Classification of dental drill wear status based on acoustic monitoring and deep learning

High-speed dental drill over-wear may cause operative site overheating and thermal injury to bone tissue. The current maintenance and replacement of dental implant drills in clinical practice still rely heavily on the experiential judgment by the operators, lacking objective standards. To develop and validate an intelligent assessment method based on audio signals and deep learning to predict dental drill wear status, we first systematically collected operational audio signals of dental drills across different usage iterations (i.e., different wear stages) and materials to construct an acoustic feature dataset. Subsequently, we extracted and compared various time-domain and frequency-domain features (including LOFAR spectrograms and Mel spectrograms) and employed multiple deep learning models (including the ResNet series and EfficientNet) to train and validate the classification of wear status. Finally, based on the selected optimal model, we developed an intelligent clinical decision support software system. Experimental results demonstrate that the frequency-domain features of audio signals exhibit a high correlation with the drill wear status. The ResNetV2_50 model based on LOFAR spectrogram features demonstrated the best comprehensive performance, achieving a classification accuracy of over 95% on single-material datasets and maintaining a robust accuracy of over 84% on the more challenging multi-material mixed datasets. This study confirms the high feasibility and precision of utilizing audio signals for the intelligent classification of dental drill wear status, which provides an objective, convenient, and real-time instrument assessment way for standardizing instrument management and preventing intraoperative thermal injury.