduces a time distribution matching (TDM) mechanism to dynamically align the distribution differences of different subsequences at the hidden layer level. Combined with adaptive adjustment of importance weights and a composite loss function design, this significantly enhances the model's adaptability to power fluctuations and changes in operating status, improving its stability and reliability in complex field environments.
[0071] This embodiment provides a specific model training method. The target time series prediction model adopts the AdaRNN network structure, which includes two core modules: Temporal Distribution Representation (TDC) and Temporal Distribution Matching (TDM). The TDC module quantifies the continuous distribution of SCADA time series data, dividing the data into k sequences with the largest differences. The TDM module dynamically minimizes the distribution differences between different sequences based on the RNN structure, learns time series features that are invariant to changes in operating conditions and power levels, and improves the model's generalization ability.
[0072] First, temporal distribution representation (TDC) and pre-training are performed, which includes two processes: time series data segmentation pre-training.
[0073] When performing time-series data segmentation, the time-series data of SCADA monitoring indicators are represented as follows: , where each subsequence Total timing length .when When representing the subsequence distribution, temporal distribution drift exists. By solving the following extremum problem, the time series data can be divided into k most dissimilar subsequences with lengths satisfying [the following condition is met]. Subsequences (with preset parameters):
[0074] in, Maximum mean difference (MMD) is used to measure sequence similarity. The upper limit of the number of subsequences is set, and the optimization problem is solved by dynamic programming algorithm to ensure that the statistical characteristics of each subsequence are significantly different, thereby improving the model's adaptability to different data distributions.
[0075] During pre-training, a transfer learning model is built based on the k segmented subsequences. The pre-training objective is to minimize the prediction loss, and the loss function is defined as:
[0076] in, For subsequence The i-th sample pair, For loss function, For the model's learnable parameters, These are the model's predicted values. During pre-training, the RNN hidden layer states and outputs are calculated as follows:
[0077]
[0078] in, The hidden inter-layer weight matrix, The input layer to hidden layer weight matrix The input vector at time step t, This is the weight matrix from the hidden layer to the output layer. These are the bias terms for the hidden layer and the output layer, respectively. This is the activation function.
[0079] The gradient is calculated and the parameters are updated through backpropagation of the error. The formulas for gradient calculation and parameter update are as follows:
[0080]
[0081]
[0082]
[0083]
[0084] in, Let t be the error term. For the loss function with respect to intermediate variables The derivative of L, where L is the loss function, and k is a time interval from 1 to t. This is the learning rate.
[0085] After pre-training, the Temporal Distribution Matching (TDM) module extracts transferable features and minimizes distribution differences by aligning the probability distributions of different subsequences. For subsequence pairs... TDM loss is defined as:
[0086] in, Subsequence at time t The importance weights of the distribution, where V is the time step. (For RNN state transition functions) and These represent the hidden layer states at time t.
[0087] The total loss function of AdaRNN is a weighted sum of the prediction loss and the TDM loss:
[0088] in, This is the balance coefficient.
[0089] Dynamic updates based on the boosting algorithm If the distribution distance of the (n+1)th learning iteration is... Greater than the nth time Then enlarge The formula is as follows:
[0090] in, This refers to the sigmoid function. The final result is obtained through multiple iterations. The values are used in the regression prediction model of various monitoring indicators of WTG gearbox.
[0091] After obtaining the trained target time-series prediction model, predictions are made using this model during actual prediction. The operating status of the wind turbine gearbox is monitored based on the deviation between the predicted value and the actual measured value of the corresponding target monitoring variable, which may include the following steps 6.1 to 6.5: Step 6.1: Calculate the residual sequence between the predicted value and the actual measured value of the target monitoring variable within the continuous time window.
[0092] The residual sequence refers to the time series consisting of point-by-point differences between the predicted values output by the model and the measured values by the sensor. For each target monitoring variable, this deviation is continuously calculated within a sliding time window (e.g., a continuous 24 hours), generating a dynamically updated residual stream. The magnitude of the residual is used to characterize the degree of fit of the model to the current operating state of the variable and to determine whether there are any abnormal disturbances.
