With the expansion of Internet of Things (IoT) device deployments and dynamic topological changes, emerging network attacks continue to proliferate, posing significant threats to critical infrastructure security. Intrusion detection systems (IDS) are recognized as effective solutions for IoT security protection, though existing methods still face challenges in detection accuracy and result credibility. To address these problems, an intrusion detection system based on convolutional neural networks (CNN) and D-S evidence theory is proposed. First, a multi-structure CNN collaborative detection framework is constructed to achieve feature fusion, which enhances the model's capability to characterize complex attack patterns. Subsequently, the learning rate at different gradients is controlled by introducing a deceleration factor to solve the overshooting and local optimum problems in the parameter optimization process. Furthermore, the outputs from CNN sub-models are fused using a reconstructed combination rule-optimized D-S evidence theory, thereby increasing the trustworthiness of detection results. The proposed model is evaluated through comprehensive experiments on standardized CICIoT2023 and ToN-IoT datasets. Experimental results demonstrate that compared with existing models, the proposed model effectively extracts attack features and accurately identifies security threats in networks.