One sensor is a narrow instrument
Every sensing element is good at one thing and blind to everything else, and each carries its own characteristic failure. The gas channels drift with age and respond to whole families of gases rather than to one. Vibration sensors trip on a passing forklift. Sound sensors hear the radio. Flame sensors respond to sunlight through a window.
The conventional response is to tune each detector's threshold until its false alarms become tolerable - which necessarily means desensitising it, which means moving its detection point later. Every single-channel detector on a factory wall is sitting at some compromise on that curve, and the compromise is invisible to whoever installed it.
What correlation adds
Two channels agreeing is different evidence from either channel alone. A temperature rise is weak evidence; equipment gets warm. A vibration change is weak evidence; machines are shaken by their neighbours. A sound change is weak evidence; a factory is loud.
The three together, on the same machine, in the same minute, are not weak evidence at all. That conjunction is what a mechanical failure actually looks like from the outside, and it is available well before any of the three individually reaches a level worth alarming on.
The same logic applies in the other direction, to suppress noise. A vibration spike with no accompanying thermal or acoustic change is far more likely to be a forklift than a failure - and a system that can see both channels can decline to wake anyone up, where a standalone vibration sensor cannot.
Why it has to be one device
Correlation only works if the channels are genuinely comparable, and that is a physical constraint more than a software one.
Channels must share a timebase. Two independent detectors with independent clocks and independent reporting intervals cannot establish that a thermal change and an acoustic change were simultaneous, and simultaneity is the whole signal.
They must also share a location. A temperature sensor on one wall and a sound sensor on another are describing different volumes of air, and correlating them is a category error dressed up as data fusion.
Putting every element on one board, sampling together, removes both problems by construction. It is the unglamorous reason multi-sensor detection tends to arrive as a single node rather than as a networked collection of existing detectors.
The eleven channels on an Alfa HazardX node
The v1 reference build carries four gas channels alongside flame, temperature, vibration and sound sensing. Each is individually ordinary; the combination is the point.
- Combustible gas - smoke and LPG
- Solvent vapour - alcohol and process solvents
- Carbon monoxide - and combustible gas
- Air quality - general contaminants and chemical vapour
- Flame - read both as a digital trigger and as an analog intensity
- Vibration - latched in firmware so a brief knock is not lost between samples
- Sound - both an over-threshold flag and a continuous level
- Temperature - ambient, and on the v2 board a second independent reference
- Humidity and pressure - context that makes the other channels interpretable
From eleven numbers to one state
Eleven channels presented as eleven dials is a worse interface than one channel, not a better one. A supervisor mid-shift cannot integrate eleven readings under time pressure, and asking them to is how a system with more information produces slower decisions.
So the readings are fused into a single risk state - SAFE, MODERATE or CRITICAL - with the per-channel detail one click away for whoever wants it. The floor sees a state. The engineer sees the channels. Neither is asked to do the other's job.
Fusion by movement, not by level
The most useful fusion in Alfa HazardX does not combine levels at all. It combines rates: how fast each channel is moving, whether that movement is accelerating, how long it has been going on. Twelve such features form a behaviour signature that can be matched against known hazard patterns.
Working in movement rather than level is what makes a signature portable between sites with different ambient baselines - the same reason a doctor asks how quickly a symptom developed rather than only how bad it is now.
