A Time-Series Data-Driven Method for Milling Force Prediction of Robotic Machining

Abstract

The machining chatter is the major factor that results in low dimensional accuracy, poor product quality, and even downtime when the industrial robots mill parts. Milling force is one of the highest responsive parameters that can depict whether machining chatter occurs and has widely been used for monitoring purposes. However, most milling force prediction methods focus on offline prediction, which cannot model and predict milling forces in real-time for complex systems with varying dynamics and poses, making it difficult to reflect the machining process of industrial robotics in time. To address the above challenge, a time series data-driven method is proposed for the milling force prediction of robotic machining, which explores two types of prediction modes based on Particle Swarm Optimization and Long Short-Term Memory Network (PSO-LSTM). The first is a sequence-to-sequence mode, called time interval prediction mode (TIP), which updates the network with actual values during deployment and can perform the next cycle prediction after one prediction step when milling starts. The second one is a point-to-sequence mode, named single-step cycle prediction mode (SCP), which updates the network with predicted values during deployment and only requires the offline optimized network model to predict milling forces at the beginning of milling. The PSO algorithm is utilized as the optimization component to determine the optimal hyperparameters for the TIP and SCP model, which is subsequently used for online prediction. Experimental validation was performed on a self-constructed robotic milling platform, and results indicate that the proposed approach performs well in predicting the milling force characteristics within the next 1 second of real-time milling.