Abnormal Sound Detection of Electrical Equipment Based on Time-Spectrum Information Fusion

In the daily monitoring and maintenance of electrical equipment, the detection of abnormal sound is a very important part, and the high accuracy and efficiency of sound to distinguish the abnormal state of the equipment is more conducive to the troubleshooting of the equipment. At present, the research of algorithms for identifying the working status of electrical equipment operation sound based on machine learning and deep learning has been more widely used, but many algorithms are mainly based on frequency domain sequence modeling, ignoring the information of time sequence. Since the time-domain information in abnormal sounds is difficult to capture compared with the frequency-domain information, a time-domain information feature extraction network is designed by combining the commonly used frequency-domain analysis methods. In this paper, we propose a time-spectrum fusion-based anomaly detection optimization model that combines frequency-domain and time-domain feature fusion to improve the detection performance of anomalous sounds. Our method is experimentally validated on the DCASE 2020 Challenge Task 2 dataset, with an average AUC improvement of 1.35% and an average pAUC improvement of 1.35% compared to the current best method.