Visual Detection Model For Arc Welding Penetration

Welding is the foundation of the manufacturing industry. With the improvement of industrial manufacturing requirements, the welding quality requirements are becoming increasingly important. Due to the mature development and broad scope of application, now arc welding technology is widely using in many fields. Therefore, people put forward higher requirements of electric arc welding quality. In today’s increasingly focused on efficiency, arc welding time monitoring becomes more and more important, and then this paper researches the penetration status monitoring model based on machine vision sensor.During real-time monitoring of welding process, it must establishes a mathematical model through the relationship between the state and the welding current in order to obtain the welding quality for real-time detection of the state of penetration. Many studies show that the weld pool formed in welding process and the penetration of the back weld are closely linked, and the penetration is close with the welding quality of contact, so they can establish the corresponding relationship between the weld pool of feature extraction and the current welding quality correspondence. Through capturing the images of the molten pool in welding process, the characteristic parameters that can represent the weld pool information are extracted by the relevant image processing technology, and then corresponds to the welding quality at this time, finally a mathematical model between the molten pool characteristics and penetration states is established. The model is tested through multiple sets of welding condition detection to prove that the mathematical model can effectively predict the establishment of the current state of penetration, and finally get the weld quality.Real-time monitoring of the state of penetration can be seen as a kind of pattern recognition, and neural network has a high prediction accuracy in this respect, so according to the characteristics of neural network, this paper designs two different training algorithms of neural network model:BP algorithm (Back Propagation Algorithm) and ICA algorithm (Imperialist Competitive Algorithm), and uses these two models to predict the state of penetration. ICA is a global optimization search algorithm which simulates imperialist competition process to get the optimal solution of the optimization problems. Firstly, three characteristic parameters are extracted from the weld pool image:the area of the weld pool, the width of the weld pool, the half-length of the weld pool. Secondly, the cut-and-trial method, which can determine the neural network hidden layer neuron number, is used to set up a neural network model with high accuracy. At last, two algorithms are used to train the neural network, and the recognition rate of the two models is obtained. The experimental results show that using the optimal choice of BP neural network model can get much higher prediction accuracy of the penetration state of different welding conditions than using the optimal choice of IC A neural network model.According to BP neural network, which has high recognition accuracy but be too dependent on the network initial weight and threshold, and the advantages of ICA algorithm in global search, this paper combines BP neural network and ICA, which is used to select BP neural network initial weights and thresholds, and finally proposes a recognition model of ICA-BP neural network for penetration. First the weight thresholds of model are taken as required solutions of optimization problems, and then according to the output of the penetration state between the actual error and the expected error, the value of the weights and thresholds are timely adjusted. Secondly the neural network uses ICA to get its initial weights and thresholds, and trains the network by taking the BP algorithm to get the highest recognition accuracy model a set of weights and thresholds. The highest accuracy of penetration recognition of neural network model is validated by using different test conditions. Welding experimental results show that ICA-BP neural network prediction accuracy of penetration status is higher than BP neural network.