Research On Data Driven Fault Diagnosis Methods For Non-Gaussian Process

Data driven fault diagnosis has been widely studied in recent years.As a common characteristic,non-Gaussian distribution widely exists in modern industrial process data.The traditional data driven method based on the hypothesis of Gaussian distribution is no longer applicable to the data with such characteristic.In this dissertation,the non-Gaussian process is taken as the research object,and the fault detection,diagnosis and identification in the data driven method and its application in an industrial process called the Electric Fused Magnesia Furnace are researched.The five main contents include:1.For the fault detection aspect,the independent component Gaussian distribution transformation method is researched and the fault detection indicator and its control limit based on this transformation are developed.By comparing the results of principal component and independent component Gaussian transformation,the characteristics of Gaussian distribution and nonlinear independent components are analyzed.The transformed independent component follows univariate Gaussian distribution independently,so the diagnostic methods based on principal component analysis(PCA)such as the calculation of control limit can be transplanted parallelly.2.For the fault detection aspect,two kinds of fault indicator based on Hilbert kernel space are developed,and the geometric interpretation of the feature mapping using radial basis kernel function is given.Based the one-class support vector machine(OC-SVM)algorithm,the distance between the point of the feature space and the hyperplane is developed as the indicator of the fault magnitude,and the upper and lower limits of the indicator are derived.Based on kernel principal component analysis(KPCA),load normalization and its new eigenvalues are derived and extended to non-Gaussian process fault detection.3.For the fault diagnosis aspect,a universal root cause analysis method is researched.The concept of data source is proposed.Different with the fact that the fault diagnosis of Gaussian distribution data is the measure of data deviation from the origin,this paper considers it to be a measure of data deviation from data source for the non-Gaussian distribution data.We combine data source with virtual scale factor method,and propose an improved generalized root cause analysis method.The method can not only be combined with the traditional T2,SPE fault indicator but also be combined with all the fault indicators mentioned in this paper to meet the requirements of the Gaussian and non-Gaussian process fault diagnosis.At the same time,we studied how to evaluate the performance of root cause analysis algorithm and proposed the evaluation criteria of fault elimination rate.4.For the fault identification aspect,this paper studies the fault identification method of fault sub-space reconstruction in the feature space.Different from the general pattern classification method,this method takes the fault direction as the classification basis,extracts the nonlinear fault subspace,and combines with the non-Gaussian process fault detection method proposed in this paper to improve the accuracy of the non-Gaussian process fault identification.5.An example of an industrial process is studied.An intelligent diagnostic framework for an electric fused magnesium furnace based on its sound is proposed.A data acquisition device is developed.Through analyzing the sound signal of an electric fused magnesium furnace,it is found that the sound energy is concentrated near the frequency of 100Hz and its frequency doubling.So,the data preprocessing method based on unsupervised Laplace score is applied to select the key frequencies that can represent the running state of the furnace.Combined with diagnostic method studied in this paper,the fault detection and identification of electric fused magnesium furnace based on sound is realized.The contents of this paper include fault detection,diagnosis and identification.In the experimental part,synthetic data and Tennessee Eastman Process(TEP)simulation data are used as data benchmarks.The performance of the algorithm is evaluated by false alarm rate,miss alarm rate,fault elimination rate and precision rate and so on.Experimental results verify the effectiveness of the proposed algorithm for non-Gaussian process fault diagnosis.