How to verify accuracy claims for industrial quality analyzers

An accuracy claim is useful only when it describes performance under conditions close to your process. When comparing industrial quality analyzers, ask suppliers to connect every claimed tolerance to a defined measurement, operating range, sample type, reference method, and test procedure. A small error measured on a clean standard may say little about performance on a variable production stream.

A defensible purchase requires three things: a clearly defined specification, credible reference measurements, and an independent application test with agreed acceptance criteria. Calibration records support that assessment, but they do not establish that an analyzer will make reliable quality decisions in your installation.

Translate the accuracy claim into a testable specification

Start by defining exactly what the analyzer must measure. Moisture reported on a wet basis differs from moisture on a dry basis. Dissolved metal concentration differs from total metal concentration. Results expressed at actual process conditions may differ from results normalized to a reference temperature or pressure. Instruments can disagree because they report different quantities, even when both operate correctly.

Next, examine how the supplier expresses the claimed error. A percentage of reading changes with the measured value; a percentage of full scale refers to the selected range. These can produce substantially different allowable errors near the bottom of the range. Specifications combining a relative term with a fixed concentration allowance require both terms to be considered.

Ask whether the figure represents a maximum permissible error, typical performance, average bias, or a statistical prediction metric. Those statements support different conclusions. “Typical” performance does not establish a guaranteed limit for the unit being purchased.

An analyzer may give tightly grouped readings while remaining consistently biased. Conversely, a small average bias can hide large positive and negative errors. Ask for both systematic error and result variability, together with their test conditions.

Check what the calibration evidence actually covers

Request records for the proposed instrument configuration, including the measurement range, sensor or optical arrangement, sample conditioning equipment, software version, and calibration model where relevant. A certificate for a component or another model does not demonstrate the performance of the complete supplied system.

Useful calibration records identify the instrument, reference standards, procedure, environmental conditions, results, and associated measurement uncertainty. Where adjustment occurred, request results before and after adjustment. The earlier results reveal how the instrument behaved when received; the later results show its performance following intervention.

Calibration establishes a relationship between indications and reference values under stated conditions. Verification determines whether specified requirements are met. An instrument can have a current calibration certificate and still fail your application's acceptance limits.

Metrological traceability requires a documented, unbroken chain of calibrations to an appropriate reference, with each link contributing uncertainty. A supplier's statement that equipment is “traceable” should therefore be supported by identifiable records.

ISO/IEC 17025 accreditation can strengthen confidence in laboratory competence, but examine the scope. It should cover the relevant calibration or test activity, range, and capability. Accreditation held by a laboratory does not automatically cover every service it offers, and it does not certify an analyzer's accuracy across all applications.

Choose references that represent the sample

Certified reference materials provide assigned property values with stated uncertainties and traceability information. Their value depends on suitability: concentration range, matrix, physical form, stability, and preparation requirements must match the intended test closely enough to support the comparison.

A clean aqueous standard can check an analyzer's response while leaving production interferences untested. Suspended solids, salts, color, bubbles, viscosity, or particle size may affect the result. For gas analysis, balance gas composition and moisture can matter. Spectroscopic measurements of solids may depend on surface condition and sample presentation.

Use suitable reference materials to check response, then include representative production samples to challenge application performance. Where an appropriate certified material is unavailable, comparison with a competent laboratory using a suitable reference method can provide evidence. Document the method's uncertainty and sample handling requirements.

The reference method must measure the same property. For example, loss on drying may include volatile substances other than water, so it may disagree with a water-specific method for a legitimate chemical reason. Resolve that difference before judging either instrument.

Reference uncertainty also limits what the test can establish. If it is comparable to the analyzer's allowable error, an apparently close result may provide weak evidence of compliance. Select the reference capability and conformity decision rule together.

Design an application test the supplier cannot pass through calibration alone

Agree the protocol before testing begins. Define sample selection, operating conditions, permitted adjustments, data exclusions, statistical treatment, and pass criteria. This prevents a favorable demonstration from becoming the acceptance standard after results are available.

For analyzers using empirical calibration models, separate calibration samples from validation samples. Testing only materials used to build the model rewards fitting performance and can conceal poor predictions on unfamiliar production batches.

Keep validation reference values hidden from the supplier until analyzer results are recorded. If the supplier adjusts the model after reviewing those values, evaluate the revised model with a fresh independent set. Otherwise, the supposed validation data have become part of calibration development.

