Research On Rolling Bearing Fault Diagnosis Of Wind Turbine Based On Compression Fusion And Multi-scale Residual Network

Due to the increasingly serious global environmental pollution problem,the development of low-carbon economy and utilization of renewable energy has become the core issue in the global energy sustainable development strategy.As a type of renewable energy with huge reserves and environment-friendly characteristics,wind power industry based on wind turbines has become the fastest growing clean energy industry with the strong support from countries around the world.Wind turbines are mostly distributed in remote and harsh areas,resulting in frequent failures and high operation and maintenance costs and difficulties.As a key component of wind turbine,bearing can play the role of supporting the shaft and reducing friction.However,due to the variable operating conditions and operating environment,the bearing is one of the most easily damaged components of wind turbines.Therefore,an accurate and effective bearing fault diagnosis scheme can improve the operational reliability of wind turbines and reduce the operation and maintenance costs of wind turbines.China’s wind power industry lags behind countries in Europe and the United States.At present,the research on fault diagnosis of wind turbines is not yet mature.In order to ensure the sustainable and healthy development of the wind power industry,it is necessary to build a high-precision bearing status remote real-time monitoring system.In the process of monitoring the rolling bearings of wind turbine,the system should be equipped with the functions of signal fusion,compression and transmission,reconstruction and analysis,and intelligent diagnosis,because the system will face the problems of massive acquisition signal and large amount of interference noise.In this paper,the research focuses on the above-mentioned problems in the process of intelligent monitoring and diagnosis of wind turbine bearing faults,deeply analyzes the correlation between multi-sensor acquisition signals,analyzes the signal compression and reconstruction theory based on compressed sensing,discusses the effect of feature differentiation based on re-encoding,and builds an intelligent network-based monitoring and diagnosis method for rolling bearing of wind turbines.(1)The sensor monitoring system based on multi-sensor co-sampling can effectively avoid the loss of critical information,but it puts great pressure on the signal storage and transmission equipment.The proposed method is based on improved empirical wavelet transform and optimized random weighting,which can effectively reduce the noise component while enhancing the fault features of rolling bearing.The method firstly uses the empirical wavelet transform as the basis for decomposition and filtering of rolling bearing signal,and introduces the concept of variance contribution rate as the basis for setting the weight of each component by combining the reflection of fault features in the time-domain.Secondly,the random weighting algorithm is used as the basis for the weighted fusion operation of multi-sensor signals,and an adaptive equalization factor based on the relative random fluctuation is constructed to obtain the best estimate by combining the correlation law between signals with time.Finally,the method can reduce the amount of signal storage and transmission and effectively suppress the noise components in the rolling bearing signal by the twice weighted fusion operation while fully retaining and enhancing the fault features,which improves the accuracy of the subsequent diagnosis process.(2)Compressed sensing-based signal compression method can break through the limitations of Nyquist’s sampling theorem to realize deep compression of rolling bearing signal,and with multi-sensor fusion operation can further reduce the amount of signal storage and transmission.The high complexity and weak sparse performance of the rolling bearing fault signal make the subsequent reconstruction process more difficult.Therefore,this paper proposes an improved two-stage matching pursuit algorithm,using the target atom projection coefficients as the screening basis in the forward selection and backward rejection process,and setting the fuzzy factors in the forward and backward stages according to the signal characteristics to control the screening threshold and increase the randomness and speed of the algorithm for screening atoms.And analyze the backward excessive problem that causes the correct atoms to be rejected in the backward stage,and improve the reconstruction accuracy of the algorithm for the target signal by synchronizing the update of the observation matrix to ensure the second selection of correct atoms.(3)After completing the pre-processing task of the target rolling bearing signal,the fault features in the signal need to be extracted to realize the task of synchronous detection and diagnosis of the target signal.Since the wind power rolling bearing signal is seriously disturbed by noise and the feature information is mostly hidden in the signal,direct extraction of features is difficult and ineffective.In this paper,an enhance binary-based 1D-TP encoding algorithm is proposed to realize the effective extraction of detailed feature information from the target signal.The method is based on the traditional1D-TP encoding algorithm,and the new encoding method that can realize mode fusion is designed by analyzing the information loss problem caused by the merit selection process under two encoding modes.On the basis of this,the characteristics of different fault state signals are analyzed,and zero removal processing is introduced to increase the differentiation of different types of fault signals.Finally,by analyzing the correlation difference between the two-dimensional image pixel points and the one-dimensional signal points,a new signal encoding sequence is set.The method improves the extraction of feature information in rolling bearing signal,increases the differentiation between different fault features,and can effectively reduce the difficulty of fault diagnosis process.(4)Different from the shallow learning method based on data preprocessing and feature extraction,the data-driven deep intelligent diagnosis network can directly explore the fault feature information implicit in the target rolling bearing signal.In view of the limitation that the deep multi-scale residual network cannot adaptively extract features at different levels of depth,this paper proposes an improved deep residual network to realize the effective extraction of target rolling bearing signal features at multiple levels.The network is based on the traditional residual network,and analyzes the extraction of signal features by different perceptual fields with consistent layer and convolution kernel,and combines the hole convolution theory to obtain different perceptual fields using the expansion rate.On this basis,the information characteristics contained between different layer features are analyzed,and deep features are fused with shallow features and used as new features based on the pyramid principle to enhance the reuse and fusion of network feature information,enrich the expression of fault information,and improve the network diagnosis accuracy.