Research On Bearing Fault Noise Diagnosis Method Based On Machine Learning

The rolling bearing is called the "heart" of the rotating machinery,and it is also one of the most vulnerable parts in the rotating machinery.In order to identify the running state of the rolling bearing accurately,efficiently and in real time,it is of great significance to diagnose the fault of rolling bearing by the means of machine learning.However,the traditional fault diagnosis method based on vibration signal requires the installation of vibration sensors,which will cause damage to the mechanical equipment.To solve this problem,the noise signal of rotating machinery is taken as the monitoring signal of the bearing state,and the advantage of the non-contact measurement of the noise signal is used to overcome the shortcomings of the fault diagnosis method based on vibration signal.The feature extraction,feature selection and fault classification algorithms of rolling bearing noise signal are studied in the paper.A noise diagnosis algorithm for rolling bearing fault based on MFCC-CDET is proposed.This algorithm employs the Mel frequency cepstral coefficients(MFCC)as the feature vector of noise signal,applys compensation distance evaluation technique(CDET)for feature selection and uses the selected feature as the input of support vector machine(SVM)classifier for fault classification.The effectiveness and superiority of this method is verified by experiments.In the early stage of the rolling bearing state monitoring,there is a problem that the supervised classification algorithm can't be applied when there are no fault samples in the sample library.To solve this problem,the Single Gaussian Model,an unsupervised anomaly detection algorithm in the machine learning is used in the paper,it regards the normal samples in the sample library as target samples to model the boundary of the normal samples for the anomaly detection of rolling bearings.The boundary modeling algorithm based on Single Gaussian Model is compared with the boundary modeling algorithm based on support vector description(SVDD),and its superiority is proved.In the medium stage of the rolling bearing state monitoring,there is a problem that the supervised classification algorithm can't recognize the novel fault when there are multiple categories of fault samples in the sample library.To solve this problem,the Gaussian Mixture Model is introduced.It regards the multi class samples as the target samples set and models the boundary of the target samples for the novel fault detection of rolling bearings.Compared with the Single Gaussian Model,the superiority of the Gaussian Mixture Model for multi class sample boundary modeling is verified.By combining the function of the novel faults recognition and the function of the classification of the Gaussian Mixture Model,this paper proposes a new algorithm for novel fault recognition and fault classification based on MFCC and Gaussian Mixture Model.The validity and feasibility of the algorithm are verified by experiments.