Fault Prognosis Of Industrial Processes Based On Multivariate Correlated Time Delaed Serias

With the rapid development of modern industry, industrial processes are developed towards the integrated and intelligent direction. Meanwhile, the demands for process safety are gradually improved. When the faults occur, the isolation and protection measures are needed. Also, before the faults occur, the process condition had been known in advance. Especially, for the incipient faults (the amplitude is small and difficult to be detected), the timely detection and elimination are needed in advance. Therefore, improving the existing fault prognosis methods and strengthening the incipient fault monitoring for industrial process have become the research focuses. Therefore, aiming at these issues, the studies of the paper are as following:First of all, aiming at the complexity and time delay among the process variables, this paper focuses on the research of correlation and time delay of process variables. A new method for constructing the multivariate correlation time delayed series is proposed. It is considered that the probability density function is difficult to obtain, the mutual information is improved by k-nearest neighbor algorithm to calculate correlated coefficient among process variables. For the time delay estimation, Bayes information criterion (BIC) is used to obtain the delay time of process variables. In this way, the multivariate correlation time delayed series is constructed by the improved mutual information approach and BIC. Secondly, combining the quantitative and qualitative method, a new fault prognosis model based on time delayed signed directed graph (SDG) and improved independent component analysis (ICA) is proposed. In this approach, the correlation information and time delay information are added on the original SDG, then the time delayed SDG is constructed with more information. It is considered that ICA cannot obtain the independent components among large data quickly. This paper proposes an improved ICA based on extreme learning machine (ELM) neural network, so as to quickly obtain the independent components. The experiments on TE process show the effectiveness of the proposed method. Thirdly, aiming at the incipient faults, a new method based on the multivariate time delayed series and statistical local kernel principle component analysis (KPCA) is proposed. In order to increase the prediction accuracy for the process variables, the correlated variables are regarded as the input of neural network. Then, the statistical local technology and moving window technology are applied to improve KPCA to increase the sensitivity of monitoring methods. Finally, the experiments on TE process show that the proposed methods can realize the incipient fault prognosis.For the research results, the introduction of correlation and time delay among process variables, the construction of time delayed SDG with more information and the improvement of multivariate statistical monitoring methods, have developed their advantages for fault prognosis and provided new ways to ensure the safety of industrial processes.