Prediction of pulsed GTAW penetration status based on BP neural network and D-S evidence theory information fusion

Abstract

This paper used multi-sensor information fusion technology in pulsed gas tungsten arc welding. Arc sensor, visual sensor, and sound sensor were used simultaneously to obtain weld current, voltage, weld pool image, and weld sound information about the pulsed gas tungsten arc welding process, and special algorithms were designed to extract the respective signal features of different sensors’ information. Then D-S evidence theory was used to fuse the different signal features to predict the penetration status about the welding process. Aimed at the difficulty of obtaining basic probability assignment in D-S evidence theory, back-propagation (BP) neural network was used to obtain the basic probability assignment. Experiments were done to obtain data for training the BP neural network and test the prediction reliability of D-S evidence theory information fusion, and comparison results showed that D-S evidence theory could effectively use the information obtained by different sensors and obtain better prediction result than single sensor.