Wireless channel propagation parameter estimation forms the foundation of channel sounding, estimation, modeling, and sensing. This paper introduces a deep learning approach for joint delay and Doppler estimation from frequency and time samples of a radio channel transfer function.

Our work estimates the 2D path parameters from a channel impulse response containing an unknown number of paths. Compared to existing deep learning-based methods, the parameters are not estimated via classification but in a quasi-grid-free manner. We employ a deterministic preprocessing scheme that incorporates a multichannel windowing to increase the estimator’s robustness and enables the use of a convolutional neural network (CNN) architecture. The proposed architecture then jointly estimates the number of paths along with the respective delay and Doppler shift parameters of the paths. Hence, it jointly solves the model order selection and parameter estimation task. We also integrate the CNN into an existing maximum-likelihood estimator framework for efficient initialization of a gradient-based iteration, to provide more accurate estimates.

In the analysis, we compare our approach to other methods in terms of estimate accuracy and model order error on synthetic data. Finally, we demonstrate its applicability to real-world measurement data from a anechoic bistatic RADAR emulation measurement.