In contemporary plasma procedures, a real-time adaptation of ignition and prompt failure detection is essential yet challenging due to the dynamic nature inherent in plasma processes. In this thesis, we address these challenges by demonstrating an online optimization approach with minimal computational cost and resource utilization while maintaining a high adaptation rate. We utilize prediction from a linear model for anomaly detection, achieved through our online optimization method for estimating the parameters of the linear model. Through the evaluation of various variants of Least Squares (LS) and Least Mean Squares (LMS) methods, we ultimately identify the Recursive Least Squares with Dichotomous Coordinate Descent (RLS-DCD) as the best-suited method. Anomalies are detected based on prediction errors exceeding a threshold determined during the initial ignition phase. Our anomaly detection performance evaluation utilizes real-world data from a plasma chamber and considers factors such as prediction (adaptation) update rate, anomaly detection accuracy, and resource utilization. Additionally, we compare f loating-point and fixed-point implementations, accompanied by an assessment of VHDL implementation. Despite utilizing fixed-point representation in the VHDL implementation, anomalies remain detectable, and performance remains comparable to floating-point RLS implementations. Our simulation and implementation results from the VHDL design demonstrate that the method employed exhibits a high adaptation rate while utilizing low resources. Consequently, the results from our evaluation show us that the method chosen is efficiently implementable on an FPGA hardware platform.