A new detection method catches industrial faults that look completely normal on every individual sensor.
Most fault detectors watch each sensor's readings and flag anything that drifts outside its usual range. That approach misses coupling faults, where the physical relationship between sensor groups breaks down even though each sensor still reports normal numbers on its own. Researchers built CMR-Mamba, a system that trains per-domain encoders on healthy data to learn how sensor groups should influence each other, then scores anomalies by how far new data strays from that learned relationship. They tested it on bearing-wear data, hydraulic system data, and a water-treatment testbed, comparing it against standard reconstruction-based detectors.
The more interesting result is not the headline accuracy number. Most methods already solve the easy faults, so the real test is the hard subset, including stealthy attacks on the water-treatment system that keep every sensor inside its normal range. There, CMR-Mamba pulled ahead while sensor-by-sensor methods performed at chance, which matters most for safety-critical systems where a slow-developing failure or an attacker can hide in plain sight.
Worth noting: the paper's own ablations credit the gain mostly to how anomalies are scored, not to any particular neural architecture, a reminder that the fix here is more statistical than exotic.