DIPPED RAIL JOINTS DETECTION BASED ON VECHICLE DYNAMIC RESPONSE USING AUTOENCODER NEURAL NETWORK

• Main Author: Yosi Adhyaksa, Andreas
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/78272
• Institution: Institut Teknologi Bandung

Rail inspection and monitoring is needed to deal with rail defects such as dipped rail joints so that the defect can be repaired as soon as possible. Method for rail inspection and monitoring based on train dynamic response e.g. vehicle acceleration have been developed recently. This method is further developed by applying machine learning so that processing of larger data can be easier and faster. Dynamic simulation is performed using Universal Mechanism software to obtain quantitative data that represents the train dynamic response in the form of vertical acceleration of the train body due to dipped rail joint defect and the data is used to train and test the machine learning model developed to detect dipped rail joint response. The machine learning model developed is an autoencoder neural network based on RNN and CNN that commonly used for anomaly detection. Both models had an accuracy of more than 90% in detecting data points that were responses due to dipped rail joint defects in the majority of the test data but the models were sensitive to changes in operating parameters as indicated by the decreasing of the recall values. The RNN model successfully detected a range of dipped rail joint locations with a depth of 5 mm and above and the CNN model successfully detected a range of dipped rail joints with a depth of 2.5 mm and above.