<p>Traditional structural damage identification methods based on convolutional neural network (CNN) require extensive data collection and effective feature extraction from engineering structures, resulting in high computational costs and low recognition accuracy. To enhance the damage detection accuracy and computational efficiency, a new damage identification method is proposed based on improved one-dimensional deep separable convolutional neural network (1D-DSCNN) model. First, the traditional convolutional layers are replaced with depthwise separable convolutional layers to create a novel neural network model. Second, residual connections are incorporated into the depthwise separable convolutional blocks to accurately capture more damage features. The effectiveness and accuracy of the proposed method are validated via a numerical case of the IASC-ASCE SHM Benchmark structure model and two experimental tests on a four-story steel frame structure and a Qatar University grandstand simulator. The results demonstrate that the improved 1D-DSCNN model not only promotes damage identification accuracy, but also accelerates model convergence and decreases the number of model parameters in comparison to the Resnet-34 and conventional 1D-CNN models. Furthermore, the proposed new model exhibits strong noise robustness in terms of damage identification for steel frame structures.</p>