(402c) An Intelligent Pca Approach for on-Line Fault Isolation | AIChE

(402c) An Intelligent Pca Approach for on-Line Fault Isolation


Liu, J. - Presenter, Institute of Chemical and Engineering Sciences
Lim, K. - Presenter, Agency for Science, Technology and Research
Doan, X., Institute of Chemical and Engineering Sciences

Modern chemical plants are susceptible to various failures resulting in abnormal situation as they are highly complex and integrated, processing large volumes of materials and operating at extremes of pressure and temperature. It is often difficult to completely rely on operators to cope with such abnormal situations. Given a large amount of real-time plant data, operators have neither the time, nor often, the expertise to effectively monitor this process information. Multivariate statistical approaches and principal component analysis (PCA) in particular, have been widely investigated to deal with this challenging problem. However, they do not possess ?fingerprint' or ?signature' properties for diagnosis, which makes the fault isolation difficult. Some techniques such as contribution charts have been proposed for fault isolation purpose. However, contribution charts can only point out the process variable dominating the error space whereas it can not differentiate whether the fault is due to sensor faults so as to isolate faulty sensors or due to process faults where further diagnosis is required.

This paper proposes a method, which is a combination of PCA-monitoring approach and expert-systems approach, to differentiate sensor faults from process faults. Unlike other methods utilizing redundant sensors, this method is based on the idea that one sensor fault for non-manipulated variables only affects this variable whereas a process fault affects several variables). The real-time on-line process monitoring and fault isolation algorithm is as follows.

Build the PCA model for process monitoring and fault detection as normal. When a fault is detected, check the square prediction error (SPE) contributions. If one variable is dominating the error (such as >50%), build another PCA model without the dominating variable and check T2 and SPE limits again. If the process is within limits now, the fault is diagnosed as a sensor fault and faulty sensors can be isolated. If the process is still out of limits, the fault is considered as a process fault and SPE contribution chart is given to help the diagnosis rather than do exact diagnosis.

A refinery process case study is used to illustrate the approach proposed in this paper. It is the reaction section of a cyclohexane plant, including 2 reactors and 1 steam drum. The process simulation is developed with gPROMS, which is a general process modelling and simulation package from Process Systems Enterprise. OLE for Process Control (OPC) is used for process data accessing so that this method can be ?plug and play? to the real plant. The process simulation is done in real time, and it sends process data to an OPC server every second. However, the intelligent PCA module in Matlab implements process monitoring and fault isolation roughly every 2 seconds as 1 second seems not enough for the application to access the process data, do the calculation and display the result. Both sensor faults and process faults are considered and simulated in real time.

Real-time simulation results show that both sensor faults and process faults can be quickly detected with the proposed approach and sensor faults can be successfully differentiated from process faults.


This paper has an Extended Abstract file available; you must purchase the conference proceedings to access it.


Do you already own this?



AIChE Pro Members $150.00
AIChE Graduate Student Members Free
AIChE Undergraduate Student Members Free
AIChE Explorer Members $225.00
Non-Members $225.00