Robust Time-Series Anomaly Detection for AGI System Monitoring: A Hybrid Neural-Statistical Approach

Keywords: Anomaly Detection, Time-Series Analysis, AGI Safety, Hybrid Model, Neural-Statistical Methods, LSTM Autoencoder, CUSUM, System Monitoring, Reproducible AI, Synthetic Data Generation, Operational Metrics, Statistical Process Control

Abstract:

Autonomous AGI systems require robust anomaly detection in continuous teleme-
try streams to ensure safe operation and early intervention. Current approaches face
critical limitations: classical methods miss subtle contextual anomalies while deep
models overfit and lack operational reliability. We present a novel hybrid pipeline
combining compact neural encoders (LSTM autoencoder with 64 hidden units)
with calibrated statistical decision rules (CUSUM) to optimize early detection
while maintaining low false alarm rates. Our approach uses synthetic telemetry
generation mimicking agent failure modes for reproducible evaluation. Experimen-
tal results demonstrate a 20.4% improvement in F1-score (0.849 vs 0.705) and
26.6% reduction in mean detection delay (23.4 vs 31.9 timesteps) compared to
the best baseline while maintaining false alarm rates below 0.01/hour. The hybrid
method achieves superior performance with statistical significance (p < 0.001,
Cohen’s d = 2.87) while providing computational efficiency suitable for real-time
AGI monitoring. This work advances AGI safety by prioritizing operational metrics
and delivering a reproducible framework for agent telemetry analysis.

Submission Number: 264