A joint prediction framework of data augmentation and
machine learning was proposed to address the challenges of high-dimensional
feature selection and imbalanced small-sample data in
diesel engine performance prediction. The BorutaShap algorithm was first used
to select key parameters from 48 features, which reduced the generalization
limitations caused by subjective feature selection. Then a Wasserstein
generative adversarial network with gradient penalty (WGAN-GP) was applied to address the small-sample problems
in brake-specific fuel consumption (BSFC) and peak
cylinder pressure (PCP). The distribution rationality of samples was enhanced
by dynamic gradient punishment. The predictive performances of data generated
by WGAN-GP, SMOTE, and VQ-VAE
models were finally compared across 4 models of XGBoost, SVR, RF, and GBDT to
validate the effectiveness of the method. The results showed that WGAN-GP improved the determination coefficient R2 of XGBoost to 0.985 (MAE=1.659 g/(kW·h))
for BSFC and 0.943 (MAE=0.006 MPa) for PCP. Accuracy improvements of 5%-8% were achieved over SMOTE and VQ-VAE. In
summary, the integration of feature selection and data augmentation effectively
addressed performance prediction challenges of diesel engine under diverse
feature characteristics and limited data samples.
LI Xiang, DENG Zhonghang, GUO Zhikun, SHI Qingjue, ZHANG Fan, LU Lili
. Diesel Engine Performance Prediction Method Based on
Generative Adversarial Network and Machine Learning[J]. Vehicle Engine, 2025
, 0(5)
: 61
.
DOI: 10.3969/j.issn.1001-2222.2025.05.009
