Go Back Research Article December, 2022

CAN AI OUTSMART THRESHOLD ALERTS? A HYBRID MACHINE LEARNING APPROACH FOR SMARTER ANOMALY DETECTION IN AZURE DATA PIPELINES

Abstract

Cloud data pipelines need better monitoring—traditional threshold-based systems often miss subtle anomalies or flood teams with false alarms. But can AI do better? This study compares hybrid machine learning (ML) models against standard threshold monitoring in Azure Data Factory (ADF), testing whether combining Azure’s Anomaly Detector with custom LSTM neural networks improves detection speed, accuracy, and efficiency. Using real-world enterprise data, we measure: Detection accuracy (precision, recall) for sudden spikes, slow drifts, and tricky contextual anomalies. Alert speed—how much faster AI spots issues vs. threshold rules in streaming/batch workloads. Compute trade-offs—does the extra ML overhead pay off in high-volume pipelines? Results show the ensemble model cuts false alerts by 32% and detects anomalies 47% earlier, adding under 500ms latency—viable for real-time use. But there’s a catch: AI’s "black box" decisions complicate compliance. We break down the pros/cons for Azure teams, balancing cost, complexity, and explainability when shifting from rules to AI.

Keywords

anomaly detection machine learning hybrid ai models azure data factory predictive alerts threshold rules lstm networks cloud monitoring real-time analytics false alarms azure anomaly detector operational efficiency.
Document Preview
Download PDF
Details
Volume 13
Issue 1
Pages 170-184
ISSN 0976-6413