Physics-Informed Diagnostics for Food Engineering Technology Assets in Kenya: Out-of-Distribution Fault Detection and Lifecycle Cost Trade-offs

Food processing assets in Kenya operate under variable supply quality, intermittent power, and maintenance regimes that diverge from manufacturer specifications, yet conventional data-driven fault diagnostics assume training and operating conditions are identically distributed. The framework formalises a lifecycle cost model in which false alarms impose unnecessary maintenance expenditure and missed detections accelerate capital depreciation, and it derives an optimal detection threshold that balances these competing costs under distributional uncertainty. A comparative analysis of purely data-driven, purely model-based, and hybrid architectures demonstrates that the physics-informed approach achieves superior fault identifiability when training data are scarce, which is the prevailing condition for many Kenyan food enterprises.