A positioning error prediction method for complex surface milling based on multi-source heterogeneous data fusion and deep interpretable learning

Abstract

The machining of thin-walled parts with complex surfaces, such as aero-engine blades, poses substantial challenges due to their intricate geometry, flexibility, and the multi-axis kinematics involved. Ensuring high precision in such milling processes is crucial, as minute positioning errors can detrimentally affect aerodynamic performance and structural reliability. In this study, we propose a novel Residual-LSTM-iTransformer Network (RLTN) framework to predict and interpret machining positioning errors for complex surfaces under dynamic milling conditions. The RLTN model fuses multi-source heterogeneous data, including low-cost sensor signals (e.g. milling torque), machining parameters (spindle speed, feed rate, depth of cut, etc.), and workpiece attributes (local stiffness and curvature radius), into a unified deep learning architecture. Through hierarchical feature extraction via residual convolutional layers, sequential LSTM units, and a parameter-conditioned transformer, the model captures both local and long-range dependencies in the cutting process. A SHAP-based interpretability module is integrated, enabling quantitative attribution of error predictions to the process and material parameters. The RLTN is evaluated on high-precision blade milling datasets encompassing various cutting conditions. Experimental results demonstrate that the proposed approach outperforms conventional physical models and baseline learning methods in prediction accuracy. Moreover, the RLTN provides deep insight into the influence of key factors on machining errors, facilitating a better understanding of the error-generation mechanism. This interpretable framework paves the way for a closed-loop “predict-interpret-optimize” strategy in high-precision manufacturing, allowing process parameters to be optimized not only for minimum error but also for reduced uncertainty and improved consistency in manufacturing thin-walled parts with complex surface.