Wang Cai , Ping Jiang , LeShi Shu , ShaoNing Geng , Qi Zhou
{"title":"Real-time monitoring of laser keyhole welding penetration state based on deep belief network","authors":"Wang Cai , Ping Jiang , LeShi Shu , ShaoNing Geng , Qi Zhou","doi":"10.1016/j.jmapro.2021.10.027","DOIUrl":null,"url":null,"abstract":"<div><p>In laser keyhole welding processes, the dynamic behavior of the keyhole and molten pool is associated with the penetration state. In this paper, a deep learning-based monitoring system is employed to monitor the penetration state by inspecting the keyhole and molten pool. An image processing method<span><span><span> based on parabola template and half molten pool is proposed to extract the boundary contour of the molten pool, which can effectively reduce the interference of metal vapor plume. 12 kinds of morphological features of the keyhole and molten pool are defined and extracted, especially 6 molten pool tail features. A data processing method is proposed to reduce the amount of data by 4/5 to avoid overfitting and remove abnormal data. The 12 variance features obtained through data processing can effectively increase the feature depth and improve the monitoring accuracy. A deep learning<span> framework based on deep belief network (DBN) is established to model the relationship between the 24 features and their corresponding penetration states. The proposed framework achieves higher accuracy and stronger robustness in monitoring penetration states by comparing with the </span></span>backpropagation neural network, </span>support vector machine, decision tree, k-nearest neighbors and random forest. The results show that this proposed monitoring method can accurately and quickly judge the penetration state according to the keyhole and molten pool images obtained by the on-line monitoring system.</span></p></div>","PeriodicalId":16148,"journal":{"name":"Journal of Manufacturing Processes","volume":null,"pages":null},"PeriodicalIF":6.1000,"publicationDate":"2021-12-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"18","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Journal of Manufacturing Processes","FirstCategoryId":"5","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S1526612521007490","RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"ENGINEERING, MANUFACTURING","Score":null,"Total":0}
引用次数: 18
Abstract
In laser keyhole welding processes, the dynamic behavior of the keyhole and molten pool is associated with the penetration state. In this paper, a deep learning-based monitoring system is employed to monitor the penetration state by inspecting the keyhole and molten pool. An image processing method based on parabola template and half molten pool is proposed to extract the boundary contour of the molten pool, which can effectively reduce the interference of metal vapor plume. 12 kinds of morphological features of the keyhole and molten pool are defined and extracted, especially 6 molten pool tail features. A data processing method is proposed to reduce the amount of data by 4/5 to avoid overfitting and remove abnormal data. The 12 variance features obtained through data processing can effectively increase the feature depth and improve the monitoring accuracy. A deep learning framework based on deep belief network (DBN) is established to model the relationship between the 24 features and their corresponding penetration states. The proposed framework achieves higher accuracy and stronger robustness in monitoring penetration states by comparing with the backpropagation neural network, support vector machine, decision tree, k-nearest neighbors and random forest. The results show that this proposed monitoring method can accurately and quickly judge the penetration state according to the keyhole and molten pool images obtained by the on-line monitoring system.
期刊介绍:
The aim of the Journal of Manufacturing Processes (JMP) is to exchange current and future directions of manufacturing processes research, development and implementation, and to publish archival scholarly literature with a view to advancing state-of-the-art manufacturing processes and encouraging innovation for developing new and efficient processes. The journal will also publish from other research communities for rapid communication of innovative new concepts. Special-topic issues on emerging technologies and invited papers will also be published.