The ongoing advancements in high-speed trains have greatly enhanced personnel mobility and societal eciency. However, the issue of polygonal wear on train wheels has garnered increased attention in recent years. Polygonal wear, a form of periodic wear occurring in the wheel’s circumferential direction, compromises the safety and operational eciency of high-speed trains, elevates operating costs, and generates wheel-rail noise, which adversely a↵ects passengers’ experience. Despite the complex nature of the physical mechanisms underlying polygonal wear, research on this subject remains in its infancy. To address the e↵ects of polygonal wear and other wear patterns, current wheel maintenance strategies for high-speed trains involve periodic maintenance, which overlooks the actual condition of the wheels and may result in excessive or premature maintenance. In order to optimize cost reduction and eciency, it is crucial to assess wheel conditions and base maintenance decisions on their state. In practice, the condition of high-speed train wheels is monitored every 48 hours. However, only parameters such as diameter and flange thickness are detectable, rendering it impossible to gauge the extent of polygonal wear. Consequently, estimating the degree of wear is essential. Wheel maintenance must consider the degra- dation processes of wheel diameter, flange thickness, and the degree of polygonal wear for all wheels on the entire train, taking into account the correlations between di↵erent degradation processes within the same wheel. Moreover, during wheel monitoring, it is necessary to determine if maintenance is required and, if so, decide on the appropriate reprofiling depth for each wheel, which further complicates the issue. Since the state of the wheels at each monitoring session depends solely on wheel diameter, flange thickness, and the degree of polygonal wear, we can model this maintenance problem using a Markov decision process. Given the complexity of state and action spaces, we employ a deep rein- forcement learning approach and optimize the calculation process by adjusting the network structure and loss function, making it more suitable for addressing such problems. Nu- merical experiments demonstrate that, compared to traditional periodic maintenance, our method significantly reduces maintenance costs and enhances the reliability of the entire vehicle. This ultimately achieves the goal of cost reduction and eciency improvement in high-speed train wheel maintenance.