Intelligent Learning Model for Anomaly Detection in Microservice Architecture

Microservice anomaly detection requires multi-modal telemetry as faults manifest heterogeneously across logs and metrics. However, severe score scale disparities complicate multi-modal fusion. We propose a lightweight dual-encoder architecture coupling a Structured Log LSTM Autoencoder with a Temporal Metric Variational Autoencoder (VAE) enhanced by 20 Planar Normalizing Flow layers. We harmonize telemetry scores through validation-calibrated normalization, Fisher Signal-to-Noise Ratio (SNR) weighting, and learned neural gating. Across five random seeds on the unrebalanced Generic AIOps Atlas (GAIA) benchmark (2844 test windows with natural 38.08% window-level and 14.90% minute-level fault prevalence), our architecture delivers robust operational fidelity. Canonical metric extraction enables the metric encoder to achieve an AUC-ROC of 0.7364, matching OmniAnomaly (0.7365). Log telemetry dominates semantic microservice failures, where multi-modal fusion matches the log-only raw F1 of 0.9251 while calibrated normalization secures an AUC-ROC of 0.9971, outperforming unnormalized score summation (0.7842) by 21%. Benchmarking against random guessing exposes substantial point-adjusted inflation (an F1 of 0.8724 versus a raw F1 of 0.0698). Ultimately, operating within 18.1 ms CPU latency and 234 MB memory, our framework satisfies production Kubernetes sidecar constraints.