{"title":"利用火焰图像的深度学习预测燃烧压力","authors":"","doi":"10.1016/j.fuel.2024.133203","DOIUrl":null,"url":null,"abstract":"<div><div>Deep learning methods provide data-driven techniques for handling large amounts of combustion data, thus finding the hidden patterns underlying these data. This study aims to predict combustion pressure from flame images, which provide more comprehensive information about the combustion process than traditional pressure sensors. The flame images were captured from a single-cylinder 4-stroke optical gasoline direct injection (GDI) engine at 1000 rpm, 5.7 bar IMEP, and stoichiometric combustion conditions using a high-speed camera. To achieve this prediction, we employed five different models: EfficientNetB4, ResNet50, Ensemble Adversarial Inception ResNet, convolutional neural network (CNN), and CNN-XGBoost. The training dataset comprised 1350 flame images captured from a single-cylinder optical GDI engine across different combustion stages. To ensure robustness, 150 images were used for validation. The models were subjected to a testing set of 4500 flame images obtained from different cycles, to evaluate how well they could perform on new, unseen data. The results showed that EfficientNetB4 achieved an impressive R<sup>2</sup> of 0.94 and a low RMSE of 0.70 compared to other tested models. Saliency analysis revealed that the model focuses on subtle flame characteristics and areas without intense flames, which suggests that it detects features invisible to the human eye. Additionally, the proposed deep learning approach is applied for the sake of monitoring cycle-to-cycle variations based on in-cylinder flame propagation where it is found that it produces high accuracy compared to those obtained through pressure sensors. Our findings are intended to advance the adoption of machine learning approaches for assisting in engine design and optimization.</div></div>","PeriodicalId":325,"journal":{"name":"Fuel","volume":null,"pages":null},"PeriodicalIF":6.7000,"publicationDate":"2024-09-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Prediction of combustion pressure with deep learning using flame images\",\"authors\":\"\",\"doi\":\"10.1016/j.fuel.2024.133203\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<div><div>Deep learning methods provide data-driven techniques for handling large amounts of combustion data, thus finding the hidden patterns underlying these data. This study aims to predict combustion pressure from flame images, which provide more comprehensive information about the combustion process than traditional pressure sensors. The flame images were captured from a single-cylinder 4-stroke optical gasoline direct injection (GDI) engine at 1000 rpm, 5.7 bar IMEP, and stoichiometric combustion conditions using a high-speed camera. To achieve this prediction, we employed five different models: EfficientNetB4, ResNet50, Ensemble Adversarial Inception ResNet, convolutional neural network (CNN), and CNN-XGBoost. The training dataset comprised 1350 flame images captured from a single-cylinder optical GDI engine across different combustion stages. To ensure robustness, 150 images were used for validation. The models were subjected to a testing set of 4500 flame images obtained from different cycles, to evaluate how well they could perform on new, unseen data. The results showed that EfficientNetB4 achieved an impressive R<sup>2</sup> of 0.94 and a low RMSE of 0.70 compared to other tested models. Saliency analysis revealed that the model focuses on subtle flame characteristics and areas without intense flames, which suggests that it detects features invisible to the human eye. Additionally, the proposed deep learning approach is applied for the sake of monitoring cycle-to-cycle variations based on in-cylinder flame propagation where it is found that it produces high accuracy compared to those obtained through pressure sensors. Our findings are intended to advance the adoption of machine learning approaches for assisting in engine design and optimization.</div></div>\",\"PeriodicalId\":325,\"journal\":{\"name\":\"Fuel\",\"volume\":null,\"pages\":null},\"PeriodicalIF\":6.7000,\"publicationDate\":\"2024-09-21\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Fuel\",\"FirstCategoryId\":\"5\",\"ListUrlMain\":\"https://www.sciencedirect.com/science/article/pii/S0016236124023524\",\"RegionNum\":1,\"RegionCategory\":\"工程技术\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q2\",\"JCRName\":\"ENERGY & FUELS\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Fuel","FirstCategoryId":"5","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S0016236124023524","RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"ENERGY & FUELS","Score":null,"Total":0}
Prediction of combustion pressure with deep learning using flame images
Deep learning methods provide data-driven techniques for handling large amounts of combustion data, thus finding the hidden patterns underlying these data. This study aims to predict combustion pressure from flame images, which provide more comprehensive information about the combustion process than traditional pressure sensors. The flame images were captured from a single-cylinder 4-stroke optical gasoline direct injection (GDI) engine at 1000 rpm, 5.7 bar IMEP, and stoichiometric combustion conditions using a high-speed camera. To achieve this prediction, we employed five different models: EfficientNetB4, ResNet50, Ensemble Adversarial Inception ResNet, convolutional neural network (CNN), and CNN-XGBoost. The training dataset comprised 1350 flame images captured from a single-cylinder optical GDI engine across different combustion stages. To ensure robustness, 150 images were used for validation. The models were subjected to a testing set of 4500 flame images obtained from different cycles, to evaluate how well they could perform on new, unseen data. The results showed that EfficientNetB4 achieved an impressive R2 of 0.94 and a low RMSE of 0.70 compared to other tested models. Saliency analysis revealed that the model focuses on subtle flame characteristics and areas without intense flames, which suggests that it detects features invisible to the human eye. Additionally, the proposed deep learning approach is applied for the sake of monitoring cycle-to-cycle variations based on in-cylinder flame propagation where it is found that it produces high accuracy compared to those obtained through pressure sensors. Our findings are intended to advance the adoption of machine learning approaches for assisting in engine design and optimization.
期刊介绍:
The exploration of energy sources remains a critical matter of study. For the past nine decades, fuel has consistently held the forefront in primary research efforts within the field of energy science. This area of investigation encompasses a wide range of subjects, with a particular emphasis on emerging concerns like environmental factors and pollution.