VPPAW penetration monitoring based on fusion of visual and acoustic signals using t-SNE and DBN model

Abstract It is a big challenge to identify the joint penetration status for obtaining high-quality weld joints during variable polarity plasma arc welding (VPPAW). This paper addresses “t-stochastic neighbor embedding” (t-SNE) and deep belief network (DBN) to perform VPPAW process monitoring and penetration status identification. The multi-source weld information under different conditions is simultaneously received by using visual and acoustic sensors. Time and frequency domains features extracting from the sensors constitute a fusing feature set to reflect the variation trend of weld penetration status. Using the obtained feature vectors, t-SNE method is proposed to acquire the intrinsic features corresponding to different penetration states and map them into a 3-dimensional space to realize visualization. The visualization result produced by t-SNE is significantly better than PCA or Isomap technique. Then the DBN classification model with the optimal structure is developed to guarantee effective identification of penetration status. Experimental verification and comparisons show that the classification performance of DBN can reach 97.62% which indicates DBN outperforms back-propagation neural network (BPNN) and support vector machine (SVM) models. The proposed methodology based on the combination of t-SNE and DBN might be regarded as a promising technique for VPPAW penetration status identification.