This thesis explores how machine learning can be applied to predictive maintenance in an industrial context, with a focus on accelerometer-based vibration analysis.

The study is grounded in the increasing digitalization of manufacturing and the need for efficient, data-driven fault detection strategies. A scoping review was conducted to examine recent literature on supervised and unsupervised algorithms used in similar contexts. Based on these insights, a prototype system was developed using an STWIN sensor to collect vibration data.

Two experiments were carried out: one in a controlled setting and one in a real industrial environment at Magna. The data were analyzed using Fast Fourier Transform (FFT), and relevant features were extracted for modeltraining. Random Forest (RF) was used as a supervised classification model, while Principal Component Analysis (PCA) was applied for unsupervised visualization. The combination of RFand PCA provided both accurate and interpretable results – an important prerequisite for implementation in production environments.

These findings suggest that hybrid strategies, tailored to data availability and industrial constraints, hold strong potential for robust predictive maintenance systems.