Control system problems are the stuff of the Advanced Control Club at Westminster – which here applies a range of methods to deliver pragmatic solutions.
Today’s tough commercial conditions are forcing all of us to consider new ways of making operations faster, or more efficient. Engineers are just having to look harder and in more detail at their processes, plants and systems for any improvements that might add that essential small percentage to the bottom line.
The following are three examples which demonstrate the value of a pragmatic approach to advanced control techniques – supporting, rather than replacing conventional controls, and providing substantial economic benefits.
The first shows how a poorly-tuned pH control loop was tuned using previously recorded operational data for modelling with a neural network. Basically, a change of automation system on a fed-batch fermenter had introduced set-up and operational difficulties on the loop which set one of the main operational parameters.
The problem was exacerbated because the process was not operational at the time, so no on-line tests were possible. All developments had to be based on off-line analysis using previously recorded data.
In brief detail, the pH had to be within +/-0.05 of the setpoint. Nutrient was an acidic feed with a pH of 2.0 and the fermentation tended to acidify the medium. To adjust pH, a base solution was fed by a fixed feed rate pump controlled by a PID loop using a pH sensor for measured value. Fermentation was fed at a rate which increases exponentially, and at regular intervals feed rate was disturbed for testing.
To overcome unavailability of the plant for tuning, while needing to fix the PID settings for the next production sequence, the process was modelled and simulation used. But although there were two batches of data available, as with most plants there were interpretation difficulties.
Firstly, while data was recorded every 5 seconds, the control loop executed every 0.5 seconds, but control was only taken every 20 seconds – and the recorded pH value held unless a significant change occurred! So the dead-band in pH data meant that for periods where both base and nutrient were fed to the process no change in pH was indicated. This was observed for periods as long as 60 seconds. Thus, data pre-processing was required, recognising that the resultant data would be at irregular time intervals.
Secondly, no data was available for the control action – base feed rate or the state of the base feed pump (running or stopped). So the average of the PID action for each time interval was used to estimate the probable base feed pump running time and hence the base feed.
Following data analysis, and exploration of the process using linear regression models, a feed-forward neural network was adopted for modelling. The neural net had two inputs – mean base feed rate and mean nutrient feed rate. The output was the mean change of pH within five seconds. One hidden layer was used with 10 process elements, and the net was trained on one data set and then validated using the second in simulation.
To tune the pH controller it was first necessary to simulate the existing controller’s control algorithm. This is an essential part of all work with controllers – especially PIDs which are implemented in so many different ways. Once the simulation was set up, the controller was tuned using standard methods, like Ziegler-Nichols step and frequency response. And, once a set of initial PID parameters was available, final tuning was done using the experience of the control engineer – for practical response and robustness.
Subsequently, on-site tuning started at the beginning of the process with the nutrient fed at a low rate. As the feed rate increased, the control result was good with no steady state error. When the first reduction in feed rate occurred the pH controller was stable – however overshoot on pH was felt still too close to tolerance. Slight re-tuning was performed and the result was satisfactory – pH overshoot was much smaller.
Moving on, the next example concerned temperature controller action on a batch process in a jacketed vessel with complex thermodynamics, largely due to process changes during a batch. In the early stages, the organism required heat for growth, but as the process continued it became exothermic. So there were heating and cooling systems, and the process control demands were also unusually severe due to the sensitive nature of the organism.
Bang bang – to smooth
Main objective of the loop was to control temperature inside the vessel, in this case to within +/-0.1srC. Control was achieved by heat exchange between the medium and the vessel jacket, in this case filled with water/glycol re-circulated via separate heat exchangers for heating and cooling. Operation of the exchangers was controlled through two valves driven by a single PID controller.
From a control viewpoint the thermal characteristics were quite complex. The first heat exchange was between hot steam/cold water and water/glycol, and was fast. The second happened between the water/glycol and the fermenter, and was relatively slow. In addition, circulation of the water/glycol through the large jacket introduced a delay.
In fact, the existing PID controller achieved everything from a process viewpoint – temperature was maintained within tolerance with no offset. But, it achieved this in a `bang-bang’ mode. So there were problems. A large amount of energy was used; heat exchanger valves opened and closed frequently, resulting in high wear; the organism could be damaged due to localised overheating; and plant capacity was low because both heat exchangers were used at all times.
Unfortunately, re-tuning the PID controller for better operation did not achieve improvement. In fact, control performance was reduced in terms of temperature accuracy. But experience from these trials provided better process knowledge.
So a further trial was set up using a simple knowledge-based program, implemented in a lap top to guide manual control of the heat exchangers. Results were impressive, particularly given the rate at which the manual changes were executed. Accurate temperature control was achieved with less valve movement and lower energy consumption.
Knowledge-based control is simply a mimic of operator control extended to provide better coverage of the control region. It has the advantage of being less dependent on a process model, and is an easy way of incorporating empirical knowledge.
Operator knowledge can be included as `if-then’ rules and implemented on existing controllers. Here, the rules were easily derived from trials. The knowledge-based controller performs a look-up table search for control action. Implemented in the existing controller, the new algorithm required minimal changes to the system – and in fact all existing tuning parameters were used.
Benefits from the project included: well controlled vessel temperature; reduced energy consumption; minimal control valve movements; minimised variation in the water/glycol circuit temperatures; and substantial improvements in plant utilisation.
Finally, lets look at a difficult flow control loop in a web coating process. Here, layers of chemical solutions were laid on top of each other to form the product. Flow control was crucial to deliver the amount of solution to the coating machine. Typically, products required up to 10 flow loops to supply the various layers, and the configuration needed to change between production run to accommodate the different solutions being produced.
These ranged from Newtonian to thixotropic fluids with various ranges of viscosity and density. The range of flow setpoints was just as wide from 0.25-8 litres/min with tolerance of +/-0.5%. In-line flow meters provided the measurements and Baumann control valves, flow control. The complete system was controlled and sequenced by a conventional PLC.
The control valves displayed exponential behaviour, so non-linear control was required. While conventional flow control can be set up to handle a single fluid, or a group with similar characteristics, the combination of a range of setpoints and fluids requires much greater sophistication.
PI-gain scheduling vs fuzzy
Previously, a PI controller with gain scheduling was used. But engineers could see room for improvement. So a fuzzy logic controller seemed appropriate. In fact, it was implemented using standard ladder logic on the existing controller. As a result, no additional hardware was needed, and all development environment facilities were available.
The fuzzy logic program was written in modular form to allow easy configuration and reuse of code. In this implementation, measurement error was used as the fuzzy input. Crisp inputs were fuzzified using seven membership functions, each with an associated coefficient to partially compensate for valve non-linearities.
The fuzzy logic controller has demonstrated its value. When applied to such difficult processes on standard controllers it can provide: accurate control of setpoints for all process configurations; no requirement for gain scheduling; easy implementation on standard controllers; reduction in waste; and significant cost savings.
* Author is with the Advanced Control Club, University of Westminster.