Data-Driven Control of Heavy- and Medium-Duty Diesel Engines: from Classical Torque Tracking to Neural-Surrogate Reinforcement Learning for Coordinated Variable Geometry Turbocharger and Exhaust Gas Recirculation Optimization

This dissertation develops and evaluates data-driven control methods for transientheavy- and medium-duty diesel-engine operation on two production engines.On a heavy-duty Cummins ISX15, advanced classical methods recover substantialtransient torque-tracking accuracy over a baseline Proportional-Integral-Derivative (PID)controller. These methods add feedforward action and gain scheduling to a tunedproportional-integral loop. The accuracy is measured by the 40 CFR 1065 regressioncriteria that decide whether a certification test is valid. A recurrent Long Short-Term Memory (LSTM) surrogate of the same engine is thenidentified from test-cell data and used as a training environment for a continuous-actionreinforcement-learning controller trained with Twin Delayed Deep Deterministic PolicyGradient (TD3). By training against the learned model rather than a manually calibratedfeedforward map, the learned controller improves further on the best classical design,roughly halving trace-tracking error on the same certification metrics.The same surrogate-plus-reinforcement-learning method is then extended to a coupled,multi-objective problem: coordinating the variable-geometry turbocharger andhigh-pressure exhaust-gas recirculation of a medium-duty Caterpillar C9 to lower nitricoxide and particulate matter while holding fuel consumption, against a hand-designedproactive actuator-schedule library studied as part of Pennington's doctoral work. On therapid load transient that most degrades emissions, the learned closed-loop policy improveson that library. Because the policy is trained against a learned model, its commands arekept within the range the engine actually demonstrated, so the predicted improvementsstay physically realistic. In each study, bench data are used to train an LSTM surrogate, and the TD3 controlleris trained inside that surrogate. The method is first used for heavy-duty torque tracking and then extended to medium-duty multi-objective emissions control, improving on theclassical and hand-tuned methods it replaces.