A novel temporal convolutional network-transformer framework with assembly data augmentation for accurate micro-milling energy prediction

Precision micro-milling is widely used in advanced manufacturing, but accurately predicting its power consumption is challenging due to the high cost and time required to collect data under industrial monitoring conditions. This limits data availability and reduces the prediction accuracy of traditional machine and deep learning models. This study addresses the data-constrained prediction problem by proposing an assembly data augmentation strategy combined with a temporal convolutional network-Transformer (TCN-Transformer) model for micro-milling power consumption prediction. The assembly data augmentation treats time-series generative adversarial networks (TGAN), amplitude adjusted Fourier transform (AAFT) surrogates and Gaussian noise injection (GNI) as complementary parallel branches: TGAN learns the nonlinear distribution of power signals, AAFT preserves key spectral characteristics, and GNI enhances local variability. Their outputs are then combined with optimized weights to enrich the temporal and frequency information contained in small-sample datasets. On top of the augmented data, the TCN-Transformer model first uses TCN layers to extract multi-scale local temporal patterns and then applies a Transformer encoder to capture global dependencies between power trajectories and machining parameters, while Bayesian optimization is adopted to tune key hyperparameters. Validation on three-axis micro-milling experiments shows that the optimal augmentation ratio (50% TGAN, 25% AAFT, 25% GNI) significantly improves prediction and that the proposed TCN-Transformer achieves an R² of 0.896, outperforming competing baselines. These results demonstrate that the proposed framework effectively mitigates data collection constraints and provides a practical tool for non-invasive micro-milling energy prediction in precision machining.