A novel transfer learning model based on ED-TCN and RSD domain adaptation for thermal error prediction of multiple machine tools

High-precision machine tools are vital in modern manufacturing, yet their accuracy is often degraded by thermal errors. Traditional models lack cross-machine generalization and rely heavily on large labeled data. This paper proposes a thermal error modeling approach combining an encoder–decoder temporal convolutional network (ED-TCN) with representation subspace distance (RSD) transfer learning for cross-machine prediction. The encoder–decoder structure captures multi-scale features via dilated causal convolutions and residual blocks, enhancing long-term dependency modeling. The RSD-based domain adaptation reduces inter-machine distribution discrepancies while preserving feature scales. Through semi-supervised transfer learning, high-precision prediction is achieved using only 20% of labeled target data, greatly reducing collection costs. Experimental results on two different machine tools under three operating conditions demonstrate outstanding performance, achieving an R 2 of 99.5%, an RMSE of 1.201 µm, and an MAE of 1.008 µm, thereby confirming the practicality and robustness of the proposed method.