Temporal Convolutional Autoencoder for Interference Mitigation in FMCW Radar Altimeters

TL;DR

This work tackles RF interference challenges in FMCW radar altimeters by introducing a Temporal Convolutional Network autoencoder that operates directly on raw time-domain IQ signals. The method denoises the received signal to preserve the beat features essential for accurate altitude estimation, outperforming LMS adaptive filtering in simulations and over-the-air tests. Key contributions include end-to-end evaluation in a realistic altimeter pipeline, a compact ~920k-parameter architecture with dilated 1D convolutions, and OTA validation demonstrating real-time potential. The results show robust performance across tone, QPSK, and 5G-like interference, enabling reliable altitude estimates even under severe interference, with significant RMSE improvements and practical hardware feasibility.

Abstract

We investigate the end-to-end altitude estimation performance of a convolutional autoencoder-based interference mitigation approach for frequency-modulated continuous-wave (FMCW) radar altimeters. Specifically, we show that a Temporal Convolutional Network (TCN) autoencoder effectively exploits temporal correlations in the received signal, providing superior interference suppression compared to a Least Mean Squares (LMS) adaptive filter. Unlike existing approaches, the present method operates directly on the received FMCW signal. Additionally, we identify key challenges in applying deep learning to wideband FMCW interference mitigation and outline directions for future research to enhance real-time feasibility and generalization to arbitrary interference conditions.

Figures (10)

• Figure 1: High-level diagram of denoising autoencoder operation on received IQ data.
• Figure 2: Diagram of TCN Autoencoder Model Architecture. The model uses dilated convolutions to expand the receptive field without reducing temporal resolution. Dimensionality reduction is achieved via the latent bottleneck, transpose convolution stride is used in the decoder to efficiently upsample the latent representation to the original signal length, a rectified linear unit function is used to introduce nonlinearity, and dropout for regularization.
• Figure 3: Model Comparison. Comparison of STFTs for the input signal, reference signal, and the output of each of the three models. Each row shows a different evaluation example from the dataset, combining QPSK and tone interference.
• Figure 4: Range Profile Evaluation Scheme
• Figure 5: Conv1D Dilation Visualized bai2018empiricalevaluationgenericconvolutional