[0093] Step 6.2: Perform time-series trend analysis on the residual series of the same monitoring variable to identify any abnormal patterns of continuous increase or periodic fluctuation.
[0094] When conducting trend analysis, slope detection, moving average filtering, or abrupt change identification can be used to determine whether the residual sequence exhibits non-stationary characteristics such as monotonically increasing, intensified oscillations, or repetitive pulses. A continuous increase in residuals may indicate that the component is undergoing a slow degradation process; periodic fluctuations at a fixed frequency may be related to mechanical resonance or control timing mismatch. Identifying abnormal patterns facilitates the detection of early signs of failure.
[0095] Step 6.3: Perform intervariate correlation analysis on the residual sequences of multiple target monitoring variables that exceed the preset threshold range, and calculate the consistency index of the changes of each residual sequence in the time dimension.
[0096] When the residual of a single variable exceeds the limit, there may be a risk of false alarms. To address this, a multivariate collaborative judgment mechanism is introduced: when multiple target monitoring variables simultaneously show significant residual deviations, further analysis is conducted to determine whether the rhythm of their residual changes is synchronized.
[0097] When calculating the consistency index of changes in each residual sequence over time, methods such as cross-correlation coefficient, dynamic time regularization distance, or covariance index can be used to quantify the synchronicity of the evolution of multiple residual sequences on the time axis. High consistency indicates that the system as a whole deviates from the normal behavioral path.
[0098] Step 6.4: If the consistency index of change exceeds the set criteria, it is determined that there is an early failure or compound failure risk in the wind turbine gearbox.
[0099] The established criteria can be derived from historical fault samples to distinguish between accidental disturbances and actual faults. If the consistency index exceeds this threshold, it indicates that multiple key temperature variables are evolving abnormally and synchronously, which may indicate potential problems such as gear wear, bearing damage, or lubrication failure, thus determining the risk of early faults or multiple faults overlapping.
[0100] Step 6.5: Output the corresponding operating status identifier and alarm level information.
[0101] Based on the judgment results, structured outputs are generated, such as operational status indicators like health, warnings, and alarms. Alarm levels are classified according to severity, such as Level I (Attention), Level II (Concern), and Level III (Urgent), and pushed to the operation and maintenance platform to assist in making maintenance decisions.
[0102] The above method, by constructing residual sequences and conducting trend analysis and multivariate correlation judgment, achieves a progressive diagnosis from single-point anomalies to systemic degradation. This mechanism significantly reduces the false alarm rate and false negative rate, improves the timeliness and accuracy of fault identification, and provides strong support for predictive maintenance of wind turbine gearboxes.
[0103] In summary, this application ensures the integrity and timeliness of input data by systematically collecting historical operational data and removing invalid samples; it enhances data representativeness by constructing ideal mapping curves to screen typical operating condition data; it strengthens feature discrimination by optimizing sensitive monitoring variables based on dynamic changes in correlation; it effectively alleviates the generalization problem caused by operating condition drift by training an adaptive prediction model using time distribution representation and matching mechanisms; and finally, it achieves state identification by combining residual trend and multivariate consistency joint criteria, significantly improving the detection capability of early and complex faults. The overall solution, while ensuring high-precision prediction, significantly improves the robustness, practicality, and engineering applicability of wind turbine gearbox condition monitoring.
[0104] Based on the above method embodiments, this application also provides a wind turbine gearbox condition monitoring system, see [link to relevant documentation]. Figure 2 As shown, the system includes the following components: The data acquisition module 210 is used to acquire the operating data of the gearbox of the wind turbine to be monitored. The operating data includes wind speed, active power and multiple temperature monitoring parameters. The data filtering and variable determination module 220 is used to filter the operating data according to the mapping relationship between wind speed and active power, retain the target data segment that meets the preset correlation mode, and determine the monitoring indicators in the target data segment that meet the preset correlation conditions as target monitoring variables based on the correlation changes of each temperature monitoring parameter and non-temperature monitoring parameters under different operating conditions. The time series prediction module 230 is used to process the time series corresponding to the target monitoring variables through the target time series prediction model to determine the predicted values of each monitoring indicator; The status monitoring module 240 is used to monitor the operating status of the wind turbine gearbox based on the deviation between the predicted value and the actual measured value of the corresponding target monitoring variable.
