When dealing with massive high-dimensional data scenarios, intrusion detection technologies still confront issues such as imbalanced data distribution and low detection accuracy. To address these problems, an intrusion detection method based on deep learning is proposed. Firstly, during the data preprocessing stage, an improved autoencoder model based on the diffusion model is utilized for feature selection and the generation of high-quality minority-class abnormal traffic samples. The SMOTE-Tomek technique is employed for further upsampling and noise removal, and outlier detection techniques are adopted to remove outlier samples. Secondly, a novel deep learning model is constructed. The preprocessed data is input into a BERT encoder with split attention mechanisms to model different feature subspaces in a more fine-grained manner. The generated vector information is transformed into feature grayscale images and then input into the improved ResNet model. By integrating pyramid pooling, depthwise separable convolutions, and focal attention mechanisms in the original ResNet network, the model hierarchically extracts features step-by-step, ultimately outputting refined feature representations that integrate multi-scale information and emphasize critical features. Finally, classification experiments are conducted based on the trained model. The effectiveness experimental results on the NSL-KDD dataset indicate that the proposed model achieves 92.92% binary classification accuracy and 86.41% five-class classification accuracy, outperforming all other comparative models. The feasibility experiments on the UNSW-NB15 dataset further validate that the proposed model possesses excellent classification performance and the capability to handle imbalanced data distribution.