Abstract: To enhance the accuracy and reliability of fault early warning of induced draft fan, an unsupervised intelligent early warning method that integrated temporal convolutional network (TCN) and stacked denoising autoencoder (SDAE) was proposed. By innovatively combining the temporal modeling capability of TCN and the feature abstraction strength of SDAE, a predictive model was constructed using normal operating data of the induced draft fan. Mahalanobis distance of the reconstruction error was used to quantify deviations from expected behavior, and a sliding time window combined with Chebyshev inequality was employed to establish dynamic multivariate warning thresholds, effectively improving precision and adaptability of anomaly detection. To enhance the robustness of identification, a consecutive threshold-exceedance mechanism was designed to suppress transient disturbances. Meanwhile, a variable-level exceedance heatmap was introduced to visualize the degree and evolution of anomalies across monitored features, significantly improving the interpretability and operational relevance of the results. Experimental validation on a 300 MW coal-fired power plant demonstrated that the method enabled timely identification of potential faults and exhibited strong robustness and adaptability. The proposed approach provides a practical solution for intelligent monitoring of auxiliary systems and supports the safe and stable operation of smart power plants, and is promising for engineering application.