Observer Control

The control methods described here as “Observer Control” falls in the broad category of “Optimal Control” and essentially aims at supplementing the functionality of a traditional controller with data derived from a model that observes what the process is doing. The observer, therefore, is generally a model of an aspect of the controlled process and can be used to derive additional information for the controller, in Figure 8.4.

The observer (in this case an ANN model) is used to determine a state variable Xout(k) of

the process, which is used as input to the process controller. The controller uses this value to determine the process control signal U(k) and in turn controls the process. This

has particular benefit because it allows extra information to be feed to the process controller that may otherwise not be possible to measure. Further, the observer can be used to determine optimal control parameters as determined by the programmer. For example, in the case of the optimal control implemented in [124], the observer is used to “punish” increments of the control signal to reduce control activity.

Figure 8.4: Normal process controller with ANN observer architecture

In the case of this investigation, the observer control architecture has particular promise for a number of reasons. Firstly, unlike inverse control, the process controller is separate from the ANN observer model. This provides the ability to focus the ANN model on the determination of relevant vehicle dynamics, and to use a traditional controller to control the process based on the observer findings. Secondly, the ANN observer can be trained off-line, which avoids all the problems of specialised inverse training without many of the drawbacks of general training. Finally, the ANN observer model is extremely flexible, and can be used in a wide variety of ways.

Of particular benefit for this investigation, the observer can be used to determine gain values of the process parameters at discrete operating conditions.

8.3.3.1 Gain Method

The procedure of determining the gain of process parameters was briefly introduced previously for inverse modeling. The goal of this procedure was to determine the

instantaneous amount that a change in a process input would affect a change in the desired process output. The gradient of this relationship is referred to as the input gain, and is the amount that the input must be multiplied by to produce the desired output.

This has the particular benefit in modeling vehicle dynamics because it enables the highly non-linear process to be broken into instantaneous linear relationships. By modeling vehicle dynamics process with ANN models and incorporating this method, it becomes possible to determine the short-term effects of altering any process input parameters. It is then possible to conduct hypothetical tests on the process by answering the question “if I alter this process input value individually, or in combination, what will the affect be on the process output?”

The ability to answer this question whenever it is posed is a very valuable tool. Further, the problem of determining vehicle stability, discussed above, can also be stated in a similar way; “if I reduce wheel slip, will the steering referenced longitudinal acceleration of the vehicle increase or decrease?”

Such a method can thus be used to answer this question. This can be accomplished by, for example, training the ANN observer to predict steering referenced longitudinal acceleration based on driven wheel slip and other state variables. The ANN observer can then be used to predict the current expected acceleration, and twice again for marginally reduced slip at each driven wheel. This entails significantly less computation than the method described for inverse model specialised training. If the acceleration decreases, the vehicle can be considered stable, with a “no control” signal sent to the process controller. Conversely, if the acceleration increases, a “control required” signal can be sent to the controller, as the vehicle can be considered unstable. In the later case the slip/acceleration gain could be supplied to facilitate determination of the control aim slip.

This method essentially can be considered to provide two pieces of information to the process controller, if the vehicle can be considered stable or unstable, and what the instantaneous slope of the slip/acceleration curve is. This is very important information for the controller and, if the model operates as expected, improved performance can be expected. There is a difficulty in this form of control, however, in that the observer does not explicitly define the aim slip. Instead the process control is expected to determine this value based on supplied data. This is a problem without a definite answer, and the

process controller must estimate the aim slip and then implement appropriate control. This requires additional control assumptions and controller tuning, which will be discussed later.

8.3.3.2 Curve Method

A way to avoid this problem was developed using a “curve” technique, which further expands the process of determining process gains. It was hypothesized that by varying a single ANN observer input through its entire range, an estimate of the effect on the process output could be observed. This estimate would be based on the “if all else is constant” principle, and would clearly lose a degree of accuracy when the current condition differed to a large extent from the test condition. Nonetheless, such a test would produce a non-linear estimate of how much changing an input parameter by any extent will affect the process output.

Applying this principle to the slip/acceleration example above would, thus, produce slip/acceleration curves for each driven wheel. As the theory suggests, the slip at the transition region of each graph (corresponding to maximum acceleration) can be used as an estimate of optimum slip. Such a method would provide all of the information to the process controller the gain method provides, as well as an ANN determination of aim slip. If such a method was found to work appropriately it would vastly increase the response of the controller as it would reduce the reliance on “close-loop” controller decisions which can take a long time to converge. It would also promote the controller’s ability to cope with process state changes that may make the aim slip vary to a large degree. It is noted, however, that the method used to produce a curve of the response of process outputs to any process input will require the ANN observer model to run consecutively a number of times (depending on the desired accuracy). This level of computation may slow the controller sample rate to unacceptable levels, or require hardware upgrades.

Many aspects of the investigation of each of these methods can be completed off-line, whereby performance predictions can be compared to actual values. However, it is noted that some of these methods require greater process feedback than others, and as such this would form an unfair comparison. Furthermore, one of the principle goals of this investigation it to actually implement ANN traction control, which requires real-time data acquisition and control at suitably high sample rates. Many aspects of the hardware

that was used have been presented in previous sections, but the investigation specific software that will enable ANN control and analysis is yet to be discussed.