CCWHE: multimodel ensemble method for modulation recognition based on weight learning

Currently, the research on Automatic Modulation Identification (AMR) mostly improves the recognition accuracy by enhancing the model network architecture. However, the resulting increase in model complexity and the consumption of computing resources have become the key factors restricting the application of the model. This paper focuses on the contradiction between the network complexity and recognition accuracy of the AMR model. By learning the idea of the Adaboost.M2 algorithm, a Category Confidence Weighted Heterogeneous Ensemble algorithm (CCWHE) is proposed. This algorithm overcomes the limitation of static weight allocation of the traditional Adaboost.M2 algorithm and dynamically learns to adjust the weights based on the confidence distribution of each base model in different modulation categories. Experimental results demonstrate that upon the application of the CCWHE algorithm to the combination of the GRU model and the ResNet model, under a low computing complexity, the average recognition rate of 94.42% (SNR > 10 dB) can be obtained for the 24 types of modulation classification. In addition to the experimental verification of optimization paths of traditional models, the proposed study sheds light on the efficient way of building a lightweight but high-accuracy modulation recognition model.