Gearbox Fault Early Warning System Based on Collaborative Reasoning Network for Wind Farm Cluster Operation and Maintenance

Abstract

This invention discloses a gearbox fault early warning system based on a collaborative reasoning network for wind farm cluster operation and maintenance, relating to wind turbine early warning technology. It addresses the problem that existing technologies are unsuitable for offshore environments by proposing this solution. A data preprocessing module converts oil monitoring data from gearboxes in several offshore wind farms into status indicators; a multi-agent cluster learning module captures complex fault patterns from the data from different perspectives through parallel learning, forming diverse prediction perspectives; a collaborative reasoning module integrates the prediction results generated by the multi-agent cluster through a meta-learning mechanism; and an early fault warning decision module transforms the collaborative reasoning results into specific early warning decisions, providing probability confidence levels and tiered early warning suggestions. The advantage lies in its ability to capture complex fault patterns from different perspectives through parallel learning, effectively integrating various prediction results generated by the multi-agent cluster, which has significant engineering implications for ensuring the safe and reliable operation of wind farms.

Description

Technical Field

[0001] This invention relates to wind turbine early warning technology, and more particularly to a gearbox fault early warning system based on a wind farm cluster operation and maintenance collaborative reasoning network. Background Technology

[0002] As a key component of wind turbines, gearboxes operate in harsh marine environments characterized by humidity, high temperatures, corrosion, and fluctuating wind speeds. These harsh marine environmental conditions lead to frequent failures and maintenance difficulties, posing a significant challenge to the safety of offshore wind turbines [1]-[2]. Oil monitoring is a method for analyzing wear particles and the changing trends of lubricating oil viscosity, and is commonly used to observe the wear inside the gearbox, thereby reflecting potential early failures. However, existing oil monitoring methods usually rely on thresholds obtained through statistical methods to analyze each gearbox in a wind farm individually. They lack collaborative analysis across different offshore wind farms and between different wind turbines within the same offshore wind farm, resulting in insufficient accuracy and generalization performance in early failure warning.

[0003] Various data-driven artificial intelligence methods, such as decision tree (DT), support vector machine (SVM), convolutional neural network (CNN), and generative adversarial network (GAN), have been applied to fault warning of key equipment such as wind turbine gearboxes [3]-[7]. For example, Yang et al. [3] enhanced the abnormal detection performance of gearboxes by integrating multi-scale decomposition and convolutional neural network. Yang et al. [8] constructed an inverted Transform model and a threshold-based fault alarm strategy, and used gearbox oil monitoring samples to realize abnormal monitoring and early warning of key feature indicators. Li et al. [9] established a Weibull linear model to analyze the performance of wind turbine units, drew the power band of reliable operation, and used a BP neural network prediction model to predict the power of wind turbines. By comparing the predicted power with the normal operating power, they decided whether to issue an early warning. Gu et al.

[10] proposed an improved Siamese neural network based on feature fusion for fault warning and key component identification of wind turbine units. This method combines fault identification, classification and labeling and ensemble learning algorithm to improve the traditional Siamese neural network. Li et al.

[11] proposed a wind turbine gearbox fault early warning method based on mutual information recurrent network. This method extracts highly correlated parameters as input features of a recurrent neural network with memory capability, and uses the early warning threshold to realize the early warning of the gearbox fault status of wind turbine units. However, the above-mentioned existing fault early warning methods mainly focus on individual mechanical equipment and fail to use the difference information between different equipment for collaborative learning, which often leads to insufficient generalization performance of the early warning system.

[0004] Recently, multi-agent systems have utilized several specialized agents to collaborate in performing various analysis and reasoning tasks, showing significant advantages in areas such as social networks, traffic scheduling, and industrial inspection. For example, Zhu et al.

[12] used an expert system to perform fault diagnosis and fault warning for wind turbines. When analyzing potential faults, the proximity of the faulty wind turbine is first calculated. If the proximity is below a certain threshold, the same type of fault can be identified. Wu et al.

[13] proposed a collaborative strategy framework based on relational graph reasoning for multi-agent systems to complete adversarial tasks. Chi et al.

[14] introduced multi-agent collaborative decision-making based on visual language models for autonomous driving, balancing computational efficiency with robust reasoning and perception. Zhao et al.

