A damage detection method based on wavelet packet-singular value decomposition (WPT-SVD) and BP neural network model optimized by genetic algorithm (GA-BPNN) was proposed aiming at the problem of damage discretization of concrete structures based on piezoelectric wave method. The damage characteristics of signal were deeply explored for structural cracking in time-frequency domain. The relationship between the signal characteristics and the damage condition was built, which can identify the damage location and the damage degree of structures. The damage signal was measured by surface-mounted piezoceramic transducer on the concrete structures. The WPT was employed to obtain the multi-dimensional time-frequency matrix. Then SVD was used to reduce the dimension of time-frequency matrix under different damage states and construct feature vectors with high damage sensitivity. The GA-BPNN model with auto-adaptive learning was established, and the identification of structural damage was realized. The experimental results show that the singular value of piezoelectric signal can be used as the characteristic parameter of damage. The singular value of the main spectrum decreases with the development of damage. There is a correspondence of three stages between the normalized singular value vector distance and the damage. The GA-BPNN better fitted the correlation between the damage signal characteristics and the damage compared with the BPNN model. The identification was more stable, and the accuracy was significantly improved. The identification accuracy of damage location and damage degree of concrete structures were 95.19% and 94.47%, respectively.