In situ monitoring and penetration prediction of plasma arc welding based on welder intelligence-enhanced deep random forest fusion. (June 2021)

• Record Type: Journal Article
• Authors: Wu, Di; Hu, Minghua; Huang, Yiming; Zhang, Peilei; Yu, Zhishui
• Is Part Of: Journal of manufacturing processes. Volume 66(2021)
• Page Start: 153
• Page End: 165
• Publication Date: 2021-06
• Subjects: Keyhole molten pool -- Multi-level features -- Convolutional neural network -- Random forest -- Penetration prediction; Production management -- Data processing -- Periodicals; Manufacturing processes -- Periodicals; Procestechnologie; Productietechniek; Production -- Gestion -- Informatique -- Périodiques; Fabrication -- Périodiques; Manufacturing processes; Production management -- Data processing; Periodicals; 670.5
• DOI: 10.1016/j.jmapro.2021.04.007
• Languages: English
• ISSNs: 1526-6125

Abstract

Graphical abstract: Highlights: A flexible front-side sensing system was designed to in-situ monitor the keyhole molten pool behavior during VPPAW process, which was found to heavily influence the process stability and weld quality. By incorporating the human welder's prior knowledge, the low-level handcrafted features were extracted to quantitatively describe the geometrical appearance of the keyhole. The pre-trained convolutional neural network(CNN) model can automatically learn the high-level discriminative features and visualize the multi-layer kernel feature maps to clearly interpret the physical meaning of the deep features. The combination of multi-level features based on random forest (RF) classifier shows significant improvement on classification performance comparing with state-of-the-art machine learning algorithms. Abstract: Online process monitoring and quality control has been a long-standing challenge for variable polarity plasma arc welding (VPPAW) due to the inherent instability and fluctuation of the keyhole molten pool. This work developed an innovative welder intelligence-enhanced deep random forest fusion (WI-DRFF) approach, aiming to describe the dynamics of front-side molten pool and accurately predict the weld penetration. Based on the human welder's prior knowledge, we firstly proposed an image processing algorithm to extract the low-level handcrafted features, which could quantitatively describe the geometrical appearance of the keyhole. Afterwards, we constructed a convolutional neural network (CNN) to learn the high-level discriminative features of weld pool and interpret the physical characteristics of the deep features with visualization. Finally, we incorporated the handcrafted keyhole features and deep features to concatenate a multi-level feature vector for predicting the weld penetration based on random forest (RF) classifier. Extensive experiments demonstrate that our proposed approach yields a remarkable classification performance comparing with state-of-the-art machine learning algorithms even with limited training data. This approach is a new paradigm in the digitization and intelligence of welding process and can be exploited to provide a feedback in an adaptive quality control system.