Deploy: integration

The sensor produces a value. The question is rarely asked at scoping, and always at commissioning: who receives it, on what clock, and who has the right to act on it.

Four conditions make a measurement act instead of being looked at. A shared data model, where the batch, the equipment state and the process phase carry the same name for everyone. A common time base between the measurement and the automation. A declared ownership of the feedback: which system commands, which actuator, under what conditions. And a validation that covers the interactions, not each tool taken separately.

A compression force reads in fractions of a second. A composition in seconds. A coating thickness per unit produced. A batch decision on the scale of a campaign. Making those coexist inside one decision logic is architecture work, not wiring.

What is specified once is reused. A common communication layer, data structures aligned on the ISA-95 standard — Enterprise-Control System Integration, published jointly by ISA and IEC under the dual designation ANSI/ISA-95 and IEC 62264, its Part 1 Models and Terminology reissued in 2025, and a written interface contract between measurement and equipment: the next line, the next site and a different automation supplier start from what exists. The question is then no longer which sensor to install, but which architecture holds when you add the second, then the third.

And it is as much a governance matter as a technical one. IT, automation, quality assurance, production and partners share neither deadlines nor vocabulary. The moment to settle the architecture is while they are all still around the table.

The reasoning does not depend on what you make. Oral solid dose, bioreactor, filling, continuous processing, a food line or water treatment: wherever several partners and several automation layers have to produce one coherent decision, the same four conditions decide.

How the roles are shared between the measurement, the SCADA and the PLC — who displays, who advises, who commands — is set out on the data and systems page.

The sensor sees what the mechanics present to it, and nothing else.

• Position and access. A free process connection, an existing sight glass or an available probe port changes the scale of the work entirely. Drilling already qualified equipment brings a requalification file that is often heavier than the installation itself, which is why the question belongs at scoping.
• Mechanical support. Vibration, cleaning shocks, expansion, forces as the product passes. A probe that moves a few millimetres changes the measurement geometry, so the signal, so the predictions.
• Working distance, where it is critical. Some non-contact measurements hold a distance to within tens of micrometres. Keeping it becomes an engineering objective rather than a setting.
• Cleaning and fouling. A deposit on the window, a crust in front of the probe, a condensate. The signal drifts slowly, which is exactly why a cleanliness indicator is designed in from the start. A two-millimetre insertion offset is enough to establish a lasting deposit.
• Temperature at the measurement point. It shifts the absorption bands and changes the properties of the medium. A model built at one temperature and run at another is corrected either by stabilising the point or by carrying temperature in the model.
• Process compatibility. Wetted materials, clean-in-place, ingress protection, explosive atmosphere. These are selection criteria rather than options, and they come first.

Getting the value up to supervision is the straightforward part. Standard industrial protocols, OPC UA foremost, do it cleanly, timestamp and data quality included. The difficult question is elsewhere.

Then there is security. A communication server that stays active outside the measurement software and accepts writes from third-party clients deserves attention in a regulated environment, because it sits awkwardly with the notion of a closed system. That does not disqualify the instrument: it is handled by network segmentation, rights management and logging. It does have to be seen. What a declaration of conformity covers.

This section covers the connection. What becomes of the data afterwards is a separate set of decisions, taken at scoping: volume produced, retention of the raw data, archive format, the choice between LIMS and MES, and the indicators to display.

The calibration burden is chosen first: decomposition onto pure components, or regression on reference values. It follows from the matrix and the quantity sought. What each route asks for.

Then come the pre-processings, where the useful rule is parsimony. Correcting an identified physical effect beats stacking treatments that improve an indicator for reasons that stay unstated. Then the latent variables, a trade between performance and robustness: a model needing many components is sensitive to small changes in sample.

The decisive work is elsewhere. A multivariate model always predicts. Show it a spectrum of air, another formula, a raw material from a new supplier, and it returns a value that is plausible. What makes a method defensible is the apparatus that lets it say the sample is unfamiliar.

Their thresholds are set on the calibration set, and crossing one produces a defined behaviour: alarm, invalidation of the value, or fallback. A written range of validity is what makes a multivariate method defensible, in front of an auditor and in front of an incident.

As soon as more than one instrument is in play, two lines, a spare, a replacement in five years, the question arrives. Two instruments differ in spectral response, resolution and geometry. The differences are invisible to the eye and large enough to shift predictions.

Standardisation methods exist and work well, on one condition: a set of common samples measured on every instrument concerned. Acquiring that set during the trials costs a fraction of assembling it afterwards, once the first instrument is in production. Why two instruments give two figures.

Installation qualification checks that the delivered system is the one specified and correctly mounted. Operational qualification checks that its functions behave as stated. Wavelength accuracy, photometric linearity, drift over several hours, rights management, event log. Performance qualification checks that the whole, in its real environment, produces what is expected.

And this is where to look closely. Many available protocols qualify the instrument at rest, while the promise bears on the measurement in motion. A standard presented to a stationary probe says little about the same instrument facing a flowing product or a rotating blender. That gap is exactly what an auditor finds.

Closing it is part of the work: repeated control spectra, repeatability of the sample presentation, periodic comparison with the reference method, acceptance criteria as figures. That protocol is specific to your installation: it is written with you.

An auditor does not ask whether the method is good. They ask to follow the reasoning: why this technology, why this measurement point, on what data the model was built, how you know it remains valid.

• The user requirement specification and the justification of the technology choice.
• The proof of concept report, reservations included.
• The model description: calibration data, pre-processing, latent variables, validity domain and out-of-domain detection thresholds.
• The qualification protocols and reports, verification under real conditions included.
• The method validation dossier, when the intended use requires it.
• The monitoring plan and the recalibration procedure, with their triggers.
• Change control: formula that evolves, probe replaced, model updated.