Fault Diagnosis Based on Signed Digraph Combined with Dynamic Kernel PLS and SVR

The signed digraph (SDG) method for fault diagnosis, which is one of the model-based methods, has been widely applied in the chemical industry in recent years. However, how to elicit appropriate thresholds for SDG is a very difficult problem. This study presents a new hybrid method combining SDG with dynamic kernel partial least-squares (DKPLS) and support vector regression (SVR) for fault diagnosis. Using the relationships of each variable in SDG, a series of DKPLS-SVR models are built to estimate the values of the measured variables in process. The difference between the estimation and the measured value can determine the qualitative status of the variable, and then the fault can be diagnosed by SDG reasoning. Therefore, the threshold of each measured variable does not need to be decided in advance. The method can also overcome the limited availability of using the KPLS method alone in identifying the root cause. To verify the performance of the proposed method, its application is illustrated on the Tennessee Eastman (TE) challenge process. Through case studies, the proposed method demonstrates a good diagnosis capability compared with previous hybrid methods.