[0105] In one feasible implementation, the data acquisition module 210 is specifically used for: Collect historical operating data of the gearbox of the wind turbine to be monitored. The historical operating data includes wind speed, active power, and measured values of high-speed shaft front temperature, high-speed shaft rear temperature, oil temperature and inlet lubricating oil temperature at multiple time points. Data records in the shutdown state are removed, and abnormal values caused by null values or sensor malfunctions are eliminated to generate preliminary cleaned operating data.
[0106] In one feasible implementation, the data filtering and variable determination module 220 described above is specifically used for: An ideal mapping curve is constructed based on a preset correlation pattern between wind speed and active power under normal operating conditions. The correspondence between wind speed and active power in the operational data is compared with the ideal mapping curve; Operational data whose actual active power deviates significantly from the expected power corresponding to the ideal mapping curve under wind speed are identified as abnormal data and removed, while target data segments that conform to the correlation pattern are retained.
[0107] In one feasible implementation, the data filtering and variable determination module 220 described above is specifically used for: Calculate the first correlation coefficient between each temperature monitoring parameter and non-temperature monitoring parameters under normal operating conditions; Calculate the second correlation coefficient between the same temperature monitoring parameter and non-temperature monitoring parameters before and after the fault occurs; If the absolute difference between the first correlation coefficient and the second correlation coefficient corresponding to a certain temperature monitoring parameter exceeds a preset difference threshold, and both the first correlation coefficient and the second correlation coefficient are in the medium intensity range, then the temperature monitoring parameter is determined as the target monitoring variable.
[0108] In one feasible implementation, see Figure 3 As shown, the system also includes a model training module 310, used for: Construct a target time series prediction model to process the time series of target monitoring variables; Specifically, the model training module 310 mentioned above is used for: The time series of the target monitoring variable is divided into several subsequences to maximize the statistical distribution difference between adjacent subsequences. The distribution distance is determined by the maximum mean difference, and the optimal segmentation strategy is solved by dynamic programming to obtain multiple subsequences with significantly different distribution characteristics. Model pre-training is performed on each subsequence to minimize prediction error loss and initialize the parameters of the target time series prediction model.
[0109] In one feasible implementation, the model training module 310 is further configured to: On the hidden layer state sequence of a recurrent neural network, the probability distribution difference between different subsequences is calculated, and the influence of each subsequence on the distribution alignment process is determined based on the probability distribution difference. The importance weights of the probability distribution of the corresponding subsequences are dynamically adjusted based on the changes in the distribution differences in consecutive training rounds. The importance weights of the probability distribution are applied to the distribution difference loss term, and the distribution difference loss term is weighted and fused with the prediction error loss term to generate a comprehensive loss function; Based on the comprehensive loss function, the gradient of the parameters is calculated layer by layer and the model weights are updated through the gradient backpropagation algorithm to complete the training of the target time series prediction model.
[0110] In one feasible implementation, the status monitoring module 240 is used for: Calculate the residual sequence between the predicted value and the actual measured value of the target monitoring variable within a continuous time window; Perform time-series trend analysis on the residual series of the same monitored variable to identify any abnormal patterns of continuous increase or periodic fluctuation. Perform intervariate correlation analysis on the residual sequences of multiple target monitoring variables that exceed the preset threshold range, and calculate the consistency index of the changes of each residual sequence in the time dimension; If the consistency index of the change exceeds the set criteria, it is determined that there is a risk of early failure or compound failure in the wind turbine gearbox. Output the corresponding operating status identifier and alarm level information.
