Why PID Loops Oscillate: A Practical Guide to Finding the Real Cause and Fixing It

Diagnose the loop from synchronized PV, controller-output, and valve-travel evidence before changing PID tuning.

Quick answer: why PID loops oscillate

A proportional-integral-derivative (PID) loop oscillates when the complete feedback path has enough gain and phase lag to sustain repeated movement, or when a nonlinearity, interacting loop, sampling effect or external disturbance forces a cycle. The PID may be creating the oscillation, amplifying it or correctly responding to a real periodic disturbance. Classify the source before retuning.

PID should not be confused with P&ID, which means piping and instrumentation diagram. A P&ID shows equipment, piping and instrumentation relationships. A PID controller uses proportional, integral and sometimes derivative action to regulate a process variable.

Use this seven-step workflow:

1. Recognize: Confirm that the movement is recurring rather than random noise or a normal transient.
2. Measure: Calculate period, frequency, amplitude and whether successive peaks decay, remain steady or grow.
3. Classify: Decide whether the evidence points toward tuning, process dynamics, saturation, valve behavior, measurement, interaction or external forcing.
4. Test: Perform only authorized, bounded checks that can distinguish competing explanations.
5. Correct: Address the demonstrated cause rather than applying a generic detuning change.
6. Verify: Repeat the same measurements under comparable operating conditions.
7. Document: Record evidence, approvals, original settings, final settings, limits, modes and restoration status.

Table T01: Mini cause-to-signature guide

Possible cause	Typical trend signature	Best next check
Excessive loop gain	Smooth, repeating PV and controller output; amplitude changes substantially when gain is reduced	Verify controller form, action and scaling before a cautious gain change
Excessive integral action	Controller output reverses only after accumulated error changes; damping improves when Ti is increased	Confirm whether the setting is integral time or repeats per minute
Valve stiction	Controller output ramps while actual travel is stationary, followed by a travel jump	Trend actual valve travel, positioner signals and supply pressure
Saturation and windup	Applied output is pinned while nominal output continues beyond the limit; recovery is delayed	Compare nominal and applied output and inspect anti-windup tracking
Derivative acting on noise	High-frequency output chatter while PV is much smoother at the process timescale	Inspect raw PV, quantization, derivative location and filtering
Dead time and multiple lags	Output changes, but PV responds late enough to produce substantial phase lag	Estimate delay and lag from an approved test or existing disturbance
External periodic disturbance	Cycling persists in manual while actual actuator travel remains constant	Trace upstream, utility, pressure, feed, composition or analyzer forcing
Loop interaction	Several loops share a period, with one signal consistently changing first	Align timestamps and inspect shared process paths and utilities

The key rule is simple: do not retune until the oscillation source has been classified.

Contents

• Quick answer
• Safety before diagnosis or testing
• What PID loop oscillation looks like
• PID control in brief
• Signals to trend
• Diagnostic decision tree
• Common causes of PID loop oscillation
• Ultimate gain and ultimate period
• Worked synthetic examples
• Corrective-action matrix
• Safe testing and management of change
• Verification and documentation
• Frequently asked questions
• Planned downloads and next steps
• References and disclosure

Safety before diagnosis or testing

WARNING — KEEP UNINTERRUPTED AND AD-FREE

Do not provoke an oscillation merely to make diagnosis easier. Online gain changes, bump tests, relay tests and ultimate-cycle tests can move pressure, temperature, level, flow, composition, inventory or equipment beyond safe limits. Dead time, stored energy and continuing reactions can keep the PV moving after an operator selects manual.

Before any active test, follow site authorization and operating procedures. Define the purpose, maximum excitation, duration, stop criteria, safe limits, recovery plan and who may start, pause, abort and restore the loop. Consider alarms, interlocks, trips, relief systems, environmental constraints, downstream units and shared utilities. Confirm actuator headroom in both directions.

Screen proposed tests and tuning changes under the site’s management-of-change or temporary-change rules. Do not bypass, suppress, retune or proof-test a safety instrumented system as part of ordinary PID troubleshooting. Safety-critical applications require qualified review.

Passive diagnosis from existing synchronized trends should normally come before active testing.

What PID loop oscillation looks like

Terminology: oscillation, cycling, hunting and instability

An oscillation is repeated movement around a value or trajectory. Operators may also call it cycling or hunting. These words describe what the trend looks like; they do not identify the cause.

Instability is more specific. In a linear analysis, instability means deviations grow rather than settle. A loop can nevertheless show a sustained cycle without being linearly unstable. Saturation, stiction, deadband, backlash, relay behavior or quantization can create a bounded nonlinear limit cycle.

Calling every repeating trend “bad tuning” skips the most important diagnostic step.

Oscillation versus a normal control response

After a setpoint or load change, a functioning loop may overshoot and cross its target one or more times. That is not automatically a sustained oscillation.

Three basic response patterns help establish urgency:

• Decaying response: Each same-side peak becomes smaller. The loop is returning toward equilibrium.
• Sustained response: Peak amplitude remains approximately constant. Linear gain and phase conditions or a nonlinear mechanism may be sustaining the cycle.
• Growing response: Successive peaks become larger. The loop may be unstable and can reach output limits, alarms or trips.
• Non-oscillatory response: The PV approaches its target without crossing it repeatedly. It may still be too slow or limited, but it is not cycling.

Classify whether oscillation decays, persists, or grows, and measure period between same-direction peaks.

A growing trend deserves prompt attention, but a bounded trend should not be dismissed. A sustained limit cycle can repeatedly move a valve, consume utilities, disturb downstream units and hide smaller process changes.

Types of oscillation and the measurements that describe them

A useful description should quantify the cycle instead of relying on “fast,” “large” or “almost stable.”

DEFINITION BOX — OSCILLATION METRICS

• Period, T: Time between comparable points in successive cycles, such as two same-direction peaks.
• Frequency, f: Number of cycles per unit time. f = 1/T.
• Angular frequency, ω: ω = 2π/T.
• Amplitude, A: For a symmetric waveform, half its peak-to-peak range. State explicitly whether a reported value is amplitude or full peak-to-peak range.
• Decay ratio, DR: The second same-side peak deviation divided by the first: DR = A2/A1.
• Quarter-amplitude decay: DR = 0.25; the next same-side peak is one-quarter of the preceding deviation. It does not mean “25% overshoot.”
• Limit cycle: A sustained periodic trajectory created or stabilized by a nonlinearity such as stiction, deadband, backlash, saturation, a relay or quantization.

Measure several cycles where possible. A single peak interval can be distorted by noise, a setpoint change or historian sampling. Record the baseline used for amplitude and whether peaks were measured from raw or filtered data.

A self-regulating process reaches a new steady value after a fixed input change. An integrating process continues to ramp while the net input imbalance remains. Their trends can look different under the same controller behavior, so process type matters when interpreting waveform shape.

Forced versus self-sustained oscillation

A self-sustained oscillation originates within the closed feedback path. Excessive gain, integral action combined with delay, valve stiction or a quantized controller-actuator path can keep the loop moving without an external periodic input.

A forced oscillation enters from elsewhere. Examples include an upstream cycle, utility pressure variation, periodic feed composition, an analyzer cycle or another controller. In this case, the PID may be doing exactly what it should: responding to a real disturbance.

A forced cycle can persist with constant actuator travel, whereas a self-sustained loop cycle stops when the feedback mechanism is opened.

Manual mode can help distinguish the categories, but it is not a complete verdict:

• If cycling persists in manual and actual actuator travel is constant, suspect external forcing.
• If cycling persists in manual and travel continues moving, inspect the positioner, air supply, actuator and valve mechanics.
• If cycling stops in manual, the feedback loop was necessary for that cycle, but this does not distinguish poor tuning from stiction or another feedback-dependent nonlinearity.

