Subsystem-oriented condition monitoring for connected vehicles using unlabelled data

Connected vehicles continuously generate large volumes of telemetry data, but it remains difficult to transform these streams into useful maintenance support when failure labels are scarce, degradation is gradual and heterogeneous, and operators require stable, interpretable and operationally manageable outputs. This paper presents a subsystem-oriented monitoring architecture for electric vehicles that is designed for this setting. Instead of collapsing the vehicle into a single global health score, the proposed approach monitors four key subsystems (battery, drivetrain, brakes and tyres/suspension) and computes for each of them a Subsystem Condition Indicator (SCI) referenced to subsystem-specific baseline behaviour. The SCI combines structural deviation, derived from reconstruction error, with short-term temporal deviation, derived from next-step forecasting error. It is then converted into stable alert states through fixed, baseline-derived thresholds, persistence logic and hysteresis. Thus, the architecture transforms unlabelled telemetry data into dashboard-oriented outputs that facilitate day-to-day monitoring, including subsystem condition, short-term trends, alarm states and concise reason-code-based cues to support initial inspections. A case study based on an electric-vehicle telemetry dataset shows that the proposed architecture preserves subsystem differentiation, produces interpretable and manageable alert behaviour, and remains stable under alternative baseline definitions and moderate missing-data perturbations, while also providing a basis for future extensions toward more prescriptive maintenance support.