Due to the common existence of interlayer debonding and separation of longitudinal continuous slab track (LCST), accurately estimating the interfacial normal cohesive parameters between the concrete track slab and the cement emulsified asphalt (CA) mortar layer is an important task. This study first carried out a full-scale vertical pull test (VPT) to study the interfacial normal bond capacity of LCST. As a result, a series of load–displacement curves were measured. Then, an artificial neural network (ANN)-based approach was developed to map the relationship between the structural response and the interfacial properties under a machine learning (ML) framework. In addition, a refined macroscale finite element (FE) model that employed the exponential cohesive zone model (CZM) was established to simulate the interfacial debonding process. Two datasets of the global load–displacement and the local stress-slip responses obtained from the FE analysis were separately used to train and verify the ANN. Our results showed that the predictions given by the ANNs and the ground truth values were in close agreement. Furthermore, by feeding the well-trained ANN with the experimental load–displacement curves of the VPT, the realistic interfacial normal cohesive parameters of LCST were identified. Afterward, a comparative analysis of the experimental results and the recovered results according to the identified parameters was carried out. The results showed that the proposed parameter determination method is accurate and reliable. The developed hybrid approach that combines experimental and FE analysis with ML methods can be a promising alternative for identifying material properties in structural engineering.