Research On Application Of Fault Diagnosis In Rolling Process Based On The PLS And Improved PLS Methods

Along with the rapid development of automation technology in steel and iron field in our country, the request of reliability and security of rolling control system is getting higher and product quality is becoming stricter. The fault diagnosis of strip mill has become one of the most significant research directions in the process automation field. The technique of fault diagnosis based on data-driven is a convenient and useful method, for it doesn't depend on precise mathematical model and the on-line process data of industrial production are easy to obtain.Partial Least Squares (PLS) analysis method is a data dimensionality reduction technology, which extracts useful process information from data to build process model using statistical theory. It can not only reduce data dimensionality and extract the data feature, but also consider the regression relative between the input and output variables. This paper studied the basic principal of PLS and its application on fault diagnosis of strip mill AGC and Looper control systems. An improved PLS method of fault diagnosis based on Relative space Transformation Partial Least Square (RT-PLS) is proposed forward in this paper, for the field data has the feature of sparsity and nonlinearity, and the effect of diagnosis is not obvious in sometimes using the traditional PLS method. The main idea of RT-PLS method is to transfer the original data space to the relative space by a relative transformation, choose the Mahalanobis Distance as the distance formula of raletive transformation, and then the PLS is employed in the new space. Mahalanobis Distance is a superior distance for PLS method, because the dimension influence doesn't need to be considered, and the contribution of variables slightly changed is amplified. The RT-PLS is a method to reduce dimension of relative space and extract the relative features which is provide with more variability and divisibility in the relative space. Finally the fault can be detected on-line by monitoring T2 and Q which are calculated by relative principal components and the fault can be recognized according to the accumulation contribution chart of every variable. Applying PLS and RT-PLS methods to detect and diagnose the three kinds of faults occurring on rolling line, which is including the sensor null drift fault, the looper fault and the monitoring AGC failure fault. The MATLAB simulation result indicate that the method based on RT-PLS can detect the three kinds of faults faster and more precise than the PLS method, besides, it can identify the leading variables causing the faults more clearly.