Please enable JavaScript.
Coggle requires JavaScript to display documents.
Applications of ultrasonic testing and machine learning methods to predict…
Applications of ultrasonic testing and machine learning methods to predict
the static & fatigue behavior of spot-welded joints
Introduction
-
-
Machine learning
Machine learning is one of the applications of artificial intelligence that allows computers to learn and improve themselves automatically via experience and without planning.
This technology focuses on the creation of computer programs that have access to data and can utilize that data to learn for themselves.
Data of each ultrasonic oscillogram as the input
parameter have been given to artificial neural and identifying the spotweld defects have been considered as the output
Artificial Neural Network (ANN) technique has been used to predict the yield strength of spot-welded joint
Methods
-
Test specimens
-
to precisely fabricate the specimens and to ensure uniformity of the welding process, an appropriate fixture was designed, manufactured
samples of different qualities were randomly prepared by changing the process parameters. 60 specimens were selected and divided into two groups
Each specimen pair was chosen so that the UT images of the paired specimens are identical. As a result, each pair was used to assess the fatigue behavior and tensile strength of spot weld
Ultrasonic test
Ultrasonic tests were carried out via Sonatest Sitescan series model D70 device with a probe type double crystal of 4 mm and frequency of 10 Hz
Tensile test
Tensile tests were accomplished on 30 specimens (group-I) using
STM-250 SANTAM universal testing machine.
-
Fatigue test
Using the servo-hydraulic axial testing machine (SAF-250 SANTAM) under the frequency of 10 Hz and the loading ratio equal to zero (R = 0).
Performed as a force-controlled constant amplitude loading in fully controlled environmental conditions
Machine learning
• ANN technique works well in modeling, predicting time series of linear and nonlinear, and optimization problems which have no certain explicit relation
A number of echoes and the domain difference between consecutive echoes (first and second echoes, second and third echoes, and between third and fourth echoes) were considered as input data of the neural network which were extracted from UT results using MATLAB image processing.
Eventually, the tensile strength and fatigue life of spot-welded joints were considered as the outputs. Also, the activation functions for all of the neurons were considered Tansig
the number of neurons in the hidden layer, the
values for momentum constant (α), and learning rate (η) were considered as variable parameters
• Non-dominated Sorting Genetic Algorithm II (NSGA-II) as a multiobjective genetic algorithm was used to optimize the structure of the neural network
Results
The results of multi-objective optimization for single and dual objective neural networks showed that the dual-objective neural net work (simultaneous prediction of tensile strength and fatigue life) has about 6 % and 2 % difference with the single-objective neural network in tensile strength and fatigue life, respectively
The predicted tensile strength by the single and dual-objective neural networks has an average error of about 3 % and 5.5 %, respectively, with the actual values. This difference for prediction of fatigue life of single and dual-objective neural networks is 2.5 % and 2 %, respectively
The most important achievement of this study states that the new approach of the dual-objective neural network can be used to the inspection of spot-welds in the automotive industry.
-
-
-
