Classification of Weld Seam Width Based on Detrended Fluctuation Analysis, t-Distributed Stochastic Neighbor Embedding, and Support Vector Machine

The droplet behavior of welding is chaotic and fractal, and thus is significant for diagnosis of weld quality. To study the long-range correlation of fractals, detrended fluctuation analysis (DFA) is introduced for current and voltage signals. The DFA curve obviously has crossover and can be expressed by a two exponent model, including a short-term exponent ( α 1) at small scale and a long-term exponent ( α 2) at large scale. However, the relationship between the weld seam width and the two exponent model is not obviously linear. A high-dimensional feature is generated on basis of all points of the DFA curve, t-distributed stochastic neighbor embedding is used for dimension reduction. Then, a low-dimensional feature combined with support vector machine is used to predict weld seam width, which achieves higher classification accuracy than the two exponent model. This study provides a new attempt about the chaotic and fractal characteristics in welding.