Predictive maintenance is essential for enhancing the reliability, safety, and operational efficiency of connected vehicles. However, traditional supervised learning approaches for fault prediction depend heavily on large, labeled datasets of failure events, which are often scarce and expensive to obtain in real-world automotive environments. This paper proposes an original self-supervised anomaly detection framework that predicts emerging failures using only healthy operational data, eliminating the need for explicit failure labels.
The approach employs self-supervised pretext tasks, such as masked signal reconstruction and future telemetry prediction, to learn robust representations of normal multi-sensor behavior (e.g., temperature, pressure, current, vibration). An unsupervised anomaly detection model then identifies deviations from these learned patterns. This integrated with data-informed predictive models enable early detection of faults across critical subsystems such as batteries, electric motors, brake systems, and cooling systems.
Experimental validation on public benchmark datasets demonstrates the framework’s ability to detect early-stage anomalies with high precision and recall, outperforming traditional threshold-based monitoring techniques. The study emphasizes the significance of leveraging healthy data to achieve scalable, adaptive, and cost-efficient predictive maintenance strategies for connected fleets. Furthermore, the proposed model enhances explainability by identifying key telemetry signals contributing to anomalies, supporting actionable and timely maintenance interventions.
This contribution presents a novel, practical pathway to proactive vehicle health management, improving fleet uptime while reducing unexpected failures and maintenance costs.