Towards Robustness in Neural Network Based Fault Diagnosis

Abstract

Challenging design problems arise regularly in modern fault diagnosis systems. Unfortunately, classical analytical techniques often cannot provide acceptable solutions to such difficult tasks. This explains why soft computing techniques such as neural networks become more and more popular in industrial applications of fault diagnosis. Taking into account the two crucial aspects, i.e., the nonlinear behaviour of the system being diagnosed as well as the robustness of a fault diagnosis scheme with respect to modelling uncertainty, two different neural network based schemes are described and carefully discussed. The final part of the paper presents an illustrative example regarding the modelling and fault diagnosis of a DC motor, which shows the performance of the proposed strategy.

Gas turbine gas path diagnostics: A review

“…To answer the question why neural networks are more popular than the other AI methods, Patan et al [38] conducted a research work on two different feed forward multilayer perceptron programs using DNN models taking into account the nonlinear behavior of the gas turbine and the robustness of a fault diagnosis scheme with respect to modelling uncertainty. It was pointed out that the fault diagnostic algorithms belonging to ANN were having smaller false alarms during detection, much more sensitive to early fault detection, high fault classification rate, and are efficient in fault identification than the other AI techniques.…”

Application properties of methods for fault detection and isolation in the diagnosis of complex large-scale processes

“…The robustness is determined by the adaptive threshold signal applied to residual. The methodology of forming the envelope of uncertainty in the time domain in respect to fuzzy and neural models is intensively developed at the University of Zielona Góra [48][49][50][51][52]. Fault signaling takes place after exceeding by the residual value upper or lower envelope of the area of uncertainty -adaptation limit.…”

“…The approach based on determination of statistical error bounds has the greatest practical importance [7,[48][49][50][51][52][53]. The identification process is carried out without consideration of its uncertainty, while, in the second step, the model uncertainty is modeled (error model) based on residual signal.…”

An unscented Kalman filter in designing dynamic GMDH neural networks for robust fault detection

“…The errors following from an inappropriate selection of the neural network architecture and those related to inaccurate estimation of the parameters have deciding impact on the quality of the neural model resulting in the occurrence of model uncertainty Mrugalski and Korbicz, 2007;Patan et al, 2008). To tackle this problem, the Group Method of Data Handling (GMDH) approach can be employed (Ivakhnenko and Mueller, 1995;Korbicz and Mrugalski, 2008;Mrugalski and Witczak, 2012).…”

“…Also only few approaches ensure the stability of neural models during the process of dynamic system identification. Moreover, there is a limited number of approaches that allow a mathematical description of neural model uncertainty and this factor has major impact on the effectiveness of FDI and FTC systems (Witczak, 2007;Patan et al, 2008).…”