Research and implementation of compressed sensing technology for deep data fusion of mine Internet of Things

The explosive growth of massive multi-source heterogeneous data in the mine Internet of Things has brought significant computing, storage and transmission pressure to the underground industrial ring network, which has seriously restricted the process of coal mine intelligence. The existing compressed sensing methods generally rely on artificial prior knowledge construction, and it is difficult to adapt to the lightweight deployment requirements of edge equipment, and can not meet the requirements of real-time compressed data transmission and reconstruction in complex mine environment. In order to realize the deep compressed transmission and reconstruction of massive sparse sensing data in the mine Internet of Things, a deep compressed sensing model (DCS net) is proposed through the research on the key technologies of mining sparse sensing data compression and reconstruction. Firstly, according to the characteristics of the coexistence of linear and nonlinear mine data, a dual link multi-level convolution architecture integrating the residual module is designed, which uses single link single-level convolution compression or dual link three-level convolution technology to process linear and nonlinear data, and solves the gradient disappearance and degradation of convolution network with the help of jump connection, so as to realize the differential and efficient compression of different types of sparse data. At the same time, the “reconstruction distortion transmission adaptation rate constraint security” quaternion composite loss function is constructed, which integrates the accuracy and robustness of Huber loss guarantee data compression and reconstruction, and introduces the transmission against robust loss to adapt to the underground high interference environment. The risk of disaster data early warning failure is avoided through the security threshold constraint, and the dynamic weight mechanism is used to balance the compression ratio and reconstruction accuracy. Secondly, edge aware devices are grouped based on Gaussian similarity clustering, and a collaborative deployment architecture of “intra cluster compression inter cluster transmission cloud optimization” is constructed. The coding network is deployed in the data aggregator and the decoding network is deployed in the edge server to improve the robustness of the system. Finally, six kinds of sparse structure data sets such as emnist data set, coal mine monitoring screenshots, and main fan vibration measured data are taken as test samples to carry out comparative experiments with OMP, TVAL3 and other six common data models. The results show that DCS-Net has stable sparse data reconstruction performance. When the compression ratio is 0.25, the normalized mean square error (NMSE) is as low as 0.042, the peak signal-to-noise ratio (PSNR) is 32.54, and the maximum error of main fan vibration data reconstruction is only 1.66 mm/s; The compressed reconstruction time is more than 30% less than that of wings model. In the engineering application of qinglongsi coal mine intelligent ventilation system, the data compression rate is 87.53%, and the compression reconstruction time is saved by 53.60% compared with the mainstream model. It solves the problems of accurate perception, real-time collection and efficient transmission of multi-source heterogeneous data in the mine Internet of Things under the condition of limited network resources, realizes the accurate compression and reconstruction of multi-type mine sparse perception data, adapts to the lightweight deployment requirements of edge equipment, and provides reliable theoretical and technical support for the real-time accurate perception of massive monitoring information in the intelligent construction of coal mines.