Novel Cascade CNN Algorithm for UWB Signal Denoising, Compressing, and ToA Estimation

In this paper, for ultra-wide-band (UWB) localization systems, we present a novel deep learning paradigm composed of two cascaded convolutional neural networks (CNNs) to 1) denoise the UWB channel impulse response (CIR); and 2) estimate the time of arrival (ToA) of the received signal. Unlike the traditional rule-based algorithms, the proposed model learns suitable information for denoising and ToA estimation automatically during the training phase. The proposed method, in the first step, compresses the noisy input signal and reconstructs its denoised version. To this end, it exploits sparsity and spatial correlation phenomena of the UWB CIR. By compressing the signal, the computational complexity of the model is reduced. In the second step, the proposed model estimates ToA of the received signal regardless of the channel environment. We investigate the effectiveness of hyper-parameters, such as the number of filters in each convolutional layer, number of the layers, and the size of the kernels in the proposed model by comparing the root mean square error (RMSE) performance evaluated against standard IEEE UWB 802.15.4a CIR model in different channel environments.