Manifold absolute pressure estimation using neural network with hybrid training algorithm

Area of Science:

• Automotive Engineering
• Artificial Intelligence
• Control Systems

Background:

• Manifold Absolute Pressure (MAP) sensors are crucial for estimating engine load in modern gasoline fuel injection systems.
• Current MAP estimation methods rely on dedicated sensors, increasing implementation costs.
• There is a need for more economical and efficient MAP estimation techniques.

Purpose of the Study:

• To develop a cost-effective approach for estimating MAP using only throttle position and engine speed.
• To evaluate the performance of a hybrid neural network model for MAP estimation.
• To validate the proposed method in both simulated and real-world retrofit fuel injection systems.

Main Methods:

• A two-stage multilayer feed-forward neural network was employed for MAP estimation.
• Hybrid algorithms combining Levenberg-Marquardt (LM), Bayesian Regularization (BR), and Particle Swarm Optimization (PSO) were utilized.
• The performance was evaluated using simulated data and validated with experimental training data from a retrofit fuel injection system (RFIS).

Main Results:

• The second variant of the hybrid algorithm demonstrated superior network performance compared to other tested algorithms (first hybrid variant, LM, LM with BR, PSO).
• The trained estimator network closely estimated simulated MAP values.
• The estimator showed excellent performance in an actual RFIS, accurately predicting MAP under steady-state and transient conditions.

Conclusions:

• The proposed hybrid neural network approach offers a more economical and effective alternative for MAP estimation in gasoline engines.
• This method reduces reliance on traditional MAP sensors, lowering implementation costs.
• The validated performance in RFIS highlights its practical applicability for improving fuel injection systems.