Abstract: A multi-level data fusion monitoring method based on unified data preprocessing, improved Kalman filter and adaptive weighting is proposed to monitor the centralized control operation parameters in real time. In the phase of state estimation, aiming at the problem of process noise covariance mismatch and the decline of tracking accuracy of ordinary Kalman filter when the unit load fluctuates greatly, the improved Kalman filter based on sliding window adaptive noise estimation is introduced to realize the optimal state estimation of key parameters such as main steam temperature; In the process of multi-sensor measurement fusion, aiming at the problem of large equal weight fusion error caused by the accuracy difference of each sensor, an adaptive weighted fusion model based on the reciprocal of variance or information entropy weight is adopted to minimize the fusion estimation variance. At the same time, a hierarchical anomaly detection mechanism based on a variety of criteria such as dynamic threshold, Mahalanobis distance, rate of change monitoring and kernel density estimation, and an overall health evaluation method based on comprehensive health degree are established. The simulation verification based on the 30 day historical data of a 600 MW supercritical unit shows that compared with the traditional Kalman filter, the fusion RMSE of main steam temperature is reduced with a decrease of 55.9%; The state convergence speed increases about 5 times when the load steps; The abnormal detection rate was 96.8%, the false alarm rate was only 2.3%, the F1 score was 0.971, and the average response time was 0.4 s, which met the requirements of the national standard centralized control monitoring system that the response time was no more than 1s, and provided effective technical support for the safe and stable operation of centralized control.