Effectiveness Analysis of the Mel spectrum features of AE signals in the detection of partial discharge faults

Acoustic emission (AE) technologies have been widely used in the detection of partial discharge (PD) faults. Currently, obtaining the time-frequency spectrum of a signal by short-time Fourier transform or wavelet transform and then extracting the practical features for fault identification has received increasing attention. This is because the time-frequency spectrum holds a wider variety of intrinsic characteristics of the signal source, which helps enhance the understanding of faults and improve identification accuracy. A high-resolution linear time-frequency analysis method can provide a more detailed signal characterization capability for processing time-frequency spectra. However, the bandwidth of AE signals is usually up to several hundreds of kHz, which means the significant correlation coefficients between spectra tend to generate high redundancy characteristics. This study proposes a method to optimize the time-frequency spectrum density distribution based on the Filter Banks features. The method achieves highly desirable results by converting the AE signal into a Mel spectrum and using ResNet50 to extract spectral features for training and fault type classification. The research results promote the development of AE detection technology applications and, in some ways, provide a valuable reference for machine learning-oriented sound signal processing methods.