Implementation of two computational techniques for the detection of abnormal conditions in safety-critical systems: an application to the simulator of a nuclear fission plant

Safety in nuclear power plants is very important and the capability of detecting possible anomalies and perturbations through computational models is encouraged by the regulatory authorities as a complementary tool in safety studies and safety assessment. Anomaly detection provides early warning of faults, by identifying deviations in behavior between real-time data from the system and the expected values produced by a predictive model. In the present work, two different methods have been implemented for the detection and isolation of outliers: the Autoencoder Method and a density estimation method, using Finite Mixture Model. In the former, a sparse autoencoder is trained with normal transactions and it will learn how to represent a normal input data. Once it tries to reconstruct an anormal data, it is expected that the model will worsen its precision. In the latter methodology, a pair of features have been selected to represent the input data, while Finite Mixture Model have been used as density estimation method. In this way, it is possible to calculate the outlyingness for each transient to isolate the outliers. The two methods have been tested using two different applications: a synthetic dataset used for comparison with literature work and a simulator regarding a MSFR (Molten Salt Fast Reactor). The precision of both the presented methods appears to be very high when the degree of contamination, the percentage of outliers in a dataset, is lower than 5%. In particular, Autoencoders are capable to reach a precision higher than 98%. Increasing the degree of contamination to values between 5% and 10% both the algorithm have a comparable accuracy.