Real-time monitoring of coated tool wear based on the DSAMM-Transformer architecture: an experimental study

This study proposes a DSAMM-Transformer architecture to address the limitations of traditional methods in early tool wear detection and the high computational costs of deep learning models. The method combines dynamic multimodal feature fusion (vibration, current, and acoustic data) and a dynamic sparse attention mechanism (\({\boldsymbol{O}}( {{\boldsymbol{n}}\sqrt {\boldsymbol{n}} } )\) complexity). Experiments on industrial datasets show stable MAE (0.027), 55% lower inference latency (9 ms), and 32% faster training and 0.75 GB memory usage, outperforming LSTM/1D CNN. Optimized for edge deployment in smart manufacturing.