Neuro-Fuzzy and Regression Techniques for CNC Thermal Error Compensation

Abstract

This paper examines the capabilities of Neuro-Fuzzy and Regression techniques for compensating thermally induced errors on Computer Numeric Control (CNC) machine tools.

Thermally induced errors occur during CNC machining and make a substantial contribution to the total error in the process. Although CNC machine tools can be designed to reduce thermal distortion, compensation for that remaining is required. The axes of the CNC machine are positioned by the controller during machining. If the errors are known they may be added to the positions in the controller thus removing inaccuracy. Absolute measurement of positional errors is difficult during machining but it has been found possible to derive positional error from temperature measurements.

The use of Neuro-Fuzzy and Regression techniques to model the thermally induced volumetric errors in CNC tools has been investigated in this work. We examine the capability of the two techniques to produce control outputs from raw data obtained from experiments carried out on a well-known CNC machine.

An Adaptive Neuro-Fuzzy Inference System (ANFIS) technique is used to derive compensation values for thermally induced errors from temperature readings. ANFIS is a method for tuning Sugeno fuzzy inference systems (FIS), which may be used to model nonlinear systems by interpolating multiple linear systems. A FIS is developed for each axis using two temperature readings and the deviation as inputs, with equivalent data from a second test being used for checking purposes. The number and type of membership functions for each axis were found by a heuristic approach, FISs being produced for a range of number and types of MF. The process of identifying the most effective FIS is described.

Multiple linear regression is also used to model the system. The generalising capability of each technique is compared in terms of the effectiveness of the model at reproducing the checking data.