In a nuclear power plant, the reactor coolant pump (referred to as the "main pump") is one of the core components of the reactor coolant system. Real-time monitoring of its operational status is critical for ensuring the safe, stable, efficient, and economic operation of the nuclear power plant. To enhance the ability of traditional multivariate state estimation methods in handling time-varying systems and to achieve real-time tracking of system state changes, a dynamic multivariate state estimation technique (DMSET) algorithm was proposed. This method used the nearest neighbors (NN) algorithm to reconstruct data, calculates the anomaly degree based on the residuals between the reconstruction results and actual values, and identifies anomalies using a set threshold. In the memory matrix update, DMSET utilized the isolation forest (IF) method to assess the importance of samples, ensuring that the memory matrix contains the most recent and representative information. Experimental results on a nuclear power plant main pump operational dataset show that DMSET can effectively detect anomalies in real nuclear data, particularly excelling in real-time data streams, maintaining high stability in changing environments. The detection accuracy and recall rate are both above 90%, and the detection performance outperforms traditional anomaly detection algorithms such as principal component analysis (PCA) and MSET, demonstrating practical value and research significance.