This content is not included in your SAE MOBILUS subscription, or you are not logged in.
Misfire Detection Technology with Deep Neural Network Based on Ignition Coil Signals
- Naoki Yoneya - Hitachi, Ltd, Japan ,
- Kenji Amaya - Tokyo Institute of Technology, Japan ,
- Kengo Kumano - Hitachi, Ltd, Japan ,
- Yoshihiro Sukegawa - Hitachi, Ltd, Japan ,
- Yoshifumi Uchise - Hitachi Astemo Hanshin, Ltd., Japan ,
- Hideo Jitsu - Hitachi Astemo Hanshin, Ltd., Japan ,
- Yukio Fujiyama - Hitachi Astemo Hanshin, Ltd., Japan
ISSN: 1946-3936, e-ISSN: 1946-3944
Published July 22, 2023 by SAE International in United States
Citation: Yoneya, N., Amaya, K., Kumano, K., Sukegawa, Y. et al., "Misfire Detection Technology with Deep Neural Network Based on Ignition Coil Signals," SAE Int. J. Engines 17(1):2024, https://doi.org/10.4271/03-17-01-0004.
For achieving high efficiency and low exhaust emissions, engines need to be operated near the limits of stable combustion, such as lean or exhaust gas recirculation (EGR) conditions. Sensing technologies of the combustion state by existing engine components are of high interest. And the utilization of voltage and current signals from ignition coils is discussed in this article. The discharge channel of an ignition spark is strongly affected by flow variation and spark plug surface conditions, and the behavior of discharge channel stretching and restrike event can vary greatly from cycle to cycle. As a result, the effects of flow velocity, temperature, pressure, and electrode surface resistance are compounded in the voltage-current response, making it difficult to accurately detect the combustion state for each cycle by a threshold judgment process using a single feature value.
In this article, a method for inductively detecting misfires from voltage and current signals of ignition coils by applying deep learning image recognition is introduced. First, post-ignition for misfire detection is performed on the engine bench during the expansion stroke in an engine cycle, when the cylinder pressure is expected to differ between the combustion cycle and the ignition cycle, and the ignition coil voltage and current are measured. Next, a two-dimensional frequency distribution of voltage and current (discharge histogram) is created as an input image for deep learning, and the AlexNet model, which has been trained with more than one million images, is trained with images of the ignition and combustion cycles as a supervised learning. The accuracy of classification is then verified using a validation dataset. In addition, to making the deep learning model more explainable, the activation score distribution on the discharge histogram was visualized when the trained model judges the images, and the discharge characteristics that provided the basis for deep learning classifications were analyzed.