Signal Analysis Based on Time-frequency Transformation and Deep Learning

This paper explores time-frequency transformation techniques, object detection and deep learning-based recognition methods for four basic electromagnetic interference (EMI) waveforms in electromagnetic compatibility (EMC) testing. A Short-Time Fourier Transform (STFT)-based feature representation is proposed to overcome the limitations of traditional spectrum analysis. The integration of the DEtection-TRansformer (DETR) model and the Multilayer Perceptron (MLP) is employed for target detection and parameter extraction of EMI basic signals in time-frequency spectral images. In the switch-mode power supply (SMPS) experiment, the parasitic parameter prediction errors of the circuit diode were all below 2%. The proposed approach improves the convenience and accuracy of EMI signal classification and parameter prediction.