A fault detection and isolation system using GMDH neural networks

This paper presents an approach to fault detection and diagnosis systems, which exploits the so-called Group Method and Data Handling (GMDH) algorithm. This algorithm can be considered as a structural identification technique or a feedforward neural network with a growing structure during the training process. Based on the GMDH algorithm, a knowledge-based fault detection and diagnosis system is proposed. The distinctive features of our approach are the insensitiveness to the influence of unknown inputs and high efficiency with lack of information regarding the structure and dynamics of the system being diagnosed. Our diagnostic system has been applied to failure monitoring tasks in the measuring electronic system of a dustmeter. A simulation study shows successful results for the proposed approach.