Abstract: Data-driven based supervised structural health monitoring methods require to collect monitoring data for various damage states of structures as training data for mathematical models or artificial neural networks. To overcome the practical difficulty of requiring a large amount of labeled training data for supervised methods, this paper proposes a novel data-driven building structural health monitoring method. This method only uses vibration data collected from the healthy state of a structure to train a convolutional autoencoder, and the trained convolutional autoencoder can reconstruct the test data obtained from the same healthy state of the structure very well. For the test data obtained from the structural damage states, certain degrees of data reconstruction loss will occur in the data reconstruction process. The experimental multi-layer structure mathematical model verifies the effectiveness of the method, and the test data from different damage states of the structure correspond to different degrees of data reconstruction loss. The novel structural health monitoring method proposed in this paper can be used to accurately and timely monitor the possible structural damage of buildings.