Spiking Neural Network Enables Continual Learning for Nuclear Plant Anomaly Detection

A novel spiking neural network system addresses catastrophic forgetting in nuclear industrial control systems, where conventional neural networks fail to retain previously learned anomaly patterns when sequentially trained on new subsystems. This safety-critical challenge is met with spike-encoded asynchronous sensor fusion, which converts heterogeneous sensor streams into sparse spike trains aligned with each sensor's natural dynamics, achieving 92.7% input sparsity. The approach evaluates five continual learning strategies including Elastic Weight Consolidation and Synaptic Intelligence. Designed for continuous, energy-efficient monitoring across multiple subsystems deployed at different stages of plant commissioning, the system represents the first SNN-based anomaly detection framework with continual learning capabilities for nuclear ICS. Research was published on arXiv under identifier 2604.18611v1 with announcement type cross.