From process event to stored value

Several processes separate a local change in the liquid from a number in a trend file. Material must reach the sensitive surface, the transducer must respond, electronics process the signal, and a control system reads and stores it. An external sample path also adds transport. Each stage can have its own timescale. Total delay cannot therefore be equated with a single datasheet number.

First define the quantity: local concentration at the probe, concentration in an extracted sample stream or a process quantity calculated from it. An off-gas analyser measures composition at the analyser. Without accounting for its line, this is not the composition leaving the reactor at the same clock time. Likewise a local pH reading may differ from an assumed vessel mean without electrode failure.

A useful measurement-chain diagram names the input, output and timestamp at every boundary. Distinguish the beginning of a physical change from a control command. A subsequently stored value may have been acquired earlier. These distinctions matter especially when comparing sensors, manipulated variables and laboratory samples. The time-reference section addresses data provenance; this chapter addresses dynamic distortion.

Time constant, dead time and t90

For a normalised input step u from 0 to 1, τdy/dt + y = u describes a linear first-order system. Without dead time its response is y = 1 − exp(−t/τ). After one time constant it reaches approximately 63.2% of the full step. For any fraction f, tf = −τ ln(1 − f); specifically t₉₀ = 2.303τ and t₉₅ = 2.996τ. These results are derived from the equation and apply only to that assumed response shape.

Dead time td shifts the whole response without itself rounding it. From the appropriate origin, the signal initially remains unchanged and then follows the lag. Measured from a valve timestamp, the time to 90% can therefore be td + 2.303τ. “Response time 20 s” is incomplete without the threshold, origin and test conditions.

Several lag stages do not generally produce a single exponential response. Two equally slow stages have a distinctly more rounded onset. A t₉₀ value alone cannot identify that model; it describes only one point of the response. For rapid concentration events, the shape, dead time and possible direction dependence are at least as important as reaching a specified fraction of the endpoint.

The MIT material on two tanks in series provides a formal treatment of coupled linear systems. The mathematical structure is useful for measurement chains, but physical parameters must come from the actual chain being investigated.

Designing a response experiment that actually tests the probe

The input at the sensor must be faster and better known than the response being determined. Slowly changing the gas supply to an entire vessel measures mass transfer and probe behaviour simultaneously. Equating observed t90 with probe t90 is then unjustified. Suitable comparison arrangements either shorten the concentration change at the sensor or measure it using an independent sufficiently fast reference.

Match test conditions to the task: medium, temperature, pressure, flow and mounting can affect the effective response. Include surface condition and documented calibration in the record. A representative experiment must not change the sensor through mechanical or chemical intervention absent from normal use. Transferability is something to test, not a consequence of using the same model name.

Apply steps in both directions if both matter for the application. Different rise and fall times can indicate nonlinear dynamics, adsorption, transport or different initial states. Repeat after a waiting period to distinguish reversible changes from persistent changes. A slow falling response alone proves neither hysteresis nor fouling.

Record the unchanged initial state, onset, curved transition and a sufficiently stable endpoint. End-normalising a trace discards offset and gain information, so always retain raw signals and physical units. Report the internal filters actually configured rather than merely the defaults quoted in a handbook.

Teaching case: two seconds dead time and an eight-second time constant

The synthetic sensor-step.csv file describes an ideal input step at t = 0, 2 s dead time and then a response with τ = 8 s. Values are recorded every 2 s. The calculated output reaches 90% at t = 20.4207 s. Relative to the onset of the actual probe response, the interval is 18.4207 s. Treating these as contradictory sensor specifications overlooks their different time origins.

The 63.2% point occurs 10 s after the input event. Reading this directly from t = 0 as the time constant overestimates τ by 25%. This simple calculation explains the importance of event logs even for apparently elementary response tests. Better analysis specifies dead time separately or fits it jointly under explicit assumptions.

The file contains no noise, allowing definitions to be checked first. A further exercise may add documented noise to the script, then compare interpolated threshold crossings with a fitted model. The result should explain differences between sampling grid, noise and model error, rather than deriving a general sensor-accuracy claim from a noise-free teaching trace.

