Compensation of thermal deformations in machine tools with neural nets

Abstract

The accuracy of a milling machine is limited by thermal deformations. Although the deformations can be minimized by a good design, they cannot be completely avoided, and they depend on operating conditions. We show that a reduction and/or a compensation of these deformations is even more important when exploiting additional degrees of freedom, which is an additional motivation for our research efforts. A neural model was used to compensate thermal deformations of a five-axes milling machine. Neural nets were used because they can approximate complex multivariable non-linear relationships. First, a computer controlled test set-up was built to collect the training and test data. Then, a feedforward neural net was trained to estimate these deformations, given the temperature measurements. A backpropagation algorithm was used, modified with momentum and an adaptive learning rate. The maximum deformation of 150 μm was reduced to 15 u,m by using a single hidden layer with four neurons. Finally, the neural model was tested under operating conditions. Two workpieces were milled, one with and one without compensation. The neural model was able to reduce machining errors from 75 μm to 16μm With the presented approach, we managed to achieve successful results in a relatively short period, after first attempts had failed to reduce the deformations by eliminating the heat source.