Automatic Controls: White Paper
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
Home | Download this White Paper as a PDF file | View as one page
<< Previous 1 2 3 4 Next >>
Control Strategies
There are many different control strategies. We can list a few examples here and relate them to the implementation strategies previously described.
Related |
|---|
|
|
Model-Free
Model-free controllers use various linear or nonlinear mappings from input values and state values without explicitly representing a system model. The most common is PID control. Some generalized mapping controllers such as neural networks are in this class.
Unless there are unusual requirements for very high rate updates, diligent monitoring, large numbers of channels, etc., Data Acquisition Processors are usually not the first choice for the simple and conventional model-free approaches such as PID. It is not unusual, however, that slight variations are required, and, while these can be difficult to implement on other platforms, they are almost trivial with DAPL processing commands.
Model-Based
Model-based controllers have control laws with special structures matched to the properties of the system at design time. On the theory that there is a better chance of controlling a known system than an unknown one, a custom-matched controller design should work better than a "generic black box" solution. The cost is primarily in the engineering analysis. The resulting control law, in whatever its special form, is typically not much more difficult to apply than PID control.
Model-Reference
Certain strategies for complex systems incorporate a simulation (sometimes reduced order) that explicitly represents the system, and then bases control action on a comparison between the simulated and actual system response to observed inputs. The simulation can also be used to project future effects of present-time control decisions. There is a close relationship to state observer methods. The time-critical control rule for using the state information is usually very simple in comparison to the background evaluation of the model. Obtaining a useful system model is not an easy problem in general.
Adaptive
For systems with known form but poorly known parameter values, or systems that have properties that change over time, numerical techniques for estimating parameters dynamically are well known. These computations can require a lot of processing over an extended time, but with no urgency and no real-time deadlines. Lower-priority processing tasks in the DAPL system work well for this.
Adaptive controllers are a complement to model-reference and observer-based controllers. Those controllers presume that a discrepancy between the observed output and the model is due to a disturbed state that must be corrected. Adaptive controllers, on the other hand, presume that discrepancies in the observed output are due to deficiencies in the model, and that better model parameters will produce improved performance in the future. Model-free controllers can use a direct-adaptive approach, adjusting gains and other control parameters directly in an attempt to improve control performance without reference to an explicit model, but the effects of controller parameters can be difficult or impossible to distinguish from unknown system parameters in a closed loop configuration, so proving convergence can be difficult.
Noisy Systems
For systems in which signal measurements have a large amount of noise, the control strategy can take variability of the measurements and resultant variability in state estimates into account. Optimal controllers based on the classic Kalman Filter and the Extended Kalman Filter are of this class. These designs iteratively solve a separate, explicit model of random noise processes and use that to guide state estimate correction. Noise increases the difficulty of obtaining suitable models and limits applicability.
Contrary to the popular myth, these controllers typically are not adaptive. The state is adjusted, but the model parameters are not. Kalman filtering techniques can be used in combination with adaptive methods, but that is relatively uncommon.
Both the state update and the state correction update require matrix mathematics. A considerable amount of processing power is required. The floating point hardware support provided by a Data Acquisition Processor main processor chip is very helpful for this.
Feedforward
Given a desired system output trajectory, pre-processing of a command signal can be very effective in driving a system along that trajectory with minimum excitation of undesirable oscillatory modes. DAPL processing tasks are particularly good at this kind of processing.
A weakness of feedforward controls is that they cannot see the effects of disturbances, so they typically must be used in combination with stabilization and feedback correction controls. That combination can be more effective than feedback controls alone.
Nonlinear
Every system is nonlinear to some degree, and many optimal controllers are nonlinear. Nonlinear control strategies can be used to compensate for system nonlinearities. Even when the system is linear, a nonlinear control strategy can be applied to improve performance.
As an example of nonlinearity in a system, thermal leakage flow in a refrigeration system is one directional and varying, from a warmer ambient environment into the refrigerated chamber. To make the refrigerated temperature colder takes effort to overcome the thermal leakage, whereas thermal leakage can raise the temperature by itself if no action is taken. A controller that responds unequally to colder and warmer temperature deviations can compensate for the imbalance, improving regulation.
