temp control problem

The two outstanding difficulties are asymmetric control and long and asymmetric thermal lags. The system is far from linear.

The rate of cooling can be controlled at the surface by the amount of air flow, but the heating isn't under your control at all. The rate of change at the thermocouple is only loosely related to the rate of change at the surface, and the time lag between the two temperatures is of the order of the oscillation period you observe. You can solve this, but not with a classical PID. To begin with, you need an additional sensor.

A second thermocouple embedded closer to the surface will allow you to infer rates of cooling. (In controllers for muffle furnaces, some derivative effect can be had by partly withdrawing the thermocouple into the firebrick, thereby moving it closer to the source of heat.)

Model the thermal properties of your block and thermocouple(s). The best results will be had by using the model directly as part of the control program, but it might provide the insight you need to do good control without it.

One observation that might be useful: thermal lag limits the frequency that your thermocouple can produce to very low values. Correspondingly limiting the rate at which the airflow can change might help stability. A second thermocouple closer to the surface will provide not only the gradient information that you need, it will be able to change faster and so make hither control frequencies useful.

Jerry

Have you yet tried to tune the loop manually?

If you put the air control valve in manual, does the temperature eventually reach an equilibrium temperature that seems to be related to the air flow?

What are the tuning variable units for your controller? E.g. Proportional band or gain, integral gain or time? If time, seconds or minutes? Derivative gain or time?

yes - open loop seemed to get closest

1. Put the controller in manual mode, set the output to a nominal operating value, and allow the PV to settle completely. Record the PV and output values.
2. Make a step change in the output. Record the new output value.
3. Wait for the PV to settle.

P (Percent) = 80*K*Td/T I (Minutes) = 2.0*Td D (Minutes) = 0.5*Td

Where K = change in output / change in PV T = time constant Td = dead time

yes - it does asymptotically approach a constant value after ~3-4 min

Your tuning is pretty aggressive. Start with I=2*process open loop time constant (it will settle slowly, but the integral will not be involved in the problem).

Start with P at something like 100% and work down by multiplying that

70% for each trial. When the output starts to wobble its way to equilibrium, instead of settling from one direction, stop and back up one step. Then start stepping I by 70% factors till the same thing happens to the output. And back up one step.

Then you can see if your estimate for Td is helpful or not. If oscillations occur at a few dead times, you have too much derivative.

Clearly, the controller's response is too prompt. Slow it down to more closely match the thermal lag from airflow change to sensor response.

Jerry

I would agree with Jerry that the system is significantly nonlinear - yes even asymetrical, but not critically so - provided that there is always some air cooling present during operation. I have some ideas on how to linearize the relationship between air valve signal and cooling effect. Once this effect is linearized, you can go further.

Jerry's comments on modeling and additional measurements also ring true. If you can't move the present measurement, then without the additional measurement you would remain dependent on high-order feedback control, being always very nervous, if tuned to be accurate under disturbance conditions.

If you implemented separate feedforward compensation based on an additional measurement, you would be able to significantly de-tune feedback control, then maybe PID would be entirely successful.

Design of the feedforward compensation would require additional linearisation of the heat transfer relationship involving furnace temperature and the new measurement. A few simple assumptions on heat transfer modes (with two or three steady-state measurements) would allow a simple model to be implemented to sort this nonlinearity out.

A lot more detail could be filled in - though not likely to be all relevant without knowing more about your specific problem. I would be happy to correspond to solve your problem - either privately, or here.

John Stiekema BSc (Eng) (Mech), GDE (Elec) j dot stiekema at ee dot wits dot ac dot za

Just to add a little to much of the sensible advice already given by others:

Darths Jordan wrote: "The system has a long dead time - on the order of about 20-25 seconds between when the air comes on and when the thermocouple detects a change in temperature of the block (thermal capacitance)."

It is important to distinguish between two types of time-delay: 'dead-time' (as mentioned) is a pure time delay, like a transit delay on a conveyor belt; while 'dynamic lag' relates to the exponential response that is characteristic of systems that can be described by differential equations, such as thermal hold-up or capacitance (also mentioned).

If the time delay is primarily one of lag, then setting the integral action time to approximately the value of that lag, then adjusting the proportional gain, is a good starting point (as others have suggested). If, however, there is significant dead-time (transit delay) - which seems unlikely from the description, then a dead-time compensating approach may be needed. However 'significant' in this context means the dead-time exceeding, say, 50% of the time-constant of the dominant lag.

Kelvin B. Hales Kelvin Hales Associates Limited Consulting Process Control Engineers E-mail: snipped-for-privacy@khace.com Web:

------------------------------------------------- Ted Rubberford. 'The Man In The Red Latex Skintight Suit'

How about trying a watlow F-4 controller with cascade control where you have two thermocouples one embedded and one on the surface . Just a thought.

Just a guess: shallow embedding will work better than surface mounting. The closer to the surface, the more the effective derivative component. There can be too much of a good thing.

Jerry

Would that be more related to the life of the thermocouple, than anything else, Jerry?

John

Life is an issue I didn't address. Moving a temperature sensor closer to the heat source introduces some "anticipation" into its indication. That's what a derivitive term does, but an actual differentiator introduces more noise. An embedded sensor sees the (leaky) integral of the heating history, and needs either a fair amount of compensating derivative or a limit on the output speed that makes the integration's inherent lag insignificant.

Jerry

You can try to overcome this delays with a Smith controller, it works well with even long delay times. You will need a model, however.

rgrds, LL