Study on momentum gradient descent on-line engine calibration algorithm combined with search space division

Abstract

Because of the control problem complexity, engine calibration becomes a multi-object nonconvex optimization problem (MONCOP). To solve this problem and shorten the calibration time, this study proposes an innovative on-line engine calibration algorithm which combines momentum gradient descent algorithm with search space division (MGD-SSD). Firstly, taking a two-stroke kerosene engine as the research object, the engine response model is built by support vector machine (SVM), and based on the analysis of the spark timing and air-to-fuel ratio calibration problem, MAPs of objective function value under three typical engine operation conditions are established. Then the design of MGD-SSD is introduced in detail. The search space division is used to find the appropriate initial point quickly and the momentum gradient descent algorithm is used to quickly find the global best point through the appropriate initial point. A comparative study on three methods including genetic algorithm (GA), genetic and gradient descent hybrid algorithm (GGD) and MGD-SSD is carried out through benchmark function test and virtual calibration test. All tests show that the MGD-SSD is able to find the global optimal solution with higher accuracy and fewer iterations. Finally, the bench test of MGD is carried out, and the validity of the algorithm is verified.