{"title":"方向编码神经网络:一种机器学习方法,用于快速预测障碍物环境中的爆炸载荷","authors":"Adam A Dennis, S. Rigby","doi":"10.1177/20414196231177364","DOIUrl":null,"url":null,"abstract":"Machine learning (ML) methods are becoming more prominent in blast engineering applications, with their adaptability to new scenarios and rapid computation times providing key benefits when compared to empirical methods and physics-based approaches, respectively. However, ML approaches commonly used for blast analyses are regularly provided with inputs relating to domain-specific parameters, restricting their use beyond the initial problem set and reducing their generality. This article presents the ‘Direction-encoded Neural Network’ (DeNN); a novel way to structure an Artificial Neural Network (ANN) to predict blast loading in obstructed environments. Each point of interest (POI) is represented by the proximity to its surroundings and the shortest travel path of the blast wave in order to prime the network to learn the underlying physics of the problem. Furthermore, a bespoke wave reflection equation creates a zone of influence around each point so that obstacles are only captured in the network’s inputs if they would alter the path of the wave. It is shown that the DeNN can predict peak overpressures with mean absolute errors ∼5 kPa for unseen, complex domains of any shape or size, when compared to the results from physics-based numerical models with ∼30 times the solution time of the DeNN. The network is used to develop maps of likely human injury following detonation of a high explosive in an internal environment, with eardrum rupture levels being correctly predicted for over 93% of unseen test points. It is therefore highly suited for use in probabilistic, risk-based analyses which are currently impractical due to excessive computational cost.","PeriodicalId":46272,"journal":{"name":"International Journal of Protective Structures","volume":null,"pages":null},"PeriodicalIF":2.1000,"publicationDate":"2023-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"1","resultStr":"{\"title\":\"The Direction-encoded Neural Network: A machine learning approach to rapidly predict blast loading in obstructed environments\",\"authors\":\"Adam A Dennis, S. Rigby\",\"doi\":\"10.1177/20414196231177364\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"Machine learning (ML) methods are becoming more prominent in blast engineering applications, with their adaptability to new scenarios and rapid computation times providing key benefits when compared to empirical methods and physics-based approaches, respectively. However, ML approaches commonly used for blast analyses are regularly provided with inputs relating to domain-specific parameters, restricting their use beyond the initial problem set and reducing their generality. This article presents the ‘Direction-encoded Neural Network’ (DeNN); a novel way to structure an Artificial Neural Network (ANN) to predict blast loading in obstructed environments. Each point of interest (POI) is represented by the proximity to its surroundings and the shortest travel path of the blast wave in order to prime the network to learn the underlying physics of the problem. Furthermore, a bespoke wave reflection equation creates a zone of influence around each point so that obstacles are only captured in the network’s inputs if they would alter the path of the wave. It is shown that the DeNN can predict peak overpressures with mean absolute errors ∼5 kPa for unseen, complex domains of any shape or size, when compared to the results from physics-based numerical models with ∼30 times the solution time of the DeNN. The network is used to develop maps of likely human injury following detonation of a high explosive in an internal environment, with eardrum rupture levels being correctly predicted for over 93% of unseen test points. It is therefore highly suited for use in probabilistic, risk-based analyses which are currently impractical due to excessive computational cost.\",\"PeriodicalId\":46272,\"journal\":{\"name\":\"International Journal of Protective Structures\",\"volume\":null,\"pages\":null},\"PeriodicalIF\":2.1000,\"publicationDate\":\"2023-06-01\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"1\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"International Journal of Protective Structures\",\"FirstCategoryId\":\"1085\",\"ListUrlMain\":\"https://doi.org/10.1177/20414196231177364\",\"RegionNum\":0,\"RegionCategory\":null,\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q2\",\"JCRName\":\"ENGINEERING, CIVIL\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"International Journal of Protective Structures","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1177/20414196231177364","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"ENGINEERING, CIVIL","Score":null,"Total":0}
The Direction-encoded Neural Network: A machine learning approach to rapidly predict blast loading in obstructed environments
Machine learning (ML) methods are becoming more prominent in blast engineering applications, with their adaptability to new scenarios and rapid computation times providing key benefits when compared to empirical methods and physics-based approaches, respectively. However, ML approaches commonly used for blast analyses are regularly provided with inputs relating to domain-specific parameters, restricting their use beyond the initial problem set and reducing their generality. This article presents the ‘Direction-encoded Neural Network’ (DeNN); a novel way to structure an Artificial Neural Network (ANN) to predict blast loading in obstructed environments. Each point of interest (POI) is represented by the proximity to its surroundings and the shortest travel path of the blast wave in order to prime the network to learn the underlying physics of the problem. Furthermore, a bespoke wave reflection equation creates a zone of influence around each point so that obstacles are only captured in the network’s inputs if they would alter the path of the wave. It is shown that the DeNN can predict peak overpressures with mean absolute errors ∼5 kPa for unseen, complex domains of any shape or size, when compared to the results from physics-based numerical models with ∼30 times the solution time of the DeNN. The network is used to develop maps of likely human injury following detonation of a high explosive in an internal environment, with eardrum rupture levels being correctly predicted for over 93% of unseen test points. It is therefore highly suited for use in probabilistic, risk-based analyses which are currently impractical due to excessive computational cost.