With the increasing scale and complexity of modern wind turbines, conventional physical sensors are no longer sufficient for high-fidelity monitoring and control. This review delivers the first in-depth synthesis of machine learning-enabled virtual sensors, offering scalable, intelligent alternatives for real-time diagnostics, predictive maintenance, and adaptive turbine control. These virtual sensors, often integrated into DT frameworks, estimate vital turbine parameters such as structural faults, gearbox loads, and wind conditions, including wind speed and vertical profiles that significantly influence turbine dynamics, without relying on direct instrumentation. We examine state-of-the-art approaches, including ANN, LSTM, Random Forests, and hybrid DT models, highlighting their application across SCADA systems, OPENFAST simulations, and condition monitoring platforms. The review maps machine learning techniques to key operational tasks such as load forecasting, wake detection, fault diagnosis, and structural health prediction. It identifies unresolved challenges in temporal resolution, model generalizability, real-time integration, and environmental robustness. Moreover, it evaluates deployment-ready innovations in edge computing, simulation-informed learning, and interpretable AI. Unlike prior studies focused on isolated technologies, this paper integrates developments across AI, DTs, and structural modelling offering a unified roadmap toward fully virtualized wind farms. Our findings underscore that virtual sensors are not just data proxies but essential enablers of resilient, autonomous, and Industry 4.0 ready wind energy systems. The paper concludes with strategic directions for future research, emphasizing the need for transferable DTs, explainable ML models, and benchmarking standards to accelerate industrial adoption.