While these results provide compelling evidence towards the ANN application, they are so far applicable to one surface type. It has been argued previously that these results can be extended to other surfaces due to the generic nature of ANN models, and forms an important potential characteristic for ANN chassis control. As such, this section presents the performance of the generic DANNSCO on-line controller on an unsealed surface.
To complete the ANN evaluation on the unsealed surface, the DANNSCO model had to be re-trained for the new surface. In this way for instance, the ANN stability controller and
the ANN surface identification method could be utilised in combination to activate different DANNSCO controllers for different surfaces. Furthermore, this method of re- training for specific roads would also be directly applicable for racing applications and provide some insight into the possibilities of DANNSCO adaptive learning. It was decided that the amount of the ANN training data that was acquired should be small. By testing in this way it would be possible to not only evaluate the performance of the DANNSCO model, but to gain some insight into the potential of racecar and adaptive learning applications (which would both require good ANN generalisations to be made with minimal training data).
With this in mind, the ANN training data was acquired by driving the test vehicle along a 4.4km section of the test track at a variety of safe wheel slips. The newly trained ANN model was transported in the DANNSCO controller, and its abilities in unsealed pavement traction control evaluated by driving along the same stretch of road. The resultant logged data for a typical stretch of this road is shown in Figure 8.67.
Figure 8.67: DANNSCO performance for unsealed road section – 0 to 600m
The ANN predicted aim slips are much higher than for the wet cement case of the previous tests. Further, the throttle position seems to have a large effect on the aim slip (as in the previous case) even though it is not used to determine this value. This is a curious effect, and is either based on transient abilities of tyres to produce increased acceleration when quickly ramped from low slip to high slip (which may or may not be true), or problems with the ANN model in predicting aim slips from one extreme to another (which has been discussed previously). Of most importance, however, maximum longitudinal acceleration is realised only when the actual slip is similar to the ANN prediction of aim slip. Lower slips, in the order of the optimum values for wet cement, simply do not produce high
acceleration. This clearly shows that the ANN model has been able to model the new surface reasonably well, and with much greater performance than if the METC or DEMTC were used. The reason for this, and the properties of the surface, can be seen in Figure 8.68 to Figure 8.71.
Figure 8.68: Average slip and tyre long. accel. DANNSCO comparison for 2.5 to
7.5m/s on unsealed surface
Figure 8.69: Average slip and tyre long. accel. DANNSCO comparison for 7.5 to
12.5m/s on unsealed surface
Figure 8.70: Average slip and tyre long. accel. DANNSCO comparison for 12.5 to
17.5m/s on unsealed surface
Figure 8.71: Average slip and tyre long. accel. DANNSCO comparison for 17.5 to
22.5m/s on unsealed surface
The most obvious feature of these graphs is that the logged data provides no indication of the critical slip. In fact, the transition zone appears to extend indefinitely with no point where increased slip will result in decreased acceleration. This is typical of unsealed surfaces, which exhibit similar properties to the “loose sandy soil” slip/coefficient of friction curve provided in the theory (Figure 2.10). It is also why traditional traction controllers have a renowned lack of ability on dirt surfaces. As such, the ANN model
should be able to identify this, and place the aim slip at a high value. The training data, which only consists of comparably low slip values makes this difficult, however, because these conditions are not represented well. Nonetheless, the abilities of ANN to learn from incomplete data come into play here, with the ANN predictions of aim slips clearly exceeding the slips within the training set. This produces a significant level of uncertainty into the determination of aim slips, which is evident in the spread of optimum slips in the 6 to 12m/s range, but is not considered a problem because these values fall outside the actual driving conditions.
Even so, Figure 8.67 still shows that situations arise where the ANN predicts there is a relatively low optimal slip (of around 5m/s in this case). It is anticipated that this is because at high slips the tyre actually “digs” through the loose gravel and makes contact with the compacted dirt, which is a feature the ANN prediction would not be able to recognise until it started to happen. In this case, a critical slip would exist, and the ANN seems able to identify it. It cannot identify it at low slips however, because the ANN curve optimisation logic is based on the “if everything else is constant” principle (which would assume a loose sandy soil).