Numerous sensor data are produced as a result of drilling operations as indications of dynamic interactions among downhole equipment, formation properties, and operational processes. It is important to detect anomalies in these data streams to avoid equipment failures, minimize non-productive time (NPT), and conduct the drilling process safely. Conventional rule-based and threshold-based surveillance systems do not represent more intricate and non-linear correlations between drilling parameters, which restricts its usefulness. This paper suggests a learning framework of real-time anomaly detection in the drilling process as an unsupervised system. The model uses clustering and autoencoders to train normal operational behaviors using multivariate sensor data such as torque, vibration, mud flow, rate of penetration (ROP), and downhole pressure. Abnormalities in the patterns are identified to raise an alert through anomalies. The models were trained on simulated drilling datasets of different formation lithologies and operational conditions and evaluated. The proposed method had F1-score of 0.91, precision of 0.93, and recall of 0.89, which is about 18 times more accurate than the baseline statistical threshold methods. It has been shown that the unsupervised learning can detect the smaller and newer anomalies that might go undetected by the conventional means, giving a scalable and adaptive process of intelligent drilling activities. Such a structure has major implications in terms of predictive maintenance, operational effectiveness, as well as designing autopilot drilling frameworks.