近年来，随着煤矿经济的发展，对煤矿安全的要求也随之提高，建立高效可靠的矿井无线通信变得至关重要。然而，煤矿特殊的地理地形导致信号大幅度衰减，造成无线信号衰落严重，使得通信质量下降。为解决当前矿井通信面临的挑战与难题，通信领域专家发现深度学习所蕴含的巨大发展潜力，因此将深度学习的算法应用于矿井通信领域。在矿井环境下，无线通信系统的接收端通过应用深度学习的方法解决信号检测误码率高、易受信道干扰等问题，有效提高了矿井无线信号的检测性能和信道估计精度。主要工作有以下几个方面：

(1) 针对复杂矿井环境中的无线通信接收信号误码率高和易受信道干扰等问题，提出一种基于双路径网络的矿井无线信号检测方法，该方法采用双路网络接收机(Dual Path Network Receiver，DPNR)对无线信号进行检测。DPNR信号检测方法优化了正交频分复用(Orthogonal Frequency Division Multiplexing，OFDM)接收端的整体性能，并解决了常规接收机的误差累积问题。仿真结果表明，DPNR在OFDM中与常规接收机相比，在信噪比大于7的矿井环境中，误码率降低1个数量级(1个数量级表示10^-1)以上。DPNR在FBMC中与常规接收机进行对比，误码率降低2个数量级以上。在不同网络接收机情况下，信噪比大于13时，DPNR的误码率低于残差神经网络(Residual Neural Network，ResNet)接收机和密集连接卷积网络(Densely Connected Convolutional Network，DenseNet)接收机的误码率。DPNR和已提出的深度接收机相比，其误码率更低。因此，DPNR优于已提出的深度接收机和常规接收机，不受信道的干扰且在信号检测方面具有鲁棒性，使信号得以更加准确地恢复。

(2) 为了解决传统信道估计和信号检测精度低、用深度神经网络代替接收端整体的深度接收机需大量数据进行训练且训练速度慢等问题，将接收端与传统通信知识相结合，提出一种基于矿井下联合剩余信道估计网络(Residual channel Estimation Network，ReEsNet)和信号检测恢复网络(Signal Detection Recovery Network，SDRNet)的信道估计和信号检测方法。仿真结果表明，在复杂矿井环境下信噪比为1~15dB的信道估计插值中，ReEsNet与最小均方误差(Minimum Mean Squared Error，MMSE)相比均方误差(Mean Squared Error，MSE)值低于9.61e-5以上，与最小二乘(Least Sqaure，LS)相比MSE值低于6.31e-3以上，ReEsNet有更好的信道估计插值效果。在基于ReEsNet和SDRNet联合的信道估计和信号检测方法中信噪比为1~15dB时，联合ReEsNet和SDRNet与MMSE相比误码率值低于7.96e-3以上，与LS相比误码率值低于1.18e-2以上，且联合的ReEsNet和SDRNet检测的误码率低于单个SDRNet模块检测的误码率。在已提出的联合信道估计和信号检测方法中，相比信道条件恢复网络(Channel Conditional Recovery Network，CCRNet)减少97%的参数，相比深度神经网络(Deep Neural Networks，DNNs)减少84%的参数，其检测效果并无显著的差异。因此本文提出基于矿井下联合ReEsNet和SDRNet的方法在复杂度低时仍有较好的信道估计精度和信号检测性能。

In recent years, with the development of coal mine economy, the requirements for coal mine safety are getting higher and higher, and the establishment of efficient and reliable mine wireless communication becomes very important. However, the special geographic terrain of coal mines leads to significant signal attenuation, resulting in serious wireless signal degradation, which makes the communication quality decline. In order to solve the current challenges and problems faced by mine communication, experts in the field of communication have found that deep learning contains great potential for development, and therefore apply deep learning algorithms to the field of mine communication. Deep learning is applied to the receiver side of the wireless communication system in the mine environment to solve the problems of high bit error rate and susceptibility to channel interference in the detection of mine wireless signals, which effectively improves the detection performance and channel estimation accuracy of mine wireless signals. The main work has the following aspects:

(1) Aiming at the problems of high BER and susceptibility to channel interference of wireless communication received signals in complex mine environments, a dual-path network-based wireless signal detection method for mines is proposed, which employs a dual-path network receiver (DPNR) to detect wireless signals. Signal detection with DPNR optimizes the overall performance of the OFDM receiver and solves the error accumulation problem of conventional receivers. The experimental results show that DPNR in orthogonal frequency division multiplexing (OFDM) reduces the BER by more than 1 order of magnitude (order of magnitude means 10^-1) compared with conventional receivers in a mine environment with a signal-to-noise ratio greater than 7. DPNR in FBMC reduces the BER by more than 2 orders of magnitude when compared with conventional receivers. The BER of DPNR is lower than the BER of residual neural network (ResNet) receiver and densely connected convolutional network (DenseNet) receiver in the case of different network receivers with signal-to-noise ratio greater than 13. DPNR has a lower BER compared to the deep receivers that have been proposed. Thus, DPNR outperforms both the proposed deep receiver and the conventional receiver, is independent of channel interference and is robust in signal detection, allowing for more accurate signal recovery.  

(2) In order to solve the problems of low accuracy of traditional channel estimation and signal detection, and the deep neural network instead of the deep receiver at the receiver side requires a large amount of data for training and slow training speed, a channel estimation and signal detection method based on a joint residual channel estimation network (ReEsNet) and a signal detection recovery network (SDRNet) for the underground mine has been proposed by combining the receiver side with the traditional communication knowledge. Simulation results show that ReEsNet is applied to interpolate the channel estimation with a signal-to-noise ratio of 1 to 15 dB in a complex mine environment, and the mean-square error (MSE) value of ReEsNet is lower than 9.61e-5 compared with MMSE, and the MSE value is lower than 6.31e-3 compared with LS, and the MSE of ReEsNet is the smallest, which is closer to the real channel estimation. In the joint ReEsNet and SDRNet based channel estimation and signal detection method with signal-to-noise ratios from 1 to 15 dB, the BER values of the joint ReEsNet and SDRNet are lower than 7.96e-3 compared to MMSE, and lower than 1.18e-2 compared to LS, and SDRNet is higher than the joint ReEsNet and SDRNet 's BER performance. Compared with the proposed joint channel estimation and signal detection methods, the trainable parameters are less compared to the channel conditional recovery network (CCRNet) with 97% fewer parameters and compared to the deep neural networks (CNNs) with 84% fewer parameters, and their detection results are not much different. Therefore, in this thesis, we propose an approach based on joint ReEsNet and SDRNet under the mine which still has good channel estimation accuracy and signal detection performance at low complexity.