Anomaly Detection in ELI-NP Front-End Laser Energy Data Using an Optimized Moving Average Method

Anomaly detection in time-series data is critical for ensuring stability in high-power laser systems, where deviations can indicate potential failures. This study optimizes a moving average-based methodology for anomaly detection accuracy by evaluating window sizes (3, 6, 9, 12, and 15) and threshold multipliers (1.0, 1.5, and 2.0). The analysis integrates Mean Squared Error (MSE), correlation analysis, and graphical evaluations, including anomaly distribution, moving average trends, and parameter sensitivity plots. Results indicate that smaller window sizes effectively detect short-term fluctuations but are more susceptible to noise, while larger windows smooth trends but may overlook minor anomalies. Threshold multipliers significantly impact detection, with lower values capturing more anomalies, potentially increasing false positives, and higher values reducing sensitivity but minimizing false alarms. MSE trends suggest a trade-off between sensitivity and robustness, where smaller windows better fit the data but risk overfitting noise, while larger windows reduce responsiveness but enhance stability. Correlation analysis scatterplots reveal a strong dependency between window size and MSE, while anomaly counts exhibit a nonlinear relationship with threshold multipliers. Anomaly detection plots and MSE vs. window size comparisons highlight detection efficiency. The study bridges statistical anomaly detection techniques with real-world laser monitoring, ensuring computational efficiency, robustness, and enhanced fault detection. These findings lay the groundwork for adaptive parameter tuning and machine learning integration in real-time anomaly detection for high-power laser systems.