ULTRASONIC NON DESTRUCTIVE TEST FOR RAIL DEFECT DETECTION USING ARTIFICIAL NEURAL NETWORK AND GENETIC ALGORITHM BASED ON ULTRASONIC WAVE PROPAGATION SIMULATION

• Main Author: Kirana Marantika - Nim: 13314021 , Winda
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/31640
• Institution: Institut Teknologi Bandung

<p align="justify">Improvements in service quality such as security and safety need to balance the acceleration of railways construction throughout Indonesia. According to Ministry of Transportation’s data, 76% of the total train accidents occurred in Indonesia from 2004 until 2008 were derailments. They most likely occurred due to defective rail. So far the method used to inspect defects in rails is only effective in detecting surface defects. Therefore, a non-destructive test system that can detect internal rail defect is needed. The simulation is conducted using pseudospectral method to observe and retrieve ultrasonic wave signals that are introduced into various types of line defects in the rail. There are two signal recording methods used, transmission and echo. After simulation, the signals recorded will be extracted by transforming them into frequency domain and then picking some signal peaks and their corresponding frequency. This data will be introduced in ANN to predict defect parameter, position, and length of the line defect. The activation function is log-sigmoid for all hidden layers and output layers. ANN learning is conducted using three methods, SGD, SGD-AG 1, and SGD-AG 2. Based on the ANN model that was successfully trained, SGD-AG 2 training was able to predict defect position and length while overcoming problems in SGD training. 80% and 100% of the test results using test data have an error of no more than 20% in predicting position and length of the line defects. The error percentage of the location parameter is 12.08% and the error percentage of the length parameter is 9.25%.<p align="justify">