基于轻量级混合神经网络的边缘设备调制识别方法

Edge devices modulation recognition method based on lightweight hybrid neural network

摘要:

由于计算能力和内存的限制,很难将传统的深度学习模型应用于物联网边缘设备, 以实现自动调制识别。为有效结合卷积神经网络和视觉Transformer网络的优势,引入了一种应用于物联网边缘设备的卷积和Transformer组合网络模型ICTNet(Internet of Things CNN Transformer network)。ICTNet不仅拥有Transformer的优势来捕捉特征长程依赖关系,还可以利用CNN的优势提取特征局部信息,在缩减模型大小的同时增加调制识别精度。ICTNet模型在RadioML2016.10a、RadioML2016.10b和RadioML2016.04c数据集中的平均识别准确率分别为61.51%、64.18%和71.96%。此外,ICTNet在典型边缘设备上处理每个信号样本的时间接近0.01 ms,且比现有的DL-AMR模型小很多,只有29 455个参数。

Abstract:

Due to the limitations of computing power and memory, it is difficult to apply traditional deep learning (DL) models to the Internet of Things (IoT) edge devices for automatic modulation recognition (AMR). To effectively combine the advantages of convolutional neural networks (CNNs) and visual Transformer networks, a combined convolutional and Transformer network model called ICTNet (Internet of Things CNN Transformer network) was introduced for IoT edge devices. ICTNet not only has the advantage of Transformer to capture feature long-range dependencies, but also utilizes the strengths of CNNs to extract local feature information, reducing model size while increasing modulation recognition accuracy. The average recognition accuracies of the ICTNet model in the RadioML2016.10a, RadioML 2016.10b and RadioML2016.04c datasets are 61.51%, 64.18% and 71.96% respectively. Moreover, ICTNet processes each signal sample on a typical edge device in nearly 0.01 ms, and is much smaller than existing DL-AMR models, with only 29 455 parameters.