SupportNet: a Deep Learning Based Channel Equalization Network for Multi-type Multipath Fading

High-speed moving receivers generate Doppler shift superimposed on multipath effects to produce serious self-interference in the signal, direct channel equalization is more difficult, often requiring channel estimation, although neural networks can perform channel estimation and channel equalization, the neural network training process requires both the corresponding channel estimation and channel equalization results of the two labels as a loss function. It is more difficult to take labels for channel estimation in realistic scenarios, there are small errors in channel estimation by various methods, and the use of a large number of channel estimation labels causes an increase in data cost. This paper proposes a channel equalisation model called SupportNet, which simulates both channel estimation and channel equalisation processes by inducing a sub-network into a model collapse state so that a part of the network acts like channel estimation without using channel estimation labels, allowing features to be separated and processed separately, and using the channel estimation results for channel equalisation to reduce BER. The property of neural networks that rely on gradient descent for training to produce pattern collapse allows the network to separate features without the need to add labels to each feature. The experimental results show that the homogeneous network can effectively reduce the impact caused by time-selective fading under the fast-fading channel generated by the physical layer emulation parameters of three mobile environment provided by the IEEE 802.11p standard, resulting in a lower BER.