AN INTELLIGENT FAULT DIAGNOSIS SYSTEM FOR MACHINING

Abstract

Modern manufacturing technology leads to the installation of a monitoring and control system in an industrial manufacturing process. This system will increase the effective machining time of a machine tool and prevent the occurrence of the faults. In this study, a framework of an intelligent fault diagnosis system (IFDS) for a turning operation is presented. Main works and conclusions are as follows: The signal features are extracted by a wavelet technique involving the physical properties of the cutting process in the time and frequency domain synchronously. Thus a multi-feature approach for diagnosis is proposed and implemented based on the wavelet technique. Comparing multi-sensor and multi-mode methods with the multi-feature method, the latter takes less diagnosis time and is more feasible for on-line diagnostic applications. A new tool life criterion depending on a pattern recognition technique is proposed. The new tool life criterion represents the overall tool wear condition and can be realized by the wavelet and neural network techniques. The experimental results show when a sensitive sensor and a valid signal processing technique are used, the multiple-feature vector extracted from the sensor can monitor tool wear according to this criterion in a wide range of cutting conditions. A supervised back-propagation neural network and an unsupervised Kohonen neural network are used in the IFDS respectively. Both neural networks have a high diagnosis success rate. A supervised neural network can monitor multiple cutting states synchronously but can not recognize new kinds of patterns which are not used for training. The network must undergo complete retraining in order to learn the new kinds of patterns. On the other hand, an unsupervised neural network can monitor the change of the cutting state but can not predict how the cutting state changes. When this main problem is solved, from the point of view of a practical application, the unsupervised neural network has greater potential in a real world setting. Thus, finally, an improved IFDS with a Kohonen neural network is presented. An additional interpretation function is added in the IFDS. Under a wide range of cutting conditions, this system can detect the change of the cutting state and the additional interpretation function can explain how the cutting state changes. It solves the main problem encountered in the application of a supervised neural network.