A Predictive Model on Percolation Theory and Machine Learning

Abstract

Early detection of vehicular faults remains a major challenge in the development of predictive maintenance strategies. Traditional diagnostic systems based on OBD2 (On-Board Diagnostics II) data are commonly used in a reactive manner, identifying failures only after they have already affected vehicle performance. To address this limitation, a predictive model based on percolation theory, OBD2 data, and Machine Learning is proposed. Real-world data were collected from an internal combustion vehicle using an ELM327 OBD2 scanner under different operating conditions. Next, a functional network was constructed in which the sensors were represented as nodes and their statistical relationships as edges. Percolation theory was applied to evaluate network robustness and estimate the critical connectivity threshold associated with system stability. In parallel, unsupervised learning techniques, particularly DBSCAN and Isolation Forest, were used to identify patterns and detect anomalous observations in the sensor data. The results indicate that the integration of network-based structural analysis and Machine Learning provides a valuable framework for the early detection of incipient vehicular faults and supports the design of data-driven predictive maintenance strategies.