This implementation of Neural Network (NN) programmed in Python + Labview shows the feasibility of Artificial Intelligence (AI) assisted Decoder, particularly in our OWC system.
In this post, some idea of NN is disclosed, and some piece of implementation results are provided for explanation.
In general, NN is very helpful for our OWC if it is properly used. In the PHY layer of OWC, the following use-cases of NN are promising to note:
- NN can be applied for reliable light sources detection and tracking. Later, we call massive-Tx detection a compressed-sensing platform.
- NN can bring amazing performance for the decoder by taking its nonlinear classification potential.
In the upper PHY or MAC layer of OWC, the following use-cases of NN are promising to note:
- NN is promising for any resource optimization. Resource here may be the communication bandwidth, the number of connectivity under mobility, or the handover mechanism.
- NN is also promising for allocating resources for several constraints.
Implementing NN-assisted PHY seems to be a smart start at this moment for a reliable communication platform in OWC.
Overall System Architecture
A simple architecture for OWC system is as Figure 1.
- The Tx has an encoder and some functional lighting feature blocks.
- The Rx has a decoder and some pre-processing blocks for light sources detection/tracking plus some post-processing blocks if needed.
NN-Assisted Detection Platform
Figure 2 draws the idea for applying CNN to detect the light sources. We can create one unique characteristic to intended light sources, such as a blinking frequency, to support the detection. In this way, our Tx has its unique characteristic compared to other noise sources, so that CNN can learn to detect and track the Tx(s) efficiently.
Here is an expected performance. Surely, the performance of NN detector should be better than typical detection based on computer vision.
The channel condition directly impacts to the communication. If the decoder is trained based on particular channel conditions, it should perform better.
A basic example of NN-assisted decoder is the XOR classifier. The nonlinear classification characteristic in NN is promising.
A great challenge for the decoder may come if the image is blurry. In this case, the intensity of neighbor LED-sources may mix together, and a typical decoder will not work. The following figure shows a simulation for an illustrative explain. Luckily, NN-assisted decoder can learn the linear mixing characteristic of our LEDs, thus, it is a rescue solution.
- Nowadays, NN is applicable everywhere, every research field.
- In OWC, NN is also promising.