A neural networkmultiple fault diagnosing framework based on dynamic characteristics for Tennessee Eastman Plant

فاقد چكيده

Fault diagnosing in the plant wide systems is a complicated problem, especially in detecting multiple faults. One of the common methods for diagnosing faults is based on the neural network. In many cases, faults considered for diagnosing are not detectable and therefore the conventional neural network approach which uses the data corresponding to the steady state behavior of the system is not adequate. In this work, two frameworks have been proposed which are based on the utilization of a feed forward neural network trained based on a hybrid set of data consists of both the dynamic characteristics and steady state behavior of the system to diagnose multiple faults. The dynamic characteristics data includes the overshoot and undershoot values in the measured variables and also the time at which the variables met these values. The difference between these frameworks is how to integrate the dynamic characteristics data with steady state data for diagnosing multiple faults. To evaluate the performance of the proposed framework, the Tennessee Eastman (TE) process was used as the plant wide benchmark. Six faults have been considered in the assessment of the proposed framework, these six faults have been occurred in various scenarios in which each of these faults was occurred in a single manner and cases at which various combination of multiple faults (from double and triple simultaneous faults up to six simultanous faults) occurred in the TE process. The proposed framework helps to establish the detectable conditions in the plant wide system. The results indicate the generality, flexibility and accuracy of the proposed frameworks in diagnosing of multiple faults in the TE process.