Machine Learning-Based Anomaly Prediction for Smart Manufacturing

hanks to the widespread availability of sensor data, it is today possible to accurately predict anomalies in machinery functioning, preventing so potential breakages, downtime, and poor quality of products. In the case of punching machine, it is important to monitor the surface of the punch tool in order to detect abnormal incipient deformations. This paper addresses the problem of model building when only few punch-tool samples are available for model training. To this end, sample data are augmented by generating synthetic deformations and then using, hybridlike, both synthetic and real data for model training. The feature extraction process relies on the new concept of Profile Integration Matrix, which accounts for punch-tool surface deformations. Using the Profile Integration features, the predictive model is based on the supervised classifier one-class Support Vector Machine. The achieved results are promising, showing accuracy rates of 97.4% with hybrid data and of 97.7% with synthetic data.