一种深度学习的立体阵波达方向估计方法

针对常规波束形成主瓣宽且目标分辨能力低的问题,提出一种基于深度卷积神经网络的波达方向估计方法。算法使用常规波束形成计算二维空间功率谱,将预处理后的空间功率谱图输入深度卷积神经网络。该文利用神经网络学习解卷积映射关系,输出主瓣宽度更窄的空间功率谱图,从而实现高分辨率二维波达方向估计。该算法对阵列结构没有限制,适用于立体阵。仿真结果表明该文方法在不同目标个数、快拍数及信噪比参数下均能准确估计目标方向。该文方法目标分辨能力优于常规波束形成方法。在低快拍情况下,目标方向估计误差低于自适应波束形成方法。     

基于深度学习的ELM实时识别研究

基于深度学习的方法,在HL-2A装置上开发出了一套边缘局域模(ELM)实时识别算法。算法使用5200次放电数据(约24.19万数据切片)进行学习,得到一个深度为22层的卷积神经网络。为衡量算法的识别能力,识别了HL-2A装置自2009年实现稳定ELMyH模放电以来所有历史数据(约26000次放电数据),共识别出1665次H模放电,其中误识别35次,误报率为2.10%。在实际的1634次H模放电中,漏识别4次,漏识别率为0.24%。该误报率和漏报率可以满足ELM实时识别的精度要求。识别算法在实时控制环境下,对单个时间点的平均计算时间为0.46ms,可以满足实时控制的计算速度要求。     

Methodology to classify hazardous compounds via deep learning based on convolutional neural networks

Compounds information such as Chemical Abstracts Service (CAS) registry number, hazards, and properties have been provided through Globally Harmonized System (GHS) based Material Safety Data Sheet (MSDS). This information can help users avoid hazardous compounds and handle chemicals in proper way. GHS specifies that hazards of compounds are categorized through animal testing (or in vivo testing) , in vitro testing, epidemiological surveillance, and clinical trials. In this study, artificial intelligence (AI) is used to replace traditional approaches in predicting the toxicity of chemicals. A database of hazardous compounds is generated by data provided by the Ministry of Environment (ME), training and learning based on convolutional neural network (CNN) are carried out following data featurization. As a result, 90% of accuracy for CNN-based model is obtained using the image dataset. In contrast to the previous methods, the classification method based on CNN-based model in this study allows for the efficient discrimination of hazard chemicals without any additional tests.     

Deep learning assisted time-varying channel estimation in multi-user mmWave hybrid MIMO systems

Hybrid analog/digital multiple input multiple output (MIMO) system is proposed to mitigate the challenges of millimeter wave (mmWave) communication. This architecture enables utilizing the large array gain with reasonable power consumption. However, new methods are required for the channel estimation problem of hybrid architecture-based systems due to the fewer number of radio frequency (RF) chains than antenna elements. Leveraging the sparse nature of the mmWave channels, compressed sensing (CS)-based channel estimation methods are proposed. Recently, machine learning (ML)-aided methods have been investigated to improve the channel estimation performance. Additionally, the Doppler effect should be considered for the high mobility scenarios, and we deal with the time-varying channel model. Therefore, in this article, we consider the scenario of time-varying channels for a multi-user mmWave hybrid MIMO system. By proposing a Deep Neural Network (DNN) and defining the inputs and outputs, we introduce a novel algorithm called Deep Learning Assisted Angle Estimation (DLA-AE) for improving the estimation of the Angles of Departure/Arrival (AoDs/AoAs) of the channel paths. In addition, we suggest Linear Phase Interpolation (LPI) to acquire the path gains for the data transmission instants. Simulation results show that utilizing the proposed DLA-AE and LPI methods enhance the time-varying channel estimation accuracy with low computational complexity.     

