Visual Neural Network Model for Welding Deviation Prediction Based on Weld Pool Centroid

Abstract

To solve the problem in the process of weld seam tracking, a new prediction model for welding deviation based on the weld pool image centroid has been proposed in the paper. First, some weld images under different weld currents were captured by a vision sensor. A composite filter system, which is composed of narrow-band and neutral filters, is used to reduce the disturbance of weld arc. So, several clear weld pool images can be obtained. Then a frontier of weld pool is chosen to be the processing region. Median filter and gray transformation operations are used to enhance the contrast of processing region. On this basis, the variation trend of centroid difference [Formula: see text] and welding deviation [Formula: see text] were analyzed. The centroid difference value [Formula: see text] and the weld current [Formula: see text] were determined to be welding status parameters. Moreover, a BP neural network was set up, which was composed of three layers. Next, elastic gradient descent method was used to be the training function. So a prediction model between the welding status parameters [Formula: see text] and [Formula: see text] and the welding deviation [Formula: see text] was set up. In the end of the paper, several experiments were performed to test the accuracy of the setup prediction model. The results showed that prediction values of welding deviation calculated by the vision model are fit to the real measured values. The final errors of the vision model under the weld current 70[Formula: see text]A and 73[Formula: see text]A were 0.033[Formula: see text]mm and 0.027[Formula: see text]mm, which showed excellent accuracy, environmental suitability and intelligence of the model.