On-line learning applied to spiking neural network for antilock braking systems

Abstract

Computationally replicating the behaviour of the cerebral cortex to perform the control tasks of daily life in a human being is a challenge today. First, … Finally, a suitable learning model that allows adapting neural network response to changing conditions in the environment is also required. Spiking Neural Networks (SNN) are currently the closest approximation to biological neural networks. SNNs make use of temporal spike trains to deal with inputs and outputs, thus allowing a faster and more complex computation. In this paper, a controller based on an SNN is proposed to perform the control of an anti-lock braking system (ABS) in vehicles. To this end, two neural networks are used to regulate the braking force. The first one is devoted to estimating the optimal slip while the second one is in charge of setting the optimal braking pressure. The latter resembles biological reflex arcs to ensure stability during operation. This neural structure is used to control the fast regulation cycles that occur during ABS operation. Furthermore, an algorithm has been developed to train the network while driving. On-line learning is proposed to update the response of the controller. Hence, to cope with real conditions, a control algorithm based on neural networks that learn by making use of neural plasticity, similar to what occurs in biological systems, has been implemented. Neural connections are modulated using Spike-Timing-Dependent Plasticity (STDP) by means of a supervised learning structure using the slip error as input. Road-type detection has been included in the same neural structure. To validate and to evaluate the performance of the proposed algorithm, simulations as well as experiments in a real vehicle were carried out. The algorithm proved to be able to adapt to changes in adhesion conditions rapidly. This way, the capability of spiking neural networks to perform the full control logic of the ABS has been verified.