~矿井巷道中的无线通信是煤矿安全生产和矿井应急救援的重要技术手段，矿井复杂环境下无线通信的可靠性是煤矿安全生产的关键。随着煤矿智能开采和智慧矿山建设的推进，井下无线通信系统已经有了大规模的部署。井下无线业务需求迅速增加，异构网络、多业务终端智能接入以及各信息系统融合等问题日益突出。井下无线电信号的识别、检测是解决该问题的关键。矿井无线信号的传输，经过信道编码、调制、波束成型后通过天线发射出去。由于矿井无线信号传播环境复杂，射频损伤、信道衰落、噪声、震动等干扰等因素的影响，到达接收机的信号会存在较严重干扰和衰落，接收机在信号识别和信息恢复检测过程中存在较高的误码率。因此，如何提高矿井复杂环境下的无线信号识别率和信息检测正确率、降低矿井无线通信系统的误码率、提高传输的可靠性是本文研究的重点。本文的主要研究内容和取得的主要成果如下：
（1）针对煤矿复杂环境下多模无线信号识别困难以及识别率低的问题，本文提出基于高阶累积量的矿井多模信号识别算法。特征参数提取采用基于高阶累积量的特征提取算法，分类器设计分别采用决策树和遗传优化支持向量机和深度神经网络。仿真结果表明，基于高阶累积量和深度神经网络的矿井多模信号的正确识别率和所有信号的平均识别率优于决策树和遗传优化支持向量机分类算法，高阶累积量算法模型简单、所需数据样本少、时效性高。
（2）针对高阶累积量值和特征参数选取对煤矿环境下多模信号识别率影响较大、低信噪比下信号识别率不佳等问题，本文利用复解析小波变换提取信号的包络特性和相位特性，构造矿井多模信号识别特征。为加强IQ信号数据空间特征和时间特征的识别度、消除OFDM信号的载波相位偏移（PO），对卷积神经网络（CNN）和长短期记忆神经网络（LSTM）网络参数进行调整、优化，设计了适合矿井多模信号识别的复合神经网络（CLDNN）。仿真结果表明，基于CLDNN的矿井多模信号识别率明显优于CNN和LSTM等单一网络模型，其中OFDM信号的识别率在信噪比为0dB时达90%。
（3）针对基于高阶累积量和小波变换矿井多模信号识别的预处理计算复杂度高、识别率上限受特征影响、低信噪比下识别率低的问题，本文将IQ格式的数据与其幅度相位格式的数据联合作为深度神经网络的输入来增加信号的特征量，利用深度学习模型主动学习信号特征，建立学习和推理的机制。随着信号种类的增加和对信号检测识别高精度的要求，本文从增加残余块、增加卷积层滤波器尺寸来增加残差神经网络（ResNet）的“宽度”对ResNet进行网络优化，提高矿井多模信号的检测识别精度。改进中心损失函数与softmax损失函数构成混合损失函数，根据输入数据进行训练和调整参数对密集连接神经网络（DenseNet）进行优化设计。实验结果表明，矿井衰落环境下的多模信号的识别率和矿井TD-SCDMA与LTE无线信号的数据流检测正确率都明显提高，基于DenseNet矿井TD-SCDMA和LTE信号在信噪比为-5dB时检测正确率高达91%以上。
（4）针对目前接收机性能受矿井环境干扰影响较大，接收性能差、误码率高等问题，本文首次提出矿井智能接收机模型。该模型利用深度神经网络整体替代目前接收机端到端的接收过程，从IQ信号到原始信息比特流的整个信息恢复环节，该深度神经网络接收机在本文简称“深度接收机（Deep Receiver）”。深度接收机采用密集连接神经网络结构，利用全局池化来适应不同的输入信号长度，采用多个二分类器实现多个比特的信息流恢复。仿真结果表明，在AWGN信道下，针对不同调制方式的OFDM接收机和不同天线模式的MIMO接收机，本文提出的智能接收机的性能大大提升；在矿井Nakagami信道下，MIMO接收机在不同衰减系数、不同调制编码方式、不同MIMO模式和不同数量的数据丢失情况下，本文提出的智能接收机性能均有良好的改善。

~Wireless communication in mine roadways is an important technical means for coal mine safety production and mine emergency rescue. Reliability of wireless communication in complex mine environment is the key to coal mine safety production. With the promotion of green mining in coal mines and the construction of smart mines, underground wireless communication system is deployed on a large scale. With the rapid increase of underground wireless business demand, the problem of heterogeneous network, intelligent access of multi-service terminals and integration of various information systems become increasingly prominent. The identification and detection of underground radio signal is the key to solve this problem. The transmission of mine wireless signal is transmitted through antenna after channel coding, modulation and beam shaping. Due to the complex mine wireless signal propagation environment, radio frequency damage, channel fading, noise, vibration and other factors, the signal arriving at the receiver will have serious interference and fading, and the receiver will have a high error rate in the process of signal recognition and information recovery detection. Therefore, how to improve the recognition rate of wireless signals and the accuracy rate of information detection in complex mine environment, reduce the error rate of mine wireless communication system, and improve the reliability of transmission is the focus of this dissertation. The main contents and achievements of this dissertation are as follows:
(1) In view of the difficulties and low recognition rate of multi-mode wireless signals in complex coal mine environment, this dissertation puts forward a mine multi-mode signal recognition algorithm based on high-order cumulants. Feature parameter extraction uses feature extraction algorithm based on high-order cumulative variables, and classifier design uses decision tree and genetic optimization support vector machine and deep neural network, respectively. The simulation results show that the correct recognition rate of multi-mode signals and the average recognition rate of all signals based on high-order cumulative and deep neural network are better than decision tree and genetic optimization support vector machine classification algorithm. The high-order cumulative algorithm model is simple, requires fewer data samples, and has a high timeliness.