[15] proposed a multi-agent collaborative reasoning framework for Human Activity Recognition (HAR) based on Large Language Models (LMMs), which not only improved the recognition accuracy but also gave the HAR task stronger interpretability. However, due to the lack of domain knowledge enhancement and the synergistic fusion of the base learner performance of multiple agents, the above methods are difficult to directly apply to the early fault warning of complex gearboxes in several offshore wind farms.

[0005] [1] Li J, He D, "Strategies of Sustainable Development in China's Wind Power Industry," 2020.

[0006] [2] MS Bridgman, “Relating failure prognostics to systembenefits,” Proceedings, IEEE Aerospace Conference, Big Sky, MT, USA, 2002, pp. 7–7.

[0007] [3] G. Yang, Y. Zhong, L. Yang, H. Tao, J. Li and R. Du, "FaultDiagnosis of Harmonic Drive With Imbalanced Data Using Generative AdversarialNetwork," in IEEE Transactions on Instrumentation and Measurement, vol. 70,pp. 1-11, 2021, Art no. 3519911, doi: 10.1109 / TIM.2021.3089240.

[0008] [4] G. Yang, Y. Zhong, L. Yang and R. Du, "Fault Detection ofHarmonic Drive Using Multiscale Convolutional Neural Network," in IEEETransactions on Instrumentation and Measurement, vol. 70, pp. 1-11, 2020, Artno. 3502411, doi: 10.1109 / TIM.2020.3024355.

[0009] [5] G. Yang, H. Tao, R. Du and Y. Zhong, "Compound Fault Diagnosis ofHarmonic Drives Using Deep Capsule Graph Convolutional Network," in IEEETransactions on Industrial Electronics, doi: 10.1109 / TIE.2022.3176280.

[0010] [6] G. Yang, H. Tao, K. Wu, R. Du and Y. Zhong, "Fault Diagnosis ofHarmonic Drives Using Multimodal Collaborative Meta Network With SeverelyMissing Modality," in IEEE Transactions on Industrial Informatics, doi:10.1109 / TII.2024.3396339.

[0011] [7] G. Yang, H. Tao, T. Yu, R. Du and Y. Zhong, "Online FaultDiagnosis of Harmonic Drives Using Semisupervised Contrastive GraphGenerative Network via Multimodal Data," in IEEE Transactions on IndustrialElectronics, vol. 71, no. 3, pp. 3055-3063, March 2024.

[0012] [8] G. Yang, H. Tao, S. He, W. Feng, R. Du and Y. Zhong, "MultimodalTime Series Forecasting for Online Oil Monitoring of Petrochemical PelletizerGearbox Using Multiscale Inverted Transform Network," in IEEE Internet ofThings Journal, doi: 10.1109 / JIOT.2024.3514081.

[0013] [9] Y. Li, Z. Ma, J. Feng, R. Zhang, N. Fu and D. Liang, "FaultWarning and Reliability Analysis of Wind Turbine Failure Based on Data-driven," 2024 8th International Conference on Green Energy and Applications(ICGEA), Singapore, Singapore, 2024, pp. 28-32.

[0014]

[10] C. Gu and S. Yin, "Improved Siamese Neural Betwork Based onFeature Fusion for Wind Turbine Fault Warning and Identification of KeyComponents," 2024 IEEE PES 16th Asia-Pacific Power and Energy EngineeringConference (APPEEC), Nanjing, China, 2024, pp. 1-5.

[0015]

[11] X. Li, Z. Cheng and Z. Xie, "Gearbox Fault Early Warning forWind Turbines Based on Mutual Information Recurrent Networks," 2025 2ndInternational Conference on Smart Grid and Artificial Intelligence (SGAI),Changsha, China, 2025, pp. 1067-1070.

[0016]

[12] Z. Guangwei, C. Sifan, R. Na and F. Shaonan, "Fault diagnosisand warning design of wind turbines based on expert system," 2021 IEEE 4thInternational Conference on Automation, Electronics and ElectricalEngineering (AUTEEE), Shenyang, China, 2021, pp. 755-758.

[0017]

[13] S. Wu, T. Qiu, Z. Pu and J. Yi, "Multi-agent CollaborativeLearning with Relational Graph Reasoning in Adversarial Environments," 2021IEEE / RSJ International Conference on Intelligent Robots and Systems (IROS),Prague, Czech Republic, 2021, pp. 5596-5602.

