Ballasted track substructure is designed and constructed to provide uniform crosstie support and serve the functions of drainage and load distribution over trackbed. Poor and nonuniform support conditions can cause excessive crosstie vibration which will negatively affect the crosstie flexural bending behavior. Furthermore, ballast–tie gaps and large contact forces at the crosstie–ballast interface will result in accelerated ballast layer degradation and settlement accumulation. Inspection of crosstie support condition is therefore necessary while very challenging to implement using current methods and technologies. Based on deep learning artificial intelligence techniques and a developed residual neural network (ResNet), this paper introduces an innovative data-driven prediction approach for crosstie support conditions as demonstrated from a full-scale ballasted track laboratory experiment. The discrete element method (DEM) is leveraged to provide training and testing data sets for the proposed prediction model. K-means clustering is applied to establish ballast layer subsections with representative ballast particles and provide additional insights on layer zoning for dynamic behavior trends. When provided with DEM simulated particle vertical accelerations, the proposed deep learning ResNet could achieve 100% training and 95.8% testing accuracy. Fed with vertical acceleration measurements captured by advanced “SmartRock” sensors from a full-scale ballasted track laboratory experiment, the trained model could successfully reach a high accuracy of 92.0%. Based on the developed deep learning approach and the research findings presented in this paper, the innovative crosstie support condition prediction system is envisioned to provide railroaders accurate, timely, and repeatable inspection and monitoring opportunities without disrupting railway network operations.
Reducing the Energy Budget in WSN Using Time Series Models