Noncoherent demodulation is an attractive choice for many wireless communication systems. It requires minimal protocol overhead for carrier synchronization, and it is robust to radio impairments commonly found in low-cost transceivers. Machine learning techniques, such as neural networks and deep learning, offer additional benefits for these systems. Practical communication systems often include nonlinearities, non-stationarity, and non-Gaussian noise, which complicate mathematical derivation of optimum demodulators. Learning approaches can optimize demodulator performance directly from simulated or measured radio data, which is often plentiful in the design and verification of today's integrated transceivers. This paper examines several candidate neural network topologies for use in noncoherent demodulation and provides a mathematical framework for their comparison. Each is based on a complex-valued feature detection layer, which may be characterized as coherent or noncoherent, followed by one or more real-valued classification layers. Backpropagation equations for the noncoherent feature layer include a synchronization term that facilitates training with noncoherent input data. The coherent layer does not synchronize training data, however a noncoherent demodulator can still be constructed by increasing the coherent layer capacity and adding a max pooling layer to marginalize the unknown signal phase. A frequency classification example highlights the differences between the topologies and confirms that optimum noncoherent demodulation can be learned in the presence of AWGN and random phase offsets. The topologies considered here are suitable for noncoherent demodulation of power-efficient modulations such as FSK and ASK, which are typical in today's short-range wireless communication systems. It is hoped that such topologies will lead to a future common architecture that can support the wide range of modulation formats in this space.

中文翻译：

---

非相干解调的复值神经网络

对于许多无线通信系统，非相干解调是一个有吸引力的选择。它需要最小的协议开销来实现载波同步，并且对低成本收发器中常见的无线电损害具有鲁棒性。机器学习技术（例如神经网络和深度学习）为这些系统提供了其他好处。实际的通信系统通常包含非线性，非平稳性和非高斯噪声，这会使最佳解调器的数学推导复杂化。学习方法可以直接从模拟或测量的无线电数据直接优化解调器性能，这在当今集成收发器的设计和验证中通常非常丰富。本文研究了用于非相干解调的几种候选神经网络拓扑，并为它们的比较提供了数学框架。每一层均基于复数值特征检测层，该特征检测层可以被描述为相干或不相干，然后是一个或多个实值分类层。非相干特征层的反向传播方程式包含一个同步项，可促进非相干输入数据的训练。相干层不同步训练数据，但是仍然可以通过增加相干层容量并添加最大池化层来边缘化未知信号相位来构造非相干解调器。频率分类示例突出显示了拓扑之间的差异，并确认可以在存在AWGN和随机相位偏移的情况下学习最佳非相干解调。这里考虑的拓扑适用于功率有效调制（例如FSK和ASK）的非相干解调，这在当今的短距离无线通信系统中很常见。希望这样的拓扑结构将导致未来的通用体系结构能够在该空间中支持各种调制格式。