Study On Intelligent Fault Diagnosis Method Of Automobile Main Reducer And Its Application

Main reducer is the key part of drive system,and its operating status has a direct impact on the comfort and safety of automobile.Due to the complex structure of main reducer and the close correlation between each part,when some functional invalidation occurs in main reducer,multiple faults may occur at the same time.The existing fault diagnosis method of main reducer mainly focuses on single fault diagnosis,and unable to recognize multiple fault pattern accurately.Recently,with the development of artificial intelligence technology,traditional fault diagnosis technology with signal processing as core is gradually converting into intelligent fault diagnosis technology with machine learning as core,and intelligent fault diagnosis technology can effectively improve the level of intelligence of fault diagnosis and obtain some achievements in the field of fault diagnosis.However,in order to achieve intelligent diagnosis of the single fault and the multiple fault of main reducer,there still are some key issues need to be addressed,including how to preserve features of the signal during denoising aiming at strong noise components that are unavoidable in the vibration signal,how to extract weak fault features which can reflect the status of main reducer aiming at the complex non-linear and non-stationary vibration signal,how to obtain accurate complete sample set to train intelligent fault diagnosis model aiming at the problem of scarcity of labeled fault samples,how to construct intelligent fault diagnosis model to simultaneously achieve intelligent diagnosis of single fault and multiple fault patterns.Based on this,an intelligent fault diagnosis method which can simultaneously recognize single fault and multiple fault of main reducer is proposed.This dissertation aims at the problem of intelligent fault diagnosis method of main reducer in the strong noise background,and then makes research on the denoising of vibration signal in strong noise,the extraction of non-linear non-stationary weak characteristics,the semi-supervised learning of unlabeled fault samples,and the intelligent diagnosis of single fault and multiple fault patterns of main reducer.The main work of this dissertation is as follows:(1)Aiming at the problem of intelligent fault diagnosis of main reducer in strong noise environment,in order to solve the problem that signal features are over-smoothing during denoising,an adaptive sparsity tree structure wavelet shrinkage denoising method is proposed to effectively improve the noise-signal ratio of the collected vibration signal.First,a set of tree structure is constructed to realize the tree structure wavelet estimation according to correlation between different scales of wavelet coefficients during the same time interval.Second,in order to preserve the signal characteristics during denoising,dual-tree complex wavelet transform is adopted to detect the significative positions of feature points from the vibration signal.Finally,regularization weights corresponding to the coefficients which cover the positions of feature points are adaptively adjusted.(2)In order to improve the accuracy of intelligent fault diagnosis method,a non-linear supervised feature extraction method based on three-stage multi-kernel learning framework is proposed to reduce dimensionality of the signal and to extract non-linear,non-stationary and weak features from the denoised signal.First,a three-stage multi-kernel learning framework is presented by optimizing traditional multi-kernel learning method in respects of basis kernel function selection,basic kernel function combination and algorithm efficiency.Second,the three-stage multi-kernel learning framework is combined with a non-linear feature extraction method based on spatial transformation to achieve non-linear supervised feature extraction.(3)In order to accumulate complete sample set for training of intelligent fault diagnosis model for main reducer aiming at the problem of scarcity of labeled fault samples,a novel semi-supervised learning model based on clustering discrimination manifold regularization framework and extreme learning machine is proposed to mark labels for the unlabeled samples,then expand the labeled sample set to improve the generalization of intelligent fault diagnosis model.First,a novel semi-supervised learning framework based on clustering discrimination manifold regularization is proposed by improving manifold regularization and combining clustering labels with paired constraint regularization.Second,a novel extreme learning semi-supervised classification model based on clustering discrimination manifold regularization framework is proposed by integrating the advantage of extreme learning machine in learning speed.Finally,in order to improve the performance of the model,a multi-objective fruit fly optimization algorithm is presented to optimize parameters of the model.(4)Considering the impossibility of collection the samples of all the multiple fault patterns,in order to simultaneously achieve intelligent diagnosis of single fault and multiple fault patterns,an intelligent fault diagnosis model based on paired sparse Bayesian extreme learning machine which is trained by single fault samples is constructed and applied to intelligent fault diagnosis of main reducer.First,based on the advantages of sparse Bayesian extreme learning machine in probability output,the classification model of paired sparse Bayesian extreme learning machine is constructed by using the paired strategy and single fault samples.Second,the optimal decision threshold value is generated by using single fault samples and multiple fault samples.Finally,the optimal decision threshold value is used to convert the probability vector resulting from classification model to fault patterns.