As an advanced and highly efficient welding method, Keyhole Tungsten Inert Gas (keyhole TIG) welding has drawn wide interests from the manufacturing industry. In order to improve its manufacturing quality and automation level, it's necessary to develop an online monitoring system for the keyhole TIG welding process. This study developed a visual monitoring system, which utilized an HDR welding camera to monitor the welding pool and keyhole during keyhole TIG welding process. A state of the art Convolutional neural network (Resnet) was developed to recognize different welding states, including good weld, incomplete penetration, burn through, misalignment and undercut. In order to improve the diversity of training dataset, image augmentation was performed. To optimize the training process, a metric learning strategy of center loss was introduced. Furthermore, visualization methods, including guided Grad-CAM, feature map and t-SNE were applied to understand and explain the effectiveness of deep learning process. This study will lay a solid foundation for the development of on-line monitoring system of keyhole TIG.