Non-linear Neural Network Equalizer for High-Speed SERDES Channels

The patent transforms the linear equalization parameters into non-linear parameters by introducing neural network weights and activation functions. The equalizer transitions from using simple linear coefficients to using learnable non-linear parameters that can adapt to channel characteristics, thereby improving signal detection accuracy while managing complexity through efficient neural network architectures.

The patent replaces the traditional mechanical/mathematical linear equalization system with a neural network-based system. Instead of using fixed linear filters, the system employs neural networks with non-linear activation functions that can learn and adapt to complex channel impairments, substituting a more sophisticated computational approach for the simpler linear method.

The patent segments the non-linear equalization task into multiple manageable components: feature extraction layers, non-linear transformation layers, and decision-making layers. By dividing the neural network into distinct functional blocks, the system achieves high measurement precision through complex non-linear processing while keeping the overall device complexity manageable through modular architecture.

The patent extends the equalization process from the traditional one-dimensional linear filtering domain to a higher-dimensional non-linear feature space. By introducing multiple layers of non-linear transformations, the system maps input signals into a more complex feature space where channel impairments can be more effectively separated, improving detection accuracy while the layered structure manages computational complexity.

The patent applies preliminary non-linear transformations to the received signal before final detection. By pre-processing the signal through neural network layers that learn to compensate for channel non-linearities, the system prepares the signal in advance for accurate detection, enabling close data levels to be distinguished despite channel impairments and thus maintaining high data transmission rates.

The patent incorporates feedback mechanisms where the neural network learns from detection errors and adjusts its parameters accordingly. Through adaptive training and refinement, the system improves its ability to distinguish close data levels over time, maintaining high measurement precision even for high-order signaling schemes with closely spaced constellation points.