MMM Defect Degree Quantification Based On Optimization Algorithms

Welded joints usually work in the environment of high temperature, high pressure, andhighly corrosive. The stress concentration districts are located on the welding andheat-affected zone, where the development of corrosion, fatigue and creep process is easy toemerge. It will give us enormous harm of the property and life. For conventional NDTmethods, it can not detect early stress concentration. Metal magnetic memory (MMM)technology is known as one of the most promising nondestructive testing. MMM can not onlydetect the macroscopic defects, but also detect the stress concentration zone and early damage.This can be avoided equipment safety hazard caused by early damage. However, MMMsignals of the welded defect have a large discrete due to different type of failures. In order todetermine the components of the actual damage, the optimization algorithm is introduced toMMM defect degree quantification basis of the experimental data.In this paper, we design the fatigue test of welded joints. The experimental materials areQ345R and Q235B, respectively. Two types of defects precuts are incomplete penetration andslag. By comparison of the two defects MMM signals, the result shows that the slag specimenMMM signals fluctuant within a certain range, and the incomplete penetration MMM signalspresent a dramatic peaks and troughs downs. By contrast between X-ray detection and MMM,MMM signal characteristic law of welded damage critical state is got. The damage degrees ofwelded joints are divided into four levels that are normal, stress concentration, hidden damageand macroscopic damage. Because MMM single feature value can not be accuratelydistinguish the different damage states, are selected the five-dimensional feature vector,peak-peak value, the gradient, metal ultimate state coefficient, area signal intensity and areasignal energy. BP neural network model and BP neural network model optimized by geneticalgorithm are built based on MMM characteristic signals. By comparing the training error andtest error, it is found that predicting results of BP neural network model are instable due to theinfluence of random initial weights and thresholds. However, after the initial weights andthresholds of BP network are optimized by genetic algorithm, predicting results have betterstability and the smaller error. So BP neural network optimized by genetic algorithm is chosento study MMM defect degree quantifying, which provides a new idea for practicalengineering weld defect evaluation.