Considering the shortcomings of current intrusion detection methods in feature extraction,classification accuracy, and generalization capability, this paper proposes an intrusion detection model that integrates multi-scale convolutional neural networks, long short-term memory networks, and residual networks (multi-scale spatial-temporal residual network, MSST-RNet). First, log1p smoothing is applied to transform highly skewed data, and the effectiveness of this transformation is validated by comparing the correlation with labels before and after the transformation and through data distribution visualization. Next, the spatial features and temporal features of the data are extracted and fused using multi-scale 1D convolution modules and long short-term memory modules, respectively. Finally, the idea of residual networks is incorporated by adding identity mappings to avoid issues such as gradient vanishing, gradient explosion, and network degradation. Experimental results on the UNSW_NB15 dataset show that the proposed method can effectively enhance the model's representation capability and generalization ability, with significant improvements in various performance metrics.