Fault diagnosis based on the quality effect of learning algorithm for manufacturing systems

Abstract

Fault diagnosis is becoming an important issue in industrial environment, and the accurate diagnosis is the most significant part in fault handling. This article discusses a fault detection and diagnosis problem for manufacturing systems taking into account rapid detection and speed performances of fault isolation with minimum ambiguity. However, in many complex real plants, it may not be possible to discover accurately the causes of probable faults. The accuracy of fault or fault detection by the traditional approaches is not adequate. Considering the quality effect of the learning algorithm, a new hybrid neural network approach is developed using the integration of a regression task for classification accuracy. Two models of neural networks: gradient descent and momentum & adaptive learning rate and Levenberg–Marquardt are investigated and compared. The performance of the proposed approach is evaluated using mean square error, convergence speed, and classification accuracy. The case study and experimental results are presented and discussed. A comparison with the Levenberg–Marquardt regression approach shows the importance of considering the proposed learning algorithm quality in the fault detection and diagnosis problem compared with those reported in the literature.