A Data-Driven Methodology for Health Index Estimation in Internal Combustion Engine Maintenance

This thesis addresses predictive maintenance for spark-ignition internal combustion engines by proposing a practical, interpretable metric to quantify cumulative mechanical usage. In place of difficult-to-obtain Remaining Useful Life labels, we define a Comparative Fatigue Index (CFI): a time-integrated proxy of cyclic loads that combines a pressure-driven term—built from the in-cylinder peak pressure at its occurrence angle—and an inertial term that scales with engine speed. The index is normalized to a full-endurance reference, so its output expresses life consumption as a fraction of a validated durability test. This framing guarantees monotonic growth, enables like-for-like comparisons across duty cycles, and avoids heavy FEM/CFD prerequisites while remaining physically grounded. Because direct pressure sensing is not always available during bench campaigns, a virtual pressure sensor to estimate the peak pressure (PMAX) and its crank-angle location (APMAX) was developed. PMAX is reconstructed via a constrained five-dimensional map indexed by engine speed, throttle position, coolant temperature, intake air temperature, and the deviation between mapped and applied spark advance; APMAX is estimated with a two-dimensional map over speed and load. The mapping pipeline blends statistically reliable anchors, temperature corrections, and calibration-friendly binning aligned with ECU tables, limiting extrapolation and preserving traceability. The methodology is implemented in Python, with data extraction utilities that unify multi-rate signals from AVL/ETAS ecosystems into synchronized data frames, enabling large-scale analysis across historical test archives. The study focuses on Ducati single-cylinder off-road engines tested in an instrumented bench environment that controls load, speed, and ambient conditions. Results show that CFI coherently ranks operating regimes and test types by mechanical severity, reflecting known stress drivers (high load, high speed, adverse thermal/ambient states). Where sensorized runs exist, the virtual sensor reproduces pressure trends under varying temperature conditions with sufficient fidelity for fatigue accounting, supporting the use of pressure-based terms even when direct measurements are intermittent. The overall framework therefore offers an interpretable, data-leveraged pathway to compare usage patterns, support calibration and test planning, and monitor life consumption at component level using signals already available in development workflows and also on-board. Limitations and avenues for extension are discussed, including broader validation across engine families and the integration of additional health cues once consistent degradation labels become available.