Enhancing engine reliability with machine learning techniques on spark plug deposition using green alcohol blend fuels on gasoline engine

• ML-based spark plugs deposit detection by contrasting CNN, FSL and ANN models. • A 200-hour engine endurance test assessed PF100, PF90E10, and PF80E20 deposits. • Compared to baseline gasoline, ethanol-petroleum mixes exhibited reduced carbon deposition 34.57% for PF80E20. • Noise pressure level was 7.9 dB higher with PF100 and 4.6 dB lower with PF80E20. There is an urgent need for more sustainable alternatives due to the negative effects of conventional petroleum fuels on global warming and climate change. Alcohol-based fuels, including ethanol, have been investigated as a potential remedy. However, using some fuel blends for an extended period of time can harm engine parts, lower efficiency and leading to excessive carbon buildup. In this study, wear, deposits, and noise emissions in single-cylinder air-cooled spark ignition engines are examined in relation to three distinct fuel samples: PF (petroleum fuel), PF90E10 (10% ethanol and 90% PF), and PF80E20 (20% ethanol and 80% PF). The PF80E20 engine's spark plug (34.57%) has less carbon deposits than engines powered by PF90E10 (51.09%) and PF100 (71.15%), according to the results. Scanning electron microscopy (SEM) and energy dispersive X-ray spectroscopy (EDX) were employed in the study for in-depth analysis. All blended fuels had lower noise emissions than pure petroleum fuel, with the fuel mix being mostly responsible for the reductions. Experimental evaluation on multi-source engine datasets showed that the FSL-integrated methodology improves remaining useful life (RUL) estimation precision by up to 20%, reduces false maintenance alarms by 18%, and boosts forecast accuracy by 15–22% when compared to existing methods. Few-Shot Learning (FSL) demonstrates a significant advantage over traditional machine learning methods in engine maintenance and reliability prediction, primarily by overcoming the critical industry challenge of data scarcity. Because of its distributed design and versatility, the proposed framework shows significant promise for practical usage in aviation, maritime, and industrial engine maintenance systems where data privacy, scalability, and high prediction accuracy are critical. The model outperformed traditional methods such as Artificial Neural Networks (ANN) and Convolutional Neural Networks (CNN), with a training loss of 0.0294 and validation loss of 0.0258, compared to ANN's training loss of 0.0054 (validation loss 0.1452) and CNN's training loss of 0.0037 (validation loss 0.0326), demonstrating its superior efficiency in fuel performance prediction. As a result, FSL provides a practical and progressively deployable solution for real-world predictive maintenance in aerospace, automotive, and smart factory environments where novel fault patterns occur on a regular basis and rapid, cost-effective model adaption is critical.