Synthetic response-test results
QuantityValueTime reference
Dead time2 sBetween input step and response onset
Time constant8 sFrom response onset
Lag t9018.4207 sFrom response onset
Time to 90%20.4207 sFrom input step
Recording interval2 sDiscrete sampling grid

Synthetic step response · Calculation · Data dictionary

How much of a rapid oscillation survives

For a settled sinusoidal input with angular frequency ω, a first-order stage has amplitude ratio 1/√[1 + (ωτ)²]. Here ω = 2π/T for period T. Dead time adds a phase shift of −ωtd, but an ideal pure delay does not change amplitude. The calculation therefore distinguishes attenuation from temporal displacement.

With τ = 8 s and T = 30 s, the teaching case retains only about 0.5125 of the original amplitude. A local oscillation of ten percentage points therefore appears as approximately 5.12 percentage points, provided the linear model applies. A visually calm trend can conceal a more variable local environment. This concerns measurement dynamics, not automatically cell damage.

Recording more values does not reverse physical attenuation. Querying an unchanged sensor every millisecond does not make it faster. Frequent recording may describe its smoothed output more accurately, but does not create missing information about brief input peaks. Conversely, infrequent recording can discard useful features of an already sufficiently fast response.

This reasoning supports instrument selection. First identify the shortest event relevant to the decision; then assess whether sensor, transport path, filtering and recording can represent it. A control loop also needs suitable actuator dynamics. A fast probe alone does not make a slow gas-delivery system respond quickly.

Internal filters and retrospective smoothing

An exponential digital smoother can be written yⱼ = αuⱼ + (1 − α)yⱼ₋₁. For a discrete first-order stage, α = 1 − exp(−Δt/τf). A small α gives stronger smoothing and a slower response. Behaviour therefore depends on both α and the sampling interval Δt. The same numerical filter factor at another update rate represents a different filter.

A moving average over past observations spreads brief events across a longer window. A centred moving average also needs future values; it is suitable for retrospective presentation but cannot be an undelayed real-time signal. Causal analysis must distinguish what a controller actually received from what was processed later.

Smoothing is useful when a noise component interferes with a decision. Its choice nevertheless depends on the question. A filter that stabilises a slowly varying endpoint may be unsuitable for measuring brief excursions. Retain raw data and identify every transformation. Compare unfiltered and filtered traces on the same time axis before describing a smoother plot as a process improvement.

The kLa analysis with probe dynamics illustrates another consequence: an unknown filter may be falsely attributed to mass transfer. If a model represents an effective overall response, label it accordingly. Separating that response afterwards into physical causes requires additional information.

Why numerical deconvolution has limits

Rearranging τdy/dt + y = x gives the formal reconstruction x = y + τdy/dt. For exactly known, noise-free signals this is valid. In measured data, differentiation amplifies rapid fluctuations. For two independent measurement errors of equal standard deviation σ, a simple finite-difference derivative has error standard deviation approximately √2 σ/Δt. A smaller interval can increase derivative noise.

This does not make every reconstruction useless. It does require a method appropriate to the problem and explicit assumptions about admissible inputs. Regularisation limits amplification of unstable components, but also changes the solution. Its strength should follow a defensible objective and independent assessment, rather than the most attractive-looking trace.

If a brief peak was strongly smoothed or clipped before storage, no algorithm can recover its exact shape from nothing. Different input histories can produce the same recorded values within noise. Bounds on duration or height may then be more defensible than a precise invented peak. A faster suitable measurement chain is often the better next step.

Time shifts also require appropriate conditions. Variable sample flow changes transport time, making a constant shift invalid across the entire record. Before aligning signals by correlation, check whether their shared dynamics already connect physically different events.

Assessing dynamics, calibration and operation together

A sensor can be correct at equilibrium yet unsuitable during transitions. A fast probe can conversely have an incorrect offset. Test these properties separately. Endpoint calibration does not correct an unknown transport path, and a favourable time constant does not establish chemical selectivity. Selection therefore starts from a measurement task with an allowable uncertainty and required time resolution.