[0111] The wind turbine gearbox condition monitoring system provided in this application has the same implementation principle and technical effects as the aforementioned method embodiments. For the sake of brevity, any parts not mentioned in the embodiments of the wind turbine gearbox condition monitoring system can be referred to the corresponding content in the aforementioned wind turbine gearbox condition monitoring method embodiments.
[0112] This application also provides an electronic device, such as... Figure 4 The diagram shows the structure of the electronic device 100, which includes a processor 41 and a memory 40. The memory 40 stores computer-executable instructions that can be executed by the processor 41. The processor 41 executes the computer-executable instructions to implement any of the above-mentioned wind turbine gearbox status monitoring methods.
[0113] exist Figure 4 In the illustrated embodiment, the electronic device further includes a bus 42 and a communication interface 43, wherein the processor 41, the communication interface 43, and the memory 40 are connected via the bus 42.
[0114] The memory 40 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 43 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 42 may be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus 42 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 4 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.
[0115] Processor 41 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 41 or by software instructions. Processor 41 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in the memory. The processor 41 reads the information in the memory and, in conjunction with its hardware, completes the steps of the wind turbine gearbox status monitoring method of the aforementioned embodiment.
[0116] This application also provides a computer-readable storage medium storing computer-executable instructions. When these computer-executable instructions are called and executed by a processor, they cause the processor to implement the aforementioned wind turbine gearbox status monitoring method. For specific implementation details, please refer to the foregoing method embodiments, which will not be repeated here.
[0117] The computer program product of the wind turbine gearbox condition monitoring method, system and electronic device provided in the embodiments of this application includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods in the preceding method embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here.
[0118] Unless otherwise specifically stated, the relative steps, numerical expressions, and values of the components and steps described in these embodiments do not limit the scope of this application.
[0119] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0120] In the description of this application, it should be noted that the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.
[0121] In the description of this application, it should also be noted that, unless otherwise expressly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.
[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such 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 this application.
Claims
1. A method for monitoring the condition of a wind turbine gearbox, characterized in that, include: The operating data of the gearbox of the wind turbine to be monitored is obtained, including wind speed, active power and multiple temperature monitoring parameters; Based on the mapping relationship between wind speed and active power, the operating data is filtered to retain target data segments that meet the preset correlation pattern. Based on the correlation changes between each temperature monitoring parameter and non-temperature monitoring parameters under different operating conditions, the monitoring indicators in the target data segments that meet the preset correlation conditions are determined as target monitoring variables. The time series corresponding to the target monitoring variables are processed by the target time series prediction model to determine the predicted values of each monitoring indicator; The operating status of the wind turbine gearbox is monitored based on the deviation between the predicted value and the actual measured value of the corresponding target monitoring variable.
2. The method for monitoring the condition of a wind turbine gearbox according to claim 1, characterized in that, Obtain the operating data of the gearbox of the wind turbine to be monitored, including: Collect historical operating data of the gearbox of the wind turbine to be monitored. The historical operating data includes the wind speed, active power, and measured values of high-speed shaft front temperature, high-speed shaft rear temperature, oil temperature and inlet lubricating oil temperature at multiple time points. Data records in the shutdown state are removed, and abnormal values caused by null values or sensor malfunctions are eliminated to generate preliminary cleaned operating data.
3. The method for monitoring the condition of a wind turbine gearbox according to claim 2, characterized in that, Based on the mapping relationship between wind speed and active power, the operating data is filtered to retain target data segments that conform to a preset association pattern, including: An ideal mapping curve is constructed based on a preset correlation pattern between wind speed and active power under normal operating conditions. The correspondence between wind speed and active power in the operational data is compared with the ideal mapping curve; Operating data whose actual active power deviates significantly from the expected power corresponding to the ideal mapping curve at the wind speed is identified as abnormal data and removed, while target data segments that conform to the correlation pattern are retained.
