To address the challenges of lithium battery state assessment and safety evaluation, this study proposes a non-invasive, online internal temperature estimation method for lithium batteries by integrating Physics-Informed Neural Networks (PINN) with edge computing. We develop PhysLiteNet, a lightweight model that synergizes physical constraints from an electro-thermal coupling model with a Multi-Scale Temporal Convolutional Network (MS-TCN). The model is quantized and deployed using the TensorFlow Lite framework, enabling low-latency, online inference on embedded edge devices. An online temperature estimation test system is developed, where the trained PhysLiteNet model is successfully migrated and experimentally validated. Experimental results demonstrate that the root mean square error (RMSE) between estimated and measured internal temperatures remains below 0.24 °C, with inference time per cell temperature estimation not exceeding 120 ms. This research provides innovative technical pathway and practical engineering solutions for edge-side temperature monitoring in large-scale energy storage systems.