To address the issues of limited sensor deployment in real-time temperature monitoring of electric motors and insufficient computational efficiency of traditional thermal models, this paper proposes a real-time motor temperature reconstruction method based on finite element model and digital twin technology. By establishing an electromagnetic-thermal coupled finite element model and constructing a dynamic field prediction framework combining neural networks and superposition theorem, the method includes three main steps: First, employing Proper Orthogonal Decomposition (POD) for reduced-order modeling of the electromagnetic field to enable real-time analytical calculation of loss parameters. Second, proposing a segmented extraction and superposition reconstruction strategy that decomposes transient temperature rise into linear superposition of loss excitation responses and temperature decay processes, effectively reducing the data extraction volume of the finite element model. Third, constructing a multilayer neural network to learn spatiotemporal characteristics of the temperature field, achieving dynamic updates of full-domain temperature cloud maps at second timescales. Simulation results on an 8-pole 48-slot permanent magnet motor demonstrate that the reconstruction error of this method is 3%-6%, and the response speed is improved by over 80% compared with traditional finite element calculations, providing a high-precision digital twin solution for motor thermal management.