The widespread digitization of critical industrial systems such as nuclear reactors has led to the development of digital twins and/or the adoption of artificial intelligence techniques for simulating baseline behavior and performing predictive maintenance. Such analytical tools, referred to as anomaly detection techniques, rely on features extracted from data that describe the underlying physical process. While these anomaly detection systems may work well with simulated data, their real-world applications are often hindered by the presence of noise. In some cases, noise may obscure subtle anomalies that may carry information about incipient stages of system faults. These subtle variations may also be the result of malicious intrusion such as so-called false data injection attack, equipment degradation causing sensor drift, or other natural disturbances in the process or the sensors. Consequently, there is a need to extract features that are robust to noise and also denoise data in a manner that aids machine-learning (ML) tools in diagnostics. In this regard, this paper presents a singular value decomposition–based statistical data–driven approach for feature extraction, denoted by randomized window decomposition, to capture the underlying physics of the system. Additionally, the features are used to denoise data to reveal subtle anomalies while also preserving relevant information for ML algorithms. The denoising algorithm is demonstrated using a RELAP5 simulation of a representative nuclear reactor with virtual noise.