Start with measurements that have meaning
A light sensor produces a number, but that number is only useful when its location, calibration, sampling interval and units are known. A reading on top of a display case may not represent the vertical face of a print inside it. Daylight can change both the intensity and distribution of illumination. The control design should state which measurement represents which surface and under what conditions that relationship remains valid.
Presence sensing adds a different kind of information. Passive infrared, radar and other occupancy technologies detect activity in different ways and with different blind spots. A visitor standing still may not trigger a movement-based device consistently. A signal from an adjacent circulation route may keep a scene active unnecessarily. Test real viewing behaviour, including seated visitors and people moving slowly, before adopting occupancy logic.
Equipment health is an input too. A stale reading, flat battery or communications failure can look deceptively like a stable environment unless the system checks timestamps and validity. Store quality flags alongside readings. Missing data should remain missing rather than being displayed as zero illumination.
Distinguish rules from prediction
Many useful adaptive functions do not require machine learning. A schedule can switch between approved scenes; a threshold can reduce electric light when daylight increases; a timer can dim a zone after a verified absence. These rules are comparatively easy to inspect and may offer a better fit for a small gallery than a model that requires ongoing retraining.
Machine learning may help forecast occupancy patterns, estimate daylight contributions or identify unusual sensor behaviour. Its role should be defined narrowly enough that teams can test whether it adds value. A model trained on one season, exhibition layout or visitor pattern may behave differently after those conditions change. Training performance alone is not evidence of reliable operation in the next exhibition.
Keep prediction separate from authority. A forecast can suggest when an approved scene is likely to be needed, but a supervisory layer should still enforce conservation constraints. Confidence scores are not permissions to exceed limits. Where uncertainty is high or the model is outside its tested conditions, the system should return to a documented fallback.
Choose what the system may adjust
Common control outputs include dimming level, scene selection, switch-off timing and gradual transitions. Some installations also control blinds or adjustable luminaires, but each added actuator creates new failure modes and maintenance requirements. Moving a beam changes geometry; altering a spectrum can affect appearance and material response. Those capabilities need separate approval rather than being treated as ordinary brightness changes.
A dimming percentage is not a measurement in lux. Drivers, optics, ageing and installation geometry affect the delivered result. Commission the relationship between commands and actual illumination at relevant object planes. If several luminaires contribute to one object, evaluate their combined effect instead of validating each in isolation.
Protocols such as DALI-2 and building-management interfaces can carry commands and status information. Interoperability does not establish conservation suitability. Ask what is available locally, what requires a gateway and whether logs and configuration remain accessible if a supplier relationship ends.
Stabilize changes for visitors and objects
A controller that reacts instantly to every fluctuation can produce distracting oscillation. Hysteresis, minimum hold times and ramped transitions can help prevent repeated switching around a threshold. Their settings must be evaluated against both accumulated exposure and viewing comfort. A long hold time can improve continuity but also keep light on after a space is empty.
Arrival behaviour deserves particular attention. Visitors should not need to wave at a sensor to reveal an object or find themselves navigating an unexpectedly dark route. Lighting transitions should be coordinated with circulation, labels and the needs of people with low vision. Emergency lighting and life-safety systems require their own appropriate design and must not depend on an experimental optimization model.
Accessibility should be tested with people and operational staff, not inferred from a simulated presence trace. Record reports of missed activation, abrupt changes and uncomfortable contrast alongside the numerical lighting data.
Verify the result and preserve the trail
The feedback loop closes only when the delivered condition is checked. A command acknowledgement confirms communication, not necessarily light output. Verification may combine independent spot measurements, representative fixed sensors and device diagnostics. The design should explain how discrepancies are detected and which person receives an actionable alert.
Log the input values, configuration version, decision, command, measured outcome and any override. Synchronize clocks and document sampling intervals so that exposure calculations can be reproduced. Keep raw readings when practical; aggregated totals alone can hide brief peaks or missing periods. If a total contains estimated intervals, show that explicitly.
Before live control, run the logic in observation-only mode. Compare its proposed actions with actual gallery conditions, then test a small bounded zone with a clear rollback plan. A successful system is not one that changes settings frequently. It is one that reduces unnecessary exposure while maintaining approved viewing conditions and leaving staff able to understand what happened.
A practical handover
At handover, staff should receive an input map, approved operating scenes, calibration records, fallback instructions and an exportable log format. A trained person should demonstrate how to disable predictive functions without losing the safe baseline. Review the human oversight requirements before operation, and use the evaluation framework to test claims under representative conditions.