ELM-based Timing Synchronization for OFDM Systems by Exploiting Computer-aided Training Strategy

Due to the implementation bottleneck of training data collection in realistic wireless communications systems, supervised learning-based timing synchronization (TS) is challenged by the incompleteness of training data. To tackle this bottleneck, we extend the computer-aided approach, with which the local device can generate the training data instead of generating learning labels from the received samples collected in realistic systems, and then construct an extreme learning machine (ELM)-based TS network in orthogonal frequency division multiplexing (OFDM) systems. Specifically, by leveraging the rough information of channel impulse responses (CIRs), i.e., root-mean-square (r.m.s) delay, we propose the loose constraint-based and flexible constraint-based training strategies for the learning-label design against the maximum multi-path delay. The underlying mechanism is to improve the completeness of multi-path delays that may appear in the realistic wireless channels and thus increase the statistical efficiency of the designed TS learner. By this means, the proposed ELM-based TS network can alleviate the degradation of generalization performance. Numerical results reveal the robustness and generalization of the proposed scheme against varying parameters.

Lightweight 1-D CNN-based Timing Synchronization for OFDM Systems with CIR Uncertainty

In this letter, a lightweight one-dimensional convolutional neural network (1-D CNN)-based timing synchronization (TS) method is proposed to reduce the computational complexity and processing delay and hold the timing accuracy in orthogonal frequency division multiplexing (OFDM) systems. Specifically, the TS task is first transformed into a deep learning (DL)-based classification task, and then three iterations of the compressed sensing (CS)-based TS strategy are simplified to form a lightweight network, whose CNN layers are specially designed to highlight the classification features. Besides, to enhance the generalization performance of the proposed method against the channel impulse responses (CIR) uncertainty, the relaxed restriction for propagation delay is exploited to augment the completeness of training data. Numerical results reflect that the proposed 1-D CNN-based TS method effectively improves the TS accuracy, reduces the computational complexity and processing delay, and possesses a good generalization performance against the CIR uncertainty. The source codes of the proposed method are available at https://github.com/qingchj851/CNNTS.