To address the challenges posed by the numerous features and the limited generalization capacity of a single model in Internet of Things(IoT) intrusion detection, we propose a dynamic weight-based integration intrusion detection model for IoT. Initially, feature selection is carried out by combining the correlation coefficient and managing multicollinearity. Subsequently, we employ random undersampling and adaptive synthetic sampling methods to rectify the issue of data imbalance. Lastly, the classification weights of the integrated model are dynamically adjusted based on the entropy of feature information and the squared classification error, effectively addressing the problem of insignificant weights within the base model. Tests have been finished on the TON_IoT dataset, the experimental results show that the improved weight updating integration model achieves 99.971% in accuracy, and the generalization of the model outperforms a single model for multi-classification tasks. The model in this paper is suitable for IoT application scenarios with multiple classifications and large-scale data.