Abstract: Accurate identification of offline faults in power intelligent IoT terminals can greatly improve operation and maintenance efficiency. But with the construction of a new type of power system, intelligent IoT terminals are gradually diversified, and the collected historical data is also gradually enriched. Strong feature extraction and analysis capabilities are needed to accurately analyze the offline reasons of terminals from historical data as much as possible. Based on the above issues, this paper proposes a generalized ensemble learning method for power intelligent IoT terminal fault classification using knowledge data fusion. Firstly, a GRU-DNN-Attention network model is constructed, which utilizes the GRU and attention layers to extract abnormal features from historical time series data, and integrates knowledge to improve the classification accuracy of the algorithm. Furthermore, a generalized ensemble learning algorithm based on dynamic weight adaptation is proposed to solve the problem of poor performance of neural networks in small datasets. Simulation experiment results show that the algorithm proposed in this paper can effectively classify and identify the causes of offline faults in power intelligent IoT terminals.