A Supervised Machine Learning Approach for Structural Overload Classification in Railway Bridges using Weigh-in-Motion Data

A Supervised Machine Learning Approach for Structural Overload Classification in Railway Bridges using Weigh-in-Motion Data

Weigh-in-motion (WIM) data provides valuable information on vehicle axle load, enabling efficient and economical railway structural safety management programs. However, the current method for assessing structural overload on railway bridges using WIM data is time-consuming and often necessitate line closure while analyses are being conducted. This paper presents the development of a novel supervised machine learning (ML) approach that can be used as an assessment tool to expedite the decision-making process and minimise economic loss. Variables for model input are carefully considered by analysing real WIM data obtained from measurement sites in New Zealand. Various supervised ML classification models are
evaluated for their capability in classifying axle load combinations (ALC) into “Normal”, meaning safe to go, or “Overload”, meaning that line closure is required for detailed inspection of the affected bridges. It is found that the model using Neural Network (NN) outperforms other candidates in this capacity and is therefore selected for detailed model development. The selected NN model is then examined to evaluate its ability to classify unseen data using the hold-out cross-validation technique. An initial investigation using a small dataset derived from real WIM measurements demonstrates that the NN model can achieve
impressive performance accuracy of 99.9%, F1-score of 99.2%, and an area under curve (AUC) index of 0.9981. Subsequently, a method for artificially generating synthetic ALC data is proposed to create extensive training datasets for comprehensive structural overload model development. It is demonstrated that with sufficient overload data in the training dataset, the model can achieve exceptional accuracy,
reaching 99.84% or higher F1-score for single overload level and 99.5% to 99.86% for multiple overload thresholds. The developed model can be integrated into the WIM post-processing systems, providing a realtime bridge overload assessment tool that facilitates more efficient and cost-effective railway structural safety management.