Railway ballast fouling detection using thermal imaging: integration of LSTM and XGBoost

This paper presents an artificial intelligence (AI)-based approach to automate the structural health monitoring (SHM) of railway ballast through the fusion of long short-term memory and XGBoost (LSTM-XGB) to surface temperature data derived from infrared thermal images. In this context, machine learning models are trained using remotely acquired surface temperature data to classify fouling index based on thermal variations within ballast aggregates captured from thermograms. The long short-term memory (LSTM) component processes sequential time-series thermal data to predict preceding values, and the XGBoost (XGB) component classifies fouled ballast conditions based on identified patterns of surface temperature variations measured via infrared thermography (IRT). The results confirm the capability of the LSTM component to capture the time-series variations of a specimen’s surface temperature in a shorter timeframe as well as the superior performance of XGBoost compared to a random forest (RF) approach, in classifying fouled ballast conditions. Therefore, the LSTM-XGB model demonstrates higher efficiency compared to the standalone XGBoost model, since the predictive nature of LSTM over time-series temperature data enables capturing shorter time window for measuring ballast surface temperature and identifying patterns. Moreover, establishing a coarser classification of ballast fouling (categorized into three groups instead of five) significantly improves the model capability for accurate assessment of the ballast fouling conditions.