Research on Fast On-line Calibration Algorithm for a Two-Stroke Kerosene Engine

Engine calibration poses a challenging multi-objective non-convex optimization problem due to its inherent complexity. In this study, we present a novel on-line engine calibration algorithm named Search Space Division-Momentum Gradient Descent (SSD-MGD), aimed at tackling this issue and reducing the calibration time. To commence, our investigation centers on a two-stroke kerosene engine, focusing on the calibration of spark timing and air-fuel ratio. We leverage experimental data to construct an engine response model utilizing support vector machines. This model serves as a virtual calibration test bench. The SSD-MGD methodology is then elaborated upon. It combines two key components: search space partitioning, which quickly identifies suitable initial points, and a momentum gradient descent algorithm, which solves for local optimal solutions from selected initial points. The two complement each other and we achieve a globally optimal solution. To evaluate the effectiveness of SSD-MGD, we performed a comparative analysis with traditional methods such as genetic algorithms in the context of virtual calibration tests. Our research results consistently show that SSD-MGD not only achieves the global optimal solution but also has significant efficiency.