Waveform shape and manual behavior are clues. Neither should be used as proof without checking actual travel and relevant disturbances.

PID control in brief

A PID controller combines:

• Proportional action, based on present error.
• Integral action, based on accumulated error.
• Derivative action, based on the rate of change of error or another selected signal.

Common signal names are:

• SP: Setpoint.
• PV: Process variable.
• CO or OP: Controller output.
• Applied CO: Output after limits, selectors, rate limits and other downstream logic.
• Valve travel: Measured physical position. It is not the same signal as CO.

Controller forms, units and action

Different PID equations can use identical parameter names while producing different behavior. Never transfer tuning numbers without identifying the controller form.

Table T02: PID controller forms and conversions

Form	Equation or representation	Parameter meaning	Important conversion
Parallel	u = Kp e + Ki ∫e dt + Kd de/dt + b	Kp: output/error; Ki: output/(error·time); Kd: output·time/error	Ki is an integral gain, not an integral time
Ideal or ISA	u = Kc[e + (1/Ti)∫e dt + Td de/dt] + b	Kc: controller gain; Ti: integral time; Td: derivative time	Kp=Kc; Ki=Kc/Ti; Kd=Kc·Td
Series or interacting	C(s)=Kc(1+1/(Ti s))(1+Td s)	Integral and derivative factors interact	Expansion adds proportional contribution, so numbers are not directly interchangeable with ideal form

CHECK YOUR CONTROLLER FORM

Before interpreting any tuning number, record:

1. Parallel, ideal/ISA, series/interacting or vendor-specific form.
2. The error equation used by the implementation.
3. Direct or reverse action.
4. Whether integral is entered as Ti, Ki or repeats per unit time.
5. Whether times are seconds or minutes.
6. Whether derivative acts on error, PV or a weighted signal.
7. Input and output scaling.
8. Output, integral, rate and selector limits.

For ideal form, a smaller Ti produces stronger integral action. If the implementation uses repeats per minute, Ri = 1/Ti when Ti is in minutes, so a larger repeats-per-minute value produces stronger action. Confusing Ti with Ki or repeats per minute can create factor-of-60 errors when seconds and minutes are mixed.

Proportional band is another source of confusion. PB(%) = 100/|Kc| applies only when signals are consistently normalized to their spans. It should not be applied blindly to engineering-unit signals.

The expected direction of control must also be documented. One convention may define error as SP−PV; another may use PV−SP with a corresponding direct/reverse setting. Verify the expected physical result: if PV rises, should CO rise or fall? Wrong action changes negative feedback into positive feedback and can cause rapid divergence.

Derivative on error reacts to a setpoint step and can create derivative kick. Derivative on PV avoids that specific kick. Two-degree-of-freedom implementations can weight setpoint terms separately, but vendor definitions vary. Derivative also magnifies rapid measurement changes approximately in proportion to ΔPV/Δt. Filtering can reduce chatter while adding phase lag; excessive filtering can therefore worsen stability. There is no universal derivative-filter value.

Related topic: derivative kick and derivative on measurement.

Where oscillation can enter the loop

A controller is only one component in the feedback path. The complete loop includes the sensor, filtering, controller calculations, limits, network or scan delays, actuator, positioner, valve and process. Disturbances and interacting controllers can enter at several points.

Cycling can originate in the controller, output limits, final control element, process dynamics, measurement path, or an external disturbance.

The practical question is not merely “Is the PID oscillating?” It is: Where does the repeating movement first appear, and what mechanism sustains it?

Signals to trend before changing tuning

Trend signals on one synchronized time base

A useful diagnostic trend should include more than SP, PV and the displayed output. Collect the following on a synchronized time base:

• SP, including remote or cascade setpoint.
• Raw PV and the filtered PV used by the controller.
• Nominal CO before limiting.
• Applied CO after limits, selectors or overrides.
• Actual valve or actuator travel.
• Controller mode and cascade status.
• Output, rate and integral limits.
• Positioner output and supply pressure where available.
• Relevant feed, utility, composition, pressure and analyzer signals.
• SP, PV and CO from suspected interacting loops.
• Alarm, override or selector states relevant to the event.

Align SP, PV, nominal and applied controller output, valve travel, mode, and limit states on one time base.

If the controller display provides only one output, determine whether it is nominal or applied. A nominal request of 140% may appear as an applied 100% after limiting. Likewise, a 55% CO signal does not prove that a valve traveled to 55%.

Related topic: how to read process trends.

Sampling for diagnosis

Sampling must capture the fastest feature relevant to the suspected mechanism. Slow historian data may show the process cycle while completely missing valve jumps or output chatter.

As an illustrative planning target—not a universal requirement—collect about 10–20 samples across the shortest feature of interest and include multiple complete cycles. A one-second valve jump requires a different collection rate from a 20-minute vessel-temperature cycle.

Record:

• Sample interval and timestamp resolution.
• Controller scan and historian collection rates.
• Clock alignment between systems.
• Missing values and interpolation.
• Compression or exception-reporting settings.
• Filter settings.
• Quantization.
• Task jitter, communication delay or execution-order changes where relevant.

Do not differentiate a heavily quantized or exception-compressed PV and then interpret the resulting spikes as process physics.

Waveform and phase clues

There is no universal rule that PV and CO must be “in phase” or “out of phase” for a particular fault. Their relationship depends on process gain sign, controller action, process type, frequency, dead time, filtering and whether the plotted output represents actual actuator movement.

For a positive-gain first-order-plus-dead-time process,

Gp(s) = K e^(−θs)/(τs+1)

the phase from CO to PV at angular frequency ω is:

phase(PV/CO) = −ωθ − atan(ωτ)

Dead time and process lag both add phase delay. At the ultimate frequency of the stated model, PV is 180° behind CO, but that result cannot be converted into a universal trend-reading rule for every implementation.

Table T03: Waveform and phase clues

Observed pattern	Consistent with	Interpretation boundary	Next check
Smooth, roughly sinusoidal PV and CO	Linear gain/phase-margin problem	Not proof of poor tuning	Compare actual travel, constraints and common-frequency signals
Square-like PV with triangular or sawtooth CO	Stiction in some self-regulating loops	Recognized clue, not proof	Look for flat travel while CO ramps, then a jump
Ramp or triangular PV	Integrating process under a cycling input	A process-dependent clue rather than a universal pattern	Confirm process type and material/energy balance
Flat CO tops or bottoms	Saturation or clipping	Display scaling can conceal downstream limits	Compare nominal CO with applied CO
CO ramps while travel is flat, then travel jumps	Stick-slip	Stronger evidence than PV/CO shape	Check positioner, air, linkage and valve mechanics
Travel fails to respond after reversal	Backlash or reversal deadband	One-direction tests can miss it	Use an authorized bidirectional test
Fixed PV–CO phase offset	Gain sign, action, process dynamics, dead time or filtering	No universal phase rule	Check implementation and estimate dynamic delay
Same period across several loops	Interaction or common forcing	Earliest signal is a candidate source, not proof	Align clocks and trace shared paths
Cycling persists in manual with constant travel	External disturbance	Confirm manual mode and actual travel	Trace feed, utility, pressure, composition or analyzer forcing
Cycling persists in manual with moving travel	Positioner, air or mechanical behavior	May be autonomous actuator cycling	Inspect the final-control element

Cross-correlation, PV–CO plots, ellipse or parallelogram methods and higher-order statistics can screen routine data, but they can produce false positives and negatives. Confirm any automated diagnosis with travel feedback or a controlled test.

