Azure Anomaly Detector仅检测峰值，无法识别异常下降问题咨询

Let’s work through how to fix your issue where Azure Anomaly Detector only picks up positive peaks and misses the critical abnormal drops you need to detect—especially with your data’s weekly seasonal pattern (expected weekend dips) and mid-week unexpected declines. Here are actionable, targeted steps:

1. Explicitly Define Seasonality and Detection Direction #

Azure Anomaly Detector might not be fully accounting for your weekly (168-hour) seasonality unless you force it to, and the default behavior can have a subtle bias toward upward anomalies. Adjust these key parameters in your request:

• Set period to 168 (7 days × 24 hours) to lock the model into learning weekly patterns. This helps it clearly distinguish expected weekend drops from unexpected mid-week declines.
• Specify direction in the anomalyDetectorConfiguration as "both" or explicitly "down" to prioritize downward anomaly detection. Example snippet:

  "anomalyDetectorConfiguration": {
    "sensitivity": 50,
    "direction": "both",
    "mode": "batch"
  },
  "period": 168,
  "granularity": "hourly"
  

2. Refine Sensitivity with Targeted Iteration #

You tested sensitivity values 20-90, but try a more granular, goal-focused approach:

• Start low (e.g., 30) to cut down on peak noise, then incrementally increase until you start capturing mid-week drops without flooding alerts with false positives.
• Remember: Lower sensitivity = fewer anomalies (less noise), higher sensitivity = more anomalies (more risk of noise). Pair this with the direction parameter to keep the model focused on your priority (drops).

3. Validate Data Quality and Alignment #

Poorly formatted or inconsistent data can throw off the model’s baseline calculations:

• Confirm all timestamps use ISO 8601 format (e.g., 2024-01-01T00:00:00Z) and are sorted chronologically.
• Check for missing hourly intervals or duplicate timestamps—gaps can skew baselines, making abnormal drops harder to spot. Use the alignTimeSeries parameter to fill or align missing points if needed.
• Ensure weekend drops are consistent enough for the model to learn as "normal"—wildly varying weekend values can blur the line between expected and unexpected declines.

4. Supplement with Post-Detection Custom Logic #

If the API still isn’t catching all critical drops, add your own validation layer:

• For each mid-week data point, compare it to the historical average ± standard deviation of the same hour on the same weekday across your 90-day window.
• Mark a point as an abnormal drop if it falls below, say, 2 standard deviations from that weekday-hour baseline. This adds specificity tailored exactly to your seasonal pattern.

5. Test with a Focused Data Subset #

Isolate a 4-week slice of data that includes both expected weekend drops and known mid-week abnormal drops. Run this through the detector with your adjusted parameters—this lets you iterate quickly without waiting for full 90-day batch processing.

By locking in your seasonality, tuning detection direction, refining sensitivity, and adding targeted post-processing, you should be able to get the model to reliably flag those critical mid-week drops while cutting down on peak noise.

内容的提问来源于stack exchange，提问作者Alex Michel