Wireless sensor networks （WSNs） are gradually applied to structural health monitoring. Due to the involved energy consumption issue， it is difficult for WSNs to achieve long-term and high-frequency data acquisition. Compressive sensing （CS） is able to use a small number of sampling points to reconstruct the original signal， which is expected to reduce the energy consumption of the WSNs. The sparsity of the measured vibration signal is limited due to the noise contamination. This causes the failure of the LASSO， a widely-used CS algorithm， in seeking the accurate sparse coefficient， which hinders the reconstruction performance of CS of the vibration signal. This paper proposes a method to optimize the sparse coefficient to effectively improve the accuracy of reconstructed vibration signal by using BP neural network. The simulated acceleration data of a three-floor frame and the monitored acceleration data of Canton Tower are both used to verify the effectiveness of the proposed CS method. The effects of regularization parameters and the number of optimization iterations are explored in detail. The result shows that the proposed optimized CS method performs better than the non-optimized one under different compressed ratios.