Audio-Driven Unsupervised Learning for Tool Wear Detection in Milling Operations

Tool wear monitoring (TWM) is a pivotal task for ensuring process efficiency, product quality, and sustainability in manufacturing. Traditional supervised approaches for TWM rely on labeled datasets that require costly and time-consuming experimentation and annotation. This work proposes an unsupervised method for anomaly detection in audio signals recorded during the milling process. The aim is to build a model capable of recognizing anomalous observations associated with high tool wear, where timely replacement is needed. Audio features were extracted and analysed using two models: an autoencoder and an Isolation Forest. The evaluation was carried out by comparing reconstruction error distributions and by adopting threshold-based strategies for anomaly detection. Results show that the autoencoder outperformed the Isolation Forest, achieving effective discrimination between normal and worn-tool conditions without labelled data. These findings highlight the potential of unsupervised learning for scalable, cost-effective TWM in industrial milling processes.