Automatic Classification and Recognition Method based on Partially-Connected Differentiable Architecture Search for ISAC Systems

Automatic classification and recognition (ACR) is considered one of promising techniques in integrated sensing and communication (ISAC) systems. Existing ACR method uses Wigner-Ville Distribution (WVD) time frequency analysis (TFA) of modulation signal to obtain its WVD images which can be recognized by convolution neural network (CNN). In this paper, we proposed an improved ACR method using partially-connected differentiable architecture search (PC-DARTS). The search space is relaxed to a continuous space, and then the performance of the validation set of the architecture is optimized by gradient descent, so as to find the optimal CNN structure for signal classification and recognition. Experimental results show that the proposed method can achieve better performance with lower complexity than the other deep learning-based methods.