A practical test should include:

• The working range: Cover routine concentrations, range boundaries, and values close to quality decision limits.
• Relevant sample variation: Include feedstocks, grades, contaminants, and physical conditions expected in production.
• Repeated measurements: Assess short-term repeatability, then repeat selected checks across days, operators, or sample preparations as appropriate.
• Operating challenges: Test relevant temperature, flow, pressure, humidity, and fouling conditions within the proposed operating envelope.
• Time-dependent behavior: Examine drift, carryover, stabilization after startup, and recovery after concentration changes.

Choose the amount of testing according to variability and decision risk. Repeated readings of one stable sample reveal little about matrix effects. A broad collection measured once may reveal application differences but provide weak evidence about repeatability. Both questions need attention.

Record failed measurements, alarms, manual interventions, and rejected results. An analyzer that reports accurate values only after frequent operator intervention may be unsuitable for unattended control, even if the retained results look excellent.

For online analyzers, verify the complete sampling path

The installed measurement includes more than the analyzer cabinet. Probe location, transport tubing, filters, pressure reduction, dilution, and sample temperature can change the material reaching the sensor.

A filter may remove particles containing the constituent of interest. A volatile component may escape during transport. Deposits can retain material and release it later, producing carryover. These effects can dominate uncertainty even when the sensing element performs well.

Compare results from samples representing the same process condition. A laboratory grab sample collected near a probe does not automatically correspond to an analyzer reading displayed at that moment. Transport delay and internal averaging can shift the analyzer's response. Establish the time alignment before interpreting disagreement.

Test response time separately from steady-state error. A slow analyzer may accurately describe an earlier process state while missing a short excursion that matters for product release or process control. Decide whether the reported result must represent an instantaneous value, a batch average, or a moving average.

Read the validation results beyond a single headline number

Request paired analyzer and reference results, sample identifiers, operating conditions, timestamps, and a record of exclusions. For each pair, calculate the difference between the analyzer result and the reference result. Examine those differences across concentration, sample type, and time.

A correlation coefficient alone is insufficient. Two methods can track concentration changes closely while maintaining an unacceptable offset or proportional error. A wide concentration range can also make correlation look impressive while hiding weak performance near a narrow release threshold.

Plot measurement differences against reference concentration. A rising error pattern can indicate proportional bias; widening scatter may suggest concentration-dependent variability. Separate clusters can expose matrix effects that disappear when all results are combined.

For prediction-based instruments, metrics such as root mean square prediction error help summarize performance on independent samples. They remain averages across a dataset. Inspect individual errors and relevant subgroups, especially when occasional large errors could release off-specification material.

Treat outlier removal as a technical decision requiring a documented cause. A transcription error differs from an unexplained analyzer failure. Removing difficult samples because they reduce apparent accuracy makes the reported performance less representative of production.

Set acceptance limits around the quality decision

Define the allowable measurement error from the process requirement before comparing suppliers. The useful question is whether measurement performance supports reliable decisions at the relevant specification limits. A better datasheet figure has little value if it applies to an irrelevant range or sample type.

Account for uncertainty in the comparison and in routine use. Reference uncertainty, sampling, preparation, repeatability, and drift can all contribute. Avoid adding overlapping terms twice; an uncertainty assessment should identify what each contribution includes.

Where results determine product acceptance, establish a decision rule explaining how uncertainty affects pass or fail judgments. A guard band can move an acceptance boundary inward to reduce false acceptance risk, while potentially increasing false rejection. The appropriate rule depends on the consequences of each outcome.

Write separate requirements where needed for bias, variability, response time, drift, and measurement availability. A single average-error limit may allow operational weaknesses to pass unnoticed. State how borderline results will be handled and what evidence permits a retest.

Make demonstrated performance part of the purchase agreement

Compare suppliers against the same application requirements and evidence standard. A slightly less impressive specification supported by representative independent testing may provide a stronger procurement basis than a tighter claim demonstrated only on ideal standards.

Include the agreed range, sample matrices, reference procedure, acceptance criteria, calibration conditions, and software or model identification in the technical agreement. Specify which checks occur during factory acceptance and which must be repeated after installation, where sampling and environmental effects become visible.

Also establish calibration checks, maintenance responsibilities, and triggers for renewed validation. New feedstocks, revised models, sensor replacement, or significant process changes can alter performance. The purchase decision should leave a clear record of what was demonstrated, under which conditions, and how that performance will be maintained.