During the operational phase of thermal power plants, coal mills, functioning as essential auxiliary devices, are prone to the occurrence of coal blockages. To effectively avoid sudden shutdowns caused by coal plugging, this paper proposes a coal mill plugging fault warning method based on a Transformer-BiLSTM model. First, through Spearman rank correlation coefficient analysis, 6 key features related to coal mill plugging are selected as inputs for the warning model. Given the characteristics of unstable coal mill operational data, wavelet denoising technology is employed to smooth the time-series data, effectively reducing noise interference while preserving critical fault-related feature information. The denoised data is input into an integrated prediction model combining Transformer and BiLSTM to obtain the residuals between predicted and actual values. Subsequently, the sliding window method is employed to determine the threshold value for fault warning, thereby enabling the early detection of potential coal mill plugging faults. The proposed method is experimentally validated using coal mill operational data from a 350 MW unit. The experimental results show that the proposed model reduces MAE by approximately 35.74% and RMSE by approximately 35.20% compared with the Transformer model. More importantly, it achieves an early detection time of approximately 4 min for coal mill plugging faults, providing sufficient response time for on-site operators to take preventive measures.