Research On Fault Diagnosis Method Of Wind Turbine Transmission System Based On Compressed Sensing Information Representation

With the increasingly serious energy crisis and environmental pollution,wind energy,as a rich,clean and renewable energy,has become an indispensable part to solve the world energy problem.And with the continuous and rapid development of wind power industry,establishing a reliable wind turbine condition monitoring and fault diagnosis system to continuously monitor the operation status of the unit in real time has become a necessary measure to ensure the safe and reliable operation of the unit.However,the massive data storage,transmission and processing problems brought by the monitoring process bring a huge burden to the information real-time monitoring and remote diagnosis system.How to improve the ability of condition monitoring and fault diagnosis system to deal with massive data has become a subject with important academic research significance and practical application value.Aiming at the urgent need of data compression for wind turbine condition monitoring and fault diagnosis system,this dissertation analyzes the mathematical model and constraints of compressed sensing theory,discusses the mapping relationship between measurement signals and feature information,and deeply studies the feature information representation form and recognition method of fault data in the change domain of compressed sensing,so as to provide technical support for fault diagnosis of compressed data.The main work is as follows:(1)The complex fault signal has high complexity and poor sparse performance,which brings a great burden to the compressed sensing reconstruction process.A composite fault diagnosis method with compressed information based on independent component analysis is proposed,which directly processes the projection compressed data,reduces the signal complexity and improves the reconstruction performance.In this method,kurtosis is used as the independence criterion to analyze the influence of the projection process of various measurement matrices on the signal Gauss,and some Hadamard measurement matrices are selected to compress and collect.The composite fault data under the projection of measurement matrices are separated using independent component analysis.Finally,the projected separated data is reconstructed and the reconstructed signal is analyzed by envelope spectrum to extract the frequencies feature of the composite fault signal for fault diagnosis.The method improves the reconstruction accuracy and separation stability of compressed fault signals,and multiple frequency features of composite fault signals can be extracted effectively from compressed data.(2)The complex structure and changeable operating environment of wind turbines seriously affect the sparse performance of vibration signals.Leading to the unguaranteed reconstruction accuracy under high data compression,and the reconstruction process pays a lot of computational cost.A fault diagnosis method based on compressed information feature extraction is proposed.In this method,the potential mapping relationship between the data features projected by the measurement matrix and the original data features is analyzed,and the feature information of the signal is directly extracted and mined from the compressed projection,so as to realize the fault diagnosis.Partial Hadamard matrix is selected as the measurement matrix.In the experiment of rolling bearing fault degradation trend,root mean square value is selected as the characteristic parameter to prove the feasibility of extracting features directly from the compressed data projected to determine the bearing degradation trend.In the experiments of rolling bearing fault type identification,the kurtosis factor,variance,and waveform factor are selected as sensitive feature parameters.And the PSO-SVM algorithm is used for pattern recognition of sensitive features to achieve fault type identification at a higher compression rate,and the fault features information is expressed with less data volume.(3)The data-driven deep sparse autoencoder network does not need to establish the mapping relationship between compress domain and original time domain data fault features,and can directly extract fault features by mining the information implied between system variables.However,the fault feature recognition effect is deteriorated by the measurement matrix and data compression.In order to improve the ability of deep sparse autoencoder network to learn data features under compressed sensing transform domain,a sparse feature autoencoder method based on compressed sensing is proposed.This method deduces the feature mapping relationship between sparse atoms and the original signal in the coding and decoding network,establishes the mathematical model of feature learning network.Analyzes the loss in the extraction process of compressed sensing feature atoms.A gradient-acreage value analysis method is proposed to evaluate the sparse performance of the signal under the basis matrix and determine the number of atoms to be screened,which is used as the iterative termination threshold corresponding to the orthogonal matching tracking algorithm to extract the feature atoms in the compressed signal.Then the deep sparse autoencoder network is used to deeply learn the feature atoms and signals of different fault types are classified and recognized.This method improves recognition rate under the condition of high fault signal compression,and achieves data compression and fault diagnosis.