Explainable Anomaly Detection in Cloud-Native Systems Through Multimodal Telemetry Fusion
Cloud-native services emit operational metrics, event logs, and request traces that describe complementary aspects of failures. This paper presents a lightweight and intrinsically explainable anomaly detector that aggregates these modalities into request windows, converts their deviations from normal behavior into robust scores, and combines them using inverse-dispersion weights. Each alert is accompanied by exact per-modality score contributions, while a separate service-level contrast ranks potential fault locations. The evaluation uses an executed three-process HTTP microservice testbed with five controlled fault mechanisms and two possible fault targets. Forty fault-injection episodes generated 1,440 HTTP requests and 240 non-overlapping telemetry windows. In four repetition-held-out folds, the weighted fusion achieved a pooled AUROC of 0.978 and F1 of 0.896; a trace-only model reached F1 of 0.934. Weighted fusion had the highest AUROC, demonstrating a useful ranking advantage without uniformly dominating thresholded classification. Paired episode bootstrap intervals, modality-pair ablations, and operating-threshold sensitivity quantify these trade-offs. The study also examines fault-family sensitivity and transparent evidence attribution.
Authors
- Linnea Sjoberg
- Erik Holmstrom
Institutions
- Lund University (SE)
Publication Details
- Journal
- International Journal of Advanced Engineering and Technology Research
- Published
- 2026-11-20
- DOI
- https://doi.org/10.54097/thtdsz51
- Primary Topic
- Software System Performance and Reliability
- Type
- article
- Field-Weighted Citation Impact
- 0.00