Directional Differentiable Architecture Search for Waveform Recognition

Waveform recognition is used to identify classes of different electrical signal waveforms. It is widely used in medical diagnosis, wireless communication, signal video processing, audio processing, and other fields. However, conventional approaches to waveform recognition perform poorly under low signal-to-noise ratio conditions and with easily confusable signals of different types. To address this challenge, we introduce automatic machine learning (AutoML) into the waveform recognition task to automatically obtain a high-performance recognition network. Among AutoML, Differentiable Neural Architecture Search (DARTS) has been widely used due to its short training time, low hardware requirements, and efficient neural networks generation ability. However, DARTS cannot guide the sparsity of the architecture parameters, leading to performance loss in discretization. To this end, we propose an end-to-end trainable evolutionary method called Directional Differentiable Architecture Search (DDAS) in this paper. First, during the architecture search, the training data is augmented with more easily confusable signals samples to enhance the model’s ability to distinguish ambiguous patterns. Second, the soft-max function delays the non-zero-is-one operation selection to a weighted sum of the different operations. This makes the operation architectural parameters differentiable and greatly reduces the training cost. Third, a directional pruning-based optimization method is used to bring the highest weight of operations closer to 1 to reduce discretization loss, where the operation with the highest weight is selected as the final operation in the generated network. Experiments on two benchmark waveform recognition datasets show that the resulting network outperforms both the traditional manually designed network and the network obtained by directly applying existing architectural search methods, achieving higher accuracy. The obtained network also has a higher ability to recognize easily confusable signals. Notably, the generated network performance performs equally well under all signal-to-noise (SNR) ratios, offering new insights for waveform recognition. Code and datasets are available on https://github.com/tju-xm/DDAS.