[0018]

[14] F. Chi, Y. Wang, P. Nasiopoulos and V. C. M. Leung, "Multi-AgentCollaborative Decision-Making Using Small Vision-Language Models forAutonomous Driving," in IEEE Internet of Things Journal, doi: 10.1109 / JIOT.2025.3624038.

[0019]

[15] Y. Zhao, Y. Chen, R. Tang and W. Zhao, "MultiagentsCR: A Multi-Agent Collaborative Reasoning Framework Based on LLM for Human ActivityRecognition," 2025 10th International Conference on Information Science, Computer Technology and Transportation (ISCTT), Nanchong, China, 2025, pp.90-95. Summary of the Invention

[0020] The purpose of this invention is to provide a gearbox fault early warning system based on a wind farm cluster operation and maintenance collaborative reasoning network, so as to solve the problems existing in the prior art.

[0021] The gearbox fault early warning system based on the wind farm cluster operation and maintenance collaborative reasoning network described in this invention includes a data preprocessing module, a multi-agent cluster learning module, a collaborative reasoning module, and an early fault warning decision module.

[0022] The data preprocessing module is used to clean, transform, and construct features for oil monitoring data of gearboxes in several offshore wind farms. Through domain knowledge-enhanced feature engineering, the original monitoring data is transformed into state indicators that comprehensively characterize the equipment degradation process.

[0023] The multi-agent cluster learning module is used to construct a multi-agent cluster composed of several heterogeneous basic learners, and capture complex fault modes in the data from different angles through parallel learning to form a diverse prediction perspective.

[0024] The collaborative reasoning module is used to integrate the prediction results generated by the multi-agent cluster through the meta-learning mechanism to achieve collaborative reasoning and optimal decision path discovery;

[0025] The early fault warning decision module is used to transform the collaborative reasoning results into specific warning decisions, providing probability confidence and graded warning suggestions.

[0026] The gearbox fault early warning system based on a wind farm cluster operation and maintenance collaborative reasoning network described in this invention has the advantage of effectively integrating various prediction results generated by a multi-agent cluster by capturing complex fault modes from data from different perspectives through parallel learning. Through a meta-learning mechanism, collaborative reasoning is performed to arrive at a final consensus, demonstrating superior performance compared to existing technologies. Applying a cluster collaborative reasoning network based on multi-agent collaborative reasoning to early fault warning in wind turbine gearbox oil monitoring technology has significant engineering implications for ensuring the safe and reliable operation of wind farms. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the gearbox fault early warning system described in this invention.

[0028] Figure 2 This is a bar chart showing the percentage of wind turbine units corresponding to three different health levels in five offshore wind farms in the embodiment.

[0029] Figure 3 This is a bar chart showing the proportion of wind turbines that triggered abnormal alarms in five offshore wind farms in the embodiment.

[0030] Figure 4 This is a radar chart of key parameters for five offshore wind farms in the embodiment.

[0031] Figure 5 This is a graph showing the iron content ratio of five offshore wind farms in the embodiment.

[0032] Figure 6 This is a graph showing the oil viscosity ratios for five offshore wind farms in the embodiment.

[0033] Figure 7 This is a bar chart showing the number of offline oil detections for five offshore wind farms in the embodiment.

[0034] Figure 8 This is a point diagram showing the health status of a single gearbox in the wind farm DD in the embodiment.

[0035] Figure 9 This is a point diagram showing the health status of a single gearbox in the wind farm GJ in the embodiment.

[0036] Figure 10 This is a point diagram showing the health status of a single gearbox in the wind farm NH in the embodiment.

[0037] Figure 11 This is a point diagram showing the health status of a single gearbox in the wind farm SH in the embodiment.

[0038] Figure 12 This is a point diagram showing the health status of a single gearbox in the wind farm SZ in the embodiment.

[0039] Figure 13 This is a graph showing the ranking of feature importance in the embodiments.

[0040] Figure 14 This is a graph showing the evolution of the health status of the wind farm DD in the embodiment.

[0041] Figure 15 This is a graph showing the evolution of the health status of wind farm GJ in the embodiment.

[0042] Figure 16 This is a diagram showing the evolution of the health status of NH in the wind farm in the example.

[0043] Figure 17 This is a graph showing the evolution of the health status of wind farm SH in the embodiment.