An example of nonlinear control of linear systems is the "California Driver" strategy. To reach a destination as quickly as possible, stomp on the fuel pedal until the last possible instant, at which time you stomp down on the brake and bring the vehicle to a halt exactly at the destination. As harrowing as the strategy might be, it works. You will spend the least time accelerating and also the least time decelerating, so you arrive as quickly as possible. In contrast, a PID control policy would gradually remove fuel or braking action as the distance to the destination decreased, allowing the vehicle to drift gradually to the final destination, but taking more time. A mixed strategy can be applied, but the mixing process is itself a kind of nonlinearity.
Fuzzy Controls
Control problems that have multiple operating regimes can sometimes use the fuzzy control formalism to apply relatively simple control strategies in combination, with the relative "membership" of fuzzy decision variables determining the mix of actions to apply.
While very general, fuzzy systems require additional mappings to convert from measurement to internal variables, and from internal variables to outputs. Fuzzy rules have the property of always applying, but to a lesser or greater degree, so every possible action is considered all of the time, possibly contributing zero to the result. Consequently, the extreme generality of fuzzy control comes at a significant cost.
If you have multiple fuzzy control commands, combining the commands in a single downloadable module allows them to share the fuzzy inference engine code, but not the rule sets or state. Each task will instantiate its own independent rule base and evaluation state.
Tuning and Monitoring
Most controls are self-contained units. Most settings must be applied physically at the device location. With Data Acquisition Processors, evaluating and adjusting settings is a software function, so the Data Acquisition Processor can assist with the measurement, application, and configuration management. The actual monitoring can be done at the local station or controlled from a remote location. While tuning and monitoring are not central to the real-time control problem, they can be central to the problem of controlling operating costs.
Passive Observation and Logging
Related |
|---|
Is the loop alive? Is the loop processing normally? Is the loop processing at a sub-optimal level?
Along with control responses in real time, Data Acquisition Processors can collect, reduce, summarize, and deliver summary data to the host system or a remote host to analyze and log. This does not take much computation, and Data Acquisition Processors can organize the data collection without much effect on the time-critical processing. Apply one DAP to multiple channels, with both control and process monitoring, and suddenly this begins to look like a significant cost improvement.
The software support for Data Acquisition Processor operation transparently includes networking support. Provided that you have communications channels available, a local processing node can transfer observed data to another location for recording, display, or interactive analysis.
Opportunistic Self-Testing
When using model-reference processing, unexpectedly large differences between what the model predicts and what the system actually does can indicate that something in the system has changed and needs attention.
There is not much that can be learned while a loop sits at a regulated level with no disturbance, but response to disturbances can provide a lot of good information. Processing can be configured to identify disturbance events and apply special analysis or actions when these occur.
Self Testing with Active Injection
While a controller holds a loop at a regulated level, it is possible for conditions in the system to change so that the control settings are no longer close to optimal, and stability may be compromised. Instead of waiting for disturbances to occur, and discovering too late that the system is in trouble, it is possible to inject a small disturbance deliberately. The injection of the disturbance signal into the control output is part of the time-critical processing, and must be integrated with the control loop update. Preparation of the injection signal can be coordinated by a lower priority process. Tests can be initiated periodically, under local control or at the request of higher control levels.
Where the system is sufficiently linear, an injection experiment can be a small disturbance of short duration, riding on top of the normal command level. Superimposition principles can be applied to isolate the disturbance's dynamic response from the static operating level, yielding information about open and closed loop characteristics. It is not necessary to remove the loop from active service for the purposes of this testing, and the testing activity can be controlled remotely.
Fail-Safe Monitoring and Recovery
If devices have failed, the process most likely needs an orderly shutdown. This kind of processing is typically enforced by a watchdog timer. Watchdog timers require very little processing, but when they do activate, their actions should be delivered at a high priority so that they are not unduly delayed by routine data processing.