Diagnostic decision tree

Use the following sequence rather than starting with a tuning change:

1. Confirm recurrence. Is the pattern repeated, or is it a single disturbance and recovery?
2. Quantify it. Record period, frequency, amplitude and whether peaks decay, remain sustained or grow.
3. Align the evidence. Trend SP, PV, raw PV, nominal CO, applied CO, valve travel, mode, limits and likely disturbances.
4. Check constraints and data quality. Look for saturation, clipping, rate limits, selectors, quantization, filtering, compression and scan-related timing.
5. Verify implementation. Confirm PID form, error equation, action, gain scaling, integral units, time base and derivative location.
6. Compare command with movement. Does actual travel follow applied CO smoothly and promptly?
7. Classify the process. Determine whether it is primarily self-regulating or integrating in the operating region.
8. Compare other loops. Search for common periods, shared utilities and upstream signals.
9. Use an authorized discriminating test if needed. Manual mode or a bounded change may help separate feedback-dependent cycling from external forcing.
10. Correct the demonstrated cause.
11. Verify under representative conditions.
12. Document evidence, approvals, final configuration and restoration.

Follow a measured, authorized workflow from recognition and classification through testing, correction, verification, and documentation.

A planned downloadable version of this decision tree is proposed for field use: download link: placeholder.

Common causes of PID loop oscillation

Excessive proportional gain

Proportional action changes CO in direct proportion to the current error. Increasing controller gain can improve responsiveness, but it also magnifies loop movement. At a frequency where the complete feedback path has nearly 180° of phase lag, sufficient gain turns corrective action into reinforcement of the preceding movement.

If a smooth cycle decays after an authorized, cautious reduction in Kc, excessive loop gain becomes a stronger hypothesis.

A smooth, approximately sinusoidal PV and CO cycle is consistent with excessive loop gain, especially if amplitude or damping changes substantially after Kc is reduced. It is not proof. Valve motion, output constraints and periodic disturbances can produce similar PV trends.

DIAGNOSTIC MINI BOX — EXCESSIVE GAIN

• Signature: Smooth PV and CO cycling; damping changes strongly with Kc.
• Confirm: Verify form, action, scaling and units; check travel, limits and disturbances.
• Fix: Reduce Kc cautiously under an approved plan and reassess disturbance rejection and robustness.
• Cannot be fixed by tuning alone: Wrong action, valve faults, external forcing or inadequate actuator capacity.

A gain reduction that makes a trend calmer does not prove the original problem was tuning. Detuning can conceal a mechanical or interaction problem while reducing control performance.

Excessive integral action

Integral action accumulates error and drives steady-state offset toward zero. It also adds phase lag. If it is too aggressive relative to process delay and lag, CO can keep moving after the process has begun an unseen response.

In ideal form:

integral contribution = Kc(1/Ti) ∫e dt

A smaller Ti means stronger integral action. In parallel form, Ki is an integral gain, and larger Ki means stronger action. If the controller uses repeats per minute:

Ri = 1/Ti, with Ti in minutes.

WARNING — INTEGRAL UNIT CONFUSION

Do not treat Ti and Ki as interchangeable. Do not enter a time in a field expecting repeats per minute, or minutes in a field expecting seconds. A factor-of-60 mistake can make a loop dramatically more aggressive or sluggish. Record the controller form and engineering units before changing the value.

A loop dominated by excessive integral may show delayed output reversals because the integrator must unwind accumulated error. Increasing Ti can improve damping and change the period, but it also affects load-disturbance recovery. The change must therefore be evaluated rather than judged from one quiet trend.

DIAGNOSTIC MINI BOX — EXCESSIVE INTEGRAL

• Signature: CO continues in one direction while error accumulation is being reversed; damping improves when Ti is increased.
• Confirm: Identify form and units; inspect the integral contribution and constraints.
• Fix: Increase Ti cautiously in ideal form, or reduce Ki in parallel form.
• Cannot be fixed by tuning alone: Saturation, stiction, incorrect limits or insufficient actuator authority.

Related topic: PID tuning methods and when each fails.

Derivative action, noise and filtering

Derivative anticipates movement by responding to a rate of change. That can add damping in suitable applications, but differentiation magnifies rapid measurement changes, quantization and communication steps.

A noisy PV may produce high-frequency CO chatter even when the physical process changes slowly. Derivative on error can also react sharply to a setpoint step. Derivative on PV avoids that specific setpoint kick but is not universally required.

Filtering derivative or PV can reduce chatter. However, filtering adds delay and phase loss. Adding a stronger filter without evaluating stability can trade visible noise for a slower oscillation. Vendor definitions of derivative filter factor N differ; there is no universal value.

DIAGNOSTIC MINI BOX — DERIVATIVE AND NOISE

• Signature: Rapid CO chatter; raw PV is noisy or quantized while the main process response is smoother.
• Confirm: Inspect raw PV, derivative input, sample rate, quantization and filtering.
• Fix: Correct measurement or sampling problems first; then review derivative location, magnitude and filter.
• Cannot be fixed by tuning alone: Sensor faults, grounding issues, communication steps or inadequate resolution.

Do not assume derivative always damps an oscillation. Poor implementation, noisy measurement or excessive filtering can aggravate it.

Dead time and multiple process lags

Dead time is the interval between an input change and the first observable effect on the measured output. During that interval, the controller receives no evidence that its last action is working.

A first-order-plus-dead-time model is:

Gp(s) = K e^(−θs)/(τs+1)

where:

• K is process gain.
• θ is dead time.
• τ is the first-order time constant.

This model is a simplification, but it illustrates the mechanism. The delay term contributes phase lag proportional to frequency, while the first-order lag adds atan(ωτ). Multiple lags, filters, valve dynamics and computation delays can add further phase loss.

Dead time and process lag delay PV response; at the ultimate frequency the accumulated phase shift can sustain a cycle.

Dead time often comes from transport, analyzer cycles, thermal propagation, mixing, communication or batch sequencing. It cannot be removed by increasing gain. Aggressive gain or integral action generally becomes less tolerable as delay becomes more important relative to the process time scale.

Before estimating dead time from a trend, check timestamp alignment and filters. A historian timestamp or slow sensor filter can appear as process dead time.

GO DEEPER — PHASE CONDITION

For a positive-gain FOPDT process:

phase(PV/CO) = −ωθ − atan(ωτ)

At a frequency where the complete loop reaches approximately −180° with unity loop magnitude, a sustained linear cycle can occur. This is a model result, not a field test instruction.

DIAGNOSTIC MINI BOX — DEAD TIME AND LAGS

• Signature: CO changes well before the PV response; repeated corrections arrive too early.
• Confirm: Estimate delay from existing disturbances or an authorized test; separate process delay from filtering and timestamp error.
• Fix: Use less aggressive settings appropriate to the dynamics; reduce avoidable measurement, communication or actuator delay.
• Cannot be fixed by tuning alone: Physical transport delay, slow analyzers or equipment lags that require process or instrumentation changes.

Related topic: dead time in process control.

Operating-point nonlinearity

A loop may be stable at one flow, level, pressure, product grade or valve position and oscillatory at another. Process gain, time constant, dead time and valve gain can all change across the operating envelope.

A tuning set chosen at one point should not automatically be claimed as suitable everywhere. Identify where stability and performance change, then determine whether operating-region tuning, gain scheduling, valve characterization or process modification is justified.