[0044] Figure 18 This is a diagram showing the evolution of the health status of wind farm SZ in the embodiment. Detailed Implementation

[0045] The gearbox fault early warning system based on the Cluster Cooperative Reasoning Network (CCRN) for wind farm cluster operation and maintenance described in this invention, such as... Figure 1 As shown, it includes a data preprocessing module, a multi-agent cluster learning module, a collaborative reasoning module, and an early fault warning and decision-making module.

[0046] The data preprocessing module primarily cleans, transforms, and performs feature engineering on the oil monitoring data of wind turbine gearboxes, preparing high-quality, high-dimensional input features for subsequent intelligent analysis. First, basic features are constructed by processing time and categorical variables (such as wind farm, gearbox ID, and oil type). Then, advanced features (such as metal wear index and oil oxidation index) are created using domain knowledge (e.g., wear principles and oil degradation trends). Finally, normalization eliminates the influence of dimensions. Its core principle lies in explicitly quantifying implicit physicochemical patterns closely related to equipment health status through feature engineering, thereby transforming simple monitoring data into comprehensive state indicators characterizing the equipment degradation process. This lays a solid foundation for accurate model decision-making.

[0047] The normalization formula is: z = (x - μ) / σ;

[0048] Where x is the original feature value, μ is the mean of the feature in the training set, and σ is its standard deviation.

[0049] In the feature engineering for domain knowledge enhancement, the formula for constructing composite features with domain knowledge enhancement is as follows:

[0050] ;

[0051] ;

[0052] ;

[0053] in, It is the metal wear index; It is the oil oxidation index; It is a comprehensive degradation index; Fe, Cu, Pb, Zn, and Ca are the contents of the corresponding metal elements in the oil, with units of mg / kg; It is the kinematic viscosity at 40°C, in mm² / s; α, β, and γ are weighting coefficients, and satisfy... .

[0054] The standardized formula for handling cross-wind farms is:

[0055] ;

[0056] ;

[0057] ;

[0058] in, It is the standardized value of the j-th feature of the i-th sample; These are the original eigenvalues; It is the global mean of characteristic j across wind farms; It is the global standard deviation of characteristic j across wind farms; It is the mean value of characteristic j of the k-th wind farm; It is the standard deviation of characteristic j of the k-th wind farm; It refers to the number of wind farms.

[0059] The multi-agent swarm learning module functions to construct a multi-agent swarm consisting of several heterogeneous champion-level base learners, capturing complex failure modes from data from different perspectives through parallel learning. The network architecture comprises seven carefully tuned base models: two XGBoost models, two random forest models, two gradient boosting trees, and one extreme random tree model, each differentiated based on hyperparameters and learning bias. In implementation, these models are trained in parallel on the same dataset as independent "agents," subsequently generating their respective probabilistic predictions through 5x hierarchical cross-validation. Its underlying principle aligns with diversity theory in ensemble learning, which posits that combining several high-performance models with uncorrelated errors expands the hypothesis space, covering more potential failure modes. This approach generally surpasses the performance of any single model, resulting in a more powerful and robust collective intelligence.

[0060] The weights of the diversity-driven agents are assigned using the following formula:

[0061] ;

[0062] ;

[0063] in, It is the weight of the m-th agent; It is the diversity score of the m-th agent; It is the correlation coefficient between the prediction results of agents m and m'; It is a diversity adjustment parameter; It represents the total number of intelligent agents.

[0064] The weighted prediction of cross-validation performance is calculated using the following formula:

[0065] ;

[0066] ;

[0067] in, It is the cross-validation average prediction probability of agent m; It is the prediction probability of agent m at the k-th fold; It is the overall performance score of agent m; It is the accuracy of the kth fold; It is the F1 score of the k-th fold; α is the AUC value at the k-th fold; α2, β2, and γ2 are the weights of the performance indicators, respectively.

