In-situ quality monitoring of laser hot wire cladding process based on multi-sensing diagnosis and machine learning model

This study aimed to build an in-situ monitoring system based on a multi-sensing diagnosis strategy combined with a machine learning model for laser hot wire cladding applications. A high speed camera and spectrometer were mounted to obtain in situ synchronization information. Eight process signal features were extracted based on the molten pool and plasma. The silica brittle inclusions result in a higher, narrower, and rougher clad layer with higher and more uneven hardness than the normal clad layer and eventually caused undercut defects. Three quality features based on a seam tracker based surface profiling technique were extracted to label the process signal features, and an empirical database was produced based on the quantified experimental results. The minimum redundancy maximum relevance algorithm was then used to rank the importance of the process signal features, through which, a new feature vector was created. Finally, the eXtreme gradient boosting (XGBoost) model was used to develop an in-situ monitoring system for the laser hot wire cladding process. This MRMR-XGBoost based system reduces the redundancy of process information and achieved the cladding quality classification with an accuracy of 0.9727 and identify time of 7.979 ms.