Research on fault sensor detection and localization method based on autoencoder

Some sensors in nuclear power plants are exposed to extreme conditions, including high temperatures, high pressures, and intense radiation. As a result, the sensors may experience varying degrees of aging or malfunction. In this paper. we investigate the problem of detecting failures and locating faults in nuclear power plant sensors and proposes a sensor failure detection and localization method based on autoencoders. This method primarily consists of three parts: anomaly detection, fault diagnosis, and faulty sensor localization. Among these, the autoencoder method is used for anomaly detection and faulty sensor localization; feature extraction methods and the random forest algorithm are employed for fault diagnosis. This study used the Fuqing Nuclear Power Plant simulator as the data source to verify the effectiveness of the proposed algorithm. The experimental results show that the autoencoder method can identify faults and normal states with 99.6 % accuracy. The feature extraction-Random Forest method used to distinguish between sensor faults and system faults can identify system faults and sensor faults with 100 % accuracy, and the feature extraction method proposed in this paper is generally applicable, improving accuracy by 2.65 %-14 %. The autoencoder method for fault sensor localization demonstrates good diagnostic accuracy in both single-sensor and multi-sensor fault localization with a low false alarm rate. The autoencoder-based sensor fault detection and localization method demonstrates good diagnostic capability, providing a universal and direct sensor fault diagnosis framework for various types of sensor fault diagnosis.