Feature Extraction for Classification of Welding Quality in RSW Systems Xing-Jue Wang 1,2 , Jun-Hong Zhou 2 , and Chee Khiang Pang 1, ⋆ 1 Department of Electrical and Computer Engineering, National University of Singapore, 4 Engineering Drive 3, 117576 Singapore 2 A*STAR Singapore Institute of Manufacturing Technology, Singapore {a0120725,justinpang}@nus.edu.sg jzhou@simtech.a-star.edu.sg Abstract. Resistance spot welding is one of the most important welding techniques widely used in manufacturing, and monitoring the welding quality draws close attention in industries. The common on-line moni- toring systems apply multiple sensors to acquire various signals to pre- dict the welding quality. In this paper, only the voltage and current data are acquired for classification of welding quality. In this way, the cost of data acquisition system is significantly reduced while acceptable accuracy is maintained. Furthermore, the past works are sensitive to the random perturbation from various sources. By extracting the key features such as RMS values of each half cycles of voltage and current, resilience against signal disorder due to noise from the environment is enhanced. By feed- ing various features into SOM neural network, classification of the welds with 92.9% accuracy is achieved, showing great potential to deal with var- ious conditions and different materials because of the fast training speed compared with BP neural networks. Keywords: current, dynamic resistance, self-organizing map, spot weld- ing, voltage. 1 Introduction Resistance spot welding (RSW) as an important welding technique invented in 1877 is extensively used in industry nowadays, especially the automotive in- dustry. Traditionally, the main quality control tests arethe off-line destructive chisel test and peel test examining the weld nugget obtained from the produc- tion line [1]. These traditional quality tests are very time-consuming and raise the labor costs as well as product wastage. To solve these problems, many on- line monitoring schemes have been proposed by making use of various com- bination of associate parameters such as electrode displacement, weld current, weld voltage, dynamic resistance, ultrasound inspection and electrode force. Among them, dynamic resistance and electrode displacement provide the most ⋆ Corresponding author. c  Springer International Publishing Switzerland 2015 307 H. Handa et al. (eds.), Proc. of the 18th Asia Pacific Symp. on Intell. & Evol. Systems – Vol. 2, Proceedings in Adaptation, Learning and Optimization 2, DOI: 10.1007/978-3-319-13356-0_ 25