The locomotive number is a crucial identifier for distinguishing locomotive types and identities, and its accurate recognition is of great significance for railway transportation safety and the construction of intelligent operation and maintenance systems. However, in real operating environments, complex illumination and background interference easily lead to truncated feature activations and loss of semantic information; significant scale variations of number regions restrict the representational capability of fixed receptive fields; moreover, degradation phenomena such as character wear and adhesion weaken both local discriminability and global sequence consistency. To address these issues, an end-to-end joint recognition method for locomotive number recognition is proposed.The proposed method is constructed based on the YOLOv11 object detection network and an improved locomotive number recognition network, T-LPRNet (Train License Plate Recognition Network). A staged recognition strategy is adopted: after locating the locomotive number region in high-resolution global images, character sequence recognition is performed on the local regions, thereby achieving decoupled yet collaborative detection and recognition.In the design of the recognition network, the ELU activation function is introduced to enhance the stability of gradient propagation under complex illumination conditions. Furthermore, to handle character scale variations and local morphological degradation, an Adaptive Receptive Fields (ARF) module is devised, which realizes context-aware dynamic feature modeling through multi-branch convolutions and a spatial selection mechanism. Finally, to address the diversity of character categories and inconsistency in number lengths, a Character and Length Dual Supervision Mechanism (CLDSM) is designed. By jointly employing a character-aware focal loss and a length-aware loss, explicit constraints are established between local character discriminability and global sequence structural consistency. In addition, curriculum learning and self-paced learning strategies are incorporated to improve training stability and generalization ability.Experiments are conducted on a self-constructed locomotive number dataset collected from real operating environments. The results demonstrate that the proposed method improves the Whole plate accuracy by 5.6% over the baseline, reaching 0.937. The Character accuracy, Precision, Recall, and F1 score achieve 0.979, 0.926, 0.908, and 0.917, respectively. These results verify the effectiveness of the proposed method in complex scenarios and provide a technically valuable solution for the construction of intelligent railway operation and maintenance systems.