Denoising MEMS data using Adaptive Wavelet Threshold in ETL Process for Predictive Railway Maintenance

As the most prevalent cause of train accidents, derailment becomes more dangerous with increasing train passengers in Indonesia. Manually prevention and maintenance used to do in recent times cannot bring any reduction in total cases. Predictive maintenance (PdM) as a data-driven technique based on the factual condition has become popular in many fields, including railway maintenance. An essential step in this technique is the ETL (Extract – Transform – Load) process, where data quality becomes its concern to support the decision-making process. One dimension in data quality is accuracy. However, the MEMS (micro-electro-mechanical systems) sensor installed in the train contains noise that influences the data accuracy. Related to that problem, [1] propose a wavelet threshold to denoise the MEMS signal. However, this technique highly depends on a threshold value. Thus, this research tries to find the best threshold estimator by comparing SNR (Signal-to-Noise Ratio) from the SURE Shrink threshold estimator and Bayes Shrink threshold estimator. According to the simulation using Matlab, the author finds that the raw MEMS output needs to be observed in a 5-second duration before denoising to get the best representation. In this duration, the author infers that the Bayes Shrink threshold estimator can give better results than the SURE Shrink threshold estimator with around 7 dB of a significant SNR value in accelerometer and about 14 dB of a significant SNR value in gyroscope.