14.1 Loop Auditing, Process Variability Analysis & Control Optimization

Loop Auditing, Process Variability Analysis & Control Optimization

In continuous and batch manufacturing facilities, distributed control systems (DCS) and programmable logic controllers (PLCs) regulate thousands of process loops. However, the initial commissioning of a control loop does not guarantee long-term performance. Without systematic auditing and active optimization, control loops degrade over time due to sensor drift, changing process dynamics, heat exchanger fouling, catalyst decay, and physical wear in final control elements. A Certified Automation Professional (CAP) must establish plant-wide loop auditing programs, detect mechanical and operational non-linearities, and maintain baseline regulatory and advanced control models.

---

The Plant-Wide Control Performance Reality

Published control-loop audits (such as those reported by Ender and by Bialkowski in the early 1990s, and later vendor benchmarking studies) have repeatedly found loops performing poorly for a few recurring reasons:

+-------------------------------------------------------------------------+
|             TYPICAL INDUSTRIAL CONTROL LOOP OPERATING STATUS             |
|                                                                         |
|   [ LOOPS RUN IN MANUAL ]                                               |
|   - Operators lost confidence due to cycling, hunting, or sluggishness  |
|                                                                         |
|   [ LIMITED BY VALVE DEFECTS ]                                          |
|   - High stiction, oversized trim, deadband, or actuator hysteresis     |
|                                                                         |
|   [ POORLY TUNED ]                                                      |
|   - Oscillatory gains, wrong integral times, or incorrect controller action|
|                                                                         |
|   [ PERFORMING WELL ]                                                    |
|   - Stable, robust rejection of load disturbances, minimal variance     |
+-------------------------------------------------------------------------+

When a control loop exhibits poor stability, control room operators invariably switch the controller from Automatic or Cascade into Manual. Operating in manual forces operators to act as human controllers, elevating fatigue, inducing process variability, and driving unit operating points away from optimal economic constraints.

Loop Mode Assessment Metrics

A comprehensive loop audit begins with an automated survey of controller operational modes across the DCS configuration database over a representative multi-week period:

$$\text{Service Factor (Operational Index)} = \frac{T_{\text{Cascade}} + T_{\text{Auto}}}{T_{\text{Total}}} \times 100\%$$ $$\text{Cascade Utilization Ratio} = \frac{T_{\text{Cascade}}}{T_{\text{Cascade}} + T_{\text{Auto}}} \times 100\%$$

Where $T$ represents the cumulative running time in hours. Typical site targets are:

• Plant-Wide Automatic Service Factor: $\ge 95\%$ for all non-batch regulatory loops.
• Cascade Utilization Ratio: $\ge 90\%$ for master/slave architectures.
• Loops with a Service Factor below about $80\%$ are flagged on the engineer's daily exception report for diagnostic troubleshooting.

---

Control Loop Benchmarking: The Harris Index

Evaluating whether a control loop is performing well cannot rely on standard deviation alone. A flow loop regulating a turbulent slurry will naturally display higher raw variance than a well-insulated thermal storage tank. Control performance must be benchmarked against what is theoretically achievable given the transport delay (dead time) of the process.

In 1989, T.J. Harris established the Minimum Variance Benchmark, which defines the absolute minimum achievable variance ($\sigma_{\text{mv}}^2$) for a linear process with dead time $d$ under feedback control.

FEEDBACK CONTROL VARIANCE BENCHMARKING:

 Actual Output Variance (sigma_y^2)
 +-------------------------------------------------------------------+
 |   Achievable Performance Improvement (Tuning / Modernization)     |
 |   +-------------------------------------------------------------+ |
 |   | Unavoidable Process Disturbance Variance (sigma_mv^2)       | |
 |   | Governed strictly by process dead time (d)                  | |
 |   +-------------------------------------------------------------+ |
 +-------------------------------------------------------------------+

Mathematical Formulation

The Harris Index ($I_H$) is defined as the ratio of minimum achievable variance to actual observed output variance:

$$I_H = \frac{\sigma_{\text{mv}}^2}{\sigma_y^2}$$

Because the actual variance $\sigma_y^2$ cannot be smaller than the theoretical minimum variance $\sigma_{\text{mv}}^2$, the Harris Index is bounded:

$$0 < I_H \le 1.0$$

Many commercial control diagnostic packages invert or rescale this metric into the Normalized Performance Index (NPI):

$$\text{NPI} = 1 - I_H = 1 - \frac{\sigma_{\text{mv}}^2}{\sigma_y^2}$$

Screening bands vary by company; a typical set is:

---

Valve Non-Linearities: Stiction, Deadband & Hysteresis

When a control loop cycles continuously, automation technicians often assume the PID gains are too aggressive and detune the controller by decreasing the proportional gain ($K_c$) and increasing the integral time ($T_i$). If the underlying cause is mechanical friction in the control valve, detuning the controller merely slows down the cycle frequency without eliminating the oscillation.

