Objective: With the continuous expansion and increasing complexity of water diversion tunnels in hydropower projects, their long-term structural safety has become a critical engineering challenge. Conventional safety evaluation methods often rely on qualitative assessments or multi-index systems, which are highly subjective and fail to adequately account for progressive material deterioration and time-dependent deformation. This study proposes an integrated quantitative framework that combines real-time inversion of mechanical parameters and deformation evolution analysis to dynamically evaluate the structural safety of tunnels. Methods: The proposed framework integrates four major components: data preprocessing, parameter inversion, deformation simulation, and safety evaluation. First, the raw deformation monitoring data are preprocessed by imputing missing values using the K-nearest neighbors (KNN) algorithm, identifying and correcting outliers with a sliding-window Z-score method, and reducing noise through logarithmic trend fitting. Second, a physics-informed inversion approach combining deep learning architectures—fully connected layers (FCL) and gated recurrent units (GRUs)—with Bayesian optimization is established to infer the current mechanical parameters of the tunnel from preprocessed deformation data. Third, an elastoviscoplastic damage creep constitutive model based on internal variable thermodynamics is employed to simulate deformation behavior under various material-degradation scenarios, represented by different strength reduction coefficients (K_r). Finally, based on the analysis of material creep behavior and deformation evolution patterns, a time-dependent "3S" safety evaluation index system is established. This system comprises the long-term deformation acceleration safety coefficient (S₁), the nonlinear deformation initiation safety coefficient (S₂), and the short-term deformation acceleration safety coefficient (S₃). The safety state of the tunnel structure is quantified according to the relevant deformation evolution characteristics using the proposed 3S safety coefficients. The physical implications of these indices are as follows: S₁ characterizes the critical point at which the structure transitions into an accelerated creep phase under continuous strength attenuation, indicating the long-term instability risk; S₂ reflects the onset of deviation from the linear response during initial deformation, marking the beginning of dominance by nonlinear mechanical behavior; and S₃ indicates the threshold for notable acceleration of deformation within a defined short-term period, serving as an indicator of potential sudden instability. Results: The proposed method was implemented in the JLL Tunnel, a 20km-long underground structure located in Hunan Province, China, which features complex geological conditions. The mechanical parameters were successfully inverted from field monitoring data, with simulated deformation curves showing high agreement with the measured values. Numerical simulations under different K_r conditions revealed distinct deformation patterns. For K_r≥0.7, deformation stabilized after initial convergence. When 0.4≤K_r≤0.6, deformation exhibited slow growth, followed by an acceleration phase. For K_r≤0.3, deformation accelerated rapidly within a short time. The computed 3S safety coefficients were S₁=2.4-3.0, S₂=3.7-4.6, and S₃=5.7-7.3, indicating that the tunnel is currently in a safe state with sufficient safety margins. These results validated the method's effectiveness in distinguishing between short- and long-term risks and in providing early safety warnings through deformation trajectory analysis. Conclusions: This study proposes an integrated quantitative framework for tunnel structural safety evaluation that effectively combines real-time monitoring data, physics-based modeling, and deformation evolution analysis. The established 3S index system provides a refined insight into structural behavior under material degradation and enables safety assessment across multiple time scales. Compared with conventional methods, the proposed framework enhances objectivity, supports the dynamic prediction of time-dependent performance, and facilitates lifecycle safety management and preventive maintenance of tunnel structures. The methodology demonstrates strong generalizability and offers remarkable practical value for risk prevention and sustainable operation in tunnel engineering.