Real-Time Anomaly Detection Across Subsystems

Watching many sensor streams at once to flag deviations from normal behavior fast enough to act on them.

The task

A fusion plant has thousands of sensors: temperatures, pressures, flows, voltages, magnetic signals. Anomaly detection means learning what normal looks like across those streams and raising a flag when the live data drift from it, ideally early enough to intervene before a small deviation becomes a fault.

Approaches

• Threshold and rule based: simple, transparent, but blind to novel patterns
• Statistical: track means, variances, and correlations, flag outliers
• Model based: predict the next value from physics or a learned model, flag large residuals
• Multivariate: detect when the joint pattern is off even if each signal looks fine alone

Why multivariate matters

Many faults show up first as a broken relationship between signals, not as any single signal crossing a limit. A pump losing efficiency might keep flow nominal while its power draw and downstream temperature drift together. A method that models the joint distribution catches this earlier than per-channel alarms.

import numpy as np

def mahalanobis(x, mean, inv_cov):
    d = x - mean
    return float(np.sqrt(d @ inv_cov @ d))   # distance in correlated space

# flag when distance exceeds a calibrated bound

The two errors

Every detector trades false alarms against missed detections. Too sensitive and operators learn to ignore it; too lax and it misses the event it was built for. Thresholds are calibrated against historical and simulated fault data, and detections are ranked by severity so attention goes where it matters.

Kronos context

In the design phase these detectors run against simulated subsystem data and digital-twin outputs to prove the monitoring concept and tune sensitivity ahead of operation. They complement, not replace, hard interlocks that act automatically regardless of what the detector thinks.