In urban rail transit, rail maintenance is critical for ensuring safety and efficiency. Accurately assessing the state of rail damage and its influencing factors can focus attention on potential vulnerabilities, significantly reducing accident rates and enhancing economic efficiency. This study aims to develop a data-driven model to evaluate rail damage state, using maintenance record data from a rail transit system in actual operation from 2010 to 2021. Inputs for the model include line features, rail duration, and train passing speeds, with rail condition level as outputs. Various machine learning models were established and compared in this research, among which the random forest and the two-branch model provide the best overall balance. Results show that rail age (duration) is the most important factor, which needs to be focused on. Additionally, a high occurrence of railhead defects suggests poor wheel-rail contact, which also leads to noise. This model could potentially guide the prioritization of inspection and maintenance tasks at locations prone to rail failures, thereby enhancing infrastructure safety.