Early Leak Detection¶

Goal¶

Detect a leak as early as possible within the transient—i.e. with minimal delay after the leak starts—while keeping false positives under control.

How the Platform Supports It¶

Window size and step: Smaller windows and smaller steps give finer time resolution; the model can fire on the first window where the leak signature appears. Trade-off: more windows, more compute and possible noise.
Features: Oscillation and coherence features (see Leak detection features) tend to respond quickly when a leak starts; we use them so the model can learn “leak present” from the earliest reliable signal.
Threshold: Detection pipelines often expose a decision threshold (e.g. on probability or score). Tuning this (e.g. via “max F1” or “min recall” constraint) lets you balance early detection (higher recall) vs false alarms (precision).
Temporal context (see Temporal context): Adding “previous window” and “delta” features helps the model notice onset of a change (e.g. sudden increase in variability) and thus react earlier in the transient.

Practical Tuning¶

In training config: Adjust detection_threshold, threshold_strategy, and optional min_recall.
In offline evaluation: Use the test-offline pipeline to plot detection rate vs time-into-transient or vs threshold, so you can choose a threshold that gives “early enough” detection with acceptable false-positive rate.

This document states the early-detection intent; the exact parameters live in the detection pipeline config and in the offline test reports.