Process Analyzer Reliability
Why in-situ measurement, optical spectroscopy, chemometrics and AI are changing the maintenance equation
For decades, discussions about process analyzer reliability have concentrated on maintenance: how often to calibrate, when to change filters, how many technicians are required and whether maintenance should be preventive, condition-based or reactive.
All of these questions remain relevant. But there is a more fundamental one: How much maintenance should the analyzer system require in the first place?
The most effective way to improve analyzer reliability is often not to create a better maintenance programme. It is to design out as many maintenance requirements and potential failure points as reasonably possible.
This matters increasingly as plants operate with larger installed analyzer populations and fewer specialist personnel. ABB, for example, reports that about 63% of analyzer technician time can be consumed by scheduled maintenance and investigation of faults that ultimately prove not to exist. It also notes that the ratio of process analyzers to assigned technicians continues to increase.
The direction is therefore clear: simpler measurement architecture, fewer mechanical components, more direct measurement and more intelligence in the data layer.
Reliability begins before the analyzer
An analyzer may contain sophisticated electronics and optics yet depend on a surprisingly complicated chain between the process and the actual measurement.
A conventional extractive installation can include a process probe, isolation valves, long sample lines, fast loops, pumps, filters, coalescers, regulators, coolers, heaters, vaporizers, flow controllers, stream-selection valves, calibration manifolds and sample recovery or disposal systems. Every component has a purpose. It is also another potential source of blockage, leakage, contamination, adsorption, condensation, pressure instability or maintenance.
International Society of Automation (ISA) specifically identifies the sample conditioning system as an important maintenance element because filters, demisters, flow regulators, heaters and similar devices must themselves be cleaned, checked or calibrated.
Sampling-system design also affects measurement quality. Transport delay, dead volume, adsorption, permeation and unintended phase changes can make the analyzer report accurately on a sample that no longer represents what is happening in the process. Swagelok identifies these as fundamental issues in process analyzer sampling-system engineering.
This leads to an important engineering principle: Every component that can safely be eliminated is one component that cannot fail, leak, plug or require maintenance.
That does not mean removing necessary protection. Difficult applications will always require filtration, pressure reduction or temperature control. It means starting with the simplest possible measurement architecture rather than automatically reproducing a conventional sample system because that is how it was done twenty years ago.
Move the measurement towards the process
Where the analytical technology allows it, direct or in-situ measurement changes the reliability equation substantially. Instead of bringing the process to the analyzer, the analyzer measures at or very close to the process.
The benefits go well beyond saving a few meters of tubing. Long sample transport lines disappear. Sample pumps may disappear. Pressure reduction can sometimes be avoided. Heating or cooling requirements can be reduced. There is less opportunity for condensation, adsorption or contamination and the measurement normally responds much more quickly to a genuine process change.
This is already well established with optical technologies. Siemens describes its in-situ laser systems as avoiding the sampling and conditioning delays associated with conventional extractive gas analysis, while also reducing maintenance requirements. Yokogawa similarly notes that direct in-situ laser measurement can eliminate sample extraction and conditioning equipment.
Of course, in-situ measurement is not automatically the right answer everywhere. Fouling, optical path length, process temperature, pressure, solids, corrosiveness and accessibility still have to be considered.
The sensible hierarchy is therefore in-situ where practical, close-coupled where possible and fully extractive only where necessary.
Correlative optical analyzers go one step further
The maintenance advantage becomes particularly interesting with correlative optical technologies such as near-infrared spectroscopy.
Traditional discrete analyzers generally reproduce, directly or indirectly, one specific physical or chemical measurement. If a refinery wants continuous information on several product properties, several separate analyzers may therefore be required.
A correlative spectroscopic analyzer works differently. The instrument measures an optical spectrum containing information about the molecular composition of the process stream. Chemometric models then establish relationships between this spectral information and reference laboratory values. One spectrum can therefore provide predictions for several properties.
For petroleum applications this can include, depending on the stream and calibration, parameters such as octane number, density, distillation points, aromatics, olefins, benzene, cetane-related properties and other composition or quality variables.
This changes the economics as well as the maintenance requirement. Instead of maintaining several mechanical or discrete analyzers, it can be possible to obtain several process properties from a single spectrometer and optical measurement.
Modcon’s Beacon-3000 Process NIR Analyzer follows this architecture. The main NIR analyzer can be located in the control room, connected through fibre optics to field measurement points. The field units contain no electrical power or moving parts, while a single instrument can serve multiple measurement locations.
The difference is not merely replacing one analytical principle with another. It is moving complexity away from mechanical hardware in the field and towards optics, software and mathematical modelling. That is an important change because software does not need its filter changed on a Friday afternoon.
The traditional weakness: maintaining the model
Correlative analysis does, however, introduce another type of maintenance. The analyzer may remain mechanically stable while the relationship between its spectrum and the required laboratory property changes.
