Comparative Analysis of Ensemble Machine Learning Algorithms For Diesel Injector Health Classification Using ECU Sensor Data

Abstract

Dysfunction of the diesel fuel injector is a major cause of reduced engine performance, elevated fuel consumption, and increased emission levels in contemporary common-rail diesel engines. Traditional time-based maintenance methods are unable to identify injector faults at an early stage, often remaining undetected until engine failures occur. This paper presents a comparative analysis of four machine learning algorithms, RandomForest (RF), eXtreme Gradient Boosting (XGBoost), Gradient Boosting (GB), and K-Nearest Neighbors (KNN), applied to binary injector health classification using Engine Control Unit (ECU) sensor data. The G-IDSS diagnostic platform was used to collect data from Isuzu 4JA1 diesel vehicles, yielding 43,835 labeled samples across nine engineered features. Hyperparameter optimization was performed using the Optuna framework with the Tree-structured Parzen Estimator (TPE) algorithm under five-fold stratified cross-validation. The optimized RandomForest achieved the highest test accuracy of 85.06% (precision: 0.87, recall: 0.77 for the faulty class), followed by XGBoost at 84.34%. Feature importance analysis identified Engine Speed as the most discriminative parameter (~27% of classification contribution).