Prediction of Welded Joint Strength in Plasma Arc Welding: A Comparative Study Using Back-Propagation and Radial Basis Neural Networks

International Conference on Advances in Materials and Manufacturing Applications 2016

Srinivas, Kadivendi^1 ; Vundavilli, Pandu R.^2 ; Hussain, M Manzoor^3 ; Saiteja, M.^1

Department of Mechanical Engineering, DVR and Dr. HS MIC College of Technology, Kanchikacherla

Andhra Pradesh, India^1

School of Mechanical Sciences, IIT Bhubaneswar, Odisha, India^2

Department of Mechanical Engineering, JNTUH College of Engineering, Telangana, Hyderabad, India^3

关键词: Back propagation neural networks;    Comparative studies;    Conventional regression analysis;    Feed-forward back-propagation neural networks;    Levenberg-Marquardt;    Radial basis neural networks;    Response equations;    Welded joint strength;   

Others  :  https://iopscience.iop.org/article/10.1088/1757-899X/149/1/012033/pdf
DOI  :  10.1088/1757-899X/149/1/012033

来源: IOP

【 摘 要 】

Welding input parameters such as current, gas flow rate and torch angle play a significant role in determination of qualitative mechanical properties of weld joint. Traditionally, it is necessary to determine the weld input parameters for every new welded product to obtain a quality weld joint which is time consuming. In the present work, the effect of plasma arc welding parameters on mild steel was studied using a neural network approach. To obtain a response equation that governs the input-output relationships, conventional regression analysis was also performed. The experimental data was constructed based on Taguchi design and the training data required for neural networks were randomly generated, by varying the input variables within their respective ranges. The responses were calculated for each combination of input variables by using the response equations obtained through the conventional regression analysis. The performances in Levenberg-Marquardt back propagation neural network and radial basis neural network (RBNN) were compared on various randomly generated test cases, which are different from the training cases. From the results, it is interesting to note that for the above said test cases RBNN analysis gave improved training results compared to that of feed forward back propagation neural network analysis. Also, RBNN analysis proved a pattern of increasing performance as the data points moved away from the initial input values.