Timely and accurate detection of defective fuel rods is critical as the release of radioactive fission products from defective fuels can lead to primary circuit contamination and radiation exposure. Due to the complexity of the physical phenomena, models for fault diagnosis can be difficult to construct and recently data driven surrogate models have being increasingly used to detect and characterize defective fuel rods: they make use of a computational database to learn from and make predictions about new unknown data. In this paper, we present a method for the elaboration of an anomaly detector based on neural networks, taking into account the fact that physical computation can be CPU intensive and thus overcome this issue. A physical model for fission products release and coolant activity calculation was built and used to generate a surrogate activity model that enables the generation of a bigger database in small amount of CPU times. Then using this bigger computational database, a recurrent autoencoder was trained for anomaly detection. The network classifies the defect status with 100% accuracy and a good time precision. A sensitivity analysis with lower activity increase at defect onset and addition of noise was conducted in order to better understand the limits of this method. Such methods can be useful for operators of the existing as well as future reactors to make timely predictions of defective fuel rods and avoid operational and economic setbacks for power plants.