Characterising domain gaps in milling tool wear prediction: the primary bottleneck for scalable Digital Twins

Scalable Digital Twin tool health monitoring requires transfer learning across heterogeneous machine configurations, yet the structural domain gaps preventing such generalisation remain poorly characterised. This study identifies three gap categories (sensor dimensionality, target variable semantics, and sensor modality) as the primary bottleneck, operating at the data harmonisation layer before model inference. These gaps are quantified through 78 classification experiments across three public milling benchmarks using convolutional and recurrent neural network architectures under three training paradigms, complemented by regression and multi-channel analyses. The results reveal a frequency-alignment asymmetry: models trained on high-frequency data (25–50 kHz) transfer effectively to low-frequency targets (250 Hz) with less than 5% accuracy loss, whereas the reverse fails catastrophically (over 40% degradation). Fine-tuning improves 67% of transfer pairs but leaves persistent 10–20% gaps to within-domain baselines. Architecture-target interactions further complicate generalisation, as recurrent models fail on datasets with globally defined indirect wear targets. This study does not implement a complete Digital Twin system; rather, it characterises the structural data harmonisation prerequisites that any scalable twin deployment must overcome.