Comparative study of ANN and ANFIS predprediction models for thermal error compensation on CNC machine tools (English)

In this paper, various traditional ANN and hybrid ANFIS models were proposed for the prediction of the thermal errors on a CNC machine tool, and the following conclusions can be drawn: The new technique GM (0, N) model is able to find the optimal temperature sensors for thermal error modelling. The advantage of using key temperature sensors is greater economic efficiency and reduces modelling time. In addition, the robustness of the models can be increased and the predicting precision of the models using the optimal combination of the temperature sensors are enhanced. Experimental results show that the thermal error in the Y direction can be significantly reduced to less than ±5 µm using ANFIS models with validation data (different conditions of rotational speeds on the machine tool). The results also show that the ANN models can reduce the thermal error to less than ±10 µm. However, BP algorithms are limited during the training procedure, in that they are sensitive to the weight's initial values. If the weights are not correctly chosen, the training process might be cornered in a local minimum or maximum. In contrast, while a PSO algorithm gives unsatisfying results, further investigation is necessary in order to overcome some of the limitations of BP. A PSO algorithm also gives fast convergence during the training stage.