Monitoring robot machine tool sate via neural ODE and BP-GA

Abstract

Tool wear during robotic polishing affects material removal rates and surface roughness, leading to erratic and inconsistent polishing quality. Therefore, a method that can predict the tool state is needed to replace the robot end tool in time. In this paper, based on the cutting-edge neural ordinary differential equations (Neural ODE) and BP neural network optimization based on genetic algorithm (BP-GA), we propose a method to identify the tool state during robotic machining: firstly, a new training method of Neural ODE is proposed to avoid the model from falling into poor stationary points, and then on this basis, Neural ODE is utilized to predict the changes of vibration signals during robot machining; secondly, the predicted vibration signals of the tool are processed using variable modal decomposition method to extract the eigen kurtosis index and envelope entropy of the modal function as the vibration signal eigenvectors, and compare them with the traditional vibration signal eigenvectors. Finally, the predicted tool states were identified using BP-GA, and numerical experiments yielded an F1 score of 91.76% and an accuracy of 96.55% for model identification.