Due to the unique power generation method used in nuclear reactors, nuclear power plants are more safety-sensitive compared to conventional power plants. Therefore, daily monitoring of the operating state of nuclear power units is critical for ensuring operational safety. Currently, status monitoring of nuclear power plants is conducted through automated alarms with preset fixed thresholds. and manual supervision. However, this method cannot detect anomalies below the alarm thresholds, which may lead to risks of underreporting. Nuclear power operational data, characterized by high-dimensional time series, faces challenges of imbalanced distributions between normal and abnormal samples as well as the lack of labeled data. These factors limit the application of supervised deep learning methods. In this paper, we propose an unsupervised deep learning model based on Variational Autoencoders (VAE) for anomaly detection in real operational data. This model learns the distribution of data in the latent space under normal operating conditions and relies on the principle that abnormal data cannot be reconstructed effectively. Anomalies are detected by evaluating the magnitude of reconstruction error. The experiment focused on the upper charging pump in the chemical and volume control system (RCV) of a nuclear power plant. It involved the validation of the model using real operational data with deliberately inserted anomalies and compared it to classical machine learning methods. The results of the experiment show that the model based on Variational Autoencoders effectively detects abnormal data segments and outliers in real nuclear power plant data. It achieves precision and recall rates both exceeding 90%. In terms of detection performance, it outperforms classical machine learning algorithms like Isolation Forest and support vector machine (SVM). This demonstrates its practical value and research significance.