The operation status of coal mills in coal-fired power plants is influenced by factors such as coal quality variations, load fluctuations, and equipment aging, which can lead to dynamic shifts. Existing monitoring methods typically rely on the “offline modeling, online deployment” approach, which is difficult to realize the adaptive, continuous and precise condition monitoring. It′s khown that knowledge distillation methods use lightweight student models to inherit the superior performance of complex teacher models, facilitating the rapid model updates and online deployment. Therefore, we propose an adaptive condition monitoring method for coal mills based on the continuous distillation. This method enables the quick adaptation to dynamic changes of coal mill during the operation process by continuously guiding the student model through the teacher model and updating the student model online. Considering the characteristics of coal mill data, we ropose the graph temporal convolutional network as the teacher model by combing the feature extraction advantages of graph convolutional networks and temporal convolutional networks. The student model is constructed based on a composite loss function, inheriting the knowledge from the teacher model via distillation loss and ensuring the monitoring accuracy with supervised loss. The new parameter fusion strategy is designed to periodically update the student model′s parameters based on real-time data, achieving the iterative optimization of the parameters. Validation with operational data of power plant shows that the proposed method outperforms comparison methods in both monitoring accuracy and adaptability. At the normal operating conditions, the standard deviation of prediction residuals under the continuous distillation approach is reduced by an average of 8.45% compared to the offline modeling method, significantly enhancing the stability of model. While in the abnormal operating scenarios, the proposed method successfully captures the fault symptoms and issues early warning signals 116 hours in advance, while maintaining a zero false alarm rate. In conclusion, the proposed method can improve the intelligence level of equipment operation and maintenance, demonstrating the broad prospects for engineering applications.