An Improved BP Neural Network Model based on High-Dimensional Feature Selection Strategy and Maxwell's Equations Derived Optimization

As systems continue to expand in scale, there is a corresponding increase in the variety and number of electronic devices within these systems, which contributes to a heightened complexity in electromagnetic emissions (EME). This complexity complicates the process of identifying electromagnetic emitters. To tackle this challenge, this paper introduces an improved back propagation (BP) neural network model based on high-dimensional feature selection strategy and Maxwell’s Equations Derived Optimization (HFS-MEDO-BPNN). The proposed model initially utilizes a feature selection approach grounded in random small samples to improve computational efficiency by streamlining the network architecture. Subsequently, an electromagnetic optimization algorithm, referred to as Maxwell's Equations Derived Optimization (MEDO), is applied to refine the parameters of the neural network, thereby enhancing the model's recognition capabilities. Ultimately, the innovative model is employed to identify a dataset of electromagnetic emission signals, which is constructed based on fundamental emission waveform theory. Experimental findings indicate this approach achieves reduced time complexity and improved recognition accuracy comparing to other existing BP neural network models.