首页|DEEPRIDGE: A HOP-LAYER DEEP LEARNING MODEL FOR DETECTING ENERGY RIDGES IN HOIST OPERATION NOISE SPECTROGRAMS
DEEPRIDGE: A HOP-LAYER DEEP LEARNING MODEL FOR DETECTING ENERGY RIDGES IN HOIST OPERATION NOISE SPECTROGRAMS
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NETL
NSTL
The variation trend of energy ridge in the noise spectrum of mine hoist is an important index of its health state. Due to the interference of background noise, the energy ridges in the hoist spectrum have unclear boundaries and poor continuity, making them hard to separate and analyze. To address this challenge, we propose a new end-to-end training deep neural network that extracts and fuses multi-scale features of spectrographs using multiple encoder-attention module-decoder blocks. The multi-scale features enable the network to capture the energy ridges in the spectrum more accurately and robustly. The small-scale feature map is used for coarse localization, and the large-scale feature map is used for fine refinement. We design and train our network, named DeepRidge, on a noise spectrum dataset of mining hoist collected from a preliminary experiment. We compared DeepRidge with other state-of-the-art methods, and the test results show that DeepRidge achieves better accuracy at smaller parameter sizes, with an average precision (AP) of 0.826. We also conduct experiments to find the optimal network configuration for the energy ridge detection task in the hoist noise spectrum.
NETL
NSTL