Definitions and Physical Mechanisms

1. Deadband: The range through which an input signal can be varied without causing a measurable change in the physical stem position. Caused by mechanical play, loose linkages between actuator and positioner, or sloppy rack-and-pinion actuators.
2. Hysteresis: The maximum difference in valve position for a given input signal when approached from opposite directions (upscale vs downscale). Caused by magnetic hysteresis in positioner coils or elastomeric diaphragm deformation.
3. Stiction (Static Friction): The phenomenon where the static friction resisting stem breakaway exceeds the dynamic kinetic friction resisting motion ($F_{\text{static}} > F_{\text{dynamic}}$). When the controller demands small corrections, the valve stem remains mechanically stuck until the integrator builds sufficient pressure in the actuator to overcome static friction. The stem then abruptly "jumps" to a new position, overshooting the setpoint. The controller then integrates in the reverse direction, sticking again, creating a permanent limit cycle.

VALVE STICTION SLIP-JUMP BEHAVIOR (Choudhury / Kano models):

 Valve Stem Position
        ^
        |                 Slip-Jump! (Overshoot)
   X_2  +-------------------->+-------------------------
        |                    ^
        |                    |  Stuck Region
        |                    |  (Actuator pressure builds)
   X_1  +--------------------+-------------------------
        +--------------------------------------------->
                                  Actuator Drive Force (Controller Output)

Diagnostic Techniques for Valve Stiction

To prove whether a cycling loop is caused by valve stiction or controller tuning, the engineer employs three diagnostic tests:

PHASE-PLANE TRAJECTORY ANALYSIS (PV vs. Controller Output):

  PV (Process Variable)
   ^
   |         +-----------------------+
   |         |                       |   Characteristic "Parallelogram"
   |         |                       |   or rectangular limit cycle
   |         |                       |   confirms VALVE STICTION
   |         +-----------------------+
   |                    
   |                    /            Linear / Elliptical trajectory
   |                   /             indicates LINEAR MIS-TUNING
   |                  /              (Aggressive PID gain)
   +---------------------------------------------> Controller Output (OP)

1. Phase-Plane Trajectory (PV vs. OP / Valve Travel):
   • A plot of the Process Variable (PV) against Controller Output (OP) or valve position feedback under closed-loop cycling.
   • Linear Controller Mis-Tuning: Generates an elliptical or circular Lissajous figure.
   • Valve Stiction: Generates a distinct parallelogram or box-like trajectory with sharp, flat edges representing the period during which the valve stem is stuck while the controller output ramp-integrates.
2. Cross-Correlation Analysis:
   • Cross-correlating the controller output error $e(t)$ with the valve position change $\Delta u(t)$.
   • Pure linear systems display smooth sinusoidal correlation functions. Stiction generates triangular wave cross-correlations with an odd harmonic structure.
3. Bimodal Error Distribution:
   • Under pure stiction limit cycles, the process variable does not follow a Gaussian normal distribution around the setpoint. Instead, it spends most of its time dwelling at the upper and lower extremes of the oscillation, yielding a bimodal (two-peaked) probability density function.

---

Process Variability & Frequency-Domain Spectral Analysis

When multiple interconnected process units oscillate simultaneously, finding the root-cause "culprit" loop requires spectral analysis. Time-domain trend inspection is deceptive because an oscillation originating in a feed pump control loop will propagate downstream through preheaters, distillation columns, and product storage.

Fast Fourier Transform (FFT) & Power Spectral Density (PSD)

Transforming time-series operational data into the frequency domain decomposes complex disturbances into constituent spectral peaks:

$$S_{xx}(f) = \int_{-\infty}^{\infty} R_{xx}(\tau) e^{-j 2 \pi f \tau} d\tau$$

Where $S_{xx}(f)$ is the Power Spectral Density and $R_{xx}(\tau)$ is the autocorrelation function of the process variable.