[0068] The collaborative reasoning module effectively integrates various prediction results generated by a multi-agent cluster through a meta-learning mechanism to arrive at a final consensus. Structurally, the predicted probabilities output by the base model clustering are used as meta-features input into the XGBoost meta-learner for training. The implementation process is as follows: First, the probability outputs generated by the base models during cross-validation are superimposed into a high-dimensional meta-feature matrix, which encapsulates all the uncertainty judgments of the base models regarding the samples. Subsequently, the meta-learner is trained on this matrix to learn how to weigh and integrate this information, and may assign weights to the outputs of the base models based on their performance metrics (such as accuracy) on the validation set. Its basic principle is stacked ensemble; the meta-model is trained to learn the complex mapping relationship between the base model predictions and the true labels, thereby discovering the optimal collaborative decision-making path and achieving a "1+1>2" reasoning effect.

[0069] The calculation formula for adaptive fusion of meta-features is:

[0070] ;

[0071] ;

[0072] MetaFeatures is the fused metafeature vector;

[0073] It is a vector concatenation operation;

[0074] It is a set of categories {Normal, Attention, Alert}

[0075] It is the prediction uncertainty of agent m.

[0076] The formula for collaborative training of meta-learners is:

[0077] ;

[0078] ;

[0079] in, It is the meta-learner loss function; It is cross-entropy loss; It is the meta-learner mapping function; KL is the KL divergence regularization term; It is the regularization coefficient.

[0080] The function of the early fault warning decision module is to transform the probability predictions output by the collaborative inference module into specific operational decisions, thereby completing the closed loop from model output to actionable measures. Structurally, it receives probability vectors P = [p_normal, p_attention, p_alert] generated by the meta-learner for three categories: "Normal", "Attention", and "Alert", and determines the final class label through the argmax decision logic: predicted class = argmax(P). In addition to outputting discrete classification results, this module also provides the probability confidence level of the prediction. When the model identifies "Attention" or "Alert" with high confidence, it triggers the corresponding warning signal to support maintenance personnel in making decisions. Its basic principle is rooted in risk decision theory, which transforms the soft output of machine learning models into executable hierarchical warning actions. This makes it possible to monitor the gearbox status throughout its entire lifecycle, from normal operation to early anomalies and severe alarms, thereby promoting timely intervention and ultimately achieving the core goal of predictive maintenance. To evaluate the accuracy of the warning decisions, the module calculates key performance indicators using the following formula:

[0081] Precision = TP / (TP + FP);

[0082] Recall = TP / (TP + FN);

[0083] F1-score = 2 × (Precision × Recall) / (Precision + Recall);

[0084] TP stands for True Positive, representing the number of samples correctly predicted as alarms by the model. FP stands for False Positive, representing the number of normal or attentional samples incorrectly predicted as alarms. FN stands for False Negative, representing the number of samples with true alarm states missed by the model.

[0085] The formula for confidence-weighted decision fusion is:

[0086] ;

[0087] ;

[0088] in, It is the final fusion prediction probability; These are the confidence weights of agent m; It is a temperature parameter used to control the smoothness of the weight distribution.

[0089] The counting formula for the adaptive hierarchical early warning mechanism is:

[0090] ;

[0091] ;

[0092] ;

[0093] in, This is the normal level threshold; It is the attention level threshold; It is the alarm level threshold; It is a trend risk assessment; It is a severity risk assessment; W is the size of the time window; All of these are risk threshold parameters.

[0094] The calculation formula for multi-dimensional performance evaluation is as follows:

[0095] ;

[0096] ;

[0097] ;

[0098] ;

[0099] Where F1Macro is the macro-average F1 score; AUCROC is the area under the ROC curve.

[0100] The offline oil testing parameters for gearboxes in various offshore wind farms mainly include kinematic viscosity, acid value, water content, and analysis of 21 oil elements reflecting the wear components of gearbox lubricating oil, as detailed in Table I. In this embodiment, to further illustrate and demonstrate the technical effects of the present invention, five actual offshore wind farms (SZ, DD, GJ, HN, SH) are used as examples.

[0101] Table I

[0102]

[0103] The network structure of the CCRN model in this invention is shown in Table II:

[0104] Table II

[0105]

[0106] The base model set includes 7 champion-level algorithms. There are a total of 35 meta-features derived from probability outputs (7 models × 5 classes). The training strategy employs independent parallel training combined with cross-validation.

[0107] Data analysis of 5 offshore wind farms:

[0108] 1) Wind Farm DD: Gearbox IDs: DD 01, DD 02, DD 03, DD 04, DD 05, DD 06, DD 07, ​​DD 08, DD 09, DD 10. Sample distribution: 80 samples in total. Health status: Normal: 11 (13.8%), Caution: 60 (75.0%), Warning: 9 (11.2%). Warning gearboxes: DD 01, DD 03, DD 04, DD 05, DD 06, DD 07, ​​DD 08, DD 09, DD 10.

