Hybrid PCA-Based and Machine Learning Approaches for Signal-Based Interference Detection and Anomaly Classification...

Area of Science:

• Signal Processing
• Machine Learning
• Data Science

Background:

• Anomaly detection is critical for safety-critical systems like sensor networks.
• Strict low-false-alarm constraints (e.g., 1% False Positive Rate) are essential.
• Existing methods may struggle with dynamic environments and varying signal-to-noise ratios (SNR).

Purpose of the Study:

• To develop a lightweight, interpretable anomaly detection pipeline for multichannel spatiotemporal data.
• To rigorously enforce bounded false alarms under low-false-alarm constraints.
• To enhance robustness against signal interference and SNR shifts in sensor networks.

Main Methods:

• A fusion ensemble of three anomaly signals: PCA Reconstruction Error, Local Outlier Factor on residual maps, and Monte Carlo Variance.
• Implementation using NumPy and scikit-learn, avoiding deep learning dependencies.
• Combination of signals via logistic regression (F*) and Neyman-Pearson optimized fusion (F**, F***) for bounded false alarms.

Main Results:

• The fusion approach demonstrated exceptional robustness on synthetic benchmarks with realistic anomalies and SNR shifts (±12 dB).
• Achieved a True Positive Rate (TPR) of ≈0.74 at 1% FPR, significantly outperforming single baselines (≈0.60 TPR).
• Maintained high performance under degraded SNR (≈0.62 TPR at -12 dB), avoiding baseline performance collapse.

Conclusions:

• The proposed interpretable, edge-ready pipeline offers a deployable solution for reliable anomaly detection in dynamic environments.
• The fusion strategy effectively mitigates performance degradation caused by SNR variations.
• This approach provides a transparent and highly effective capability for critical monitoring applications.