FREQUENCY-DOMAIN SPECTRUM (PSD) FOR ROOT-CAUSE IDENTIFICATION:

 Power Spectral Density
   ^
   |             Primary Disturbance Peak (f_0 = 0.012 Hz, Period = 83s)
   |                     |
   |                   +---+
   |                   |   |         Harmonic (2 * f_0)
   |                   |   |               |
   |                   |   |             +---+
   |    Noise Floor    |   |             |   |
   +-------------------+---+-------------+---+------------------> Frequency (Hz)
                       Loop A: Culprit (Sharp, non-linear fundamental peak)
                       Loop B: Downstream victim (Filtered, attenuated peak)

Spectral Troubleshooting Rules of Thumb

---

Advanced Process Control (APC) Maintenance & Model Degradation

Many refining, petrochemical, and chemical plants operate multivariable Model Predictive Control (MPC) applications (such as Dynamic Matrix Control) supervising underlying regulatory PID loops. While MPC generates substantial economic benefits by driving the process to physical and economic constraints, MPC controllers suffer from model degradation over time.

MPC SUPERVISORY ARCHITECTURE & MODEL MAINTENANCE:

 +-------------------------------------------------------------------------+
 |                     MPC MULTIVARIABLE CONTROLLER                        |
 |                                                                         |
 |   Manipulated Variables (MVs)              Controlled Variables (CVs)   |
 |   - Heavy Naphtha Draw Rate  ---------\    - Distillation Top Temp      |
 |   - Column Reflux Ratio      -----\    \-> - 95% Cut Point              |
 |   - Reboiler Steam Flow      --> [ Dynamic Matrix ] -> - Column Bottom Level |
 |                                    [ Model (A) ]                        |
 |                                          ^                              |
 |                                          | Model Verification / PRBS    |
 |                             +--------------------------+                |
 |                             | Dynamic Model Maintenance|                |
 |                             +--------------------------+                |
 +-------------------------------------------------------------------------+
                                            |
                                            v (Setpoints to Base PID Loops)
                              [ Base Regulatory PID Layer ]

Causes of Dynamic Model Degradation

An MPC application relies on an empirical linear dynamic model representing the step response between Manipulated Variables (MVs), Disturbance Variables (DVs), and Controlled Variables (CVs):

1. Heat Exchanger & Reboiler Fouling: As organic scale, wax, or biological growth accumulates on heat exchanger tubes, the overall heat transfer coefficient ($U$) drops. The process gain ($K_p$) between steam valve position and process temperature decreases, while the effective thermal time constant ($\tau$) increases.
2. Catalyst Deactivation: In catalytic reactors (hydrocrackers, reformers), catalyst activity declines over months of operation. Lower conversion kinetics alter the steady-state gain between space velocity, reactor bed temperature, and product yield.
3. Distillation Column Wear / Weeping: Damaged valve trays, fouling in structured packing, or foaming changes the hydraulic residence time, modifying the apparent dead time ($\theta$) between reflux changes and overhead composition.
4. Feedstock Composition Shifts: Switching crude oil slates from light sweet crude to heavy sour crude completely invalidates linear models derived from step testing under sweet crude conditions.

MPC Model Maintenance Protocols

When MPC model degradation occurs, the multivariable controller begins "fighting" the process, causing severe cycling across multiple manipulated variables, constraint violations, and frequent operator shedding of the controller.

• Passive Model Residual Auditing: Continuously calculate the model prediction error vector: $e(k) = y(k) - \hat{y}(k)$, where $y(k)$ is the measured CV and $\hat{y}(k)$ is the model-predicted CV. If the integrated prediction error exceeds statistical confidence thresholds ($> 3\sigma$), the model matrix is flagged as invalid.
• Closed-Loop / Open-Loop Re-Identification: Historically, plants performed invasive open-loop step testing, moving valves up and down across shifts. Modern best practice employs Pseudo-Random Binary Sequence (PRBS) signals or small sinusoidal perturbations injected into MVs during closed-loop operation. This allows identification algorithms to extract updated step response coefficients without interrupting commercial production.
• Gain Scheduling & Adaptive Control: For processes with predictable non-linearities (e.g., pH neutralization or throughput-dependent dead times), the automation engineer implements gain scheduling, adjusting the dynamic matrix based on production rate: $\theta(F) \propto \frac{V}{F}$, where $V$ is pipe volume and $F$ is volumetric flow.

---

The Economic Payoff of Loop Optimization

Why invest engineering labor into continuous loop auditing and retuning? The fundamental objective of regulatory control optimization is to reduce process variability (narrow the standard deviation $\sigma$).