[0109] 2) Wind Farm GJ: Gearbox IDs: A01, A02, A03, A04, A05, A06, A07, A08, A09, A10, A11, A12, A13, A14, A15. Sample Distribution: 125 samples in total. Health Status: Normal: 79 (63.2%), Caution: 43 (34.4%), Warning: 3 (2.4%), Alert: 3 (2.4%). Gearboxes: A01, A02, A05.

[0110] 3) Wind Farm HN: Gearbox IDs: HN 01, HN 02, HN 03, HN 04, HN 05, HN 06, HN 07, HN 08, HN 09, HN 10. Sample distribution: 86 samples in total. Health status: Normal: 18 (20.9%), Caution: 59 (68.6%), Warning: 9 (10.5%). Warning gearboxes: hn02, hn03, hn04, hn05, hn06, hn08, hn09.

[0111] 4) Wind Farm SH: Gearbox IDs: SH F01, SH F02, SH F03, SH F04, SH F05, SH F06, SHF07, SH F08, SH F09, SH F10, SH F11, SH F12, SH F13, SH F14, SH F15. Sample distribution: 29 samples in total. Health status: Normal: 8 (27.6%), Caution: 17 (58.6%), Warning: 4 (13.8%). Warning gearboxes: SH F05, SHF11, SH F12, SH F13.

[0112] 5) Wind Farm SZ: Gearbox IDs: F01, F02, F03, F04, F05, F06, F07, F08, F09, F10, F11, F12, F13, F14, F15. Sample distribution: 34 samples in total. Health status: Normal: 26 (76.5%), Caution: 8 (23.5%), Warning: 0 (0.0%).

[0113] Integrated visual comparison of 5 offshore wind farms Figures 2 to 7 As shown, a comprehensive statistical comparison of five offshore wind farms was conducted. Figure 2 This indicates the percentage of wind turbines with three different health levels in five different wind farms. Figure 3 This indicates the proportion of wind turbines triggering abnormal alarms in five different wind farms, using medium and high risk level thresholds to assess the overall health of the wind farm. Figure 4 In a radar chart, each axis represents a parameter, with the center point (zero value) representing the average value across all wind farms. Outward extension indicates values ​​above the average, while inward contraction indicates values ​​below the average. For example, the radar chart of the DD wind farm (red curve) exhibits significant characteristics: the iron content (Fe) axis extends considerably, while the acid value (Acid) axis bulges noticeably. Figure 5 The data represents the iron (Fe) content in five different wind farms, and the wear level of the gearbox is assessed using medium wear and severe wear levels. Figure 6This indicates the oil viscosity in five different wind farms, and the degree of deterioration of the gearbox lubricating oil is evaluated based on the ideal range. Figure 7 These figures reflect the number of offline oil level tests conducted at each of the five different wind farms, as well as the number of wind turbines in each wind farm.

[0114] also, Figures 8 to 12 The data on gearboxes for a single wind farm is displayed. Gearboxes requiring urgent attention are marked with warning labels and their current health status is indicated by color coding.

[0115] Random forest feature importance is used to measure the "contribution" of features in a predictive model. Its calculation method is based on the splitting quality of the decision tree, assessing the ability of features to reduce uncertainty. Using random forest feature importance analysis, the top 8 most relevant features are selected from the feature importance ranking to analyze the impact of feature indicators on oil sample test results, such as... Figure 13 As shown. From Figure 13 It can be observed that the kinematic viscosity (40°C, mm² / s) value is 0.1388, indicating that this feature contributes 13.88% of the information in predicting health status. Furthermore, the Pearson correlation coefficient was used to quantify the "strength of the linear relationship" between the top eight most relevant features and health status. This coefficient measures the degree of linear correlation between two variables; see Table III for details.

[0116] Table III

[0117]

[0118] As shown in Table III, kinematic viscosity showed a moderate positive correlation with health status, with a correlation coefficient of 0.491. Combined with... Figure 13 The results showed that the feature importance of kinematic viscosity was 0.1388, indicating that it is the most critical feature in the prediction model.