DIAGNOSTIC MINI BOX — OPERATING POINT

• Signature: Stability changes repeatably with load, grade or valve position.
• Confirm: Compare dynamics and constraints by region.
• Fix: Review region-specific tuning or gain scheduling under site procedures.
• Cannot be fixed by one tuning set alone: Strong process or actuator nonlinearity across the full envelope.

Wrong action, scaling or units

A direct/reverse-action error can create positive feedback: rising PV produces a CO movement that drives PV further upward. Scaling errors can make moderate-looking tuning numbers extremely aggressive. Common mistakes include using normalized-span relationships on engineering-unit signals, confusing gain with proportional band and mixing seconds, minutes and repeats per minute.

DIAGNOSTIC MINI BOX — ACTION, SCALING AND UNITS

• Signature: Immediate divergence, unexpectedly aggressive behavior or a factor-of-60 timing error.
• Confirm: State the error equation, expected physical response, process-gain sign, spans and time units.
• Fix: Correct the configuration through authorized change control.
• Cannot be fixed by retuning alone: A wrong feedback sign or incorrectly scaled signal path.

Do not change action simply because CO and PV appear to move in the same direction. Expected direction depends on the process and valve arrangement.

Saturation and integral windup

Saturation occurs when the requested output exceeds what the applied output or actuator can deliver. Output limits, selector stations, rate limits, override control and physical actuator limits can all create a difference between nominal and applied CO.

Integral windup occurs when the integrator continues accumulating error while the requested action cannot be delivered. When the process finally becomes reachable, the stored integral contribution can delay output re-entry and produce a large recovery excursion.

During windup, nominal CO can continue beyond the limit while applied CO remains saturated, delaying recovery after the setpoint returns.

Compare these signals:

• Nominal CO before limiting.
• Applied CO after limits and selectors.
• Actual travel.
• Integral state or contribution where available.

Anti-windup methods include:

• Clamping or conditional integration: Stop or condition integration while saturation would drive the output farther into the limit.
• Back-calculation: Feed the difference between nominal and applied output back to the integrator. A tracking time constant Tt controls correction speed.
• External reset or tracking: Make the integrator track an actual actuator or selected signal, useful around overrides, selectors and cascade limits.

A representative back-calculation expression is:

I-dot = Ki e + (u_applied − u_nominal)/Tt

Signs depend on the implementation. Feedforward contributions must be included consistently in tracking.

No anti-windup method creates missing actuator capacity. Clamping and back-calculation also recover differently, and back-calculation adds another setting.

DIAGNOSTIC MINI BOX — WINDUP AND SATURATION

• Signature: Applied CO is pinned while nominal CO continues beyond the limit; recovery remains delayed after the constraint clears.
• Confirm: Trend nominal CO, applied CO, integral state, limits, selector states and travel.
• Fix: Correct limit handling and evaluate conditional integration, back-calculation or external-reset tracking.
• Cannot be fixed by tuning alone: Inadequate actuator capacity, unreachable setpoint or persistent process constraint.

Related topic: integral windup and anti-windup.

Valve stiction, deadband, backlash and hysteresis

The controller can calculate the correct output while the final-control element fails to deliver it. The strongest practical evidence comes from comparing applied CO with actual travel.

Table T04: Valve and actuator nonlinearity definitions

Term	Practical definition	Distinguishing feature
Stiction	Resistance to the start of motion: static friction holds the valve until breakaway, then travel slips or jumps	CO can ramp while travel remains stationary, followed by a jump
Deadband	Input range over which no observable output change occurs	Describes observed input/output behavior; backlash can be one cause
Backlash	Lost motion after the direction of input reverses	Travel response is delayed specifically around reversal
Hysteresis	Output depends on previous excursions and direction	The same input can correspond to different outputs depending on history
Dead zone	Input region that produces no output response	Not necessarily associated with reversal; distinct from deadband

In a stick-slip cycle, integral action changes CO while the valve remains stuck. Once the command exceeds breakaway resistance, the valve jumps. The error reverses, the command changes direction, and the sequence repeats.

Valve stiction is investigated by comparing controller output with measured travel: travel remains fixed, then slips to a new position.

A square-like PV with triangular or sawtooth CO can be a stiction clue in some self-regulating loops. It is not proof. Flat travel while CO changes and then jumps is much stronger evidence.

A useful way to state the evidence:

Valve stiction is a mechanical nonlinearity, so ordinary retuning does not remove the underlying fault. With integral action present, a stick-slip limit cycle can persist over a broad range of tuning. Changing gain or integral time often changes the cycle period and may change its amplitude, but the result depends on process dynamics, process type, operating point, stiction model, and controller structure. A calmer trend after detuning is not proof that stiction has been repaired.

No single rule describes how tuning changes affect a stiction cycle: amplitude and period responses depend on the model and the loop.

A one-direction bump may miss backlash. An authorized bidirectional test is more informative because reversal exposes lost motion. Positioner output, air supply, actuator pressure, linkage and packing condition may also be relevant.

DIAGNOSTIC MINI BOX — VALVE PROBLEMS

• Signature: CO ramps while travel is flat, followed by a jump; or travel is lost after reversal.
• Confirm: Use synchronized travel feedback, positioner and supply-pressure data; perform a controlled bidirectional test only when authorized.
• Fix: Correct air, positioner, actuator, linkage or valve mechanical problems; repair the underlying fault.
• Cannot be fixed by tuning alone: Stiction, backlash, damaged linkage or inadequate actuator performance.

Related topic: valve stiction explained.

Measurement, sampling, filtering and quantization

A loop can oscillate because the controller is acting on an inaccurate or poorly timed representation of the process.

Potential mechanisms include:

• Sensor noise or intermittent connections.
• Quantized signals that move in coarse steps.
• Historian compression that hides fast changes.
• Aliasing when the sample rate is too low.
• Communication delay or jitter.
• Controller tasks executing in an unexpected order.
• Filters adding more phase lag than expected.
• A control update tied to a slow analyzer or network scan.

A small limit cycle spanning one or a few digital counts can come from quantization. A cycle tied closely to scan or update timing suggests computation or sampling behavior, but it must be checked against physical periodic disturbances at the same timing.

DIAGNOSTIC MINI BOX — MEASUREMENT AND SAMPLING

• Signature: CO chatter follows raw-PV steps, or cycle timing is tied to scan/update intervals.
• Confirm: Check sensor quality, raw and filtered values, resolution, timestamps, task period, jitter, communication and compression.
• Fix: Repair measurement issues and correct sampling or filtering design.
• Cannot be fixed by tuning alone: Faulty sensors, insufficient resolution, network timing or unsuitable analyzer update behavior.

Adding a filter is not a neutral repair. It may suppress noise while adding phase lag. Reassess stability after any filter change.

Interaction, cascade loops and split-range systems

A loop can cycle because another controller changes the same process path. Several loops may share a manipulated stream, pressure header, heat source, vessel inventory or upstream disturbance.

When several loops show the same period:

1. Align timestamps accurately.
2. Compare waveform distortion as well as period.
3. Identify which signal changes earliest and most cleanly.
4. Trace physical and control-system connections.
5. Treat the earliest signal as a candidate source, not proof.

In cascade control, the primary controller sets the setpoint of a faster secondary loop. If the secondary loop is not sufficiently faster, is saturated or has poor actuator control, the primary may repeatedly demand corrections that the inner loop cannot deliver.

In split-range control, one controller output drives two or more final elements across different ranges. Cycling can concentrate around the split point if there is a gap, overlap, discontinuity, incorrect characterization or selector problem.

Cascade interaction can reveal a common period between linked loops, while split-range cycling may originate near the valve crossover.

