Current systems for network intrusion detection (IDS) are essential for mitigating network attacks, yet they often struggle with accuracy issues and high rates of false positives, especially when confronting diverse types of attacks. To tackle these shortcomings, we have designed and evaluated a new model for detecting network intrusions. This model utilizes depthwise separable convolutions to isolate and analyze the spatial characteristics of attack traffic, followed by the use of gated recurrent units (GRU) to capture temporal features. After fusing spatial and temporal features, position encoding and a self-attention mechanism are applied to assign different weights, allowing the model to adaptively achieve precise intrusion identification. Evaluations performed using the NSL-KDD and UNSW-NB15 datasets demonstrate the proficiency of our model in multi-class tasks. Notably, with the UNSW-NB15 dataset, our model achieved a 95.22% accuracy, a 96.11% detection rate, and successfully maintained the false positive rate at 1.93%.