HYBRID QUANTUM NEURAL NETWORK AND SHAPELY ADDITIVE EXPLANATIONS IN RAIL TRACK GEOMETRY DEFECT PREDICTION

Abstract

This study applies quantum machine learning, which brings together quantum computing and machine learning, for predictive maintenance in the railroad industry. Rail companies invest substantially in their networks to achieve their objective of a safe and efficient railroad. However, derailments caused by track geometry defects pose significant risks to this objective. Issues such as profile deviations and alignment irregularities often lead to train derailments, which can disrupt operations and increase repair costs. Therefore, there is a need for predictive strategies. To address this challenge, this research develops a quantum neural network (QNN) to predict track profile and alignment defects based on subsurface conditions, utilizing data obtained from Ground-penetrating Radar (GPR), which includes the ballast fouling index, ballast thickness index, layer roughness index, and moisture likelihood index. The developed QNN consists of a ZZ feature map layer and an EfficientSU2 ansatz layer, trained using the COBYLA and LBFGS optimizers. Additionally, a classical neural network (CNN) was developed to establish a baseline for comparison. Furthermore, Shapley Additive Explanations (SHAP) were employed to analyze the QNN’s decision-making process, ensuring interpretability for practical applications. The study’s findings indicate that the QNN, trained with the COBYLA optimizer, outperformed the CNN in predicting alignment defects, while the CNN performed better for the left profile. Both models exhibited similar results for the right profile. SHAP analysis revealed that the layer roughness index and ballast fouling index were the most influential factors in predicting defects in track geometry, followed by the ballast thickness index, with the moisture likelihood index being less significant. These results align with engineering principles, confirming the model’s suitable performance. This work demonstrates that QNNs can improve predictive maintenance, with SHAP providing a clear understanding of the model’s predictions for industry use.