Diagnosing a solar volumetric receiver combining NN-based modelling with online parameter identification and rule-based techniques RAMON FERREIRO GARCIA ∗ , Departamento de Ingenieria Industrial, University of A Coruna, ETSNM, Spain JOSE LUIS CALVO-ROLLE † and FRANCISCO JAVIER PEREZ CASTELO ‡ , University of A Coruna, Departamento de Ingenieria Industrial, EUP, Spain Abstract The lack of detectability when model-based techniques are applied to intelligent fault detection and isolation tasks, which mainly occur in most non-linear processes so far is an unresolved problem. Generally, all types of non-linear processes will also suffer from lack of detectability due to the inherent ambiguity in discerning faults in the process, sensors and/or actuators. The aim of this work is to develop and apply a strategy to detect and isolate process and/or sensor faults by combining a neural network-based functional approximation procedure associated with an online identification algorithm, both processed by recursive rule-based techniques using parity space approaches. A case study dealing with the supervision of a solar volumetric receiver was performed using the proposed intelligent techniques. The conducted study produced reliable and acceptable intelligent fault detection and isolation results on the basis of heuristic knowledge-based rules. Keywords: Conjugate gradient, fault detection, fault isolation, neural networks, parity space, residual generation. 1 Introduction 1.1 Preliminary: model based on fault detection methods During the past three decades, some model-based approaches applied on fault detection characterized by using mathematical models have been developed. Thus, in [1–6], relevant contributions to this topic have been proposed. The used strategies consist of the detection of faults occurred in the processes, actuators and sensors on the basis of some dependencies between different real -time measured signals, where such measured dependencies are described by mathematical process models. Briefly, the strategy consists of generating residuals, parameter estimates or state estimates based on the measures of the inputs and outputs for the detection of features. The changes of features are then detected by comparing the normal pattern of features with the measured features which conduct us to the analytical symptoms [6–8] are also among the classic bibliography dealing with model-based fault detection and isolation (FDI) methods, which have also been previously published. ∗ E-mail: ferreiro@udc.es † E-mail: jlcalvo@udc.es ‡ E-mail: javierpc@udc.es Vol. 23 No. 3, © The Author 2015. Published by Oxford University Press. All rights reserved. For Permissions, please email: journals.permissions@oup.com doi:10.1093/jigpal/jzv004 Advance Access published 29 March 2015 Downloaded from https://academic.oup.com/jigpal/article-abstract/23/3/379/723000 by guest on 18 October 2019