非高斯环境下的深度学习脉冲信号去噪与重构*

侧扫声呐进行沉底小目标探测时,底混响是主要背景干扰。底混响通常是一种非平稳、非高斯的带限噪声,它使得白噪声条件下的滤波器性能受到限制。在混响背景下常利用自回归模型对接收信号进预行白化处理,但对于实际侧扫声呐应用,白化后直接匹配滤波的处理效果不甚理想。针对此问题,在自回归模型预白化的基础上,提出采用一种次最佳检测与多分辨二分奇异值分解相结合的改进方法。该方法首先对接收信号进行分段处理,利用改进Burg算法估计每段数据自回归模型的系数及阶数;然后构造白化滤波器对分段数据预白化,并对白化后的数据进行多分辨二分奇异值分解;最后应用ostu方法对原始声图和处理后的声图进行目标检测。仿真与实验结果表明,该方法明显提高了信混比,改善了侧扫声呐沉底静态小目标的成图质量,有利于后期实现基于图像的目标自动检测。     

深度学习在超声检测缺陷识别中的应用与发展*

深度学习（Deep Learning）是目前最强大的机器学习算法之一,其中卷积神经网络（Convolutional Neural Network, CNN）模型具有自动学习特征的能力,在图像处理领域较其他深度学习模型有较大的性能优势。本文先简述了深度学习的发展史,然后综述了深度学习在超声检测缺陷识别中的应用与发展,从早期浅层神经网络到现在深度学习的应用现状,并借鉴医学影像识别和射线图像识别领域的方法,分析了卷积神经网络对超声图像缺陷识别的适用性。最后,探讨归纳了目前在超声检测图像识别中使用CNN存在的一些问题,及其主要应对策略的研究方向。     

A multi-channel and multi-user dynamic spectrum access algorithm based on deep reinforcement learning in Cognitive Vehicular Networks with sensing error

In this paper, a spectrum access problem is proposed to improve the spectrum access rates of secondary vehicles in Cognitive Vehicular Networks, where the channel capacity mathematic model is established under the conditions of spectrum sensing errors rates and the dynamic occupancy spectrum rates. Meanwhile, an improved Q-learning method is proposed to conform the dynamic communication under the different conditions of the reward functions. In this function, a Deep Q Network method with a modified reward function (IDQN) is proposed to deal with the situation of multi-vehicle in multi-channel. In order to verify the effectiveness of the IDQN method, the Myopic method, the improved Q-learning method, and the traditional DQN method are compared on Python. The simulation results shown that the proposed IDQN method not only outperforms the compared methods in terms of channel utilization and channel capacity but also improves the ability that the vehicle adapts to the dynamic communication environment.     

基于双向循环神经网络的汉语语音识别*

当前基于深度神经网络模型中,虽然其隐含层可设置多层,对复杂问题适应能力强,但每层之间的节点连接是相互独立的,这种结构特性导致了在语音序列中无法利用上下文相关信息来提高识别效果,而传统的循环神经网络虽然做出了改进,但是只能对上文信息进行利用。针对以上问题,该文采用可以同时利用语音序列中上下文相关信息的双向循环神经网络模型与深度神经网络模型相结合,并应用于语音识别。构建具有5层隐含层的模型,其中第3层为双向循环神经网络结构,其他层采用深度神经网络结构。实验结果表明:加入了双向循环神经网络结构的模型与其他模型相比,较好地提高了识别正确率;噪声对双向循环神经网络汉语识别有重要影响,尤其是训练集和测试集附加噪声类型不同时,单一的含噪声语音的训练模型无法适应不同噪声类型的语音识别;调整神经网络模型中隐含层神经元数量后,识别正确率并不是一直随着隐含层中神经元数量的增加而增加,神经元数量数目增加到一定程度后正确率出现了降低的趋势。     

