In system state recognition and prognostics, state transition probability matrix of hidden semi-Markov model (HSMM) is constant and the predicted life value shows stepladder change, which is different from the actual residual life of the system. To solve this problem, an HSMM with time varying state transition probability matrix was proposed. Based on the analysis of three typical degradation states of the system, three different state transition coefficients were given. Combined with initial state transition matrix, a time varying state transition matrix was obtained, the estimation accuracy of residual life of the system under current healthy state was increased, and a more accurate overall residual life prediction value can be obtained. Experiment results show that, compared with traditional HSMM, HSMM based on time varying state transition probability matrix can increase the accuracy of residual life prediction and can be used in life prediction with high precision.