TEA-net: A multimodal deep learning framework for tool wear classification in biomedical machining

Abstract

In biomedical machining, proper tool wear monitoring is required as surface integrity and dimensional accuracy have a direct impact on the way of the implants to work and the safety of patients. Traditional monitoring methods that are based on single-modality data do not usually identify the presence of the subtle wear development in a complex cutting environment. In this research, the authors introduce a deep-learning methodology based on multimodality as Thermo-Edge Attention Network (TEA-Net) consisting of the fusion of thermal images, line maps, and statistical wear detection to obtain accurate tool-wear separation. The model uses multi-head attention to focus on local wear areas and combines similarities of complementary cues of various data modalities using a common fusion point. Experimental analyses on Austenitic Stainless Steel 316L (SS316L) and Zirconia (ZrO₂) machining tools show that TEA-Net can be evaluated at 82.5 and 88.4 classification accuracy on direct comparison with conventional machine-learning models, and with standard convolutional networks, respectively, with 15-percent higher accuracy. This framework also has a high capability of discrimination whose Area Under the Receiver Operating Characteristic Curve (AUC) exceed 0.97 therefore showing the reliability in both ductile and brittle materials. The findings show that multimodal integration has a significant positive effect on interpretability and prediction stability despite the scanty data quantity. TEA-Net is therefore an effective and timely solution to intelligent tool-wear saving and predictive maintenance in biomedical manufacturing sector.