Chemical Process Fault Diagnosis Algorithm Based On Fisher Discriminant Analysis

Modern chemical production is pretty important for the national economy, has some insecure factors about the production. This makes the requirements for the using and developing of fault diagnosis higher. Fault diagnosis technology can make a quick discriminant, separation and elimination of the faults by monitoring the operation state of the production process, monitoring the changes of the process, for preventing of the happening of any catastrophic accidents. The chemical process fault diagnosis based on multivariate statistical methods have been paid more and more attention and become the most popular research field of the fault diagnosis, because it is easy to use and only relys on the control system process data,dosen't need any other additional devices.This paper is based on the theory of multivariate statistical analysis and aims to the practical engineering application, studies some important aspects of the multivariate statistical methods.This paper's work is based on Fisher Discriminant Analysis, makes the complicated chemical process as the research platform, and has a further research and thinking.The traditional Fisher discriminant analysis is a kind of very good discriminant method; it can be applied to general fault diagnosis field effectively. But, when deal with a chemical process diagnosis, due to chemical process data has nonlinear and high noise, make it hard to apply the fisher discriminant analysis to the linear discriminative field; Even if we can make diagnosis, the widespread, strong noise, also can make a bad influence to the correctness of the judgments. According to these characteristics of chemical process data, this paper in kernel Fisher discriminant method, and the basis of wavelet transform and method and combining with the research. First,some method based on the wavelet do de-noises with the original signal;then, diagnoses the nonlinear data with the Kernel Fisher Discriminant Analysis method. By verificating, the proposed method can overcome the disadvantages of the chemical process data, and its effect is well.Although the introduction of kernel function can make the traditional linear Fisher discriminant method has the ability to deal with the nonlinear problem, but the kernel function is introduced, the existence of nuclear mapping data space will push higher space. In the actual problem, especially in the face of itself is huge chemical process data, digits disaster is likely to occur, which leads to the nuclear matrix calculated spend a lot of time, even can not calculate the possible. On the other hand, multivariate statistical analysis is a basic prerequisite to be measured data meet the need of chemical process, gaussian distribution and often cannot be met, the data of the correctness of discrimination created a lot of influence. Aiming at these problems, this paper puts forward the nuclear Fisher neighbour boundary based on judging method. The advantage is neighbour boundary method without any hypothesis treat processing data, at the same time because the role of with dimension reduction, more advantages and nuclear Fisher discriminant method combined. Through the simulation platform, TE process proved the feasibility and effectiveness of this method.