Unsupervised Anomaly Detection using Vibration Signals for Milling Processes

Tool condition monitoring has become essential to achieve high-quality machining as well as cost-effective production. Moreover, cutting tool degradation may vary considerably under different operating conditions and material behaviours. Therefore, real-time identification of the tool state during machining is critical. In this study, vibration responses of metal pieces during milling processes were measured with the help of a piezoelectric accelerometer. Milling experiments were conducted under different conditions, and the features were then calculated in time and frequency Although all the measures have been labelled, in this study we focus on unsupervised machine learning approaches based on non-labelled time series data, specifically on anomaly detection approaches, which are of high importance in industrial production. Therefore, this work concerns the automatic identification and classification of tool defects. The performances results obtained with different machine learning models are compared, depending on the sliding time window lengths. Unlike other studies described in the literature, more kinds of tool defects are identified in this work.