Fault Detection And Diagnosis Of Non-gaussian And Multimodal Process Based On Multivariate Statistical Analysis

As the scale of modern industrial system increases and the system becomes more and more complex,people put forward new requirements for the safety and reliability of industrial production process.The complexity of industrial processes is mainly manifested in the large coupling between process variables,the non-linearity and non-Gaussian of process data,and the operation process is multimodal.Based on this,fault detection and diagnosis of industrial processes comes into being.Multivariate statistical analysis,as an important branch of data-driven method,does not need to build an accurate physical model of the process.It only completes process monitoring by analyzing the correlation between variables.It has been widely used in metallurgical,chemical and pharmaceutical industries.This paper studies the fault detection and diagnosis of industrial processes with non-linear,non-Gaussian and multimodal characteristics based on the multivariate statistical analysis method,and mainly completed the following aspects of work.Taking Independent Component Analysis and Fisher Discriminant Analysis as examples,the basic principles of their application in industrial processes fault detection and diagnosis are introduced.A fault detection and diagnosis model of TE process is established based on this method.By analyzing the simulation results,the main reasons affecting its performance are pointed out,and the solutions are drawn.Aiming at the non-Gaussian characteristics of industrial processes data,a fault detection method based on fault feature Selected Kernel Independent Component Analysis and Support Vector Data Description is proposed to improve the performance of fault detection.Under normal working conditions,the independent elements are extracted by using the Kernel Independent Component Analysis,and the corresponding statistics are calculated.According to the deviation degree of the independent component statistics of the process data at the time of fault occurrence from that of the normal condition,the important independent elements are selected and given higher weights.The weighted independent component statistics are selected as the input of Support Vector Data Description,and fault detection is realized by calculating new statistics and control limits.The simulation of numerical process and TE process verifies the feasibility and validity of this method.Because actual industrial processes usually operate in multiple modes,the realization of fault detection and diagnosis for multi-modal process is mainly divided into modal identification,fault detection and fault classification The probability density of the on-line independent element approaching the normal condition independent element is estimated by the Kernel Density Estimation method,and the modal is identified.The fault detection is completed based on the improved Kernel Independent Component Analysis under the corresponding modes.The Kernel Local Fisher Analysis has advantages in dealing with the classification of data classes with different distributions,and realizes the classification of normal data and fault data in multi-mode.The experimental results of numerical process and TE process simulation verify the superiority of the model in modal identification,fault detection and fault classification.