Neural Network Diagnosis System for Multiple Sensor Failures in Stationary Internal Combustion Engines

E. Leon, H. Malm, and M.P. Mintchev (Canada)


Neural Networks, Real-Time Systems, Sensor Validation.


Stationary internal combustion engines (SICE) are an important component in many industrial plants and have been used in various production processes for decades. In addition to the monitoring and control that is needed to ensure a good SICE performance, in the recent years industrial plants are facing strict environmental regulations that demand enhanced engine control. Since control actions are based on sensor readings, it is imperative that a sensor failure does not lead to confusion or misinterpretation of the actual operating state of the SICE. In addition, when critical sensor failures (real or apparent) are detected, the control system initiates a safety action usually resulting in a total shutdown. Avoiding unnecessary shutdowns caused by corrupted sensor readings is important to ensure high availability. In this paper we introduce a software-based sensor validation system that utilizes neural networks as an estimation mechanism, and employs logic operation and information exchange policies that are based on flexible Knowledge-Based System architecture. The developed sensor validation system is able to detect in real time potential failures in sensor values, and can provide a virtual sensor response, which replaces corrupted sensor values before they reach any process monitoring system (e.g. parametric emission estimation, or torque estimators).

Important Links:

Go Back