Driver-assistance systems are essential to reduce accidents and improve efficiency through efficient convoying. To improve their effectiveness, it is important that vehicles are able to communicate with each other, informing each other of their intentions, and warning each other of obstacles or cross traffic that only some of the vehicles might see. However, since such communication must be done wirelessly, reliability and latency of the communication can be an issue. A METRANS funded research project lead by USC researcher Andreas Molisch aimed to improve these aspects of communication by means of improved radio channel prediction.

APPROACH

A key challenge for wireless vehicle-to-vehicle (V2V) communication is that resource allocation for communication must rely on the current/future propagation channel state, but vehicles only have past measurements. Therefore, it is important to find effective methods for predicting what the channel will look like in the future. Traditional methods using simplified models and classical tracking/extrapolation perform poorly in real-world environments. This motivates the use of Machine Learning (ML), which handles complex data well but faces challenges like limited channel measurement and mismatched neural network structures for V2V channels, therefore posing a need to find new ML-based prediction methods that alleviate these problems.

This study focused on the holistic prediction of V2V channels through deep predictive learning. The PI introduced a novel network, SE-LSTM, which combines a Squeeze-and-Excitation (SE) module and an attention mechanism within a long-short-term memory architecture, managing dependencies within and between sequences. It is tailored to simultaneously model CSI (channel state information) sequences across Time, Doppler, Delay, Angular, and Geometry domains. Enhancing the model's flexibility for various geometrical setups, researchers integrate the meta Pseudo-Label learning method, substantially boosting the generalization ability of their approach across diverse scenarios.

RESULTS

The new approach is capable of predicting the channels in future timesteps considerably more accurately than existing methods. The importance of this lies in the ability to make a communication system more robust and efficient by means of this prediction. In particular, resource allocation must always be done for future timesteps, i.e., we need to decide now which users get resources for data transmission several timesteps (typically tens ofmilliseconds) from now. Only by knowing what the channel will be at that time can we have efficient and reliable communication.

IMPACT

Reliable V2V communication will be a central challenge in a future with driverless vehicles. This research addresses a gap in existing communication technology, by developing methods for improved channel prediction, allowing more reliable V2V communication. This allows a more efficient, safer transportation system. For example, notifying a car driving behind oneself of an imminent emergency braking can avoid a rear-end collision, but of course such information has to be given in a timely manner; this is helped by our channel prediction, because the algorithm more correctly anticipates and corrects conditions that make communication more difficult. Furthermore, these improved channel predictions also help in situations where there are a large number of communicating vehicles, e.g., in a traffic jam, by allowing more efficient communication and thus leaving more spectral resources for other cars. This will allow more effective use of both driver assistance systems and future connected autonomous vehicles.