Complex-Valued Neural Networks for Signal Classification in Sensor Networks Based on Multidomain Feature Fusion

The processing of complex-valued data and multidomain feature fusion is essential in sensor networks, especially for applications like communication modulation recognition and signal classification. However, traditional neural networks often overlook the interactions between real and imaginary components in complex-valued data and lack comprehensive integration of features from diverse domains, limiting performance in complex signal tasks. This article presents a complex multidomain neural network (CMDNet), a multidomain feature fusion network that unifies information from the time, frequency, and Wigner-Ville distribution (WVD) domains. CMDNet also introduces specialized neural modules, including complex convolution, maximum magnitude normalization (MMN), and the Cardioid ReLU activation function, which are specifically designed to extract complex-valued features and retain phase information. Extensive experiments reveal that CMDNet offers significant improvements in accuracy, robustness, and efficiency in complex signal classification, highlighting its potential for diverse applications in sensor networks.