Selectors and override controllers can also cause nominal and applied outputs to differ. An integrator that does not track the selected signal may wind up even though the displayed controller output appears reasonable.

DIAGNOSTIC MINI BOX — INTERACTION, CASCADE AND SPLIT RANGE

• Signature: Common periods across loops, slow inner-loop response, selector switching or cycling around a split point.
• Confirm: Trend all related SP/PV/CO signals, selected outputs, valve travel and shared disturbances.
• Fix: Correct inner-loop performance, tracking, selector logic, split-point gap/overlap or process interaction.
• Cannot be fixed by one-loop tuning alone: Shared process coupling, discontinuous handoff or inadequate secondary-loop authority.

Related topic: cascade control.

Related topic: split-range control basics.

External periodic disturbances

The PID may be responding correctly to a disturbance that repeats. Sources can include upstream cycling, utility-pressure variation, composition changes, analyzer updates or another unit.

A strong diagnostic observation is a cycle that persists in manual while actual travel remains constant. That points away from the controller calculation and toward external forcing. If travel keeps moving in manual, investigate the positioner, actuator, air supply and valve instead.

DIAGNOSTIC MINI BOX — EXTERNAL DISTURBANCE

• Signature: PV cycles in manual while verified actual travel is constant.
• Confirm: Correlate PV with upstream, utility, feed, pressure, composition and analyzer signals.
• Fix: Correct or compensate for the disturbance at its source where practical.
• Cannot be fixed by tuning alone: A continuing periodic disturbance outside the loop.

Ultimate gain and ultimate period

The Ziegler–Nichols closed-loop method uses:

• Ultimate gain, Ku: Proportional-only gain at which the idealized loop reaches a sustained cycle.
• Ultimate period, Tu: Period of that cycle.

For the ideal PID form Kc[1 + 1/(Ti s) + Td s], the traditional settings are:

• P: Kc = 0.5Ku
• PI: Kc = 0.45Ku, Ti = Tu/1.2
• PID: Kc = 0.6Ku, Ti = Tu/2, Td = Tu/8

Parallel equivalents are:

• PI: Kp = 0.45Ku, Ki = 0.54Ku/Tu
• PID: Kp = 0.6Ku, Ki = 1.2Ku/Tu, Kd = 0.075KuTu

Do not label Ki as integral time. Do not transfer these values into series or vendor-specific forms without conversion.

WARNING — ULTIMATE-CYCLE TESTING

Ultimate-cycle testing deliberately approaches the stability boundary. If Ku is exceeded, the oscillation can grow. Saturation, dead time, stored energy and process reactions can keep the PV moving after an operator changes mode. A clipped cycle is not a valid linear ultimate-cycle result because its amplitude is set by output or process limits.

Do not perform an ultimate-cycle or relay test without an approved, bounded test plan, defined stop criteria, independent monitoring where consequences justify it and a verified recovery route. Relay feedback bounds the requested excitation but still forces cycling and must never be described as inherently safe.

See Safe testing and management of change.

The method is historically associated with a quarter-amplitude-decay objective and can be aggressive for many applications. It does not adequately account for noise, hysteresis, constraints, valve wear, interaction or safety. It should not be called optimal, and no single tuning rule is universal.

Relay results can be misleading with stiction, backlash, deadband, saturation, rate limits, integrating or open-loop-unstable plants, asymmetry, noise chatter, selectors and interacting loops.

Textbooks or qualified training can help engineers understand the method’s assumptions, but no book, course or product can make an unsafe plant test safe without site-specific review and authorization.

Worked examples using scripted synthetic data

All four examples below use scripted synthetic data, not plant, field or laboratory data. Model parameters were chosen for teaching. Generated signals are illustrative, and reported metrics are calculated from those signals. They are not universal tuning recommendations.

Example 1: measuring ultimate period, decay ratio and clipping

SCRIPTED SYNTHETIC DATA — NOT PLANT DATA

A normalized loop uses the convention e = SP−PV, positive process and controller gain, and signals expressed as percent of span. A synthetic sustained record has:

• Sample interval: 5 seconds, or approximately 0.0833 minute.
• Record duration: 72 minutes.
• Baseline PV: 50%.
• Candidate ultimate gain: Ku = 4.0 %CO/%PV.

Eight same-side PV peaks occur at approximately:

12.0, 19.8333, 28.0, 35.9167, 43.9167, 52.0, 60.0833 and 68.0833 min

The peak values are:

55.0239, 54.9969, 55.0100, 54.9934, 55.0116, 55.0199, 55.0042 and 55.0065%

The seven measured periods are:

7.8333, 8.1667, 7.9167, 8.0, 8.0833, 8.0833 and 8.0 min

Using the first and last peak:

Tu = (68.0833−12.0)/7 = 8.011904761904761 min

For practical reporting, Tu ≈ 8.0 min. Peak-time uncertainty is at least one sample interval.

The corresponding values are:

• fu = 0.12481426448737 cycles/min
• ωu = 0.78423115275347 rad/min
• Full-record PV range: 10.060964631539314 percentage points
• Half-range amplitude: 5.030482315769657 points

The full-record range includes noise, so it slightly exceeds the nominal five-point peak amplitude.

Table T05: Ziegler–Nichols calculations for synthetic examples

Dataset	Controller form	Calculated settings
Ex1 nominal: Ku=4.0, Tu=8.0 min	P	Kc=2.0
Ex1 nominal	PI ideal	Kc=1.8; Ti=6.666666666666667 min
Ex1 nominal	PI parallel	Kp=1.8; Ki=0.27 %CO/(%PV·min)
Ex1 nominal	PID ideal	Kc=2.4; Ti=4.0 min; Td=1.0 min
Ex1 nominal	PID parallel	Kp=2.4; Ki=0.6 %CO/(%PV·min); Kd=2.4 %CO·min/%PV
Ex1 measured: Ku=4.0, Tu=8.011904761904761 min	PI ideal	Kc=1.8; Ti=6.676587301587301 min
Ex1 measured	PI parallel	Kp=1.8; Ki=0.26959881129271923 %CO/(%PV·min)
Ex1 measured	PID ideal	Kc=2.4; Ti=4.0059523809523805 min; Td=1.0014880952380951 min
Ex1 measured	PID parallel	Kp=2.4; Ki=0.599108469539376 %CO/(%PV·min); Kd=2.403571428571428 %CO·min/%PV
Ex2 analytic: Ku=4.251212494222509, Tu=7.441522726018012 min	P	Kc=2.1256062471112545
Ex2 analytic	PI ideal	Kc=1.913045622400129; Ti=6.201268938348344 min
Ex2 analytic	PI parallel	Kp=1.913045622400129; Ki=0.308492607145281 %CO/(%PV·min)
Ex2 analytic	PID ideal	Kc=2.550727496533505; Ti=3.720761363009006 min; Td=0.9301903407522515 min
Ex2 analytic	PID parallel	Kp=2.550727496533505; Ki=0.6855391269895132 %CO/(%PV·min); Kd=2.3726620791666386 %CO·min/%PV

These calculations illustrate the importance of stating controller form beside every number.

The separate synthetic decay-ratio dataset is:

PV(t)=50+8e^(−αt)[cos(ωt)+(α/ω)sin(ωt)]

where α=ln(4)/8 and ω=2π/8. Same-side peaks at t=0, 8, 16 and 24 min are 58, 52, 50.5 and 50.125%. Their amplitudes from the 50% baseline are 8, 2, 0.5 and 0.125 points.

Therefore:

DR = |52−50| / |58−50| = 2/8 = 0.25

That is quarter-amplitude decay.

