In order to support the student performance evaluation in rail transit teaching, this paper constructed a student performance prediction model using multi-source teaching data. The dataset contains 5240 valid samples from the theoretical course platform, driving simulation system, scheduling training terminal, signal disposal module and stage assessment records, and the behavior and evaluation characteristics are established around learning time, operation accuracy, response delay and alarm disposal. After label construction, normalization and feature selection, the model combines behavior weighted mapping, weekly scale aggregation, stability index and structure score with gradient boosting training and temperature calibration to output three types of results and corresponding confidence levels: excellent, standard and early warning. Experimental results show that the accuracy of the proposed model on the test set reaches 94.6%, the macro-average F1 reaches 93.8%, the AUC reaches 0.962, and the ECE is reduced to 0.028. The model maintains stable performance under different modules and repeated tests, which can provide a computable basis for hierarchical identification, process tracking and teaching feedback in rail transit teaching.
