Real-time fault detection and isolation in industrial machines using learning vector quantization

This paper presents a real-time approach to the problem of fault detection in industrial machines. In this study learning vector quantization (LVQ) of adaptive neural networks has been used to identify faults by relating patterns of fault signatures to their causes. The fault detection system is designed to operate based on steady state and transient responses obtained by monitoring sensitive parameters of an industrial machine. The steady state signal acts as a stimulus. When the signal exceeds a predetermined threshold it will initiate a non-destructive test on the machine during which a transient response of a second sensitive parameter will be captured. This transient pattern will be compared with the database of the patterns of fault signatures. The closest match will determine the cause of fault. The technique has been applied to a computer numerically controlled (CNC) machining centre. The diagnostic system is shown to be capable of first deciding whether the system is healthy or faulty; if faulty, it then decides whether one of the known faults or a novel fault, not seen before, is occurring. Having made the decision that one of the common faults is occurring it is then capable of deciding, from four different levels, the approximate severity level of the fault. In this approach, the use of LVQ significantly reduces the training time period of the network by about 90 per cent as compared to other learning methods such as back propagation (BP).