Deep Learning-Driven Framework for Automated Electromagnetic Modulation Identification and Optimization in IoT Networks

Abstract

With the rapid development of IoT technology, the connection and data transmission of a large number of devices have put forward higher requirements for the identification and optimization of electromagnetic modulation signals in communication networks, which are often difficult to meet by traditional methods. This study proposes a deep learning-based automatic identification and optimization design method for electromagnetic modulation signals of IoT communication networks, which aims to solve the problems of complex signals and serious interference in the Internet of Things environment. In this study, the multi-source data sets such as RadioML2016.10b, MATLAB analog signals and air port acquisition signals are integrated, and the training/verification/test sets are divided into 7:1:2 after 300MHz-3GHz bandpass filtering denoising, STFT time-domain to two-dimensional time-frequency graph, normalization and data enhancement. The "CNN-LSTM-Attention" hybrid architecture is constructed as the core recognition model, in which CNN extracts local features of the signal through 4-layer convolution and 2-layer pooling, captures the timing dependence of bidirectional LSTM (64 hidden units in two layers), strengthens the key features by the attention mechanism, and finally outputs 16 types of modulated signal probabilities from the fully connected layer. In the signal optimization process, LSTM combined with Q-learning output power commands of 0.5-2dBm step to adjust the signal strength, DBN and PSO are used to achieve dynamic frequency optimization, and the interference suppression ratio is increased by 15-20dB through CBAM-GRU and LMS filtering. The model training is based on the PyTorch framework, using the Adam optimizer (learning rate 1e-3) and cross-entropy loss function for 100 rounds, adding learning rate scheduling and early stop mechanisms, and fine-tuning hyperparameters through Bayesian optimization to innovatively construct a deep learning optimization model. Based on the model recognition results and relying on the above signal optimization scheme, the stability and efficiency of the Internet of Things communication network can be significantly improved. By training the deep learning model on 10,000 sets of electromagnetic modulation signals in different scenarios, the experimental results show that compared with the traditional signal recognition method, the signal recognition accuracy of the proposed method is improved by 30%, and the average accuracy of the test set reaches 98.2% (still 92.5% under 5dB low signal-to-noise ratio), which reaches a high level overall. At the same time, in terms of signal optimization design, the network communication efficiency is improved by 20%, and the bit error rate is as low as 0.0005 at 20dB signal-to-noise ratio, which significantly enhances the performance of IoT communication networks.