Real-time abnormality detection and classification in diesel engine operations with convolutional neural network

• Published in: Expert systems with applications 2022-04, Vol.192, p.116233, Article 116233
• Main Authors: Shahid, Syed Maaz, Ko, Sunghoon, Kwon, Sungoh
• Format: Article
• Language: English
• Subjects: Abnormalities; Abnormality detection; Artificial neural networks; Classification; Convolutional neural network (CNN); Cylinders; Damage detection; Damage prevention; Diesel engines; Efficiency; Engine load; Feature extraction; Internal combustion; Misfire; Multi-class classification; Neural networks; Pattern recognition; Real time; Reliability; Signal processing
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2021.116233

In this paper, we propose a real-time diagnostic method using a convolutional neural network (CNN) to detect cylinder misfires and engine load conditions in multi-cylinder internal combustion (IC) diesel engines. To enhance engine efficiency and reliability, it is necessary to detect and classify multiple irregularities in engine operation, such as misfires and changes in load conditions. Optimized combustion parameter setting increases engine efficiency, which depends on the engine load conditions. Similarly, misfire detection prevents engine damage and increases fuel efficiency and reliability of the engine. Previous works investigated misfire detection while ignored the change in engine operating conditions that affects the engine efficiency. This work presents a real-time CNN-based method to detect misfires and load changes together in engine operations by analyzing the IC process. First, the sensor signal is pre-processed to extract a primary signal and translate it into a crank angle degree (CAD) signal, which represents a complete cycle of the IC process in the engine. After that, a CNN is designed for multi-class classification and trained using one-dimensional CAD vectors, which combines a feature extraction capability and pattern recognition in a single learner. The proposed CNN-based method detects and classifies multiple abnormalities in engine operations with an accuracy of more than 99% and low generalization error. The designed CNN uses a single convolutional layer for feature extraction resulting in more efficient systems in terms of performance and speed. •Fluctuations in the rotational speed of prime mover are translated in to a CAD signal to observe the behavior of the engine IC process.•Combined detection of cylinder misfires and engine load conditions is an important tool in advanced engine control systems.•A CNN-based classifier extracted the feature from CAD signals and detected multiple irregularities in each IC cycle of the engine operation in real-time.•Low computation complexity and a lower prediction time by proposed shallow CNNbased classifier.