PREDICTING SEVERITY LEVEL AND POSITION MAPPING OF WHEEL FLAT WITH TRACK ACCELERATION BASED MACHINE LEARNING MODEL

• Main Author: Althaafa Sukmaputra, Muhammad
• Format: Final Project
• Language: Indonesia
• Subjects: Teknik (Rekayasa, enjinering dan kegiatan berkaitan)
• Online Access: https://digilib.itb.ac.id/gdl/view/83088
• Institution: Institut Teknologi Bandung

Wheel flat results in high impact force on rail-wheel contact and can cause significant damage to railway infrastructure. Wheel flats occur due to excessive braking force, causing the wheels to slide and form flat spots on the wheel treads. This research proposes a method for predicting the severity and location of wheel flats using machine learning with rail acceleration response features. A dynamic model of the train and rail, along with simulations using Universal Mechanism (UM), was used to collect rail acceleration data from the railway track. This data was then processed using Fourier Transform and Wavelet Transform for feature extraction in the frequency and time-frequency domains. Machine learning algorithms, Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM) were used for severity classification and location prediction of wheel flats. The results of the research indicate that the developed CNN model, which uses time-frequency domain as the feature input, can classify the severity of wheel flats better than the ones using frequency domain only. The model can also map the location of wheel flats with reasonable accuracy. However, the study found that the model has limitations in detecting rounded wheel flats and difficulties in mapping train passes when there are two adjacent wheel flats.