ニューラルネットワークによるMIG溶接における突出し長さ推定

Abstract

Measuring of arc length is important to obtain good welding quality in spite of variation of torch height. Therefore, it is necessary to detect arc behavior in the transient state in addition to the steady state. For this purpose, this paper proposes neural network models which output the present wire extension from the data relating to wire melting, such as welding current, current pickup voltage and wire feed rate in every sampling period. Since performance of the neural network model depends on threshold functions, authors investigate the performance of the neural network models based on both sigmoid function and radial base function.To confirm the validity of these systems, fundamental experiments were carried out. The arc was directly observed and recorded as image data using a high speed camera. The output data from the neural network were compared with the measured data which were obtained from every captured image. It was found that the neural network model based on the radial base function is useful than the sigmoid function to estimate the wire extension length in MIG welding because of better responses in the transient state and smaller steady state error.