A contour error prediction method for tool path correction using a multi-feature hybrid model in robotic milling systems

Abstract

Achieving high precision in robotic milling presents significant challenges due to inherent errors caused by various factors such as robot stiffness deformation and uneven machining allowances in large workpieces. Traditional error corrected methods often fall short in effectively addressing the complexity and dynamic nature of such errors. To address these challenges, a contour error prediction model has been proposed by using a combination of Gaussian Processes and a CNN-BiLSTM architecture. Firstly, extract the potential error features, including the robot's posture and stiffness information, as well as the workpiece's machining allowance during the milling process. Then, process these features to create a uniformly structured training set. Subsequently, develop a CNN-BiLSTM neural network model to realize an accurate contour error prediction, where the CNN layers are responsible for extracting hidden local features from the structured data, while the BiLSTM layers capture temporal correlations and hidden features related to tool path. Finally, validate on a saddle-shaped workpiece with surface features similar to those found in aero-engine casing cavities. The results demonstrate that the fusion-based error prediction model effectively reduces the maximum contour error from 0.9629 mm to 0.4881 mm, and decreases the mean absolute contour error from 0.7171 mm to 0.3048mm, representing reductions of 49.30 % and 57.40 %, respectively. These reductions well validate the effectiveness of the proposed method.