Research On Pitch Fault Diagnosis Of Wind Turbines Under Variable Operating Conditions Based On Deep Small-World Neural Network

In recent years,the rapid development of the wind power industry has made its domestic market share nearly saturated.At present,most of the domestic wind turbine equipment is in the embarrassing situation of being insured,but still insisting on service;frequent failures,low efficiency,poor reliability,and high operation and maintenance costs are the serious challenges that the wind farm is facing.In addition,under the government's catalysis,the “de-subsidy” transformation of the wind power industry has entered a peak period in 2020.It can be seen that the benefits under cost control drive the entire wind power industry to pay attention to the health management and safe operation and maintenance of wind turbines.However,achieving effective condition monitoring and fault detection is not accomplished overnight.Affected by the wind speed fluctuations,the wind turbine is operated in the variable operating mode for life,and its own complex structure makes it have complex nonlinear coupling characteristics on the operating status monitoring data.Among them,the variable operating mode of wind turbine itself has always been an important factor that restricts the existing diagnostic methods from being practical.The operation process and status information of the wind turbine are recorded in the SCADA(Supervisory Control And Data Acquisition)system.How to use these SCADA data to extract and mine the effective characteristic information of the wind turbine failure under the variable operating conditions,undoubtedly for the diagnosis of wind turbine failure It has important practical significance and academic value.This dissertation is based on solving the fault diagnosis needs of wind turbines under variable operating conditions,starting from three aspects: fault statistical analysis of variable operating conditions,data processing and feature selection of variable operating conditions,and fault detection of variable operating conditions.A comprehensive fault diagnosis method covering all operating conditions of the wind turbine is proposed.The main research work is shown as follows:(1)Analysis and research on the fault characteristics of the pitch under variable operating conditions.In-depth study of the operating characteristics of the pitch system under variable operating conditions,and the classification and summary of the five major types of common faults in the pitch system and mechanism analysis were determined,and nine typical faults were identified as the diagnostic goals of this article;further combined with the actual wind field The alarm data is statistically analyzed for 9 kinds of faults,and it is found that the pitch faults will show different distribution rules according to the changes of the wind turbine operating conditions.In addition,in order to find the relationship between the pitch faults and the wind speed fluctuations,the wind speed is defined for the first time in the article According to the concept of jump value,it is found through statistics that 9 types of pitch failures are most likely to occur when the wind speed jump value is ±4m/s.(2)Study on the selection method of wind turbine characteristic parameters under variable operating conditions.Due to the variable operating mode of the wind turbine,the monitoring parameters that can reflect the operating state information of the equipment under different operating conditions will also be different,so it is necessary to select the most suitable characteristic parameters under different operating conditions.Aiming at the problem that the existing methods can not be used for adaptive quantitative measurement of continuous data,a new method based on Adaptive Neighborhood Rough Mutual Information Entropy(ANRMIE)is proposed to achieve quantitative measurement of multi-dimensional monitoring parameters.Comparative classification experiments verify the accuracy and reliability of the proposed method.(3)Theoretical research on deep small world neural networks.In order to solve the problem of insufficient ability to manually extract features and the scarcity of label samples in SCADA monitoring data,a new Deep Small-world Neural Network model(DSWNN)based on semi-supervised learning is proposed,which realizes the small-world neural network from supervised.After a new design,a multi-layer restricted Boltzmann machine(RBM)with efficient self-learning capabilities is added to the network structure of DSWNN,which effectively improves the characteristics of the network for unlabeled data Extraction ability;network training has changed from a single BP training to a combination of multi-step training methods such as unsupervised training,small world transformation and supervised fine-tuning,which can effectively solve the under-fitting and over-fitting in feature learning;Cross-layer edge-connected,DSWNN network can effectively avoid the problem of gradient disappearance caused by too many layers.Through the analysis of the characteristics of the small world,it is found that the DSWNN model with appropriate edge-added probability can show excellent learning ability.The proposed DSWNN network can realize nonlinear fitting of multi-dimensional complex data,and is suitable for feature extraction of wind turbine operation status and fault information from multi-dimensional SCADA data of wind turbines.(4)Research on fault diagnosis method based on deep small world neural network.Aiming at the existence of strong nonlinear coupling and spatio-temporal correlation among multiple input parameters,a Sliding-Window Deep Small-World Neural Network method(SL-DSWNN)based on dynamic sliding window is proposed.This method first uses sliding window and small-scale filtering to process dynamic data It is processed to capture the time-series feature information of the input parameters,and then the DSWNN network is used to learn the spatial correlation features between multiple input parameters;FAST simulation experiments and wind turbine measured data experiments show that: Compared with DNN,SWNN,DBN models,the SL-DSWNN method has higher accuracy and reliability in fault detection.In addition,the comparison of feature visualization clustering results verifies the excellent performance of the method in network learning and feature extraction.(5)Research on fault detection method based on multi-model dynamic selection integration under variable operating conditions.Aiming at the needs of fault diagnosis in wind turbines with variable operating conditions,a DSWNN based Selection Ensemble method(SE-DSWNN)is proposed.The method adopts distributed structure,and an independent diagnosis unit is established under each working condition.Based on the idea of dynamic selection integration,the SE-DSWNN method first uses the feature parameters selected by ANRMIE as the data source under each operating condition,and proposes to use the data division method considering the wind speed jump value to reconstruct the alternating overlapping distributed training set.Then construct multiple homogeneous and heterogeneous DSWNN sub-classification models in different wind speed intervals;in terms of dynamic selection integration,a global phase relationship algorithm is proposed to dynamically select the best sub-classifier,and weighted probability fusion is used to achieve online fault diagnosis.Finally,the SE-DSWNN method is verified by the variable-pitch fault classification experiment under variable operating conditions.The results show that the SE-DSWNN method considering the variable operating conditions can effectively divide the online data into the variable operating conditions and variable wind speed intervals accurately,and achieve accurate status monitoring and abnormal identification.