Recently water injection for engine has been developed quickly for F1 racing cars since 1990's. It has been shown experimentally that engine knocking can be reduced to a minimum and emission is cleaner with less petrol consumption and more torque can be generated since there is no retarded ignition angle and reduced boost pressure. However, they all adopted fixed water mapping in their controllers without considering the dynamic driving environment. Thus the drivability is not smooth under unexpected conditions.
This research proposes a new architecture of Self-Learning Fuzzy-Neural-Network(SLFNN) for water injection control in turbo-charged automobile. The major advantage of SLFNN is that no off-line training is needed for initialization. The SLFNN will initialize itself with a random set of initial weighting factors.
In this research, we adopt the FNN architecture to create the new SLFNN. The SLFNN will pick up the knocking signal from engine ECU and initiates a dynamic optimal training to guide the weighting factors toward a maximum error reduction direction. In SLFNN architecture, Manifold Absolute Pressure and RPM signals will be fuzzified to produce the fuzzy rule. The SLFNN controller will output final water injection ratio signal to control the water injection in a turbo-charged engine.
Real implementation of SLFNN for water injection control has been done in Saab 900 NG (1994~1998). With the SLFNN control, the experimental results show that car performance(horsepower and torque, speed, acceleration of Gravity) increase.