Fault Early Warning And Identification Of Key Components Of Wind Turbine Based On Data Mining

Wind power generation is a well-established method for producing energy that has gained popularity in various countries over the past few decades.As wind turbines become more advanced and intelligent,the number of sensors and monitoring parameters has increased,leading to fluctuations in daily operation under different conditions.It’s challenging to spot potential hidden faults manually,especially in critical components such as wind turbine generators and gearboxes.Equipment damage can be expensive to fix,and shutdowns can result in significant power loss.Therefore,identifying critical components,predicting equipment health,and warning of faults are crucial challenges that must be addressed in wind power generation.Solving these issues would have a significant impact on reducing the failure rate and power loss of wind turbines,making it a valuable opportunity for both theoretical research and practical applications.This study focuses on the important components of doubly-fed wind turbines and utilizes time series data from the supervisory control and data acquisition system(SCADA)and vibration data from the condition monitor system(CMS).The goal is to develop fault early warning and identification models for critical wind turbine components by integrating SCADA and CMS data through feature fusion techniques.The methods include SCADA-based fault early warning modelling,multi-parameter feature fusion in the timefrequency domain feature fusion integrated learning for fault early warning and identification of essential components,and an improved siamese neural network for fault early warning and identification of crucial components.The ultimate aim is to achieve precise fault warning and identification of gearboxes,generators,spindles,and other critical components through data mining.A real-world dataset from a large wind power company is used to validate and analyze the proposed methods.According to the findings,the warning and identification of critical component faults have been significantly enhanced in terms of accuracy,reliability,and practicality.The specific research details are as follows:(1)Aiming at the data feature selection of critical components of wind turbines,operating conditions,and long-term characteristics of the gradual process of fault degradation,a combined modelling method for fault early warning based on SCADA data is proposed.A new method has been developed that utilizes three algorithms:improved principal component analysis(PCA),comprehensive learning particle swarm optimizer Gaussian mixture model(CLPSO-GMM),and improved attention-based bidirectional long short-term memory neural network(Att-BiLSTM).The enhanced PCA algorithm includes the determination of the Spearman correlation coefficient based on traditional PCA output.which allows for validation of dimensionality reduction data,reduces algorithmic complexity,and mitigates the loss of critical feature extraction information.The CLPSOGMM algorithm uses CLPSO to optimize GMM parameters,which improves the convergence speed and accuracy.The bidirectional long short-term memory neural network introduces an efficient channel attention(ECA)module based on traditional attention to balance model training performance and complexity.To verify the efficacy of this combined modeling method,actual SCADA data from a large wind farm was analyzed to detect faults in the main shaft and generator of the wind turbine.Using this method,wind turbine component failures can be predicted up to one month in advance.(2)In view of the limitations and deviations of using single SCADA data or single vibration data to establish wind turbine fault warning and identification model,a CMS vibration data reconstruction algorithm for wind turbines based on improved AC-GAN is proposed.The algorithm utilized is built on the classic GAN algorithm and incorporates real-time operation data from wind turbines to enhance the relationship between generator output and input signal.Through a reconstruction algorithm,the challenge of mismatch between SC AD A time series data and vibration data of wind turbines is resolved.By fusing residual characteristics of SCADA time series data and the time-frequency domain characteristics of vibration data,it effectively captures and analyzes multi-source information for fault warning and identification of critical components of wind turbines.(3)An integrated learning algorithm for optimal natural gradient boosting(NGBoost)under feature fusion is proposed,taking into account the issue of decision-making basis in defect warning and identification of important components of wind turbines.The SFO algorithm has been improved with the NGBoost hyperparameter optimization process based on feature fusion.Its performance has been compared with four other ensemble learning algorithms,namely random forest,NGBoost,XGBoost,and LightGBM,using various evaluation indexes.In addition,the fault warning and identification results under ensemble learning and its optimization algorithm are determined using the voting probability output of the model,which eliminates the need for manually setting thresholds in fault warning decision-making.Moreover,the fault category can be accurately identified based on fault early warning,thus expanding the algorithm model’s application scope.The optimized NGBoost ensemble learning algorithm under feature fusion has been tested and verified with actual SCADA and CMS vibration data from a large wind farm,demonstrating its effectiveness in improving the accuracy and efficiency of fault warning and identification of crucial wind turbine components.(4)A fault warning and identification algorithm based on feature fusion and improved siamese neural network is proposed to address the issue of overfitting the fault warning of critical wind turbine components and identification model caused by small and mediumsized sample data.Through maintaining the similarity training of the traditional siamese neural network,the algorithm has been designed to improve the feature extraction subnetwork,known as the siamese neural subnetwork.An analysis has confirmed that this algorithm,which utilizes feature fusion and an enhanced siamese neural network,significantly improves the accuracy of fault warning and identification of critical components in wind turbines,even when dealing with small sample data.This conclusion is based on the use of actual SCADA data and CMS vibration data from a large wind farm.