Automatic modulation recognition in low-computation edge devices under noisy environments has become increasingly critical. To address the dual challenges of lightweight design and noise resistance, this paper proposes WCTFormer, a novel lightweight network that integrates discrete wavelet transform for frequency-domain noise suppression and a Transformer for global feature extraction, enabling robust performance in low SNR conditions. Experiments on open-source datasets demonstrate that WCTFormer achieves superior recognition accuracy, with 92.40% accuracy at 0 dB SNR, requiring only 60K parameters. WCTFormer combines high recognition performance and computational efficiency, making it suitable for deployment in resource-constrained edge devices.
25 August 2026 in Signal, Image and Video Processing
The generalized modeling of core loss aims to predict the power loss of magnetic components, which is crucial in fields such as electronic industry design. However, the influence relationships of operating condition variables are complex. Empirical models, such as the Steinmetz Equation, cannot fully characterise such nonlinear relationships, because the used priori equation has insufficient variables and their coefficients are all fixed. Although deep learning alleviates this issue, most existing methods still only consider limited variables and are not applicable to cross-material modeling. Therefore, we combine the prior knowledge of empirical models and the powerful fitting ability of deep learning, and propose a Deep Self-Representation (DSR) module that can either function as an independent model or be integrated into existing empirical formulas. Specifically, the integrated model can dynamically adjust the coefficients according to the input variables, resulting in better accuracy and generalization. In addition, based on the log-linear property provided by Steinmetz, we propose a flexible approach that can conveniently add more variables. To evaluate the effectiveness of the proposed method, we conducted a series of comparative and simulation experiments on the open-source dataset MagNet. The experimental results show that DSR, used alone as the core loss model, outperforms baseline methods in accuracy, with 0.875% and 1.229% errors in single-material and multi-material cases, respectively. Furthermore, the integrated model has even lower errors of 0.657% and 1.01% and better robustness. Moreover, our model can perform training and prediction on an inexpensive GPUor CPU, providing an efficient solution. We release our code online: https://github.com/.