Fault Diagnosis And Performance Degradation Assessment Of Rolling Bearing Based On Feature Optimization And Adaptive Learning

Fault diagnosis for rotating machinery has great significance to practical engineering safety,whose key is the validity of fault feature extraction.The traditional feature extraction methods mainly depend on expert knowledge and signal processing technology.However,the feature sensitivity based on tradational methods to faults is unstable and the adaptive optimazation of features is lacking when dealing with mechanical vibration signals under complex working conditions.Adaptive feature extraction is a new feature extraction method.Different from the traditional feature extraction methods,this novel method converts the original signal multiple times and the converted data can be optimally approximated to the original data.Typical features are extracted from the converted data as the adaptive features of the signal.However,the research on the model selection and the physical meaning o f adaptive feature extraction methods is still insufficient when the methods are used for different diagnosis targets.Meanwhile,the theory of the methods needs to be further improved.In addition,during the actual service of the rotating machinery,the working condition is often variable in speed and load.The influence of the operating conditions on the featuredomain is easy to submerge the feature changes caused by the fault,which makes fault feature extraction difficult.In summary,the theoretical research and practical application of different feature extraction methods under different working conditions are still need to be further exploration.Funded by the Natural Science Foundation of China(No.51575168 and No.51875183),this paper further studies and improves the methods of fault diagnosis and performance degradation assessment of rotating machinery under constant and variable working conditions based on feature optimizationand adaptive learning methods.Then the proposed methods are applied to the fault diagnosis and performance degradation assessment ofrolling bearings.The main researches and innovations of this paper are shown as below.(1)The LFSS method is proposed to extract early fault-sensitive samples sincethe samples in the healthy state of bearing whole-life data are over redundant,which affects the analysis of feature validity.Moreover,the FSBBA algorithm is proposed to overcome the shortcomings of original BBA algorithm,which is easy to fall into local optimum.The global searching ability of the FSBBA is effectively improved by increasing the effect of hunting quality on bat speed.Based on LFSS and FSBBA based feature selection method,an online condition assessment model of rolling bearings is constructed and it is applied to two rolling bearing whole life datasets.The results of assessment index analysis show the effectiveness of the proposed method.(2)Aiming at solving the problem of traditional features cannot accurately represent the fault under the varible conditions,a projection method based on NAP is proposed.Firstly,the simulation signals of bearings at different rotational speeds are analyzed to verify that NAP can effectively remove the operation condition information in the feature domain.Secondly,the measured signals of different fault modes and simulation signals of different fault degrees under different rotational speeds are analyzed to verify that the project ed features of NAP still retain the fault pattern information and fault degree information.Finally,the whole life experiment of rolling bearings under variable operating conditions is analyzed to verify the ability of NAP to remove operation condition information again.Furthermore,by comparing and analyzing the influence of the RMI-based feature selection method and the application sequence of the NAP method on the evaluation results of rolling bearing performance degradation,it can be inferred that the NAP method should be used before the feature selection method in fault diagnosis.(3)Aiming at the application of deep learning model in performance degradation assessment of rotating machinery,a performance degradation assessment method based on CSCL is proposed.The convolution sparse coding method in deep learning is combined with combination learning method to extract features of rolling bearing under constant operating conditions.By analyzing the changes of sparse decomposition and reconstruction signals of CSCL in bearing simulation signals with different fault degrees,it is revealed that the activation energy of fault-related kernels in the sub-dictionary increases with the deepening of fault degree,while the reconstruction error based on sparse so lution decreases.Combining the above changes of the two indexes,a performance degradation index based on kernel activation energy and reconstruction error is proposed.CSCL algorithm is applied to rolling bearing whole life data,and the evaluation results show that CSCL has more advantages than traditional methods.Meanwhile,the comparative experiments show that combination learning can improve the real-time performance of model computing and optimize the performance degradation assessment results.(4)Aiming at the application of deep learning model in rotating machinery fault diagnosis,an intelligent recognition method of rolling bearing fault based on MC-CNN is proposed.MC-CNN is an improved convolution neural network algorithm which enhances the input information.The model structre of MC-CNN is to add a multi-scale cascade layer before the first layer of the original CNN.By convoluting the original signal with convolution kernels of different lengths and concatenating the output results,a cascade sig nal with more distinguishable information is constructed.By classifying four patterns of bearing fault signals in normal and noisy environment,the effectiveness of MC-CNN is verified and the recognition effect is better than that of original CNN.It is r evealed that the function of different scale convolution kernels is to find the differences of different fault patterns in frequency domain under different frequency resolutions by deep analysis of convolution kernels in multi-scale cascade layers.Finally,the validity of the proposed method is further verified by comparing the T-SNE clustering results of original CNN and MC-CNN.(5)Aiming at the application of deep learning model in fault diagnosis of rotating machinery under variable operation conditions,a fault diagnosis method of rolling bearings under variable working conditions based on wavelet time-frequency diagram and two-dimensional MC-CNN is proposed.Since the sensitive time-frequency band of the same fault pattern under variable working condition s will shift with the change of speed and its energy will change with the load,but its shape is basically unchanged.Therefore,the rolling bearing fault diagnosis under variable operation conditionsmethod based on time-frequency diagram and two-dimensional convolution neural network with translation invariance is reasonable.Consistent with the one-dimensional MC-CNN,in order to enhance the information of time-frequency image,the convolution kernels of different scales are convoluted with the time-frequency image,and then the multi-scale cascade imagesare joined together as the input of the convolution layer.Experiments show that the fault diagnosis method based on wavelet time-frequency diagram and two-dimensional MC-CNN cannot only extract fault-related features,but also has higher generalization ability and robustness than other methods.