[0119] To visually demonstrate and compare the differences in the health performance of wind turbines in different wind farms, the evolution of wind farm health status results are presented. Figures 14 to 18 A comprehensive visual comparison of five offshore wind farms is presented, showcasing the time-series results of oil monitoring tests recorded in the historical database for all wind turbines within each farm. (Comparison) Figures 14 to 18 Wind farm SZ exhibited the best overall turbine performance, with the vast majority of its gearboxes remaining in normal condition during historical oil level monitoring tests. In contrast, the other four wind farms all experienced fault alarms triggered by specific turbines. For example, turbine #04 in wind farm HN triggered fault alarms multiple times, intermittently displaying "attention" and "alarm" states. Furthermore, it was shown that even within the same wind farm, there were significant differences in the operating conditions of different turbines.

[0120] To assess the results of different early fault warning methods, this invention employs multiple methods to compare the early fault warning performance of all gearboxes in five wind farms based on oil monitoring data. The results for accuracy, F1-Macro index, and AUC-ROC curves are shown in Table IV.

[0121] Table IV

[0122]

[0123] As shown in Table IV, compared to other methods, the CCRN model proposed in this invention demonstrates superior overall performance in terms of accuracy, F1-Macro, and AUC-ROC. This indicates that its collaborative inference mechanism—integrating the cluster collaboration of multiple base models through cross-validation and performance-based weighted meta-feature generation—achieves excellent classification performance. Furthermore, Table V presents detailed performance metrics of the CCRN model for accuracy, recall, and F1 score in the early fault warning task for all wind turbine gearboxes in five wind farms.

[0124] Table V

[0125]

[0126] As shown in Table V, the CCRN model effectively detected faults in all abnormal wind turbines across five wind farms, achieving an accuracy, recall, and F1 score of 1.00, thus enabling early warning of gearbox faults. This invention also demonstrated effective detection capabilities for wind turbines in both normal and critical states.

[0127] In summary, this invention proposes a clustered collaborative reasoning network (CCRN) for early fault prediction of wind turbine gearboxes in several offshore wind farms. The constructed data preprocessing module, through feature engineering, clearly quantifies implicit patterns closely related to equipment health status, transforming raw data into comprehensive status indicators to fully characterize the equipment degradation process. The designed multi-agent clustering learning module establishes an agent cluster composed of several heterogeneous basic learners, capturing complex fault patterns from different perspectives through parallel learning, thus surpassing the performance of any single model. The collaborative reasoning module employs Stacking ensemble learning, training the meta-model to learn the complex mapping between basic model predictions and true labels, thereby discovering the optimal collaborative decision path and improving inference performance. Experimental results show that CCRN not only outputs discrete classification results but also provides probabilistic confidence for prediction, exhibiting superior overall performance compared to methods such as decision trees and gradient enhancement. This invention has significant engineering implications for ensuring the safe and reliable operation of wind farms.

[0128] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

Claims

1. A gear box fault early warning system based on wind farm cluster operation and maintenance collaborative reasoning network, characterized in that, It includes a data preprocessing module, a multi-agent cluster learning module, a collaborative reasoning module, and an early fault warning and decision-making module; in the feature engineering of domain knowledge enhancement, the formula for constructing composite features with domain knowledge enhancement is: ； ； ； wherein, is a metal wear index; is an oil product oxidation index; is a comprehensive degradation index; Fe, Cu, Pb, Zn, Ca are the contents of corresponding metal elements in the oil, in units of mg / kg; is a kinematic viscosity at 40°C, in units of mm² / s; α1, β1, γ1 are all weight coefficients, and satisfy ; The data preprocessing module is used to clean, transform, and construct features for oil monitoring data of gearboxes in several offshore wind farms. Through domain knowledge-enhanced feature engineering, the original monitoring data is transformed into state indicators that comprehensively characterize the equipment degradation process. The multi-agent cluster learning module is used to construct a multi-agent cluster composed of several heterogeneous basic learners, and capture complex fault modes in the data from different angles through parallel learning to form a diverse prediction perspective. The collaborative reasoning module is used to integrate the prediction results generated by the multi-agent cluster through the meta-learning mechanism to achieve collaborative reasoning and optimal decision path discovery; The early fault warning decision module is used to transform the collaborative reasoning results into specific warning decisions, providing probability confidence and tiered warning suggestions; the counting formula for the adaptive tiered warning mechanism is: ； ； ； in, This is the normal level threshold; It is the attention level threshold; It is the alarm level threshold; It is a trend risk assessment; It is a severity risk assessment; W is the size of the time window; All of these are risk threshold parameters; among them, It represents the final fusion prediction probability; Normal, Attention, and Alert are all categories; It is a temperature parameter; It is a set of categories {Normal, Attention, Alert}.

