Anomaly detection in drilling using neural networks

Abstract

With increasing competitive pressures, manuf acturing systems in the automotive industry are being driven more and more aggressively. The pressures imposed on the processes and lack of system 'slack' have led to increased use of Tool Condition Monitoring systems. In parallel, there has been wide-ranging research in academ ia. However, a closer examination shows that there has been very little migration of this re search into industrial pr actice. Furthermore, the success of industrially deployed monitoring systems has been poor. It has been suggested that a significant factor behind both these phenomenon ha s been the 'difficult' environment in which such systems must operate; an en vironment where they are subject to many stochastic influences, ranging from ambient conditions, to user input, to workpiece consistency. Neural networks have found increasing favour in manufacturing systems research because of their ability to perform robustly in noisy envi ronments. Almost all the applications of this technology in tool condition monitoring have been in the detection/prediction of tool wear. From an academic standpoint, it may be speculated that the lack of focus on breakage and missing tool detection has been due to the relatively trivial nature of detecting such anomalies in the laboratory environment. However, detection in the production environment is compromised by a wide range of factors, which can give rise to fa lse alarms when such strategies are transported from laboratory conditions. In this paper, data from a real manufacturi ng process is used to demonstrate the potential application of neural networks to the task of anomaly detection in the production environment.