4. The method for monitoring the condition of a wind turbine gearbox according to claim 3, characterized in that, Based on the changes in the correlation between various temperature monitoring parameters and non-temperature monitoring parameters under different operating conditions, monitoring indicators in the target data segment that meet preset correlation conditions are determined as target monitoring variables, including: Calculate the first correlation coefficient between each temperature monitoring parameter and non-temperature monitoring parameters under normal operating conditions; Calculate the second correlation coefficient between the same temperature monitoring parameter and non-temperature monitoring parameters before and after the fault occurs; If the absolute difference between the first correlation coefficient and the second correlation coefficient corresponding to a certain temperature monitoring parameter exceeds a preset difference threshold, and both the first correlation coefficient and the second correlation coefficient are in the medium intensity range, then the temperature monitoring parameter is determined as the target monitoring variable.
5. The method for monitoring the condition of a wind turbine gearbox according to claim 1, characterized in that, The method further includes: Construct a target time series prediction model to process the time series of target monitoring variables; The construction of the target time series prediction model includes: The time series of the target monitoring variable is divided into several subsequences to maximize the statistical distribution difference between adjacent subsequences. The distribution distance is determined by the maximum mean difference, and the optimal segmentation strategy is solved by dynamic programming to obtain multiple subsequences with significantly different distribution characteristics. Model pre-training is performed on each subsequence to minimize prediction error loss and initialize the parameters of the target time series prediction model.
6. The method for monitoring the condition of a wind turbine gearbox according to claim 5, characterized in that, Building a target time series prediction model also includes: On the hidden layer state sequence of a recurrent neural network, the probability distribution difference between different sub-sequences is calculated, and the influence of each sub-sequence on the distribution alignment process is determined based on the probability distribution difference. The importance weights of the probability distribution of the corresponding subsequences are dynamically adjusted based on the changes in the distribution differences in consecutive training rounds. The importance weights of the probability distribution are applied to the distribution difference loss term, and the distribution difference loss term is weighted and fused with the prediction error loss term to generate a comprehensive loss function; Based on the comprehensive loss function, the gradient of the parameters is calculated layer by layer and the model weights are updated through the gradient backpropagation algorithm to complete the training of the target time series prediction model.
7. The method for monitoring the condition of a wind turbine gearbox according to claim 1, characterized in that, The operating status of the wind turbine gearbox is monitored based on the deviation between the predicted value and the actual measured value of the corresponding target monitoring variable, including: Calculate the residual sequence between the predicted value and the actual measured value of the target monitoring variable within a continuous time window; Perform time-series trend analysis on the residual series of the same monitored variable to identify any abnormal patterns of continuous increase or periodic fluctuation. Perform intervariate correlation analysis on the residual sequences of multiple target monitoring variables that exceed the preset threshold range, and calculate the consistency index of the changes of each residual sequence in the time dimension; If the consistency index of the change exceeds the set criteria, it is determined that the wind turbine gearbox has an early failure or compound failure risk. Output the corresponding operating status identifier and alarm level information.
8. A wind turbine gearbox condition monitoring system, characterized in that, include: The data acquisition module is used to acquire the operating data of the gearbox of the wind turbine to be monitored. The operating data includes wind speed, active power and multiple temperature monitoring parameters. The data filtering and variable determination module is used to filter the operating data according to the mapping relationship between wind speed and active power, retain the target data segment that meets the preset correlation pattern, and determine the monitoring indicators in the target data segment that meet the preset correlation conditions as target monitoring variables based on the correlation changes of each temperature monitoring parameter and non-temperature monitoring parameters under different operating states. The time series prediction module is used to process the time series corresponding to the target monitoring variable through the target time series prediction model to determine the predicted values of each monitoring indicator; The status monitoring module is used to monitor the operating status of the wind turbine gearbox based on the deviation between the predicted value and the actual measured value of the corresponding target monitoring variable.
9. An electronic device, characterized in that, The method includes a processor and a memory, the memory storing computer-executable instructions that can be executed by the processor, the processor executing the computer-executable instructions to implement the wind turbine gearbox condition monitoring method according to any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the wind turbine gearbox condition monitoring method according to any one of claims 1 to 7.