2. The gearbox fault early warning system based on a wind farm cluster operation and maintenance collaborative reasoning network according to claim 1, characterized in that, The data preprocessing module first constructs basic features by processing time variables and categorical variables; Then, domain knowledge is used to create advanced features; finally, normalization is used to eliminate the influence of dimensions.

3. The gearbox fault early warning system based on a wind farm cluster operation and maintenance collaborative reasoning network according to claim 2, characterized in that, The normalization calculation formula for the data preprocessing module is: z = (x - μ) / σ; Where x is the original feature value, μ is the mean of the original feature value in the training set, and σ is its standard deviation; The standardized formula for handling cross-wind farms is: ； ； ； in, It is the standardized value of the j-th feature of the i-th sample; These are the original eigenvalues; It is the global mean of characteristic j across wind farms; It is the global standard deviation of characteristic j across wind farms; It is the mean value of characteristic j of the k-th wind farm; It is the standard deviation of characteristic j of the k-th wind farm; It refers to the number of wind farms.

4. The gearbox fault early warning system based on a wind farm cluster operation and maintenance collaborative reasoning network according to claim 3, characterized in that, The multi-agent cluster learning module allocates weights to agents driven by diversity, using the following formula: ； ； in, It is the weight of the m-th agent; It is the diversity score of the m-th agent; It is the correlation coefficient between the prediction results of agents m and m'; It is a diversity adjustment parameter; It represents the total number of intelligent agents; The weighted prediction of cross-validation performance is calculated using the following formula: ； ； in, It is the cross-validation average prediction probability of agent m; It is the prediction probability of agent m at the k-th fold; It is the overall performance score of agent m; It is the accuracy of the kth fold; It is the F1 score of the k-th fold; α is the AUC value at the k-th fold; α2, β2, and γ2 are the weights of the performance indicators, respectively.

5. The gearbox fault early warning system based on a wind farm cluster operation and maintenance collaborative reasoning network according to claim 4, characterized in that, The collaborative reasoning module first superimposes the probability outputs generated by the base model during the cross-validation process into a high-dimensional meta-feature matrix, which encapsulates all the uncertainty judgments of the base model on the samples. The meta-learner is then trained on this matrix to learn how to weigh and integrate information, and outputs weights based on the performance metrics of the base model on the validation set.

6. The gearbox fault early warning system based on a wind farm cluster operation and maintenance collaborative reasoning network according to claim 5, characterized in that, In the collaborative reasoning module, the calculation formula for adaptive fusion of meta-features is as follows: ； ； MetaFeatures is the fused metafeature vector; It is a vector concatenation operation; It is a set of categories {Normal, Attention, Alert} It is the prediction uncertainty of agent m; The formula for collaborative training of meta-learners is: ； ； in, It is the meta-learner loss function; It is cross-entropy loss; It is the meta-learner mapping function; KL is the KL divergence regularization term; It is the regularization coefficient.

7. The gearbox fault early warning system based on a wind farm cluster operation and maintenance collaborative reasoning network according to claim 6, characterized in that, In the early fault warning decision module, the formula for calculating key performance indicators is as follows: Precision = TP / (TP + FP); Recall = TP / (TP + FN); F1-score = 2 × (Precision × Recall) / (Precision + Recall); TP stands for True Positive, representing the number of samples correctly predicted as alarms by the model; FP stands for False Positive, representing the number of normal or attentional samples incorrectly predicted as alarms; FN stands for False Negative, representing the number of samples with true alarm states missed by the model. The formula for confidence-weighted decision fusion is: ； ； in, It is the final fusion prediction probability; These are the confidence weights of agent m; It is a temperature parameter used to control the smoothness of the weight distribution; The calculation formula for multi-dimensional performance evaluation is as follows: ； ； ； ； wherein F1 Macro is the macro-averaged F1 score; AUC ROC is the area under the ROC curve.

Images