| As the core component of the vast majority of mechanical equipment,rotating machinery is widely used in industrial production.Once the failure occurs,it will often cause significant economic losses or even casualties and other serious consequences.Therefore,in order to improve the safety and reliability of rotating machinery during operation and avoid the occurrence of major accidents,it is necessary to carry out a lot of research on rotating machinery fault diagnosis technology.For most of the existing fault diagnosis algorithm is based on the thought of supervision,and need to use a priori knowledge about the data,that is to have a tag data,and the parameters of the algorithm is mostly involved need to manually,the lack of practical operability is not strong,this paper proposes a data without a label oriented unsupervised algorithm,parameter from the specialisation of rotating machinery fault diagnosis(Sm-DLLOF-AFCM),without any priori information(tag)can be the completion of the sample set of adaptive fault diagnosis work,has a good accuracy and strong applicability and practical significance.The main work is summarized as follows:1.The typical faults of rotating machinery were analyzed.The acquisition and pretreatment technology of rotating machinery condition monitoring data are studied.Finally,the data acquisition scheme based on Lab VIEW and the pre-processing method of wavelet threshold denoising are determined.2.The feature extraction and feature selection techniques are studied.An improved empirical wavelet transform(SEWT)algorithm based on scale-space theory is used in feature extraction of vibration signals of rotating machinery.The SEWT algorithm solves the problem that the traditional empirical wavelet transform(EWT)algorithm needs to assign the number of frequency band segmentation artificially.It can adaptively determine the number of frequency band segmentation and the retained mode component(IMF).In feature selection,an improved Laplace score algorithm(m RMR-LS)based on MRMR is used.The m RMR-LS algorithm takes the correlation and redundancy of features into consideration,and it can be used for feature selection in an adaptive and unsupervised way.3.The pattern recognition method is studied,and the unsupervised pattern recognition method of anomaly recognition and clustering is adopted.First in abnormal recognition work,adopted the second to identify ways to improve recognition accuracy and efficiency,the first to identify the improved DBSCAN algorithm based on kernel density estimation method(KDBSCAN),KDBSCAN algorithm can determine the traditional DBSCAN algorithm of adaptive two parameters of Eps and Minpts,less output noise and class as the first to identify abnormal sample cluster sample.The improved local outlier factor algorithm(LLOF)based on the natural nearest neighbor search method is used for the secondary identification,which can adaptively determine the neighborhood parameter k in the LOF algorithm,and output the abnormal samples and their abnormal score value(LOF)of the secondary identification.Then the advantages of the two algorithms are verified by real data sets.Finally,the adaptive fuzzy C-means clustering algorithm(AFCM)is used to classify the abnormal samples and their scores.4.Determine the final fault diagnosis scheme and verify it.Firstly,Sm-DLLOF-AFCM rotating machinery fault diagnosis algorithm is proposed,and its general process is given.Then using the laboratory motorized spindle condition monitoring test system and the built motorized spindle comprehensive test software platform,the rotor data set was obtained by the way of artificial fault simulation.Finally,the Sm-DLLOF-AFCM algorithm was applied to the laboratory rotor data sets and the open bearing and gear data sets to verify the results.The results show that the SM-DLLOF-AFCM algorithm has achieved good results on the three data sets.While ensuring the efficiency,it can successfully identify and cluster the vast majority of abnormal(fault)samples. |