Now apply artificial PV limits of 48–52%. The detected period remains 8 minutes, but the half-range becomes 2.0 points and peak-to-peak range becomes 4.0 points. This clipped waveform is invalid for ultimate-cycle inference because the amplitude is set by the limits rather than by an unconstrained linear cycle.

A usable ultimate-cycle estimate requires a sustained, unclipped cycle; quarter-amplitude decay means successive same-direction deviations have a decay ratio of 0.25. (Scripted synthetic data, not plant data.)

Example 2: how dead time changes stability margin

SCRIPTED SYNTHETIC DATA — NOT PLANT DATA

Consider the illustrative process model:

Gp(s) = 2e^(−2s)/(10s+1)

Thus:

• Process gain K=2.
• Time constant τ=10 min.
• Dead time θ=2 min.

At the ultimate frequency, the process phase is −π radians:

2ω + atan(10ω) = π

A simple bisection check gives:

• At ω=0.8, 2ω+atan(10ω)=1.6+1.4464=3.0464, below π.
• At ω=0.9, 2ω+atan(10ω)=1.8+1.4601=3.2601, above π.

Refining the interval gives:

ωu = 0.8443413449792343 rad/min

At that frequency:

• Phase: −3.1415926535897927 rad
• ωuτ = 8.443
• sqrt(1+(ωuτ)^2) ≈ 8.502
• |Gp(jωu)| = 0.23522700908482508
• Ku = 1/|Gp(jωu)| = 4.251212494222509 %CO/%PV
• Tu = 2π/ωu = 7.441522726018012 min

These results are exact for the stated mathematical model within the numerical solution shown. They do not include transmitter dynamics, filters, valve behavior, computation delay or process nonlinearity.

A P-only synthetic simulation uses:

• e=SP−PV
• Time step dt=0.005 min
• A 400-sample first-in-first-out dead-time buffer
• Exact zero-order-hold lag update
• No saturation
• Duration 120 minutes
• Unit SP step

The responses are:

Example 2 P-only synthetic gain sweep

P-only setting	Kc	Median same-side amplitude ratio	Synthetic result
0.5Ku	2.1256	0.1046 across 13 peaks	Strong decay
0.9Ku	3.8261	0.7563 across 16 peaks	Decay
Ku	4.2512	1.0093 across 15 peaks	Nearly sustained
1.1Ku	4.6763	1.2962 across 15 peaks	Growing

At Ku, the ratio differs slightly from exactly 1.0 because of discretization drift relative to the continuous analytic model.

At 1.1Ku, the no-saturation model reaches a maximum PV of 55.86, and the PV is still swinging at 6.56 at 120 minutes. The mathematical output is unbounded. A real plant would encounter physical limits and might alarm, trip or be damaged before reproducing this theoretical behavior.

Synthetic P-only gain sweep for Gp=2e^(-2s)/(10s+1): the response approaches a sustained cycle at Ku=4.25, with Tu=7.44 min. (Scripted synthetic data, not plant data.)

This example explains why ultimate-cycle testing is not an invitation to increase a real controller to Ku. Even a simple model grows rapidly when gain exceeds the boundary.

Example 3: saturation and four anti-windup responses

SCRIPTED SYNTHETIC DATA — NOT PLANT DATA

Consider:

Gp(s)=e^(−s)/(5s+1)

The illustrative ideal PI controller has:

• Kc=2.0
• Ti=4.0 min
• Parallel equivalent Ki=0.5 %CO/(%PV·min)
• Output limits 0–100 %CO
• Time step dt=0.02 min

The setpoint schedule is:

• 20% until 5 minutes.
• 120% from 5 to 25 minutes.
• 50% from 25 to 80 minutes.

The 120% setpoint is deliberately unreachable. With process gain 1 and maximum CO of 100%, the PV can only approach approximately 100%.

Four illustrative methods are compared:

1. No anti-windup.
2. Conditional integration.
3. Back-calculation with Tt=1.0 min.
4. Instantaneous external-reset tracking using I(k+1)=u_applied−Kc·e.

The external-reset expression is illustrative and not a vendor algorithm.

Table T06: Synthetic windup comparison

Method	Total saturation duration	Nominal CO re-entry after t=25	Maximum PV above new 50% SP	±2% settling after return	Maximum absolute integral state
None	30.46 min	10.46 min	49.82 points	25.92 min	456.70 %CO
Conditional integration	19.74 min	2.18 min	48.53 points	11.78 min	56.58 %CO
Back-calculation, Tt=1.0 min	22.0 min	2.0 min	48.53 points	10.88 min	70.58 %CO
External reset	10.32 min	1.04 min	48.51 points	8.44 min	100.0 %CO

The “maximum PV above new SP” values are large because the PV is still approximately 98–99% when the SP returns to 50%. This metric describes the process state left by the unreachable high setpoint; it is not a small conventional overshoot.

Without anti-windup, nominal CO reaches approximately 448.7% at 20 minutes and 360.3% at 25 minutes while applied CO remains at 100%. The integral state reaches 456.7% at 25 minutes. Applied CO does not leave its upper limit until approximately 40 minutes.

In the scripted comparison, nominal-CO re-entry after the SP return is 10.46 min with no treatment, 2.18 min with conditional integration, 2.0 min with back-calculation, and 1.04 min with external reset. (Scripted synthetic data, not plant data.)

The ranking is specific to this schedule, model, Tt, limits and algorithm semantics. It does not establish one anti-windup method as universally superior. Most importantly, anti-windup cannot make the unreachable 120% target reachable.

Example 4: a phenomenological stiction model

SCRIPTED SYNTHETIC DATA — NOT PLANT DATA

The illustrative process is:

Gp(s)=e^(−s)/(5s+1)

An ideal PI controller has:

• Kc=1.2
• Two Ti cases: 1.5 min and 3.0 min
• Sample time: 1 second, or 1/60 min
• SP: 50% initially, then 55% at 5 minutes
• CO limits: 0–100%

The two-parameter stiction model is phenomenological. It is not a friction-force or positioner model.

Its rules are:

• While stuck, if |c−v|<S, valve travel v does not change.
• If |c−v|≥S, breakaway occurs and travel jumps by J in the command direction.
• During continuing motion, v=c−d(S−J), where d is the direction.
• Command reversal returns the valve to the stuck state.

Illustrative parameters are:

• Stiction threshold S=2% span.
• Jump J=1% span.

The calculated results are:

Example 4 synthetic stiction results

Ideal PI setting	Mean late peak period	Late PV half-range	Moving samples	Largest stationary-run CO range
Kc=1.2, Ti=1.5 min	20.4 min	0.769 points	2317	Approximately 3.0 points
Kc=1.2, Ti=3.0 min	35.27 min	0.636 points	896	Approximately 3.0 points

In this model, doubling Ti increases period by 14.87 min and reduces PV half-range by 0.133 points. This is model-specific. The continuing stick-jump behavior shows that stiction was not repaired.

The synthetic manual-mode case sets constant CO to 50%. Travel remains at 50%, PV settles at 50%, and no autonomous cycle occurs. This simulated result shows only that feedback was necessary for this model’s cycle. It does not distinguish stiction from poor tuning by itself, and the simulation contains no external disturbance or autonomous positioner fault.

An illustrative bidirectional step sequence begins with travel at 50%:

1. Command 50→51%: no movement because the change is below S=2%.
2. Command 51→53%: breakaway; travel progresses 50→51→52%.
3. Command 53→52%: travel remains at 52%.
4. Command 52→50%: reverse breakaway; travel moves 52→51% and re-sticks.

