Data-driven Based Engine Air-fuel Ratio Predictive Control

Air-Fuel Ratio (AFR) Control is one of the basic control in engine control system,which afects the engine performance signifcantly. AFR impacts on engine power, econo-my and emission performance directly, so it obtains many attention of experts. Howeverthe control performance under engine transient operation still represents a challengingproblem for researchers. The AFR control is achieved by regulating the fuel injectionwith respect to the air fow mass into the cylinder. We can consider it as a Air-Fuel“Master-Slave”control. The air intake system satisfes the engine output requirement,while the fuel control as the “Slave”to air-intake control. Due to the complexity andthe parameter uncertainty of the engine model, the workload of engine modeling is heavy,which requires many calibration experiments. What’s more, the control performance ofmodern control algorithms is dependent on the precision of the plant model. That’s maybethe reason for why the AFR control accuracy is relatively not good in the engine transientconditions. Fortunately, by using the data-driven approach，the modeling problem canbe avoided in the controller designing process. By combining the subspace identifcationwith Model Predictive Control(MPC), a data-driven control algorithm is created in thisthesis. Finally, the simulation experiments is conducted to verify the efectiveness of thealgorithm.This thesis mainly includes the following three parts:(1) Design a controller for the Port Fuel Injection Spark Ignition (PFISI) gasolineengine. Although the data-driven approach doesn’t require a precise mathematical plantmodel, in order to excite the system sufciently, and get the proper dynamic and stableresponses of the system, we need to understand the principles and mechanisms of ourobject sufciently, so that the exciting inputs can be chosen properly. In this paper,the air-path subsystem and fuel-path subsystem of the AFR model are studied and theircharacteristic are analyzed respectively, and then the dynamics of whole AFR system isanalyzed. In addition, By taking the difculty of the acquisition of real engine benchdata, and the rationality and feasibility of the exciting inputs into account, the excitinginputs are chosen fnally. Then by utilizing these inputs work on the enDYNA high-precise virtual engine, the system I/O data was obtained for designing controller.(2)Based on the of-line exciting data, the data-driven controller is designed. Byusing the subspace identifcation method, the system output predictive equation is derivedfrom the exciting data directly. Further, using the concept of model predictive control,a data driven model predictive controller is proposed. As the element of the optimalproblem solution of MPC is correspond to the exciting inputs one-to-one, however theexciting inputs are composed of control input and system states. In this thesis twomethod is proposed to deal with this problem. The frst method is achieved by extendingthe constraint conditions of MPC optimization problem to handle with system stateselements of the exciting inputs. The second one is achieved by improving the controlalgorithm. The system states are extracted from the exciting inputs before calculatingthe predictive equation, so the control algorithm can deal with the system states (throttleangle, engine speed). Finally, the control action is obtained by solving the quadraticprogramming (QP) problem using Particle Swarm Optimization(PSO) algorithm. For itseasy implementation, fast convergence and high accuracy, PSO attracted many academicattentions.(3) In order to verify the proposed controller, several of-line simulations with thevirtual enDYNA engine plant in diferent engine operations are conducted. Finally, basedon the XPC-target and dSPACE, the efectiveness of controller is validated in realtimeby the Rapid Control Prototyping (RCP) experiments.Due to the complexity of the engine model, the data-driven approach in this thesis arechosen, and it provides a new novel approach for engine control. Some work needs to becarried out for improve this method further. Although the object is a virtual engine, theexciting inputs should be feasible and some signal processing is required for real enginebench application. fnally, for optimal problem solving, the computational speed andaccuracy will be a challenge to the real ECU, that requires powerful and high-speed chip.