基于深度学习的船舶辐射噪声识别研究

为了改善船舶辐射噪声识别系统的性能,进一步提高船舶辐射噪声识别的正确率,该文提出采用一种基于深度学习的船舶辐射噪声识别方法。该方法首先提取了船舶辐射噪声的频谱、梅尔倒谱系数等特征,将提取特征后的图像样本分别用于训练卷积神经网络和深度置信网络,再对船舶辐射噪声进行识别。通过文中所给实例,将深度学习和支持向量机两种识别方法的性能进行比较,得出深度学习方法可以有效地提高船舶辐射噪声识别正确率的初步结论。     

Enhancing network lifespan in wireless sensor networks using deep learning based Graph Neural Network

Inadequate energy of sensors is one of the most significant challenges in the development of a reliable wireless sensor network (WSN) that can withstand the demands of growing WSN applications. Implementing a sleep-wake scheduling scheme while assigning data collection and sensing chores to a dominant group of awake sensors while all other nodes are in a sleep state seems to be a potential way for preserving the energy of these sensor nodes. When the starting energy of the nodes changes from one node to another, this issue becomes more difficult to solve. The notion of a dominant set-in graph has been used in a variety of situations. The search for the smallest dominant set in a big graph might be time-consuming. Specifically, we address two issues: first, identifying the smallest possible dominant set, and second, extending the network lifespan by saving the energy of the sensors. To overcome the first problem, we design and develop a deep learning-based Graph Neural Network (DL-GNN). The GNN training method and back-propagation approach were used to train a GNN consisting of three networks such as transition network, bias network, and output network, to determine the minimal dominant set in the created graph. As a second step, we proposed a hybrid fixed-variant search (HFVS) method that considers minimal dominant sets as input and improves overall network lifespan by swapping nodes of minimal dominating sets. We prepared simulated networks with various network configurations and modeled different WSNs as undirected graphs. To get better convergence, the different values of state vector dimensions of the input vectors are investigated. When the state vector dimension is 3 or 4, minimum dominant set is recognized with high accuracy. The paper also presents comparative analyses between the proposed HFVS algorithm and other existing algorithms for extending network lifespan and discusses the trade-offs that exist between them. Lifespan of wireless sensor network, which is based on the dominant set method, is greatly increased by the techniques we have proposed.     

UAV path design with connectivity constraint based on deep reinforcement learning

Cellular networks are expected to communicate effectively with unmanned aerial vehicles (UAVs) and support various applications. However, existing cellular networks are primarily designed to cover users on the ground; thus, coverage holes in the sky will exist. In this paper, we investigate the problem of path design for cellular-connected UAVs, taking into account the interruption performance throughout the UAV mission to minimize the completion time. Two types of connectivity constraints requirements are assumed to be available. The first is defined as the maximum continuous time interval that the UAV loses connection with base stations (BSs) below a predefined threshold. For the second, we consider the sum outage of UAV is limited during the entire UAV mission. The UAV is tasked with flying from a starting location to a final destination while minimization the mission time, satisfying the two constraints, separately. The formulated path design problem which involves continues variables and a dynamic radio environment, is not convex and thus is extremely difficult to solve directly. To tackle this challenge, a deep reinforcement learning (DRL) based trajectory design algorithm is proposed, where the Dueling Double Deep Q Network(Dueling DDQN) with multi-steps learning method is applied. Simulation results demonstrate the effectiveness of the proposed DRL algorithm and achieve a trade-off between the trajectory length of the UAV and connection quality.     