Feedstock composition can move outside the original calibration space. Process temperature may change. A new crude or blend component may be introduced. Instrument characteristics can drift. An optical component may be replaced. The result is model drift rather than mechanical failure.
Recent research describes calibration-model maintenance as one of the central practical challenges for long-term deployment of spectral sensors. Changes in sample composition, instruments or measurement environment create what machine-learning specialists call dataset shift.
Historically, correcting this required a skilled chemometrician to collect suitable samples, identify outliers, rebuild models and validate the revised calibration.
That is precisely the sort of specialist work that becomes difficult when analyzer teams are smaller and one engineer is supporting dozens or hundreds of measurement points. Modern chemometrics and machine learning are changing this as well.
From manual chemometrics to automated model management
A modern spectroscopic system does not need to treat calibration as a model built once during commissioning and then forgotten. Laboratory reference data, process spectra and analyzer diagnostics can continuously provide information about model performance.
Machine-learning tools can help select the most informative samples, detect abnormal or unrepresentative observations, identify developing bias, evaluate candidate models and determine when recalibration should be considered. Modern calibration-transfer techniques can also reduce the work required when instruments or measurement conditions change.
Modcon has been applying this principle through model-maintenance software that links process NIR measurements with plant and laboratory information. More recently, Modcon.AI development has included automated selection of suitable samples for spectroscopic model development while identifying and excluding outliers.
The objective is not simply to build a more complicated mathematical model. It is to reduce the amount of specialist manual work needed to keep that model useful throughout the analyzer’s life. There is an important qualification. Automatic model updating should not mean uncontrolled self-learning.
For measurements affecting product certification, safety or closed-loop control, proposed model changes should remain subject to defined statistical acceptance criteria, independent reference data, version control and appropriate engineering approval. Automation should remove repetitive work, not engineering responsibility.
AI can monitor the analyzer as well as the process
Once the analyzer becomes digitally connected, considerably more information becomes available than the reported property alone.
Spectral residuals, optical intensity, analyzer diagnostics, temperature, pressure, validation results, laboratory differences and associated process variables can all contribute to understanding measurement health.
AI and statistical monitoring can then identify gradual deterioration before the measurement becomes unusable. For example, a steadily increasing laboratory bias may indicate model drift. Falling optical intensity could suggest window contamination. An abnormal combination of pressure, flow and analyzer response may indicate a sample-system problem rather than an analyzer fault.
This is where condition-based maintenance becomes much more practical. Instead of opening an analyzer every three months because a calendar says so, maintenance can increasingly be triggered by evidence that something is actually changing.
Remote connectivity extends the same concept. Modcon Remote Service, MARS, allows diagnostics, software support and many corrective activities to be carried out without first sending a specialist to site. Modcon reports that remote monitoring can identify a significant proportion of deviations before they disrupt normal operation.
For plants spread across several locations, offshore installations and facilities where experienced analyzer specialists are scarce, that capability is becoming increasingly important.
Traditional and emerging analyzer architecture
Designing for low maintenance
A practical reliability-first project should follow seven priorities:
- Start with the process requirement, not the analyzer catalogue. Define what needs to be measured, required accuracy, acceptable response time and how the result will actually be used.
- Consider direct or in-situ measurement first. Every metre between the process and analyzer should have a technical reason for being there.
- Simplify unavoidable sampling systems. Use the shortest practical lines, minimize dead volume and avoid unnecessary filters, regulators and switching components.
- Consider multi-property optical measurement. Where the chemistry supports it, one NIR or other spectroscopic analyzer may replace several individual measurements.
- Design the model lifecycle together with the hardware. Laboratory validation, calibration transfer, model drift detection and updating are part of the analyzer system, not an afterthought.
- Collect diagnostics as well as analytical results. A process value without information about measurement health gives only half the picture.
- Move towards condition-based and remote maintenance. Service should increasingly be performed because the equipment indicates that intervention is required, rather than simply because another month has passed.
Reliability is increasingly an architectural decision
No industrial analyzer is literally maintenance-free. Optical windows can foul. Light sources age. Electronics eventually fail. Calibration models need supervision and some process applications genuinely require sophisticated sample conditioning.
The aim should therefore not be the marketing promise of zero maintenance. The practical target is a maintenance-light analyzer architecture in which unnecessary hardware has been removed, the remaining equipment is observable and diagnosable and much of the intellectual maintenance work is handled by software.
For Modcon Systems Ltd., this means combining several developments: direct and in-situ measurement where practical, robust photonic instrumentation, NIR and other correlative technologies, simplified sample handling, automatic and semi-automatic chemometric model management, remote diagnostics and AI-enabled analysis of both process and analyzer data.
The result is a shift in emphasis. Instead of asking only: How can we maintain this analyzer more efficiently?
the better question is: How can we design the measurement so that there is much less to maintain?
That is likely to become one of the defining principles of the next generation of process analysis.