BENEFIT OF REDUCING VARIABILITY (SHIFTING THE OPERATING MEAN):

 Frequency
   ^
   |                               Upper Operating Constraint / Safety Trip
   |                                           |
   |           Old Operating State             |
   |           Wide Spread (Large 3-sigma)     |
   |               +---------------+           |
   |              /                 \          |
   |             /                   \         |
   |   ---------+-------+-------------+--------|
   |                    |                      |
   |             Old Mean (mu_1)               |
   |                                           |
   |                       New Operating State (Optimized)
   |                       Narrow Spread (Tight 3-sigma)
   |                             +----+        |
   |                            /      \       |
   |                           /        \      |
   |   -----------------------+----+-----+-----|
   |                               |           |
   |                       New Mean (mu_2)     |<-- Operating Shift Benefit
   +------------------------------------------------------------------------>
                                                           Process Value (PV)

When a process variable (such as reactor temperature, product moisture, or column pressure) exhibits wide standard deviation ($3\sigma_{\text{old}}$), plant operators are forced to maintain the operating setpoint ($\mu_1$) far away from the physical constraint or safety trip limit to prevent off-spec production or emergency shutdowns.

By auditing loops, eliminating valve stiction, and retuning PID parameters, variability shrinks to $3\sigma_{\text{new}}$. The plant can safely shift the mean operating target closer to the economic optimum:

$$\Delta \mu = \mu_2 - \mu_1 = 3 \left( \sigma_{\text{old}} - \sigma_{\text{new}} \right)$$

In a catalytic reformer, shifting the average reactor temperature closer to the metallurgic constraint by just $1.5^\circ\text{C}$ without increasing trip risk can generate millions of dollars annually in high-octane gasoline yield.

---

Realistic Industrial Scenario: Hydrotreater Fractionator Optimization

An automation engineer at an oil refinery audits the heavy gas oil hydrotreater fractionator unit. Operators complain that overhead pressure controller PIC-104 and reflux accumulator level controller LIC-102 are constantly "fighting each other," causing the diesel flash point to oscillate outside commercial sales specifications.

The Audit Findings

1. Mode Assessment: LIC-102 had an automatic service factor of only $42\%$ over the preceding quarter; operators repeatedly dumped the level controller into manual mode during night shifts.
2. Benchmarking: Passive calculation of the Harris Index on PIC-104 yielded $I_H = 0.11$ (NPI $= 0.89$), indicating massive excess variance over the theoretical minimum.
3. Spectral Sweep: Fast Fourier Transform (FFT) of pressure PIC-104, reflux flow FIC-105, and level LIC-102 revealed a single dominant spectral peak at $f = 0.0055\text{ Hz}$ ($T_{\text{cycle}} = 182\text{ seconds}$) present in all three loops with multiple odd harmonics.
4. Valve Diagnostic: Plotting the phase-plane trajectory of FIC-105 controller output vs. reflux valve positioner feedback revealed a classic rectangular stiction pattern with an apparent deadband of $3.8\%$.

The Resolution Protocol

• The mechanical maintenance team inspected the reflux control valve FV-105 during a planned unit mini-outage, discovering over-torqued PTFE packing that had scored the valve stem.
• The valve packing was replaced with a low-friction live-loaded graphite assembly, and the smart digital positioner was re-calibrated. Post-repair deadband measured $< 0.3\%$.
• The automation engineer retuned LIC-102 using averaging (surge-absorbing) level control designed to absorb flow surges rather than violently stroking the reflux valve.
• The overhead pressure loop PIC-104 was retuned using the Lambda tuning method to match the closed-loop time constant of the furnace firing controls.
• Outcome: The $182\text{ s}$ limit cycle completely vanished. The standard deviation of the column overhead pressure dropped from $\pm 4.2\text{ psig}$ to $\pm 0.4\text{ psig}$. Diesel flash point giveaway was reduced, yielding an estimated $340,000 annual economic benefit.

---

Common Exam Traps

1. Assuming Controller Retuning Fixes All Oscillations: A widespread exam trap asks how to eliminate a sustained, constant-amplitude limit cycle. Candidates often select "reduce proportional gain" or "increase integral time." If the oscillation is driven by valve stiction, retuning the controller will not stop the limit cycle; the physical valve friction must be remediated or compensated.
2. Misinterpreting the Harris Index: Exam questions often test whether a high or low Harris Index signifies poor performance. Remember: $I_H = \sigma_{\text{mv}}^2 / \sigma_y^2$. A value close to 1.0 indicates excellent (near minimum variance) control. A value close to 0.0 indicates poor control with severe excess variability.
3. Overlooking Process Non-Linearity in APC Degradation: Candidates frequently assume MPC model errors stem from software communication bugs. In reality, model degradation is almost always caused by physical equipment degradation: heat exchanger scaling, catalyst deactivation, or changing feed compositions that alter steady-state gains and time constants.