Modelling of adhesive joints and predicting their strength with the use of neural networks

Abstract

This paper presents the results of the analysis of suitability of artificial intelligence for processing of experimental information related to strength of adhesive joints. The efficiency of neuron artificial network was compared with the efficiency of typical methods of statistical analysis such as linear and polynomial regress. The research was conducted, based on a statical, determinated multifactorial program. The list of arguments comprised length of overlap, thickness of adhesive layer, thickness of joined materials, geometrical parameters of surface. The output result was the load capacity as a force needed to destruct an adhesive joint. As a result of the conducted efficiency analysis, it was stated, that the best network enabling the modeling the joints and prediction of strength of overlap adhesive joint in the creation conditions was the perceptron many-layered structured 4:4-11-1:1. This network was learned with the usage of Conjugate Gradient Descent algorithm. The RMS error for the neuron artificial network does not exceed 10% of the average force value and amounts 739 N. This is a very good results for the strength of adhesive joints. To compare, this error for multiple regression is 1413 N, and 1166 N in case of polynomial regression.