Tool Condition Monitoring Based on an Adaptive Neurofuzzy Architecture

Metal cutting operations constitute a large percentage of the manufacturing activity. One of the most important objectives of metal cutting research is to develop techniques that enable optimal utilization of machine tools, improved production efficiency, high machining accuracy and reduced machine downtime and tooling costs. Machining process condition monitoring is certainly the important monitoring requirement of unintended machining operations. A multi-purpose intelligent tool condition monitoring technique for metal cutting process will be introduced in this paper. The knowledge based intelligent pattern recognition algorithm is mainly composed of a fuzzy feature filter and algebraic neurofuzzy networks. It can carry out the fusion of multi-sensor information to enable the proposed intelligent architecture to recognize the tool condition successfully. Introduction Traditional tool change strategies are based on conservative estimates of tool life from experience or past tool data. In the unmanned machining environment, this will result in frequent tool changes and high production costs. Automatic machining operations rely greatly on the ability to monitor the tool wear states. In an unmanned production environment, this function can be performed by an integrated system composed of sensors and intelligent signal processing algorithms. The research work of S. C. Lin and R. J. Yang [1] showed that both the normal cutting force coefficient and the friction coefficient could be represented as functions of tool wear. An approach was developed for in-process monitoring tool wear in milling using frequency signatures of the cutting force [2]. An analytical method was developed for the use of three mutually perpendicular components of the cutting forces and vibration signature measurements to monitor tool wear[3]. A tool condition monitoring system was then established for cutting tool-state classification using a single wear indicator [4]. In another study, the input features were derived from measurements of acoustic emission during machining and topography of the machined surfaces [5]. X. Li, etc. showed that the r.m.s. of the different frequency bands of vibration measured indicates the tool wear condition [6]. Tool breakage and wear conditions were monitored in real time according to the measured spindle and feed motor currents, respectively [7]. Advanced signal processing techniques and artificial intelligence play a key role in the development of modern tool condition monitoring systems. Sensor fusion is also attractive since loss of sensitivity of one of the sensors can be compensated by other sensors. A new on-line fuzzy neural network (FNN) model was developed with the functions of classifying tool wear [8]. A advanced approach for online and indirect tool wear estimation in turning using neural networks was developed [9]. Two methods using Hidden Markov models, as well as several other methods that directly use force and power data were used to establish the health of a drilling tool [10]. In this study, a unique neurofuzzy network based pattern recognition algorithm has been developed to accomplish multi-sensor information integration and tool state classification. The established monitoring system provides accurate and reliable tool wear classification results over a range of machining environments. Materials Science Forum Vols. 471-472 (2004) pp 196-200 online at http://www.scientific.net © (2004) Trans T ch P blications, Switzerland Online available since 2004/Dec/15 All rights reserved. No part of contents of this paper may be reproduced or transmitted in any form or by any means without the written permission of the publisher: Trans Tech Publications Ltd, Switzerland, www.ttp.net. (ID: 130.203.133.33-14/04/08,13:51:04) Materials Science Forum Vols. *** 197 Tool Condition Monitoring System and Signal Feature Extraction The tool wear monitoring system is composed of four types of sensors, signal conditioning devices and the main computer, as shown in Fig.1. This is a multi-purpose condition monitoring system. By properly combining the four sensors, the system can effectively monitor the tool condition for turning, drilling, milling processes. Fig.1 The tool condition monitoring system The original signals have large dimensions and can not be directly used to estimate tool wear value. The purpose of feature extraction is to greatly reduce the dimension of the raw signal but at the same time maintain the tool condition relevant information in the extracted features. This step is the foundation for the pattern recognition process. A typical group of features extracted from the time domain and frequency domain for the further pattern recognition are as follows. Power consumption signal : mean value; AE-RMS signal: mean value, skew and kutorsis; Cutting force, AE and vibration : mean value, standard deviation and the mean power in 10 frequency ranges. Fuzzy Clustering Feature Filter Up to four kinds of signals are used to describe tool condition comprehensively, so that the tool wear recognition results can be more robust. But this comprehensiveness also means a comparatively large number of input features for the recognition neurofuzzy networks. In order to improve the efficiency and reliability of the neurofuzzy pattern recognition algorithm, redundant features should be removed. Here the fuzzy clustering technique is applied to develop an effective fuzzy feature filter. A set of n data samples to be classified is defined as { } n X X X X ,...., , 2 1 = . Here { } im i i i x x x X ,...., , 2 1 = in the universe X is a m dimensional vector of m features. Those m features have different units so each of them must be normalized to a unified scale before classification. Bezdek [11] suggested using an objective function approach for clustering the data into hyperspherical clusters. In this study a fuzzy clustering feature filter has been developed to remove the redundant features. Using fuzzy clustering method, signal features of an object (the tool being monitored) should be classified into one of the models (tools with known wear values) they actually belong to. Those features whose variations of cluster centers for different models are smaller than a threshold 1 ε , or whose variation of practical values corresponding to their own cluster center are bigger than a threshold 2 ε should be removed. This is because they don't have recognizable variation along with the development of tool wear values or they don't have stable values. After some of the redundant features have been removed, the clustering process is carried out again. The value of the objective function should drop and the membership value with which an object is classified into the model to AE Sensor KISTLER 9257B Dynamometer