This paper proposes an evolution and prediction model of composite structural damage based on the approach of combining local tangent space alignment (LTSA) with relevance vector machine (RVM). In light of the characteristics of composite structural damage, we used a vibration fatigue test to predict structural damage. First, we acquired the health information of the composite structure via multi-sensor decomposed into intrinsic mode function (IMF) by ensemble empirical mode decomposition (EEMD). We then applied Hilbert-Huang transform(HHT) to IMF and obtained corresponding Hilbert marginal energy as the characteristics of each sensor. Second, we made feature fusion in order to obtain the characteristic energy using LTSA, and defined the resulting characteristic energy as the structure of the health index using the DMS method. Finally, regarding the structure of the health index as modeling data, we established a relevance vector machine (RVM) model, and completed the composite structural damage prediction research by predicting the structure of the health index. Results showed that this model can be used effectively to predict structural damage.