Sensor Fusion for Health

Combining many imperfect measurements and the twin's prediction into one trustworthy health estimate, robust to any single sensor failing.

One estimate from many sources

No single diagnostic is fully trustworthy: each has noise, bias, blind spots, and its own failure modes. Sensor fusion combines redundant and complementary measurements with the digital twin's model prediction into a single best estimate of plant health, weighted by how much each source can be trusted right now. It is what lets control degrade gracefully when a sensor fails rather than trip.

Weighted fusion

The standard tool is a Bayesian filter that weights each measurement by its inverse variance and blends it with the model forecast. A sensor flagged as suspect by the diagnostics FMEA gets its weight driven toward zero, so the estimate leans on the remaining sensors and the twin - a virtual measurement.

def fuse(measurements, variances, prior_mean, prior_var):
    # Bayesian precision-weighted fusion (scalar form)
    prec = 1.0 / prior_var
    num = prior_mean / prior_var
    for z, v in zip(measurements, variances):
        if v is None:      # sensor declared suspect -> ignore
            continue
        prec += 1.0 / v
        num += z / v
    return num / prec, 1.0 / prec    # fused mean, fused variance

Trust is dynamic

Each sensor's variance is not fixed; it grows when the sensor disagrees with its peers and the twin, shrinks when it agrees. This makes the fusion self-healing: a drifting sensor is progressively down-weighted before it corrupts the estimate. The fused health estimate feeds FDI, RUL, and control.

For the burner plug, where the 166-830x regime means the twin's own prediction is low-confidence, fusion widens the fused variance honestly rather than reporting false certainty - the resulting wide interval correctly drives conservative operation.