Extracting sensitive damage features from structural response signals is crucial for damage identification methods based on pattern classification. To this end， a hybrid network that combines a deep belief networks （DBN） and a long-short term memory （LSTM） network is proposed through a hybrid learning mechanism to utilize the merits of both networks in the aspects of extracting high-order abstract features and considering data sequence correlations. First， transmissibility data from response signals are sequentially input into the DBN to achieve the initial data compression and feature extraction， reducing the redundant information in the responses. Then， the extracted feature sequences are input into the LSTM network to consider the correlation between the different responses for acquiring the relevant sensitive damage features. Finally， a classification layer with the Softmax function is used to classify the features output by the LSTM network. Thereby， different structural damage patterns can be identified. The damage identification results on a three-dimensional experimental steel frame demonstrate that the hybrid learning mechanism can better train the network parameters， and the fine-tuning on the whole hybrid network contributes to the subsequent damage feature classification. Under the pollution of numerical or measured noises， the hybrid network can still effectively perform the data compression， feature extraction and classification. The various damage scenarios of the experimental frame are well identified.