Unsupervised Machine Learning Methods for Predictive Maintenance

The growing complexity of industrial systems makes predictive maintenance increasingly difficult, especially in the absence of labeled failure data. Unsupervised learning addresses this challenge by modeling normal behavior and flagging deviations as potential faults. This work, Unsupervised Machine Learning Methods for Predictive Maintenance, explores both statistical and deep learning-based anomaly detection approaches using high-frequency, three-axis vibration data collected from an industrial production machine. We first extract statistical features from short vibration segments and apply Gaussian Mixture Model (GMM) clustering to identify typical patterns and detect deviations based on likelihood scores. In parallel, we develop a deep learning-based forecasting model using the recently proposed Mamba architecture. The model predicts future vibration signals in a sequential manner, and anomalies are detected through a smoothed forecasting error signal, combined with both static and dynamic thresholding strategies. To evaluate the performance of these methods, we use a calibration-validation-degradation (test) split that simulates real-world deployment, where only unlabeled data is available during training and testing. The results show that forecasting-based deep learning offers improved adaptability and robustness in detecting evolving machine faults, while the statistical clustering approach provides interpretable baseline performance. Together, these methods form a practical and scalable framework for predictive maintenance in industrial environments.