This direction-dependent pattern resembles deadband or backlash behavior in the synthetic model. Real tests require authorization and may produce different travel paths.

In this synthetic stiction model, increasing Ti from 1.5 min to 3.0 min changes the cycle period from 20.4 min to 35.27 min without removing the modeled valve nonlinearity. (Scripted synthetic data, not plant data.)

The lesson is not that Ti should be doubled. It is that tuning changes can alter a stiction cycle without removing its mechanical cause.

Corrective-action matrix

Table T07: Master corrective-action matrix

Evidence	Likely category	Verification	Corrective direction
Smooth cycle changes strongly when Kc is reduced	Excessive loop gain	Verify form, action and scaling	Reduce Kc cautiously and reassess robustness
Response improves when Ti is increased	Integral too aggressive	Confirm ideal-form Ti versus parallel Ki	Increase Ti cautiously; check disturbance rejection and constraints
CO chatters with noisy or quantized PV	Derivative/noise	Inspect raw PV, derivative location and filter	Fix measurement first; then review derivative and filtering
Nominal CO exceeds applied CO	Windup	Check limits, selectors, tracking and integral state	Evaluate clamping, back-calculation or external reset
CO ramps while travel stays flat, then jumps	Stiction	Inspect travel, pressure, air and mechanics	Repair final-control element
Travel is lost after reversal	Backlash/deadband	Use authorized bidirectional test	Inspect linkage, actuator and positioner
Cycle concentrates around split point	Split-range discontinuity	Check overlap, gap, characterization, selector and handoff	Correct split-range design or configuration
Many loops share a period	Interaction or common forcing	Compare timestamps, spectra and physical paths	Address shared utility, upstream source or interaction
Manual mode plus constant travel	External disturbance	Trace feed, pressure, composition, utility and analyzer signals	Correct disturbance source or add justified compensation
Manual mode plus moving travel	Positioner, air or mechanical fault	Inspect positioner, supply, actuator and valve	Repair or service final-control element
Stability changes with load	Operating-point nonlinearity	Identify dynamics and constraints by region	Review gain scheduling or region-specific strategy
Cycle tied to scan/update period	Sampling or computation	Inspect task period, jitter, execution order, communications and filtering	Correct timing, sampling or implementation

Related topic: control loop performance monitoring.

Safe testing and management of change

WARNING — ACTIVE TEST AUTHORIZATION

Before any online tuning, bump, relay or forced-cycle test, document:

• Test purpose and decision to be made.
• Maximum test amplitude, duration, cycles and recovery time.
• Safe PV, CO, actuator, pressure, temperature, level, composition and equipment limits.
• Stop criteria and who may call a stop.
• Alarm, interlock, trip, relief and operator-response implications.
• Inventory, stored energy, reaction and environmental hazards.
• Downstream propagation and shared-utility effects.
• Actuator headroom in both directions.
• Who may start, pause, abort and restore the test.
• Independent monitoring where consequences justify it.
• A backed-up original configuration.
• A tested route to manual or an approved safe mode.

Avoid startup, shutdown, grade transition, abnormal utilities, maintenance periods or impaired safeguards unless those conditions have been specifically assessed. Manual mode is not an instant safety barrier: delayed and stored process effects may continue.

Testing should begin with a written question. For example: “Does actual travel remain stationary while applied CO changes?” A narrowly designed test is safer and more informative than changing several tuning parameters and observing whether the trend looks calmer.

A test/change record can use these fields:

Table T08: Test and change record template

Field	Required entry
Loop and equipment	Tag, service, final-control element and affected units
Diagnostic question	Competing hypotheses the test will distinguish
Baseline	Operating point, product or grade, mode, limits, disturbances and current metrics
Original configuration	PID form, Kc/Kp, Ti/Ki, Td/Kd, action, scaling, filters, limits and anti-windup
Risk review	Process limits, stored energy, alarms, trips, downstream effects and safeguards
Authorization	Named roles permitted to start, pause, abort and restore
Test bounds	Maximum amplitude, duration, cycles and recovery time
Stop criteria	Numeric or clearly observable conditions
Data plan	Signals, sample rate, clocks and historian settings
Result	Evidence, calculations and unresolved questions
Restoration	Final settings, modes, limits, tracking, anti-windup and backup status
Follow-up	Maintenance, engineering review, temporary-change expiry and closeout

Online tests and tuning changes must be screened under the site’s management-of-change (MOC) and temporary-change procedures. Legal requirements are jurisdiction- and process-specific.

For processes covered by the United States Environmental Protection Agency Risk Management Program, 40 CFR 68.75 requires written procedures for changes other than replacements in kind and consideration of technical basis, safety and health impact, operating-procedure modifications, duration and authorization. OSHA Process Safety Management, 29 CFR 1910.119, applies only to covered highly hazardous chemical processes and includes management of change.

Center for Chemical Process Safety guidance notes that control-system programming or sequence changes may require MOC rather than treatment as replacement in kind, depending on facility procedure. Temporary tuning changes need expiry, restoration and closeout. These United States examples do not mean every tuning change worldwide is legally subject to the same rule. Confirm jurisdiction, regulatory coverage and corporate policy locally.

A PID test can increase demand on alarms, interlocks or safety instrumented functions. Identify shared sensors, trip setpoints and process safety time. Do not alter trip logic, voting, setpoints or bypasses as an incidental part of PID work; those are separate safety-system activities.

Training materials and textbooks may help a team plan a test, but they do not replace competent review, local procedure or operating authority.

Related topic: management of change for control changes.

A planned test/change-record template and MOC/test-plan checklist are proposed: download link: placeholder.

Verification and documentation

A correction is not complete when the trend first looks smoother. Verify the mechanism and the operating result.

Use this checklist:

• Repeat the same period, amplitude, peak and recovery measurements used in the baseline.
• Use comparable operating points and disturbance conditions.
• Verify SP, PV, nominal CO, applied CO and actual travel remain synchronized.
• Confirm the actuator follows applied CO without unexplained stationary periods or jumps.
• Test recovery from relevant constraints without deliberately violating safe limits.
• Check performance after normal load or setpoint disturbances.
• Review representative operating regions rather than only one steady condition.
• Confirm controller form, action, scaling and time units.
• Verify final modes, cascade status, selectors, output limits, rate limits, tracking and anti-windup.
• Confirm filters and sample settings.
• Restore temporary instrumentation, trends and test logic as required.
• Record original and final parameters.
• Record evidence supporting the diagnosed cause.
• Close maintenance, MOC or temporary-change actions.
• Note remaining limitations and conditions not tested.

Do not permanently detune a loop merely to avoid valve maintenance. Do not present a synthetic trace as field evidence. If the correction is valid only within one operating region, document that boundary.

Frequently asked questions

Why do PID loops oscillate?

A PID loop oscillates when gain and phase lag sustain repeated movement, or when a nonlinearity, saturation, interaction, sampling effect or external disturbance forces a cycle. The PID may create the movement, amplify another source or correctly respond to a real periodic disturbance. Trend nominal and applied output, actual travel and disturbances before deciding to retune.

How can I tell valve stiction from poor tuning?

Compare applied CO with actual valve travel. A smooth PV and sawtooth CO can suggest stiction, but it is not proof. Stronger evidence is CO ramping while travel remains flat, followed by a sudden travel jump. Use positioner and supply-pressure information and, when authorized, a bidirectional valve test to confirm the mechanical behavior.

What does it mean if the oscillation continues in manual?

If the PV cycles in manual while actual actuator travel remains constant, suspect an external periodic disturbance. If travel continues moving, inspect the positioner, air supply, actuator and valve mechanics. Confirm that the loop is genuinely in manual. Persistence does not automatically exclude the valve, because a positioner or actuator can cycle independently.

