Contexto Físico

The experiments and control structures presented in the previous chapter did not necessarily provide demonstrations of the best possible solution to the control of the an active laser-scanning inspection system. Rather they were implanted in order to test the mechanisms and ideas detailed in chapter 7 which in turn resulted from the demonstration of behaviour based control reported and discussed in chapters 5 and 6. Nevertheless, it has been found that the solutions arrived at are still of a usefully robust nature and can provide a serviceable system for use in a real industrial situation. There is of course much remaining scope for improvement and further development, but first it is necessary to discuss some important factors that affected the experiments and implementations of this work.

9.1.1. Factors Affecting Performance

Test Surfaces

One of the most important factors affecting the development and results of this work was the range and nature of sample test surfaces available for experimentation. In chapter 5.1.1 tlie provision of target inspection surfaces was detailed, and again in chapter 8 the variety of samples was shown (in figure 8.2). While only a subset of these was used for the reported experiments, there were in fact several other desirable surface types, each with varying characteristics selected to provide some significant obstacle to the laser-scanner control system. The surface types that appear in the experiments reported in this thesis were chosen carefully in order to provide a manageable and representative range of situations and characteristics.

The surface samples were also somewhat limited in their size and extent. The pattern repeated every 490mm (as the surface rotated beneath the laser-scanner). Although the surface rotation unit had variable speed, effectively providing a varying resolution of feature size, it was found that for most work the slowest speed available should be used in order to ensure that sufficient feature variation was available within the imposed surface-length-time constraints.

Low-Power Laser

Another factor that affected the extent of the experiments was the comparatively low power of the laser light source in the sensor. It was found that under many circumstances the laser light would be saturated by ambient light. This had several effects, both good and bad, on the experiments. For example, the low power of the laser meant that the photomultiplier devices tended to be mnning at the limits of their operational range. This was useful in that it increased the tendency of the outputs to go into the foldback states thereby, increasing the requirement for active control. A disadvantage was that it limited the variety and types of inspection surfaces that could be used.

Sensor Input Space

For experimentation with BBAI control techniques, a suitable balance is required in terms sensory complexity and practical manageability. Too simple a system leads to an insufficient richness of problem, experiments become contrived and the outcome is often inconclusive, while complex systems can tend to lose the focus of the experiment. At first sight we thought that the laser scanner test-bed might fall into the former category of over-simplicity since the range of available actuation of the system is, in the final analysis, only a one-dimensional problem space. But the non-linear nature of the control problem, in conjunction with the rich and critical input space, leads to what we can make an argument for calling an almost optimal level of complexity given the time constraints on this project work. Each photomultiplier controller input space consisted of an array of up to 1000 pixels each with a value in the range 0 - 255 representing the intensity of detected laser light from the inspected surface. For the task in hand, this provided an interesting and yet manageable set of sensors and was probably one of the main contributing factors to the focus and timely completion of this work.

MuJii-Channel Seiutur System

The experiments reported here illustrate the use of only one sensor channel while the test-rig provided two. Work has been conducted on a two-channel system which was implemented successfully by simply copying the first channel's software to the second and providing a set of high-level user interface AFSMs as implemented in chapter 5. It was found that the adaptive AFSM processes were able to adjust for the different characteristics of the second photomultiplier channel (no two systems are alike) - which, in the case of the original reactive AFSM, would have required considerable hand-tuning of critical parameters. This illustrates a significant payoff for the cost of developing the first adaptive enhanced AFSM control stmcture. It is possible to envisage the transfer of these software modules to as many different photomultiplier channels as necessary, and in this aspect the results are certainly encouraging. However, for the purposes of continued enhancement of this application, the second channel was not used extensively due mainly to the bottleneck of the top server-level transputer and its inability to handle the necessary storage of large amounts of mntime data conceming the trace of the system's status.

Industrial Situation

The implementation of an industrial laser-scanning inspection system test-bed for experiments in behaviour based control provided an unusual problem domain. As such, the industrial framework contributed towards the shaping of the solutions reported in this thesis. The nature of the control of the sensor required a system that was able to analyse a signal that reflected characteristics of the physical environment and then to make changes to maintain the quality of the signal. The problem of differentiating between changes resulting from external influences and those from the controller itself was real and in fact characterised in a straightforward way one of

the fundamental problems that are the focus of much of the research effort on the control of autonomous systems.

9.1.2. Further Application-Specific Development and Enhancement

Utilising the Temporal Dimension

Take-sample Scan-data Take-sample Net-data Pedestal level EHT-level EHT automatic Scan Signal pixel data ' Scan Sample Edge Monitor BallPark Rigtit-edge Left-edge Pedestal Level

Set Automatic mode

E H T - S t e p s iz e l_ ( ^

EHT-level

EHT Output

Photomultilpier EHT voltage level

^ Initiate scan sample

Figure 9.1. A third subsumption layer using temporal information to inhibit the lower levels of EHT voltage stepping during "recognised" periods of surface disappearance.

Further development of the photomultiplier control strategy for a single channel may be achieved by adding a third subsumption level onto the existing control structure as shown in figure 9.1. The idea is that this enhanced AFSM learns to override and inhibit the output of EHT voltage step commands during periods when there is no surface present to inspect. This would prevent the hunting of EHT voltage that is characteristic of this transition period as the controller continuously ramps up the EHT voltage in a search for any detected laser light. When the surface actually returns, the EHT voltage setting is typically far from the right value, and a period or readjustment is needed while the signal set-point is re-acquired. By learning surface characteristics over time, this layer should be able to recognise and predict regularly-occurring breaks and hold the EHT voltage at a fixed level until the surface reappears. Obviously a strategy is required both for the occasions when mistakes are made and when the surface does not reappear as expected. After this the hunting strategy would take over again. The implementation of this expansion into the domain of temporal activity could be implemented in a number of ways, either through the simple maintenance of a shifting array of scan-level values that is updated at each characteristic time step (see [Pebody94] for an implementation of a TimePixels AFSM) or through the use of recurrent or feedback artificial neural networks such as Hopfield networks and

bidirectional associative memory nets (see [Wasserman89] for an overview and further references). The incorporation of other mechanisms is discussed in more detail in section 9.2.2 below.

Inter-Channel Communication

Finally, it is worth suggesting here that a large area of interest might be opened up through the idea of experimenting with inter-channel communication. As one possibility, the agent-like properties at the level of the photomultiplier control subsystems might be utilised in order that they "compare notes" on activities. For example, if one channel is spending large amounts of time in a foldback state it may benefit from "looking" at the state of the other channel. This may result in either the recognition of some component failure or a change in control strategy to compensate for the new situation. A possible scheme is illustrated in figure 9.2. The mechanisms of associative learning, including a temporal aspect, could prove useful here in an arrangement similar to the original work that was reported in [Pfeifer & Verschure 91J (although in this case the temporal association was a result of the physical properties of sensor devices). This communicaiion layer may well serve to bring the channels of the system together to some extent as a higher-level agent system, presenting a combined front to other parts of the system.

C h annel 3 S ta te C h ann el 2 S tate S c a n level EHT level C h annel 1 State N*arLbi - BallPark r z l l ü S c a n level EHT level EHT Output NaaiUn - , BallPark iStato Monitor S c a n level EHT level ip EHT Output NMitln —1 BallPark EHT Output C h a n n e l 1 C h a n n e l 2 C h a n n e l 3

Figure 9.2. Showing a possible inter-channel communication between photomultiplier controllers.