High-Relative-Bandwidth Multiband LNA Distortion Elimination With Neural Network

This article presents an innovative approach that harnesses neural networks (NN) to eliminate non-linear distortion in low-noise amplifier (LNA) within multi-channel direct sampling receivers (MDSR). Current mainstream methods, such as those relying on memory polynomial or Volterra models, face challenges in effectively addressing the demand for LNA linearization modeling, particularly in scenarios with high fractional bandwidth and stochastic inputs. The proposed method incorporates two key technologies to support NN. First, using input signals with specific frequency properties to derive ground truth values by isolating distortion components from the LNA output signals in the frequency domain, simplifies the process of acquiring training samples and enhances accuracy. Additionally, by utilizing mathematical characteristics of LNA output signals, such as instantaneous rate of change, magnitude, and non-uniformly sampled memory points, it performs feature engineering to simultaneously reduce the complexity of the NN and enhance its generalization capabilities. Evaluation with the LNA (ZFL-500LN+) demonstrates outstanding performance in suppressing multiple harmonics and inter-modulation components, which approaches the quantization noise floor of the analog-to-digital converter, especially harmonic reduction of up to 46 dB in the worst distorted channel. These results show the potential of this method to enhance the performance of MDSR.