Can too much integral action cause oscillation?

Yes. Integral action adds phase lag and continues accumulating error. In ideal form, a smaller Ti means stronger integral action; in repeats-per-minute form, a larger value is stronger. Excessive integral can delay output reversal and worsen windup during constraints. Verify controller form and units before increasing Ti or reducing parallel-form Ki.

Can derivative action cause or worsen oscillation?

Derivative can amplify noise, quantization and rapid sample-to-sample changes, producing CO chatter. Derivative on error can also react to setpoint steps. Filtering may reduce chatter but adds phase lag, so a poorly chosen implementation can aggravate stability. Inspect raw PV, derivative location, sample rate and filter definition before assuming derivative provides damping.

Why does dead time make a loop difficult to tune?

During dead time, the controller sees no measured evidence that its previous action is affecting the process. It may therefore add more proportional or integral correction before the delayed response arrives. Dead time adds phase lag without advance warning in the PV, reducing the gain and integral aggressiveness that the complete loop can tolerate.

What is integral windup?

Windup is integral accumulation while the requested controller action cannot be delivered because of output limits, selectors, rate limits or actuator capacity. Nominal CO can continue beyond the applied limit, causing delayed recovery when the constraint clears. Compare nominal and applied CO, then evaluate clamping, back-calculation or external-reset tracking for the implementation.

Are ultimate-gain and ultimate-period tests safe?

No active forced-cycle test is inherently safe. An ultimate-gain test deliberately approaches the stability boundary and can grow if Ku is exceeded. Saturation also invalidates linear ultimate-cycle inference. Testing requires a reviewed, bounded plan with limits, stop criteria, authority, monitoring and restoration steps. Manual mode does not immediately remove delayed or stored process effects.

Can a PID oscillation be fixed without retuning?

Yes. Retuning is unnecessary when the root cause is wrong action, scaling, time units, measurement noise, sampling, valve mechanics, limits, interaction or external forcing. Repairing a sensor, restoring actuator obedience, correcting split-range handoff or removing a periodic disturbance can solve the problem while preserving otherwise appropriate controller settings.

How do I verify that the real cause was fixed?

Repeat the same period, amplitude, recovery and constraint metrics under comparable conditions. Verify that actual travel follows applied CO and that recovery from normal limits no longer shows delayed integral release. Check representative operating points, final modes, tracking and filters. Document the original configuration, evidence, approvals, final settings and any untested conditions.

Planned downloads and next steps

The following supporting resources are planned. They are not presented as currently available files:

• Diagnostic decision tree — download link: placeholder
• Trend data-collection sheet — download link: placeholder
• Oscillation period and amplitude worksheet — download link: placeholder
• Test/change-record template — download link: placeholder
• MOC/test-plan checklist — download link: placeholder
• Synthetic CSV datasets and Python script — download link: placeholder
• Corrective-action matrix PDF — download link: placeholder

Start with passive trend collection. Measure the cycle, compare nominal CO with applied CO and actual travel, then identify the earliest credible source.

Primary CTA: Use the planned decision tree and data-collection sheet to classify the oscillation before anyone changes tuning.

Soft workbook CTA: Record each diagnostic step, test boundary, result and final configuration in the planned troubleshooting workbook.

The practical sequence remains:

Recognize → Measure → Classify → Test → Correct → Verify → Document.

References and disclosure

The article is based on the supplied research pack and the following registered sources:

1. ISA-TR5.9-2023 Realizing and Achieving Best PID (ISA): https://www.isa.org/intech-home/2023/june-2023/features/isa-tr5-9-2023-realizing-and-achieving-best-pid
2. Fundamentals of PID Control (ISA): https://www.isa.org/intech-home/2023/june-2023/features/fundamentals-pid-control
3. A Practical Guide to PID Controller Implementation (Sundström, Bauer, Guzmán, Hägglund, Soltesz): https://arxiv.org/pdf/2604.15918
4. Anti-Windup Control Using PID Controller Block (MathWorks): https://www.mathworks.com/help/simulink/slref/anti-windup-control-using-a-pid-controller.html
5. PID Control (Caltech Feedback Systems wiki): https://www.cds.caltech.edu/~murray/FBS/PID_Control.html
6. Different PID Equations (Control.com): https://control.com/textbook/closed-loop-control/different-pid-equations/
7. Quantitative PID Tuning Procedures (Control.com): https://control.com/textbook/process-dynamics-and-pid-controller-tuning/quantitative-pid-tuning-procedures/
8. Ziegler–Nichols Method (Michigan Tech): https://pages.mtu.edu/~tbco/cm416/zn.html
9. Ziegler–Nichols Controller Tuning Example (Colorado School of Mines): https://people.mines.edu/jjechura/wp-content/uploads/sites/120/2019/02/CHEN403_14_ZieglerNicholsExample.pdf
10. Modern PID Control (Bhattacharyya/UC Berkeley): https://msc.berkeley.edu/assets/files/PID/modernPID1-overview.pdf
11. Industrial Process Models (Iqbal/LibreTexts): https://eng.libretexts.org/Bookshelves/Industrial_and_Systems_Engineering/Introduction_to_Control_Systems_(Iqbal)/01%3A_Mathematical_Models_of_Physical_Systems/1.05%3A_Industrial_Process_Models
12. Stiction: Definition and Discussions (Choudhury, Shah, Thornhill; Springer): https://link.springer.com/content/pdf/10.1007/978-3-540-79224-6_11.pdf
13. Benchmarking Control Loops with Oscillations and Stiction (Horch; Springer): https://link.springer.com/content/pdf/10.1007/978-1-84628-624-7_7.pdf
14. Managing the Performance of Control Loops with Valve Stiction (Patwardhan et al.): https://www.researchgate.net/profile/Rohit-Patwardhan/publication/284446000_Managing_the_performance_of_control_loops_with_valve_stiction_An_Industrial_Perspective/
15. Analysis and Compensation of Control Valve Stiction-Induced Limit Cycles (Fang, Wang, Tan, Shang; IEEE): https://ieeexplore.ieee.org/document/7577013
16. Closed-loop Control Troubleshooting (Smuts, ISA): https://www.isa.org/intech-home/2018/march-april/departments/closed-loop-control-troubleshooting
17. Control Valves and Cycling Control Loops (Monsen, Valve World): https://valve-world.net/control-valves-and-cycling-control-loops/
18. CCPS Key Principles of Process Safety: Management of Change: https://www.aiche.org/sites/default/files/html/ccps/3028311/V1/files/downloads/MOC%20final%20Jan%202024.pdf
19. 40 CFR 68.75: https://www.ecfr.gov/current/title-40/chapter-I/subchapter-C/part-68/subpart-D/section-68.75
20. 29 CFR 1910.119: https://www.ecfr.gov/current/title-29/subtitle-B/chapter-XVII/part-1910/subpart-H/section-1910.119

The original 1942 Ziegler–Nichols paper, Optimum Settings for Automatic Controllers, was not directly inspected by the editorial team. Its bibliographic details and coefficients are therefore treated as partially verified through technical reproductions. Paid standards are not reproduced.

All numerical examples and worked-example figures are based on scripted synthetic data, not plant, field or laboratory results.

Disclosure: This page may display third-party advertising. No affiliate links or product endorsements are included in this article. Textbooks and training resources mentioned are educational aids and do not guarantee that a control-loop problem will be fixed.

This article is educational and is not a substitute for site operating procedures, hazard review, management of change, competent engineering judgment or local regulatory requirements.