Neural network modeling of engine brake thermal efficiency based on heat flow evolution

To address the degradation - induced brake thermal efficiency (BTE) declines in internal combustion (IC) engines, an investigation is undertaken into the mechanistic linkage between heat flow evolution patterns and BTE degradation in IC engines. Physics - based efficiency mapping is established via critical thermal parameter analysis. A neural - network - based BTE prediction model for steady - state operating conditions integrating adaptive particle swarm optimization with multilayer perceptron (APA - PSO - MLP) is developed. Experimental data are acquired through diesel engine bench tests and then preprocessed using steady - state screening and interquartile - range - based (IQR) outlier cleansing. A tri - method correlation analysis (Pearson/Spearman/Kendall) is employed to identify optimal input variables. The Kendall method demonstrates superior performance, achieving R2 = 0.9982 and RMSE = 0.0015, which represents a 28.6–31.8 % reduction in errors compared to the other two methods. The proposed three - layer APA - PSO - MLP model combines particle swarm optimization for determining initial weights and hyperparameters (learning rate, hidden nodes, regularization) with adaptive inertia - adjustment mechanisms to enhance global search capability. Validation results indicate significant improvements. The APA - PSO - MLP model attains R2 = 0.9964, surpassing PSO - MLP and standalone MLP by 1.3 % and 2.7 %, respectively, while also reducing parameter optimization time by 59 %. This high - precision prediction model enables real - vehicle energy efficiency optimization and intelligent maintenance strategies, providing a computationally efficient solution for in - service engine performance management.