Prediction of Thermally Induced Time-Variant Machine Tool Error Maps Using a Fuzzy Artmap Neural Network

Abstract

A novel approach to on-line learning and prediction of time-variant machine tool error maps is proposed. These error maps are measured using a fast calibration device called the laser ball-bar (LBB) that directly measures the total positioning errors at the cutting tool using trilateration. The learning and prediction of these error maps is achieved using a Fuzzy ARTMAP neural network by treating the problem as an incremental approximation of a functional mapping between thermal sensor readings and the associated positional errors at each location of the cutting tool. Experimental measurements of the positional errors for a two axis turning center were performed using the LBB over two separate thermal duty cycles. The Fuzzy ARTMAP was trained on-line using the data collected over the first thermal duty cycle, which simulated machining of large workpieces with several hours of machining, inspection and set-up time. The network was made to predict the error map of the machine for a new thermal duty cycle that simulated machining of a range of short and long workpieces with shorter machining and set-up times. Results of these predictions show that the LBB and Fuzzy ARTMAP combination is a fast and accurate method for real-time error compensation in machine tools. This method overcomes drawbacks in currently methodologies including high cost and excessive downtime to calibrate machine tools. Application of the Fuzzy ARTMAP to continuous process improvement is discussed.