Research On Multivariate Statistical Analysis For Machine Condition Monitoring And Diagnosis

With the research aim of precision diagnosis of machine conditions, this paper addresses on the application studies of multivariate statistical analysis for machine condition monitoring and diagnosis. Through analyzing the research literature of four developing multivariate statistical analysis methods including principal component analysis (PCA), independent component analysis (ICA), kernel principal component analysis (KPCA) and blind sources separation (BSS) in this research area, three sub-systems were built and studied in depth, respectively, as follows:The first is higher-order statistical information extraction, which focuses on extracting the higher-order statistical information from the one-dimensional or multidimensional measured signals to effectively represent the machine conditions based on the ICA theory. This paper mainly introduced and developed the extraction of the higher-order statistical information of one-dimensional vibration signal, which has excellent potential applications since it reveals the inherent characteristics of vibration signals. A novel ICA-based transient detection method was proposed in this paper and showed the good effect outperforming the other traditional methods. In addition, a new ICA filtered correlation feature parameter was extracted to effectively represent the class information of machine conditions.The second is redundant multivariate features fusion, which addresses on extracting the novel and more sensitive and more stable statistical structure from the original time-domain, frequency-domain and time-frequency domain features to represent and classify machine conditions, based on information compression and dimension reduction of PCA, ICA, KPCA, and so on. This paper mainly developed nonlinear feature extraction technique by using KPCA. The extracted nonlinear features have the excellent clustering effect in the feature space. Then the KPCA-based nonlinear feature subspace models were constructed to effectively represent and classify the machine conditions. This paper still emphasized the research of the feature evaluation and selection in multivariate statistical feature extraction, in which some new ideas were proposed to solve the problem of the maximum efficiency when the multivariate statistical features are used to recognize the machine conditions.The third is multidimensional measured signals separation, which is to process the multidimensional measured signals by the idea of blind sources separation to obtain the signals reflecting the information of some or all machine components, separate and extract some signal components, or only eliminate the noise. This paper explored the vibration component separation of multidimensional machine vibration signals based on the BSS technique. The linear ICA method was validated to be not ideal when applied to the complex vibration signal separation. Then the BSS model for convolution mixtures was considered to mainly study the separation and extraction of transient components, including cyclostationary components and transient impulses, from the noisy signals.Moreover, the research works above were all validated by the experiment analysis, in which two experiments were applied. One is the automobile gearbox condition monitoring by vibration signal analysis; the other is internal-combustion engine wear diagnosis by sound signal analysis. The study of this paper indicates that the multivariate statistical analysis can extract the sensitive and stable features that well represent machine conditions, which is very significant for precision diagnosis.