As a concrete reading example, the Vernier manual for an educational optical probe describes response time, temperature, installation and display units together. The manufacturer explicitly excludes industrial process-control use. This link illustrates how to read a specification; none of its values are transferred to sterilisable process probes.

An operational comparison may evaluate a reference before and after a campaign. If zero, endpoint and response shape change together, do not assign every deviation to one cause. Treat surface condition, calibration, medium, temperature and software settings as separate hypotheses. Intervention at an aseptic process boundary must follow the procedure provided for that equipment.

A useful final report answers four questions: which events can the chain represent under the tested conditions? Which are attenuated or delayed? Which numerical corrections were applied? What information is needed for stronger conclusions? The result is defensible measurement capability for a defined task, rather than a general manufacturer rating.

What sampling does to a rapid signal

Consider the purely numerical signal sin(2π·0.8t), with t in seconds. Sampling once per second at integer times produces the same values as −sin(2π·0.2t). A true 0.8 Hz oscillation thus appears as a slower 0.2 Hz oscillation with different phase. The samples alone cannot distinguish the possibilities.

Tenfold faster storage can resolve this particular ambiguity if the measurement chain actually transmits the rapid content. If the sensor has already strongly attenuated it, faster recording cannot recover its original amplitude. Physical response, analogue or digital filtering, and sampling are therefore described separately.

The theoretical sampling theorem concerns an ideally band-limited signal and ideal reconstruction. Real processes require appropriate anti-aliasing filters and margin. A fixed ratio to an expected process time does not suit every task: event detection, accurate phase measurement and smooth trend display can require different choices.

Alarm-event resolution does not automatically equal stored-value resolution either. A controller may evaluate more frequently and log events separately, or a historian may hide short peaks through averaging. Reconcile event logs, raw-value storage and displayed trends during an investigation.

Missing data need a separate status. A carried-forward last valid value looks like a stable process even when communication has failed. Exports should distinguish interpolation, held values and genuine new observations. Only then can constant-condition durations be interpreted meaningfully.

Distinguishing timestamps from transport time

A timestamp may identify measurement start, measurement completion, arrival at the control system or storage in the historian. These events need not coincide. Associating a laboratory result with its release time can place it many minutes after the state represented by the sample.

Dynamic comparison therefore benefits from an event chain: command, effective boundary change, local measurement, transmission and storage. Document known offsets. Unknown variability in an offset is not a fixed dead time that can reliably be removed by shifting an entire trace.

A shared brief marker can compare two chains in time if both observe it adequately. First establish that they observe the same physical event. A temperature change in the liquid and an off-gas change after a condenser may share a command but not their transport delays.

Cross-correlation can identify a numerically plausible shift. It does not prove causal ordering if the signals share control action or have different filters. Repeated periodic patterns can create several similar peaks. Understanding the measurement chain remains part of interpretation.

Transport delay in sampling tubing can change with flow. A fixed offset from an earlier test then ceases to apply. Changes in volume, temperature or gas holdup may also warrant renewed response assessment. A measured delay belongs to a stated configuration and conditions, not permanently to an instrument name.

A transparent filter calculation

A causal moving average of five values sampled one second apart is yt = (xt + xt−1 + xt−2 + xt−3 + xt−4)/5. For uncorrelated equal-variance noise, standard deviation decreases by a factor 1/√5. If input samples already have temporal dependence, that simple reduction no longer applies unchanged.

For a slowly changing ramp, the average corresponds approximately to the value two seconds earlier, because the five samples have mean age two seconds. An abrupt step instead produces five successive output steps. Its response shape differs from an exponential first-order low-pass filter.

An offline symmetric filter using two earlier and two future values can reduce visible shift. It needs future information and therefore is not an immediately available online value. Such a smoothed laboratory plot cannot demonstrate that a real-time controller achieved the same apparent response speed.

For a particular task, compare raw and filtered data around a deliberately generated event. Assess noise, amplitude loss, timing and preservation of the integral. A filter improving trend display may still be unsuitable for estimating a rapid derivative. The intended benefit determines assessment.

