State Estimation Based Fault Diagnosis And Prediction For Petrochemical Processes

Petrochemical processes are characterized by high temperature and pressure, flammable, explosive, toxic and harmful. If faults can not be detected and removed in time, they may lead to large personnel and economical losses, even ecological crisis. Thus, fault diagnosis and prediction techniques are critically important for industrial processes. In this paper, for multivariate, nonlinear and non-Gaussian characteristics of continuous processes, some novel fault diagnosis and prediction methods are proposed based on state estimation. The effectiveness of the proposed methods is demonstrated using simulation cases.Firstly, in order to consider the serial correlations and cross correlations in multi-dimensional observed data and to capture process dynamic behaviors with lower-dimensional characteristics, a fault detection method based on state residual of canonical variate analysis is proposed. Canonical variate analysis model can be established using observed data, and process state sequences can be computed via the model. Then, State space matrices are estimated using the state sequences, and the subspace model is obtained. By comparing the states estimated by the canonical variate analysis model and those estimated by the subspace model, residuals are generated. Then, a multivariate statistic is designed to monitor the variations of the residuals. Simulation results on a continuous stirred tank reactor show that: the proposed method can detect faults more quickly and more sensitively via lower-dimensional characteristic information.Secondly, two fault diagnosis methods based on innovation characteristics of unscented Kalman filter are proposed for nonlinear multivariate processes: a multivariate sequential probability ratio test method and an information divergence method. Unscented Kalman filter is performed to generate innovation sequence, and then further investigations are carried on into the innovations. The multivariate sequential probability ratio test method uses log-probability likelihood ratio statistic and decision rules to monitor the process. In order to consider the non-Gaussian characteristics in prediction innovations, the information divergence method applies kernel density estimation algorithm to obtain the probability distribution of the multivariate innovations. Then an information divergence statistic is designed to monitor the process. A symmetric information divergence is applied to isolate the fault by measuring the distance between the monitored process and the fault process in fault database. Simulation results on a continuous stirred tank reactor show that the proposed methods can monitor process changes effectively and distinguish the fault type correctly.For fault prediction problems of unknown nonlinear system, an improved Kalman predictor with innovation predicted by support vector machine is presented. Canonical variate analysis based subspace identification algorithm is applied to obtain the local linearization model of the unknown nonlinear system. Kalman filter is performed with the model to track and predict the process. Support vector machine time series algorithm is then used to forecast the future innovations. One-step ahead and multi-step ahead Kalman predictions are corrected with future innovations. Simulation results on a continuous stirred tank reactor show that the proposed methods can predict the future dynamic variations accurately and provide early fault alarm.Lastly, a fuzzy-adaptive unscented Kalman filter predictor is proposed to improve the tracking and forecasting capability for failure processes. The state error covariance and Kalman gain are regulated according to filter performance to strengthen the tracking capability for failure processes. For serious tracking fluctuation caused by over regulation, a Takagi-Sugeno fuzzy logic system is designed to smooth the regulation. Simulation cases show that the fuzzy-adaptive filter can guarantee the strong tracking ability and achieve smooth tracking simultaneously. The fuzzy-adaptive filter based predictor can forecast the future output accurately and provide early fault alarm.