The safety assessment is of great significance for ensuring long-term safe operation of high arch dams. The machine learning algorithm of Continuous Hidden Markov Model (CHMM) has strong capabilities in processing multi-sourced time sequential continuous data, establishing relationship between complex data, and real-time accurate evaluation. Therefore, it has great applicable potential in high arch dam safety evaluation. However, the whole-process safety evaluation of overloading failure of high arch dams via CHMM-based multivariate indices analysis has not been carried out yet. Therefore, this paper carried out an overloading test on a physical model of a high arch dam, monitored the time series of cracks and deformations throughout the dam's failure process, and constructed a high-quality CHMM dataset with low noise and stability. Subsequently, CHMM was applied to the safety assessment of high arch dams, and a new rule for state transfer matrix ordering was proposed to give meaning to the state labels, which improved the readability of the results of CHMM. Finally, the whole-process safety variation of arch dam model was evaluated by the CHMM optimized state sequence, and validated with observed results. The results indicated that CHMM-based state sequence classified the safety status of tested high arch dam into seven levels, which coincide with the time-sequential evolution of deformation and crack, and further proposed the featuring overloading safety. The failure process of tested high arch dam was categorized into quasi-linear, nonlinear of large deformation, arching effect action and complete failure stages with respective overloading safety degrees of 2, 6, 10 and 14. This study provides the new insight into safety evaluation of high arch dam and makes reference for early-warning of high arch dam.