The NIST explanation of autocorrelation supports evaluation of temporal dependence. Noise reduction and mean age in the five-point example are independently calculated consequences of the explicitly stated filter. They are not measured performance claims for a data-acquisition system.

A brief excursion can fail to cross a signal threshold

A step test keeps the input at its new level for a long time. Many process events are brief pulses instead. Their response can also be calculated exactly for the first-order linear model used here. Let the normalised input rise from 0 to 1 for a duration T and then return to 0. Starting from y = 0, the signal rises as y = 1 − exp(−t/τ) while the pulse is present. Its maximum at the end of the pulse is therefore ymax = 1 − exp(−T/τ); it subsequently decays exponentially towards zero.

With τ = 8 s, a two-second pulse produces a maximum of only 0.2212. A twelve-second pulse produces 0.7769. A threshold of 0.5 is never reached in the first case, although the true input reached the full pulse height. Without dead time, the upward threshold crossing occurs at −8 ln(0.5) = 5.545 s. The pulse must persist for at least that long to reach it. A negative concentration excursion follows the same relationships after suitable normalisation, with the direction of the signal change reversed.

Adding pure dead time shifts the whole response. Within this model, it changes neither the maximum nor the duration above a specified threshold. An additional filter can change both. This separates two different observations: an event may retain its shape but appear later, or it may be attenuated enough that the stored signal never crosses a chosen threshold. Those observations place different requirements on characterising the measurement chain.

In the twelve-second teaching pulse, the response remains positive after the input has already returned to its baseline. An apparently broad signal excursion therefore does not prove an equally long physical exposure. Interpretation requires the assumed input duration, response model and analysis threshold to be considered together. This calculation explains measurement-chain behaviour; it does not set a process alarm threshold or a tolerable biological exposure. Those choices require the actual process question and a separate assessment supported by appropriate evidence.

Normalisation must preserve evidence about gain and direction

Normalised curves make response shapes easier to compare, but they can conceal another important difference. Suppose two measurement chains experience the same physical step and have the same time constant, while the second reports only 90% of the actual change. Scaling each curve independently from its own initial and final values onto 0 to 1 makes the curves coincide completely. Matching normalised dynamics therefore does not demonstrate matching steady-state gain.

A report should retain values in physical units alongside the normalised plot. It records the starting level, applied input step, measured change and conversion factor. Dynamic characterisation does not replace static calibration, and successful calibration does not guarantee adequate dynamics. If the reported variable is calculated using temperature, pressure or other signals, their update rates and time references also matter. Otherwise, the dynamics of a compensation calculation may be mistaken for the dynamics of the sensing element itself.

A second original comparison uses upward and downward steps between the same two conditions. For a linear model with constant τ, suitably normalised responses should have the same shape. A reproducible difference indicates that this model is insufficient for describing both directions together. The cause must then be investigated separately: the input may have been established at different speeds, transport towards the surface may vary, or an internal algorithm may behave differently by direction. The discrepancy between the curves alone does not identify which explanation is correct.

Repeating directional comparisons in a different order helps interpret them. A test that always occurs second may be affected by drift, deposits or a changing sample. With only one upward and one downward transition, direction is confounded with sequence in time. The required number of repetitions depends on variability and the intended decision; there is no defensible universal number for this purpose. What matters is whether the structure of the experiment can actually test a plausible alternative explanation.

For questions about fast process events, these considerations produce a concrete record: raw measurements with units, normalised response, independently observed input, direction, repeat runs, selected model and remaining discrepancies. Together they make a dynamic specification reviewable. They also allow a later assessment of whether changing the tubing, filter or probe installation remains covered by the original characterisation. The connection with endpoint selection in kLa analysis is direct: in both cases, a convenient normalisation can remove a physical difference from the display without resolving that difference in the underlying measurement.

Supplementary teaching calculations: Python calculation and results as CSV. The data description records assumptions, units and execution.

Terms in this chapter

29 terms explained in the glossary