Hybrid Physics-Infused One-Dimensional Convolutional Neural Network-Based Ensemble Learning Framework for Diesel Engine Fault Diagnostics

Abstract

Due to their high thermal efficiency and long functional life, diesel engines have become ubiquitous in automobiles. Diesel engines are vulnerable to component failure and sensor faults. New cognitive fault diagnosis algorithms are crucial for the safe operation of equipment. Conventional model-based approaches are limited in their capabilities owing to the approximations made during the development of these models. In comparison, the efficacy of most of the data-driven approaches depends on the quantity of data. Additionally, the existing data-driven algorithms do not consider the system’s physics and are susceptible to overfitting issues. To address the aforementioned issues, we propose an end-to-end autonomous hybrid physics-infused one-dimensional (1D) convolutional neural network (CNN)-based ensemble learning framework combining a low-fidelity physics-based engine model, autoencoder (AE), 1D CNNs, and a multilayer perceptron (MLP) for fault diagnosis. The system used to demonstrate the capabilities of the devised model is a 7.6-l, 6-cylinder, 4-stroke diesel engine. The physics model guarantees that the estimations produced by the framework conform to the engine’s actual behavior, and the ensemble deep learning module overcomes the overfitting issue. Empirical results show that the framework is efficient and reliable against data from a real engine setup under various operating conditions, such as changing injection duration, varying injection pressure, and engine speed. Besides, the framework is tested against noisy data, reaffirming the model’s robustness when subjected to actual working conditions where acquired noise is a norm.