(2) In view of the problems that the selection of high-order cumulative value and characteristic parameters has a great influence on the recognition rate of multimode signals in coal mine environment and the poor recognition rate of signals under low signal-to-noise ratio, this dissertation uses complex analysis wavelet transform to extract the envelope and phase characteristics of signals, and constructs the recognition characteristics of multimode signals in coal mine. In order to enhance the recognition of spatial and temporal features of IQ signal data and eliminate carrier phase offset (PO) of OFDM signal, the parameters of convolution neural network (CNN) and long-term and short-term memory neural network (LSTM) network are adjusted and optimized, and a composite neural network (CLDNN) suitable for mine multimode signal recognition is designed. The simulation results show that the recognition rate of mine multimode signals based on CLDN is significantly better than that of single network models such as CNN and LSTM, in which the recognition rate of OFDM signals reaches 90% when the signal-to-noise ratio is 0 dB.
(3) In order to solve the problem of high computational complexity, upper limit of recognition rate affected by characteristics and low recognition rate under low signal-to-noise ratio of mine multimode signal recognition based on high-order cumulative and wavelet transform, this dissertation combines IQ format data with its amplitude-phase format data as input of deep neural network to increase the signal's characteristic quantity. The mechanism of learning and reasoning is established by using the active learning signal characteristics of deep learning model. With the increase of signal types and the requirement of high accuracy for signal detection and recognition, this dissertation optimizes ResNet by increasing residual blocks and convolution layer filter size to increase the "width" of residual neural network (ResNet) to improve the detection and recognition accuracy of mine multimode signals. The center loss function and the softmax loss function are improved to form a mixed loss function. DenseNet is optimized by training and adjusting the parameters according to the input data. The results show that the recognition rate of multi-mode signals in mine fading environment and the detection accuracy rate of data flow of mine TD-SCDMA and LTE wireless signals are significantly improved. The detection accuracy rate of mine TD-SCDMA and LTE signals based on DenseNet is over 91% when the signal-to-noise ratio is -5dB.
In view of the problems that the current receiver performance is greatly affected by the mine environment interference, poor reception performance and high error rate, this dissertation puts forward for the first time the mine intelligent receiver model. This model uses a deep neural network to replace the end-to-end receiving process of the current receiver as a whole, and recovers the whole information from IQ signal to the original information bit stream. This deep neural network receiver is referred to as Deep Receiver in this dissertation. The depth receiver uses a densely connected neural network structure, uses global pooling to accommodate different input signal lengths, and uses multiple binary classifiers to recover multiple bits of information flow. The simulation results show that, in AWGN channel, the performance of the proposed smart receiver is greatly improved for OFDM receivers with different modulation modes and MIMO receivers with different antenna modes. In the mine Nakagami channel, the performance of the MIMO receiver proposed in this dissertation is improved with different attenuation coefficient, different modulation and encoding modes, different MIMO modes and different number of data loss.