To use watchdog monitoring, an independent task can watch a data source. The task requests a task_pause interval, using the DAPL system timer. While inactive, the task causes no scheduling delays. When the task next executes, it expects a non-zero code. If it sees one, all is well. It discards or clears the value, then activates the next task_pause interval. But if it fails to see the expected value, there must be a fault because otherwise a sending task would have provided a timely update.
Fault indications will sometimes require signal channels separate from ordinary data logging. They can be hardware devices such as digital outputs, but they can also be messages sent on a communication channel. In Data Acquisition Processor systems, these communication channels are "virtual" in the sense that the actual channels pass through the same physical network medium as any other data transfers.
Special Applications
This section mentions a few examples of applications with characteristics that make them particularly challenging, yet not terribly difficult for a Data Acquisition Processor.
Distributed Process Control
Related |
|---|
|
|
For large scale systems, it is important to subdivide the plant control objective into subproblems that can be addressed individually at local processing nodes. At these local processing nodes, the host software can further distribute the controller commands to multiple Data Acquisition Processors, each controlling a cluster of control channels.
It is unusual to find an architecture that can scale up to manage many control channels almost without limit. DAP processing actually gets more efficient when supporting arrays of samples rather than just one, yet each channel can have an independently adjusted configuration.
Hydraulics
Advanced hydraulics controls such as shaker table actuators drive high forces and respond over a frequency band much too wide for most ordinary controllers. The hydraulics systems have multiple servo stages, with nonlinear seal friction effects and a tendency to "ring" at around 5 to 15 Hz. Having the speed of a DAP system for sampling intervals of 1 millisecond or shorter is important for controlling these systems to full bandwidth. DAPs have the processing power to implement effective nonlinear control methods with multi-variable feedback.
Spatially Distributed Processes
Processes that are physically distributed, such as a rolling mill, drying conveyor, or distillation column, need arrays of sensors, and they must operate arrays of control devices. Decisions made early in the process propagate through the system and directly affect downline decisions, while requirements of downline processing can feed back to influence the process decisions at the start of the line. A control strategy will involve timed and coordinated actions. Data acquisition processors can provide effective measurement and management of multiple-channel data, the ability to schedule delayed actions, and packaging of local and global processing levels as different tasks.
Conclusions
We have seen that Data Acquisition Processors offer a fixed architecture solution that, while originally intended for classic data acquisition, also has application for advanced control systems. The fixed architecture means that the hardware development problems are reduced to establishing and configuring good interconnections. The software problem is reduced primarily to implementation of the control strategy, with concerns about the distribution of critical and non-critical processing times. The disadvantage is that you must cope with some unavoidable timing delays that are always present, along with some additional delays that your control processing will introduce. The net delay depends on the amount of data traffic and the competition for processor time. If processes can be merged to perform similar processing on multiple channels in parallel by one task, the delay times are not much different from the delay times for one channel alone.
Real-time response depends on being fast enough, not just on being fast. We have consistently discussed worst-case timing and meeting response deadlines under loaded operating conditions. Much of what is called "real time on your PC" is in fact concurrent processing with unspecified delays that can be a factor of 100 to 1000 longer than delays that you might typically experience on a Data Acquisition Processor. The Data Acquisition Processor is not a replacement for a workstation system but rather an extension to it. Even so, there are limits, and response times of Data Acquisition Processors are either sufficient for the application or they aren't. If you can't meet the real-time deadlines, the platform isn't the right one and you will have to consider a higher performance, higher complexity solution.
Data Acquisition Processors are best suited for those applications with special requirements for operating speed, multiple channels, novel or complex control algorithms, or supporting computations. What seems like an expensive solution at first, after applying the solution to multiple channels and incorporating important monitoring, testing, and logging features, can turn out to be a very competitive solution. The flexibility of the high-level development environment that does not require complex and specialized embedded system tools means a low-cost infrastructure for supporting development. Effort can be concentrated on the control problem, rather than on making the hardware components work together. Methods that might ordinarily be unsupportable can become feasible, opening a new range of opportunities for process improvements and long-term cost savings.
A Data Acquisition Processor can be an excellent platform for technologies and techniques beyond the scope of this paper. Visit the Control Page and consider the possibilities.
<< Previous 1 2 3 4 Next >> View as one page