This paper explores a strategy for fault detection and recognition in a high-production horizontal broaching machine used for manufacturing lock components. The experimental campaign involved inducing artificial faults and measuring vibration signals by means of four triaxial accelerometers. The main challenges included the unknown elastodynamic response of the machine and limited data availability due to time constraints. To overcome these issues, statistical parameters were computed from shorter portions of measurements, and datasets were expanded by combining these portions through computation of cross-correlation functions. Feature vectors were extracted and labelled according to the tool’s health state. Supervised classification algorithms, including K-NN, SVM, ANN, and Random Forest, were applied. Despite the severe artificially induced faults, the method successfully distinguished between healthy and faulty states, though further investigation is needed to enhance accuracy for realistic faults.
Rosa, S., Martini, A., Rivola, A., Troncossi, M. (2024). Condition monitoring of a broaching machine through machine learning classification of a limited dataset vibration signals. Leuven : KU Leuven, Departement Werktuigkunde.
Condition monitoring of a broaching machine through machine learning classification of a limited dataset vibration signals
Rosa S.;Martini A.;Rivola A.;Troncossi M.
2024
Abstract
This paper explores a strategy for fault detection and recognition in a high-production horizontal broaching machine used for manufacturing lock components. The experimental campaign involved inducing artificial faults and measuring vibration signals by means of four triaxial accelerometers. The main challenges included the unknown elastodynamic response of the machine and limited data availability due to time constraints. To overcome these issues, statistical parameters were computed from shorter portions of measurements, and datasets were expanded by combining these portions through computation of cross-correlation functions. Feature vectors were extracted and labelled according to the tool’s health state. Supervised classification algorithms, including K-NN, SVM, ANN, and Random Forest, were applied. Despite the severe artificially induced faults, the method successfully distinguished between healthy and faulty states, though further investigation is needed to enhance accuracy for realistic faults.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.