An adaptive artificial neural network (ANN) framework is proposed for functional failure identification of passive safety systems in nuclear reactors. Applied to a passive steam generator cooling system, the method integrates ANNs with Query-by-Committee strategy to dynamically select highly informative samples for iterative model updating. This process effectively mitigates problems associated with small sample sizes and class imbalance. In contrast to traditional Monte Carlo methods, which typically require simulation runs on the order of 104, the proposed approach reduces the number of required thermal–hydraulic simulations as few as 360, representing an approximate 96.4% reduction in computational cost. After 26 iterations, the model achieved a validation precision of 0.8421, a recall of 0.8005 and a PR-AUC of 0.8118, demonstrating its ability to effectively capture low-probability failure events and achieve performance convergence across training, test and validation sets, confirming the effectiveness of the adaptive ANN methods in selecting high-information samples and optimizing model performance over successive iterations.