Mechanical Properties Predication System For Welded Joints Based On Neural Network Optimized By Genetic Algorithm

The mechanical property of weld joints is one of the key indicators to weld products quality, At present, enterprises at home and abroad mostly adopt the destructive testing, such as stretching, impacting and bending etc. on the mechanical properties of welding joints of products. These traditional methods have put forward a higher demand to personnel, materials and equipments which increase the cost of the product. With the development of the computer-aided technology, more and more computer technology has been applied to the weld field. In this paper, the mechanical property of weld joints has been predicted by neural networks, and genetic algorithm has been applied to optimize back-propagation neural networks.First of all, this paper has collected and collated a large number of weld procedures and mechanical property of weld joints. The corresponding database and management system has been established. From data analysis, material composition, welding parameters as well as heat treatment factors are found to be a very influential factor of the mechanical property of weld joints.Secondly, this paper has focused on analysis of characteristics of BP neural network model to study its gradient descent training algorithm. Meanwhile a mechanical property predication system of weld joints has been developed by making use of some programming and database software. The system is able to predicate the mechanical property of weld joints of carbon steel, low-alloy high-tensile structural steel and non-corrosive steel. The predicting results mainly include Tensile Strength, Yield Stress, Elongation percentage as well as Reduction ratio of area. The system can also feedback the relationship between parameters variation and the mechanical property of weld joints.In addition, BP neural network has been optimized by genetic algorithm in this paper. A new BP neural network has been developed, by confirming coding scheme and studying the fitness function and selection cross and mutation strategy. On the basis of studying the rules, topology and connection weight function of new arithmetic, a predication system has been presented to predicate the mechanical property of weld joints, using BP neural network optimized by genetic algorithm. The ultimate tensile strength of TIG welding joints, as well as the elongation of manual metal arc welding joints has been predicted by this system. A comparison, between old BP models and new BP models, shows that the latter has smaller relative error and higher efficiency.The system set up in this paper can not only achieve model training, testing and predicting mechanical property of weld joints, but also own a relatively perfect management system which is able to add and remove models. With the accumulation of the professional samples, other materials predicted models can be added into the system in order to fully enhance the scalability of this system. The study indicates that this system can increase productivity and reduce erection time as well as curtail expenditures.