This paper introduces a novel physics-informed neural network (PINN) approach for inferring annual reliability parameters of multi-segment distribution lines from system reliability index observations. The proposed methodology seamlessly integrates domain-specific governing principles that characterize the relationships between section-wise failure rates and repair rates with observed system reliability indices during the neural network training process. The governing physical laws are mathematically formulated as a mixed-integer optimization problem, in doing so, a unified framework that combines the rigorous optimization theory with deep learning capabilities is established. The trained PINN enables superior parameter estimation performance while maintaining the inherent physical consistency observed in historical operational data. Comprehensive experimental validation demonstrates significant improvements of the proposed approach in both prediction accuracy and computational efficiency in comparison with conventional optimization-based methods. The results highlight the method’s potential for practical deployment in power system reliability analysis and maintenance planning applications.