Auto Tuners and Adaptive Control Perspective Tips
Automating any process can yield big improvements by eliminating human error and adding repeatability and predictability. The benefits are greatest when the best technology and practices are automated, the novice is protected against mistakes, and the specialist is enabled to capitalize on creativity and expertise. The same is true for PID tuning.
In this case, the tool must also minimize the disruption to the process. Process dynamics and tuning settings cannot be identified unless a known change is introduced into the process. I learned early on from Bob Otto (Monsanto Senior Fellow) that software cannot sort out the effect of PID response and the unknown external changes in closed loop operation despite claims to the contrary. Thus, auto tuners and adaptive control tuners are triggered by a change in setpoint or output. The change in output is normally a step made in manual or remote output mode. This simulates a step load disturbance.
The tuning is quite different for a setpoint and load disturbance unless a two degrees of freedom structure or a setpoint lead-lag is used to prevent overshoot. Therefore it is wise to check the load disturbance performance by putting the PID in manual, making a change in the PID output, and then immediately returning to the automatic mode. For a fast integrating processes (e.g. phosphorous furnace pressure) or runaway process (e.g. polymerization reactor temperature), the controller can only be in manual for a dead time. If the loop requires a variable speed drive or hydraulic actuator instead of an air actuated control valve to be fast enough, the loop cannot go in manual at all. Loops for equipment, environmental, and personnel protection and safety (e.g. surge and over pressure prevention) should also stay closed.
A pulse can be automatically injected into a PID output for loops that must stay closed (e.g. PID in automatic). The software required to identify the dynamics from a pulse is more vulnerable to noise and unknown disturbances. Consequently, more pulses than step changes are required and the time to achieve good model quality may be extended.
Pattern recognition techniques using heuristic rules can be used to determine when relative contributions of the proportional, integral, and derivative modes are out of balance. These can be a useful back up but care must be taken that the rules are not fooled by noise, periodic disturbances, limit cycles, and the windows of allowable gain and reset time for integrating and runaway processes. Normally, the safer thing to do is to allow for an automatic increase in reset time but require review and approval for an automatic decrease in reset time. An increase in reset time will always reduce loop oscillations and overshoot. The detrimental effect of a larger than necessary reset time are a slow return to setpoint for a load upset and a slower rise time for a setpoint change. Most of the time the reset time is too small. A notable exception is a highly dead time dominant self-regulating process where the minimum reset time is an order of magnitude smaller. Lambda tuning is suitable if a low limit of ¼ the dead time is imposed.
If the auto tuner and adaptive control tuner identify the open loop dynamics and allows the user to choose the tuning rules, the user is empowered. The knowledge of the open loop dynamics enables a better understanding and monitoring of automation and process dynamics. For example an increase in a secondary time constant or dead time would be indicative of slower heat transfer (e.g. surface fouling and frosting) or slower measurement (e.g. electrode aging or coating). The flexibility in tuning rules enables the expertise of the specialist to be effectively used. For example, a transition from lambda tuning for self-regulating processes to lambda tuning for integrating processes when the time constant to dead time ratio exceeds four indicating a near-integrating process, provides a better load disturbance response. Also, the setting of lambda (closed loop time constant for self-regulating processes and arrest time for integrating processes), enables the user to optimize the tradeoff between maximizing disturbance rejection, variability absorption, loop decoupling, and loop coordination.
A faster test time and the use of integrating process tuning rules are possible if the tuner allows lag dominant self-regulating and runaway processes to be evaluated as near-integrating processes. The open loop integrating process gain can normally be identified within about five dead times. If the software is capable of identifying the secondary time constant, the test identification time should be extended by at least two secondary time constants.
Plants have standardized on tuning rules for maximum load disturbance rejection (e.g. Shinskey), setpoint response (e.g. Internal Model Control or Simplified Internal Model Control, or some specific process objectives (e.g. lambda). The flexibility to use alternate tuning rules honors user preference and expertise.
The ability to simulate within seconds the response of tuning settings and identified dynamics for a load disturbance and setpoint response affords rapid experimentation. The simulation enables the user to evaluate performance metrics such as peak error and integrated error for load disturbances and rise time and overshoot for setpoint response.
Regardless of tuning method, the step size must be chosen to provide a PID output and input signal change larger than the deadband, resolution limit, and threshold sensitivity limit in the final control element (e.g. control valve or variable speed drive) and measurement (e.g. sensor and transmitter), respectively. While this step size is necessary to see the effect of the change on the process variable (PV), the increase in the closed loop dead time is not identified for steps in the PID output or setpoint when the immediate change in the PID output is larger than the deadband or resolution/sensitivity limit. Pattern recognition or a ramp rather than a step is needed to find this additional dead time. The ramp rate of PID output needs to be about the same as experience seen in the plant for a typical disturbance. Since this depends upon the tuning there is a chicken and an egg scenario. The tuning is computed based on step tests and then revaluated based on closed loop responses and the use of a typical ramp rate for open loop responses.
The step size must also be larger than process and measurement noise. A step size that is as large as possible without excessively upsetting the process shortens test time and makes sure the response is not washed out by the filtering action of well mixed volumes or confused by changes in feed streams and the action of other loops. Trend charts can show typical changes in PID output employed to deal with disturbances as a guide. In general, the higher the PID gain the larger the output change needed to prevent test time out and response wash out. For example, a level loop or batch temperature loop may have a PID gain of 10 or more. In this case the step size is much larger. The PV change may not be barely noticeable unless an output change of 20% or more is used.
The use of signal characterization to account for operating point nonlinearities can significantly reduce variation in the process gain identified with step size and direction.
The PID should be tuned with a judiciously minimized signal filter. The signal filter will add dead time or create a secondary time constant. The signal before and after filtering should be historized. The PID should be tuned with the analog output (AO) block and lower loop PID setpoint rate limits set if used for directional move suppression and the PID option external reset feedback (e.g. dynamic reset limit) is enabled.
A notable exception is a highly dead time dominant self-regulating process where the minimum reset time is an order of magnitude smaller. Lambda tuning is suitable if a low limit of ¼ the dead time is imposed.