tion status (Fig. 1c) of a tunnel. With the ageing and usage of an engineering system such as tunnels, historic damage data for a given failure mode (e.g., crack or seepage) are usually collected from the field inspection periodically (solid dots in Fig. 1a). These damage data points generally follow a statisticaldistribution(e.g., Weibullorlognormaldistribution,seethe dashcurvesinFig.1a)atany given inspection time. The damage indicator at the same location inasectionorsegmentofthetunnelshouldincreaseovertheoperating time, assuming no maintenance activities conducted prior to itsservicelimit.Assuch,alldamagedatahistoricallyobservedfora given failure mode should display a monotonically increasing trend as the operating time increases due to the damage accumulation. The percentiles of damage at various inspection intervals can be estimated from the inspection data, shown as Px curves in Fig. 1a, where the subscript x represents the percentage of the damage along the operating time. If a damage limit is predefined for a given failure mode based on probabilistic limit design theory (the dash line in Fig. 1a), the time-to-failure or the residual useful life of the tunnel segment can be estimated from the damage accumulation curve at different percentages. The obtained results can be then used to schedule next inspection time in order to ensure the tunnel in a safe operating condition. In the condition based predictive maintenance of a shield tunnel, the service conditions of its prefabricated lining structure can be categorized intro three levels: initial degradation, serviceable, and repairable. Given a performance indicator such as structural deflection, the service conduction of the tunnel structure can be defined and evaluated based on the probabilistic theory, considering uncertainty (Fig. 1c).Their corresponding service limits may be determined according to the probability of structural damage, which is calculated from service condition based reliability or performance model of the structure, as shown in Fig. 1b and c. Different services conditions as well as their deviation from the design condition will determine the structural performance of a tunnel in operation. The initial degradation limit may be decided when the probability of failure reaches up to an acceptable small level. In addition, the structure does not need to be repaired before its failure probability reaches to an acceptable service limit. After then, the structureshould be repaired any time before its probability of failure reaches to the repairable limit. If the damaged structure is not repaired before repairable limit, the structure may not berepairableanymore.Notethatthedurabilitymaybeconsidered as another category in the condition assessment, as discussed in Yuan et al. (2012), which is beyond the scope of this study. In the following sections, the framework and architecture of condition based predictive maintenance is first presented at a system engineering perspective. The components of maintenance framework are then interpreted in details as nine necessary parts in a predictive maintenance strategy, such as failure modes and effective analysis, data pre-process, reliability modeling, and system-level lifinganalysis. The empirical lifinganalysis involves both risk prediction and damage accumulation models for service limit determination, system-level risk analysis, and system-level conditional risk for maintenance schedule. Next, the proposed methodology is demonstrated with the inspection data collected for six typical defects observed in real-world shield tunnels. Finally, concluding remarks are provided.