A deep-learning-based time of arrival estimation using kernel sparse encoding scheme.

• A deep-learning-based time of arrival (TOA) estimation method is proposed with a new sparse encoding scheme based on the sinc-kernel function. This scheme can effectively reduce the quantization errors and facilitate the CNN network to explore more off-grid features. • The proposed method designs the CNN framework with residual block and SE block, to improve the estimation efficiency of different off-grid distributions. • The superior performance of the proposed method is shown by comparing with other CNN based estimation methods and classical off-grid estimation methods. This letter proposes a deep-learning-based method for time of arrival (TOA) estimation with a new sparse encoding scheme, aiming to solve the problems caused by quantization errors and off-grid effects. The proposed method utilizes a convolutional neural network (CNN) to learn the relationship between the training signals and the TOA parameters encoded by the proposed kernel sparse representation. It encodes true TOA values into an approximate sparse representation by the sinc-kernel function. This scheme can reserve more off-grid features of the estimated TOA parameters to enhance the optimization efficiency of the CNN-based network, so it can effectively minimize the quantization errors and improve the estimation performance of different off-grid distributions. Numerical simulations show that the proposed encoding model is superior to the existing encoding models under the same CNN framework, and the proposed method achieves better TOA estimation accuracy than the competitive end-to-end CNN-based estimation methods and outperforms in accuracy and speed compared with other existing off-grid estimation methods. [ABSTRACT FROM AUTHOR]