DRL-M4MR: An intelligent multicast routing approach based on DQN deep reinforcement learning in SDN

Traditional multicast routing methods have some problems in constructing a multicast tree. These problems include limited access to network state information, poor adaptability to dynamic and complex changes in the network, and inflexible data forwarding. To address these defects, the optimal multicast routing problem in software-defined networking (SDN) is tailored as a multiobjective optimization problem, and DRL-M4MR, an intelligent multicast routing algorithm based on the deep Q network (DQN) deep reinforcement learning (DRL) method is designed to construct a multicast tree in a software-defined network. First, combining the characteristics of SDN global network-aware information, the multicast tree state matrix, link bandwidth matrix, link delay matrix and link packet loss rate matrix are designed as the state space of the reinforcement learning agent to solve the problem in that the original method cannot make full use of network status information. Second, the action space of the agent is all the links in the network, and the action selection strategy is designed to add the links to the current multicast tree in four cases. Third, single-step and final reward function forms are designed to guide the agent to make decisions to construct the optimal multicast tree. The double network architectures, dueling network architectures and prioritized experience replay are adopted to improve the learning efficiency and convergence of the agent. Finally, after the DRL-M4MR agent is trained, the SDN controller installs the multicast flow entries by reversely traversing the multicast tree to the SDN switches to implement intelligent multicast routing. The experimental results show that, compared with existing algorithms, the multicast tree constructed by DRL-M4MR can obtain better bandwidth, delay, and packet loss rate performance after training, and it can make more intelligent multicast routing decisions in a dynamic network environment. Code and DRL model are available at https://github.com/GuetYe/DRL-M4MR.     

Deep learning and expert knowledge based underwater acoustic OFDM receiver

In this paper, a deep learning and expert knowledge based receiver is proposed for underwater acoustic (UWA) orthogonal frequency division multiplexing (OFDM). Different from the existing deep learning based UWA OFDM receivers, the proposed receiver combines deep learning with the classical expert knowledge of block-based signal processing in UWA OFDM to improve system performance and interpretability. It performs joint channel estimation and signal detection by designing skip connection (SC) convolutional neural network (CNN) cascaded attention mechanism (AM) enhanced bi-directional long short-term memory (BiLSTM) network, abbreviated as SC-CNN-AM-BiLSTM network (SCABNet). Specifically, the channel estimation subnet is designed with SC-CNN to utilize the thought of image super-resolution to reconstruct the entire channel frequency response of all subcarriers. The signal detection subnet is designed with AM-BiLSTM to extract the correlations of received sequential data for signal detection. Especially with the AM, the signal detection subnet can focus more on effective information of the received distorted signal to train the optimal network weights to improve the accuracy of data recovery. The proposed SCABNet is evaluated by experimental data, and the results have demonstrated that the SCABNet has the lowest BER and robust performance compared to the traditional linear algorithm, deep learning based black-box receiver, and ComNet receiver. And the proposed SCABNet is effective and robust when multiple nonideal factors co-exist.     

水下目标多模态深度学习分类识别研究

水下目标的分类识别对于水声探测具有重要意义。提出一种水下目标多模态深度学习分类识别方法。针对水声信号的一维时域模态和二维频域模态特征建立一种多模态特征融合的深度学习结构,结合长短时记忆网络和卷积神经网络的优点,对一维时域信号和二维频谱信号分别进行并行处理,对输出进行典型相关分析,形成特征融合表示,并利用相邻帧的相关性进行参数优化。利用实测水声信号对算法进行了验证。结果表明：提出的算法对于水下目标识别的精度有显著的提高。     

应用深度学习识别法兰螺栓连接状态

针对常规诊断方法对螺栓的连接状态识别效果差、鲁棒性和抗噪性弱等问题,提出了基于深度学习理论的螺栓检测新方法.首先以4种预紧力状态下的法兰螺栓结构产生的声发射信号为研究对象,借助于自适应噪声的完整集成经验模态分解理论以及梅尔频率倒谱系数特征提取方式,实现了声发射信号的自适应消噪和最优模态函数分量组的选取,提取到了可以较好...     

