Noise Separation Techniques for Accurate Substation Anomaly Detection: An Intelligent Methodology

To better monitor and characterize sounds produced by substations, this study aims to separate sounds produced by the equipment from environmental ambient noise as a means of improving the relevancy (and ultimately reliability) of the power grid. To do so, we propose a deep learning-based noise monitoring system in an end-network-cloud architecture that enables remote data collection, analysis, and management. This is achieved by developing a deep learning-based noise monitoring system, enabling remote data collection, processing, and management. The proposed method consists of two basic components: a self-designed Panel Response Acquisition device that can collect sufficient acoustic information, and a refined Deep Belief Network (DBN) that is trained with a Dynamic version of the Dwarf Mongoose Optimizer (DDMO) to improve the accuracy of the noise separation process. The performance of the DBN/DDMO model is 13.1 dB for SI-SDRi and 15.7 dB for SDRi, which are large improvements for SI-SNRi and SDRi over AlexNet and CNN-VGG19. This approach minimizes SPL deviations, as shown by a thorough computation regarding several data sets; therefore, it guarantees precise noise quantification under disturbing sounds. By allowing for proactive identification of unusual noise levels, this research supports predictive maintenance methods that can avoid sudden failures and improve the overall reliability of substations.