RAILWAY VERTICAL IRREGULARITY DETECTION BASED ON CARBODY DYNAMIC RESPONSE USING MACHINE LEARNING METHOD

• Main Author: Deo Alfian, Stefanus
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/69008
• Institution: Institut Teknologi Bandung

Railway track irregularities are the deviation of rail from their nominal geometry. Severe track irregularities can reduce driving comfort and safety. Currently, a particular measurement vehicle must be used to measure track geometry for the railway assessment, which has substantial investment costs and stops train operations. One of the sources of the vehicle dynamics response is the rail geometry, but the two have a poor correlation. Some methods were used to map the correlation between vehicle dynamics response and track geometry, one that is currently under development is machine learning. Nowadays, some machine learning research on railway irregularity is limited to classifying a track section into discrete classes. In addition to discrete classification, this study will present the probability of an irregularity occurrence. Dataset is generated through dynamics simulation with predefined irregularities. Vehicle response is represented in form of carbody vertical and roll acceleration. Input features for machine learning are velocity as well as standard deviation and peak value of carbody vertical and roll acceleration, while the output is based on the acceptance level of standard deviation and peak value of irregularity. Logistic regression and neural network algorithms are chosen for their ability to describe the classification probability of each track section as well as threshold shifting. Both models managed to provide 90% accuracy with 90% precision and 93% recall. That performance shows that both models were suitable for this case. Further study is done by varying the threshold value between 0.2 and 0.8. Threshold shifting was useful to optimize the precision and recall value according to the needs.