Intelligent prediction of diesel engine dynamic performance and emissions via multi-signal fusion and Bayesian optimization

An intelligent prediction framework integrating multi-signal feature learning and Bayesian optimization is proposed to address the strong nonlinearity, multivariable coupling, and dynamic evolution of diesel engine performance and emissions under transient operating conditions. A full-engine simulation model of a four-cylinder turbocharged intercooled diesel engine was developed, and a performance–emission dataset was generated over the FTP-75 driving cycle. Pearson correlation analysis, regression random forest, and recursive feature elimination were first employed for input evaluation and selection, with recursive feature elimination providing the best overall performance. Principal component analysis, one-dimensional convolutional neural networks, and autoencoders were then compared for feature fusion and dimensionality reduction. The autoencoder showed superior capability in preserving nonlinear interactions among multi-source signals. A multi-output feedforward neural network was subsequently constructed to jointly predict brake efficiency, torque, power, NOx, CO, and HC, with its architecture and training parameters optimized through Bayesian optimization. The optimized model reduced the average test RMSE by 10.22% and increased the average R 2 by 0.34%. Under 10% input noise, the average R 2 remained 0.9652, demonstrating strong robustness to signal disturbances.