Information Technology Reference
In-Depth Information
board subsystem actuates suitable countermeasures, such as emergency breaking. The
evaluation of the above-mentioned distance values requires the knowledge of the breaking
parameters and of the actual train speed: a correct estimate of this last variable even in poor
adhesion conditions (i.e. when one or more train wheels are sliding on the rails and, thus,
the axle angular velocity is not proportional to the train speed) is crucial.
Fig. 4. Working principle of an ATP
Allotta et al. (Allotta et al 2001, Allotta et al 2002) developed a series of algorithms for
estimating the actual train speed on the basis of the information collected concerning two
axles of the locomotor. A first set of such algorithms have been developed according to
expert personnel specifications and following the traditional “crisp” reasoning, which
exploits different simple deterministic formulas for calculating the train speed depending on
the condition of adhesion of the wheels to the rails. In fact, among a huge number of state
variables that are considered in the procedure, there are two binary variables indicating the
adhesion condition of each axle. The technical personnel of the train society formalised the
reasoning that leads the human operators to a correct determination of the adhesion
conditions. Then two identical fuzzy systems have been developed, which take two inputs
each, namely the difference between the velocities of the two controlled axles and the
acceleration of the axle whose adhesion condition is estimated, and return the degree of
adhesion of one axle. The design of two fuzzy systems have been refined by means of a
training procedure exploiting a great quantity of the available data and, finally, they have
been merely substituted to the old crisp algorithm for adhesion condition estimation, by
leaving the rest of the speed estimation procedure unchanged. As an alternative, the
standard rule-based system merely implementing the human operators' reasoning has been
implemented and its own parameters (such as thresholds) have been tuned by means of
Genetic Algorithms (GA) exploiting the available experimental data and adopting as fitness
function to minimise the error between the actual and estimated train speed. A second set of
algorithms that have been tested for this application perform a direct estimate of the train
velocity taking as inputs some of the available state variables (in particular axles velocities,
accelerations and acceleration variations). Both neural networks and fuzzy inference
systems have been tested to this purpose.
From a comparison among all the tested approaches it turned out that the algorithms purely based on
AI techniques (and, in particular, the neural network (Colla et al. 2003)) outperform the rule-based
ones and have also a simpler structure. However, these systems also present the following non
negligible disadvantages: i) a difficult physical interpretation of fuzzy rules or of the neural network;
