<p>This study presents a cost-effective predictive maintenance framework for CNC milling machines using vibration data captured from the MPU6050 sensor. The system relies on an unsupervised Autoencoder neural network to model the machine’s normal operational behavior. Vibration data, collected across three axes (X, Y, Z) using Arduino UNO and stored in CSV format via Python scripting, was used to train the model. The Autoencoder reconstructs the input signals and identifies anomalies based on reconstruction error. A statistical threshold was applied to detect deviations that may indicate wear, imbalance, or early-stage faults. The model achieved high precision with minimal false alarms, although the recall was relatively lower, highlighting the challenge of detecting all fault conditions with limited data. This approach enables proactive fault detection without requiring labelled fault data, making it ideal for real-world deployment in small and medium-sized manufacturing setups.</p> Graphical abstract <p></p>