Relay selection scheme based on deep reinforcement learning in wireless sensor networks

Cooperative communication technology has realized the enhancement in the wireless communication system’s spectrum utilization rate without resorting to any additional equipment; additionally, it ensures system reliability in transmission, increasingly becoming a research focus within the sphere of wireless sensor networks (WSNs). Since the selection of relay is crucial to cooperative communication technology, this paper proposes two different relay selection schemes subject to deep reinforcement learning (DRL), in response to the issues in WSNs with relay selection in cooperative communications, which can be summarized as the Deep-Q-Network Based Relay Selection Scheme (DQN-RSS), as well as the Proximal Policy Optimization Based Relay Selection Scheme (PPO-RSS); it further compared the commonly used Q-learning relay selection scheme (Q-RSS) with random relay selection scheme. First, the cooperative communication process in WSNs is modeled as a Markov decision process, and DRL algorithm is trained in accordance with the outage probability, as well as mutual information (MI). Under the condition of unknown instantaneous channel state information (CSI), the best relay is adaptively selected from multiple candidate relays. Thereafter, in view of the slow convergence speed of Q-RSS in high-dimensional state space, the DRL algorithm is used to accelerate the convergence. In particular, we employ DRL algorithm to deal with high-dimensional state space while speeding up learning. The experimental results reveal that under the same conditions, the random relay selection scheme always has the worst performance. And compared to Q-RSS, the two relay selection schemes designed in this paper greatly reduce the number of iterations and speed up the convergence speed, thereby reducing the computational complexity and overhead of the source node selecting the best relay strategy. In addition, the two relay selection schemes designed and raised in this paper are featured by lower-level outage probability with lower-level energy consumption and larger system capacity. In particular, PPO-RSS has higher reliability and practicability.     

基于迭代学习的离散切换系统故障估计

针对一类具有任意切换序列的离散切换系统的故障估计问题,提出了一种新的故障估计算法.该算法利用引入的虚拟故障信号构建出故障估计器,并利用残差信号通过迭代学习方法对引入的虚拟故障进行逐次修正.使虚拟故障随着迭代次数的增加逐渐逼近实际故障.利用压缩映射方法严格证明了算法在各个子区间上的收敛性,给出了算法的收敛条件.理论分析表明,所提算法能够在有限区间上精确估计出切换系统发生的不同类型故障.最后通过仿真实验进一步验证了所提算法的有效性.     

A deep learning ball tracking system in soccer videos

Increasing interest, enthusiasm of sport lovers, and economics involved offer high importance to sports video recording and analysis. Being crucial for decision making, ball detection and tracking in soccer has become a challenging research area. This paper presents a novel deep learning approach for 2D ball detection and tracking (DLBT) in soccer videos posing various challenges. A new 2-stage buffer median filtering background modelling is used for moving objects blob detection. A deep learning approach for classification of an image patch into three classes, i.e. ball, player, and background is initially proposed. Probabilistic bounding box overlapping technique is proposed further for robust ball track validation. Novel full and boundary grid concepts resume tracking in ball_track_lost and ball_out_of_frame situations. DLBT does not require human intervention to identify ball from the initial frames unlike the most published algorithms. DLBT yields extraordinary accurate and robust tracking results compared to the other contemporary 2D trackers even in presence of various challenges including very small ball size and fast movements.     

基于NPU的实时深度学习跟踪算法实现北大核心CSCD

深度学习在检测领域高速发展,但受限于训练数据和计算效率,在基于嵌入式平台的边缘计算领域,尤其是实时跟踪应用中深度学习的智能化算法应用并不广泛。针对这一现象,同时为满足现阶段国产化、智能化的技术需求,提出了一种改进的孪生网络深度学习跟踪算法。在特征网络加入微调网络,解决了网络模型无法在线更新的问题,提升了跟踪的准确性;在IoUNet损失函数中加入中心距离惩罚项,解决了IoUNet当IoU相同时位置跳跃,存在收敛盲区和收敛速度慢的问题;将训练后的网络通过通道剪枝,缩减网络模型尺寸,提升了模型加载和运行的速度。在华为Atlas200NPU平台上实现了实时运行,算法准确率高达0.90(IoU>0.7),帧率达到66 Hz。