This paper presents a comprehensive review of anomaly detection techniques within Cyber-Physical Systems (CPS), emphasizing both advancements and limitations of current methods. We identify that hybrid models, such as the integration of Long Short-Term Memory (LSTM) networks with Seasonal Auto-Regressive Integrated Moving Average (SARIMA), significantly improve anomaly detection by reducing false positives by up to 30%. Traditional statistical and knowledge-based approaches are evaluated alongside machine learning and deep learning models, highlighting challenges such as handling noise in sensor data and ensuring scalability in large-scale systems. Special attention is given to emerging hybrid models and real-time detection frameworks, which address these challenges effectively. Additionally, the integration of edge computing is shown to reduce latency by 25% in real-time detection scenarios, enhancing performance in large-scale CPS applications. We also explore the role of edge computing in improving data processing efficiency and propose federated learning as a strategy for better collaboration across edge devices. The review concludes by identifying research gaps and proposing a roadmap for developing adaptive, resilient, and scalable anomaly detection frameworks, supported by practical use cases such as predictive maintenance and real-time traffic management in smart cities.