Abnormal noise detection of electric machines based on HPSS-CIS and CNN-CBAM

• Published in: Acta acustica Vol. 9; p. 39
• Main Authors: Zhao, Qingsong, Wang, Xiufeng, Luo, Kun, He, Dan, Liu, Xiang
• Format: Journal Article
• Language: English
• Published: EDP Sciences 2025
• Subjects: abnormal noise detection; continuous interphase sampling; convolutional block attention module; harmonic percussion sound separation
• ISSN: 2681-4617; 2681-4617
• DOI: 10.1051/aacus/2025023

For a long time, the traditional motor manufacturing industry relies on the artificial hearing method to identify whether there is abnormal noise in the motor, thus leading to low efficiency and poor accuracy consistency. To solve these problems, a new prediction method based on the algorithm of harmonic percussion sound separation (HPSS) and continuous interphase sampling (CIS) of cochlear implants and the CNN-CBAM (Convolutional neural network based on Convolutional Block Attention Module) model, is proposed in this paper. Firstly, the original sound signals are separated into harmonic and percussive components by the HPSS algorithm, and then each component is processed by the CIS algorithm of cochlear implant to obtain electrode stimulation signal that can simulate human hearing. Subsequently, the classification task of motors are achieved by a deep learning model that combines CNN and CBAM. The proposed method is verified that the highest accuracy of 99.27% is achieved in the motor data set. Afterward for feature extraction, the results of ablation experiments with HPSS-CIS show that the average accuracy of this method is more than 4.5% higher than that of any single component. In addition, for the human auditory feature extraction method after HPSS processing, the CIS method is compared with the widely used Mel filter bank, and shows better performance.