Label Design-based ELM Network for Timing Synchronization in OFDM Systems with Nonlinear Distortion

Chaojin Qing, Shuhai Tang, Chuangui Rao, Qing Ye, Jiafan Wang, Chuan Huang

Abstract

Due to the nonlinear distortion in Orthogonal Frequency Division Multiplexing (OFDM) systems, timing synchronization (TS) performance is inevitably degraded at the receiver. To address this, we propose an Extreme Learning Machine (ELM)-based network featuring a novel learning label design. By exploiting the prior information of the inter-symbol interference (ISI)-free region, two types of learning labels—midpoint-based and ISI-free-based—are developed to refine symbol timing offset (STO) estimation. Our hybrid approach first utilizes a classic TS scheme for coarse timing metric (TM) capture, followed by an ELM network for refinement. Experimental results demonstrate significant improvements in TS performance and robust generalization across varying channel scenarios.

The Challenge of Nonlinear Distortion in OFDM

Orthogonal Frequency Division Multiplexing (OFDM) serves as the foundation for modern wireless communication standards, including 5G and Wi-Fi. Its efficiency, however, relies heavily on precise Timing Synchronization (TS)—the ability of the receiver to accurately identify the start of each symbol.

In practical hardware implementations, components such as High Power Amplifiers (HPAs) and Digital-to-Analog Converters (DACs) introduce nonlinear distortion. This distortion "muddies" the signal, causing traditional synchronization algorithms, such as the widely used Schmidl & Cox (SC) method, to suffer from significant performance degradation. While standard Deep Learning (DL) models have been explored as alternatives, they often face challenges regarding high computational costs and poor generalization across diverse channel conditions.

Proposed Hybrid ELM Architecture

To overcome these limitations, we propose a hybrid receiver architecture that synergizes classic signal processing with the efficiency of an Extreme Learning Machine (ELM).

System architecture of the proposed ELM-based timing synchronization scheme.

The architecture follows a three-step process:

  1. Timing Preprocessing: The receiver first employs a classic Schmidl-based synchronizer to generate a coarse "Timing Metric" (TM). This step acts as a "knowledge discovery" phase, providing the neural network with a structured starting point rather than requiring it to learn from raw, noisy data.
  2. ELM Refinement: The coarse TM is then fed into an ELM—a single-hidden-layer neural network. Unlike traditional deep networks that rely on slow back-propagation, only the output weights of an ELM are trained, making it exceptionally fast and suitable for real-time communication systems.
  3. STO Estimation: The refined output from the ELM provides a more accurate estimation of the Symbol Timing Offset (STO), specifically targeting the "safe" regions of the signal.

Innovative Label Design for Robust Synchronization

The core innovation of this work lies in the redesign of training labels. Standard "one-hot" labels, which target a single "perfect" sample point, are often too brittle in the presence of nonlinear distortion. We propose two alternative label designs that leverage the ISI-free region:

  • Midpoint-based label ($T_{mid}$): This label targets the exact center of the ISI-free region, providing a more stable target than the very first sample of a symbol.
  • ISI-free-based label ($T_{ISI-free}$): This label sets the target value to '1' for the entire duration of the ISI-free region. By creating a wider "tolerance zone," the network becomes significantly more robust to noise and hardware-induced distortion.

Performance Evaluation and Key Results

The proposed system was evaluated using simulated data with significant nonlinear distortion (Saleh model for HPA, targeting a 40% EVM). The results highlight a clear advantage for the label-based ELM approach.

Comparison of timing synchronization error probability across different methods.

Key findings include:

  • Superior Error Reduction: The "ISI-free-based" ELM significantly outperformed both the classic Schmidl & Cox baseline and end-to-end deep learning approaches ("TS_Learn"). At an SNR of 12dB, the error probability for the proposed method dropped below $0.5 \times 10^{-2}$.
  • The Power of Hybrid Learning: The failure of the "TS_Learn" model (which attempts to learn timing directly from raw signals) underscores the importance of the hybrid approach. Providing the network with a coarse TM from a classic algorithm is essential for achieving high accuracy.
  • Generalization and Robustness: The model demonstrated strong generalization capabilities. It maintained high performance even when testing conditions (such as channel length $L$ or distortion intensity $\eta$) differed from the training environment. While performance naturally decreases as the gap between training and testing conditions widens, the proposed labels consistently provided a more robust "safe zone" for synchronization.

Conclusion

By integrating classic signal processing with a specialized ELM framework and innovative label designs, we have developed a robust solution for timing synchronization in the face of nonlinear distortion. The transition from point-based targets to region-based targets (ISI-free) proves to be a critical step in making neural networks practical for real-world wireless hardware. Future work will focus on further enhancing label designs and exploring more advanced ELM architectures to bridge the generalization gap in even more extreme channel scenarios.

Summary updated

Abstract

Due to the nonlinear distortion in Orthogonal frequency division multiplexing (OFDM) systems, the timing synchronization (TS) performance is inevitably degraded at the receiver. To relieve this issue, an extreme learning machine (ELM)-based network with a novel learning label is proposed to the TS of OFDM system in our work and increases the possibility of symbol timing offset (STO) estimation residing in intersymbol interference (ISI)-free region. Especially, by exploiting the prior information of the ISI-free region, two types of learning labels are developed to facilitate the ELM-based TS network. With designed learning labels, a timing-processing by classic TS scheme is first executed to capture the coarse timing metric (TM) and then followed by an ELM network to refine the TM. According to experiments and analysis, our scheme shows its effectiveness in the improvement of TS performance and reveals its generalization performance in different training and testing channel scenarios.
BibTeX
@inproceedings{qing2021label,
  author = {Chaojin Qing and Shuhai Tang and Chuangui Rao and Qing Ye and Jiafan Wang and Chuan Huang},
  title = {{Label Design-based ELM Network for Timing Synchronization in OFDM Systems with Nonlinear Distortion}},
  booktitle = {2021},
  year = {2021},
  doi = {10.1109/vtc2021-fall52928.2021.9625189},
}
Figure