The safe operation of nuclear reactors depends on steady monitoring and clear interpretation of large streams of sensor data. These signals reflect complex thermal, mechanical, and chemical interactions, and even small deviations can indicate early signs of abnormal behavior. Traditional fault detection approaches rely on fixed limits or simplified physical assumptions, which often break down when operating conditions shift. Data-driven and deep learning methods improve detection accuracy but usually work as closed systems that offer limited insight into how decisions are made, which is a major concern in regulated environments. Generative models help address the shortage of rare fault examples by producing realistic synthetic sequences that broaden the range of events available for training and evaluation. Explainable learning methods complement this by identifying which variables or time intervals influence an alert, making model behavior easier to understand. Analysis of regulatory event records adds another dimension by capturing human, procedural, and operational factors that do not appear in sensor data alone. This review brings these strands together and presents an integrated view that links synthetic data generation, interpretable anomaly detection, and text-based event reconstruction. It examines progress in model design and evaluation practices, highlights gaps in interpretability, data variety, and regulatory consistency, and outlines directions for building transparent and verifiable intelligence suited for day-to-day reactor monitoring and long-term maintenance planning.