Thermal error compensation of high-speed spindle system based on a modified BP neural network

Abstract

The accuracy, convergence performance, and robustness of the thermal error model based on traditional artificial neural networks (ANNs) are poor because the model is sensitive to the training data. To improve these performances, the genetic algorithm (GA) and particle swarm optimization (PSO) were used to optimize the parameters of ANNs with back propagation (BP) algorithm, such as the number of neurons in the hidden layer, initial weights, and thresholds. Moreover, the fuzzy cluster grouping and correlation analysis were combined to group and optimize the typical temperature variables to guarantee the robustness of the thermal error models based on BP, GA-BP, and PSO-BP neural networks. Then, thermal error compensation experiments were conducted on the spindle system of a precision jig borer, and the machining accuracy was increased from 67 to 78 % for the GA-BP model and 89 % for the PSO-BP model, respectively. To validate the effectiveness of the thermal error measurement, modeling, and compensation methods, the test samples were machined and the surface quality of the machined parts was measured. By measuring the dimension error and the surface quality of the machined parts, the results showed that the machining error can be reduced and that the surface quality can be improved by the thermal error compensation. Moreover, the accuracy, convergence performance, and robustness of the thermal error model based on the traditional BP neural network can be improved greatly by GA and PSO.