<aside> đź’ˇ
An implementation inspired by Dr Wang’s 2021 paper (https://doi.org/10.1145/3491315.349133). The project link can be found here.
</aside>
Classically, data transferred between transmitters and receivers rides on carrier waves that are usually from the electromagnetic spectrum. Unfortunately, a change of medium especially one that is denser (water) provides significant challenges to these signals that suffer from attenuation. This makes them ineffective over long distances. Taking inspiration from nature, scientists and researchers found a novel way to communicate wirelessly underwater; sound.
Underwater acoustic communication leverages sound waves to “talk” to underwater vehicles, sonar systems and under-ice sensors. But this comes with its own set of challenges. Underwater environment for signals is subject to highly variable conditions caused by temperature, salinity, reflection, refraction, noise, multipath propagation and Doppler effect just to name a few.
To build robust and resilient communication systems, we need to create accurate models of these acoustic channels. This is where deep learning comes in specifically Variational Auto Encoders (VAEs), which can be leveraged to model and simulate Channel Impulse Responses (CIRs).
The Channel Impulse Response (CIR) represents how a signal interacts with the environment as it travels between a transmitter and receiver.
Essentially, Channel Impulse Response refers to a transmitted signal’s interaction with the medium as its travelling through to the receiver. In UWA, the acoustic signal takes multiple paths besides the clear direct path which is known as the line-of-sight (LoS) signal. It bounces off of the sea surface, scatters through thermoclines, reflects from the sea floor and thus, creates multiple delayed and distorted copies of the transmitted signal at the receiver. This is called multipath propagation.
For UWA and other ranging applications, it’s essential to isolate the direct signal (the LoS signal) that is received at the receiver’s end. This is, usually, the earliest signal with the highest strength represented by a sharp peak in a CIR. Isolating this peak helps in time-of-arrival estimations, robust communication and positioning.
Below is an example of a CIR from a field experiment in Lake Tuscaloosa in July 2019.
**Figure 1:** A sample CIR from the Lake Tuscaloosa field experiment
As the figure shows, there is a impulse with a strong impulse in the early sample and then a smaller peak followed by even smaller noisier peaks. The sharp peak shows the direct path that the transmitted signal takes. The peak that follows shows the signal that has been received from an indirect path.
Modelling CIRs is crucial to understand how the signal moves in the underwater environment and thus has implications for underwater communication. A well-modeled CIR captures various effects introduced by the communication channel such as distortions, multipath effects or delays. This helps in simulating realistic communication conditions and in turn, helps make us better choices for system design.
To simulate these CIRs through any means, we need to have real world data as a reference. The Lake Tuscaloosa field experiment conducted in July 2019 provides just that (link).
In this experiment, eight hydrophones were recording transmissions from a single acoustic transmitter at the center of the lake. The signal was transmitted at a center frequency of 28kHz. The recordings were made in two modulation schemes; binary phase-shift keying (BPSK) and orthogonal frequency-division multiplexing (OFDM).
One of the dataset’s most valuable features is that it contains processed impulse responses, which makes it perfect for this project.
Each CIR from the dataset contains 2400 taps that represent discrete-time impulse response of the channel. There are 187 CIRs collected by each hydrophone. The resulting CIR is a complex-valued signal meaning it has a real and an imaginary part.