Process Fault Diagnosis Approach based on Neural Observers

Abstract

This paper presents an approach to process fault diagnosis (FDI) in nonlinear dynamical systems, based on a bank of neural observers. Each neural observer is tuned to a particular fault and predicts, using its embedded model, the expected values for the sensor readings. The residuals, the difference between the sensor readings and the predicted readings, are used as fault indicators. Each neural observer is based on a multi-layer perceptron feed-forward neural network with externalfeedback connections, and an adjustable gain.The focus of this work is on diagnosis ofparametric faults on process components, by means of analyzing the residuals. The proposed FDI technique has been implemented on a simulation model ofa DC motor under closed-loop control. Results from experiments are presented and discussed.

Banks of Gaussian Process Sensor Models for Fault Detection in Wastewater Treatment Processes

“…In WWTPs these schemes have previously been applied using physics-based state estimators configured in banks of observers (Nejjari et al, 2008;Nagy-Kiss & Schutz, 2013). However, the complexity of the process makes data-driven state estimation attractive -as in other fields (Palma et al, 2005;Sina Tayarani-Bathaie & Khorasani, 2015). An issue which arises in using these schemes for FI is that the sets of residuals that need to be monitored for FD become large (one set per observer).…”