An optimized tree-based model with feature selection for efficient fault detection and diagnosis in diesel engine systems

Abstract

Diesel engines play a pivotal role in transport and industrial operations, but remain a significant source of pollution. Timely fault detection and diagnosis (FDD) in such systems can help mitigate emissions and improve operational safety. This paper proposes a novel, computationally efficient bi-phase framework for diesel engine FDD, leveraging a Mendeley-based dataset and traditional machine learning (ML) techniques. The system is designed in two sequential phases: fault detection, which distinguishes between normal and faulty conditions, and fault diagnosis, which identifies the specific fault type among three predefined categories. A key innovation lies in the feature importance aggregation technique that integrates outputs from six tree-based classifiers, providing robust and interpretable feature selection. To address convergence challenges often encountered in multiclass problems, the proposed framework decomposes the task into two simpler problems, reducing model complexity and enhancing convergence speed to approximately 4.55 × 10 − 4 seconds per sample. Our extensive analysis shows that the system achieves 100% accuracy in both phases across most classifiers, with Random Forest outperforming others in training and convergence speeds. A feature-wise iterative analysis further reveals that only one feature is required for fault detection and nine for accurate diagnosis, underscoring the method's efficiency. Compared to existing approaches, including deep learning and entropy-based models, the proposed solution achieves faster convergence with minimal computational resources, making it suitable for real-world deployment and scalable applications. This is the first study to offer a convergence-optimized and modular tree-based approach for diesel engine fault analysis. • A bi-phase fault detection and diagnosis system improves modularity, accuracy, and convergence speed. • Only 1 feature is needed for fault detection and 9 for diagnosis, with 100% accuracy. • Feature selection combines multiple tree-based models for robust, interpretable ranking. • The framework is fast (under 5 ms/sample), scalable, and suitable for real-time deployment. • Explainable AI is integrated to ensure transparency in diesel engine fault decisions.

Topics & Concepts