Presiksi Umur Kelelahan Komposit Yang Dipengaruhi Variable Amplitude Loading Menggunakan Artificial Neural Network Dengan Algoritma Training Metaheuristik

Rohman, Muh Nur (2018) Presiksi Umur Kelelahan Komposit Yang Dipengaruhi Variable Amplitude Loading Menggunakan Artificial Neural Network Dengan Algoritma Training Metaheuristik. Masters thesis, Institut Teknologi Sepuluh Nopember.

[img] Text
02511550012004-Master_Thesis.pdf - Published Version
Restricted to Repository staff only

Download (7MB) | Request a copy

Abstract

Gradient-based search sudah banyak digunakan sebagai algoritma optimasi pada Artificial Neural Network (NN) berbasis Multilayer Perceptrons (MLP) untuk prediksi umur kelelahan komposit berbasis polimer. Akhir-akhir ini metode optimasi yang lebih baru yaitu metaheuristik telah banyak digunakan untuk optimasi NN, namun aplikasinya masih di luar bidang prediksi umur kelelahan komposit. Pada penelitian ini telah dikembangkan model NN berbasis MLP dengan satu hidden layer untuk prediksi umur kelelahan komposit berbasis polimer dengan variable amplitude loading. Bobot dan bias didapatkan dengan tiga metode optimasi metaheuristik secara murni, yaitu Genetic Algorithm (GA), Differential Evolution (DE) dan Particle Swarm Optimization (PSO). Jumlah data kelelahan yang terbatas digunakan untuk training. Tiga parameter pembebanan siklik yaitu tegangan maksimum dan minimum dan stress ratio (R) digunakan sebagai input NN dan log-umur kelelahan sebagai output. Kinerja model NN ditunjukkan dengan MSE, R2 dan persentase jumlah data yang bisa digunakan dengan mempertimbangkan kurva P-S-N hasil eksperimen. Setiap prediksi umur kelelahan didapatkan dengan mengambil nilai rata-rata dari tigapuluh kali percobaan yang dilakukan secara independen. Simulasi dengan E-glass fabrics/epoxy (layups [(±45)/(0)2]S) menggunakan 33% dari jumlah data eksperimen untuk training dan 67% untuk testing. Model NN terbaik adalah MLP-DE dengan jumlah hidden nodes 25-30 berdasarkan tiga ukuran kinerjanya, yaitu sebanyak 93.75% dari jumlah data prediksi bisa digunakan, serta nilai MSE dan R2 secara berurutan sebesar 0.138 dan 0.9532. Dengan nilai rata-rata R2 0.9370-0.9532, ketiga model NN dalam penelitian ini menunjukkan akurasi yang tinggi. Simulasi dengan E-glass/polyester (layups [90/0/±45/0]S) menggunakan 22% dari jumlah data eksperimen untuk training dan 78% untuk testing. MLP-GA dengan jumlah hidden nodes 25-30 lebih disukai karena menunjukkan jumlah data terbanyak yang bisa digunakan, yaitu 74.76%. Rata-rata MSE dan R2 dari ketiga model NN dalam penelitian ini menunjukkan nilai yang hampir sama, secara berurutan sebesar 0.2445-0.2490 dan 0.8556-0.8575. Ketiga model NN dalam penelitian ini menunjukkan akurasi yang cukup tinggi dengan mempertimbangkan nilai rata-rata R2. Akurasi model NN-metaheuristik pada penelitian ini secara umum comparable dengan NN-gradient-based, kecuali MLP-PSO pada E-glass fabrics/epoxy yang menghasilkan akurasi yang lebih rendah. =========== Gradient-based search technique have been used in many research as an algorithm to optimize Multilayer Perceptrons (MLP) based Artificial Neural Network (NN) for fatigue life prediction of polymeric-base. The later optimization method of methaheuristics have been used to optimize the NN, however the applications are still beyond the field of composite fatigue life prediction. In present research, we have developed NN model based on MLP with single layer to predict fatigue life of polymeric-base composite under variable amplitude loading. The weights and bias were obtained by utilizing three methods of methaheuristic optimization purely, i.e. Genetic Algorithm (GA), Differential Evolution (DE) and Particle Swarm Optimization (PSO). Limited number of fatigue data were used in trainings. Three parameters of cyclic loading i.e. maximum and minimum stress and stress ratios (R) were used as NN input, on the other side, log-fatigue life were used as output NN. The performance of NN models were quantified by MSE, R2 and percentage of number of fatigue data that can be used with respect to P-S-N curve which generated from experimental results. A single prediction of fatigue life was taken from the average value from thirty times of independent trials. 33% of number of experimental fatigue data of E-glass fabrics/epoxy (layups [(±45)/(0)2]S) were used for training and the 67% for testing. From simulation results found that MLP-DE with 25-30 hidden nodes is the best NN model with respect to those three performance indicators, i.e., 93.75% of total number of prediction data are can be used, with MSE and R2 of 0.138 and 0.9532, respectively. With respect to average of R2 of 0.9370-0.9532, the three of NN models in this research exhibit high accuracy. On the other side, 22% of number of experimental fatigue data of E-glass/polyester (layups [90/0/±45/0]S) were used for training and 78% for testing. MLP-GA with 25-30 hidden nodes is preferred with respect to that 74.76% of the total number of prediction data are can be used. The average of MSE and R2 of NN models exhibit almost similar values, i.e. 0.2445-0.2490 and 0.8556-0.8575, respectively. With respect to average of R2, the three NN models in this research exhibit accuracy of reasonably high. Generally, the accuracy of NN-metaheuristic models in this research are comparable with NN-gradient-based, with exception of MLP-PSO for E-glass fabrics/epoxy which have lower in accuracy.

Item Type: Thesis (Masters)
Uncontrolled Keywords: Composite; Fatigue life; variable amplitude loading; neural networks; metaheuristic
Subjects: Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
T Technology > TA Engineering (General). Civil engineering (General) > TA418.38 Materials--Fatigue.
T Technology > TP Chemical technology > TP1140 Polymers
Divisions: Faculty of Industrial Technology > Material & Metallurgical Engineering > (S2) Master Theses
Depositing User: Rohman Muhammad Nur
Date Deposited: 23 Feb 2018 08:14
Last Modified: 23 Feb 2018 08:14
URI: http://repository.its.ac.id/id/eprint/50255

Actions (login required)

View Item View Item