End-effector contact force estimation is a critical technology for high-precision force control tasks in industrial robots, and its accuracy directly impacts the safety and compliance of operations. Traditional generalized momentum observers (GMOs) have shown limitations in dynamics model accuracy and adaptability to varying operational conditions. This paper introduces an enhanced GMO approach utilizing a physics-informed neural network (PINN) dynamics model to overcome these challenges. First, the PINN method is employed to establish a precise robot dynamics model, enabling efficient inference of the inertia matrix derivatives. Second, an adaptive gain adjustment mechanism is designed to dynamically optimize observer performance based on robot joint velocities, significantly enhancing the robustness and adaptability of the observer under disturbances across different frequencies. Finally, the effectiveness of the proposed method is validated through simulation experiments. The results demonstrate that the proposed method achieves effective contact force estimation under disturbances of varying frequencies and outperforms other GMO-based methods.