Wavelet forced de-noising algorithm is suitable for denoising of unsteady drilling fluid pulse signal, including baseline drift rectification and two-stage de-noising processing of frame synchronization signal and ins...Wavelet forced de-noising algorithm is suitable for denoising of unsteady drilling fluid pulse signal, including baseline drift rectification and two-stage de-noising processing of frame synchronization signal and instruction signal. Two-stage de-noising processing can reduce the impact of baseline drift and determine automatic peak detection threshold range for signal recognition by distinguishing different features of frame synchronization pulse and instruction pulse. Rising and falling edge relative protruding threshold is defined for peak detection in signal recognition, which can make full use of the degree of the signal peak change and detect peaks flexibly with rising and falling edge relative protruding threshold combination. A synchronous decoding method was designed to reduce position uncertainty of the frame synchronization pulse and eliminate the accumulative error of time base drift, which determines the first instruction pulse position according to position of the frame synchronization pulse and decodes subsequent instruction pulse by taking current instruction pulse as new bit synchronization pulse. Special tool software was developed to tune algorithm parameters, which has a decoding success rate of about 95% for the universal coded signals. For the special coded signals with check byte, decoding success rate using the automatic threshold adjustment algorithm is as high as 99%.展开更多
Automatic modulation classification(AMC)is an essential technique in both civil and military applications.While deep learning has surpassed traditional methods in accuracy,distinguishing high-order modulations remain ...Automatic modulation classification(AMC)is an essential technique in both civil and military applications.While deep learning has surpassed traditional methods in accuracy,distinguishing high-order modulations remain challenging.Current efforts prioritize complex network designs,neglecting the integration of deep features and tailored feature engineering to reslove high-order ambiguities.Therefore,a multi-feature extraction framework is proposed,which directly concatenates the deep feature extracted by a newly designed lightweight neural network and the proposed spectrum secondary features or de-noised high-order statistical features.The proposed features and lightweight network both demonstrate superior overall accuracy than other competing features or networks.Furthermore,the effectiveness of the feature extraction framework is also validated.The average classification accuracy on high-order modulation sets reaches 67.39% on the RadioML2018.01A dataset,increasing more than 2%compared with the other competitive networks under the framework.The results indicate the effectiveness of the proposed feature extraction framework for its representational ability by combing the deep features with the proposed domain features.展开更多
为了解决自动文本摘要任务存在的文本语义信息不能充分编码、生成的摘要语义冗余、原始语义信息丢失等语义问题,提出了一种融合知识和文本语义信息的双编码器自动摘要模型(dual-encoder automatic summarization model incorporating kn...为了解决自动文本摘要任务存在的文本语义信息不能充分编码、生成的摘要语义冗余、原始语义信息丢失等语义问题,提出了一种融合知识和文本语义信息的双编码器自动摘要模型(dual-encoder automatic summarization model incorporating knowledge and semantic information,KSDASum)。该方法采用双编码器对原文语义信息进行充分编码,文本编码器获取全文的语义信息,图结构编码器维护全文上下文结构信息。解码器部分采用基于Transformer结构和指针网络,更好地捕捉文本和结构信息进行交互,并利用指针网络的优势提高生成摘要的准确性。同时,训练过程中采用强化学习中自我批判的策略梯度优化模型能力。该方法在CNN/Daily Mail和XSum公开数据集上与GSUM生成式摘要方法相比,在评价指标上均获得最优的结果,证明了所提模型能够有效地利用知识和语义信息,提升了生成文本摘要的能力。展开更多
基于采样的快速扩展随机树(Rapidly-exploring Random Tree,RRT)算法在面向智能车复杂障碍环境中进行运动规划时,普遍存在采样冗余、规划效率低及路径质量欠佳等问题。鉴于此,提出一种将深度学习与RRT算法相结合的DFNN-RRT最优运动规划...基于采样的快速扩展随机树(Rapidly-exploring Random Tree,RRT)算法在面向智能车复杂障碍环境中进行运动规划时,普遍存在采样冗余、规划效率低及路径质量欠佳等问题。鉴于此,提出一种将深度学习与RRT算法相结合的DFNN-RRT最优运动规划方法。通过自动点云编码器实现复杂障碍物场景的拓扑特征提取与高维数据编码,将预处理后的环境点云数据嵌入到障碍物空间表征;将障碍空间特征与智能车起始位姿、目标位姿等状态空间变量进行多模态融合,构建时空联合输入表征以驱动规划网络训练。在前向传播过程中,网络通过渐进式优化生成具有目标导向性的知情采样样本;以规划点时间间隔为成本函数核心参数,量化路径生长过程的动态代价。同时构建后验预测状态与目标状态间的加权均方误差损失函数,通过梯度优化实现模型高效收敛。试验结果表明:损失率阈值为0.3时,全局最优解概率最高;5 Hz规划频率下,系统耗时最短;采样规模与计算耗时呈显著正相关。在相同环境配置下,DFNN-RRT在保持RRT算法概率完备性的同时,路径质量指标显著优化:相较于基准RRT算法,计算耗时、单位路径长度、路径曲率及采样密度分别优化38.5%,7.8%、42.85%和52.5%;横向对比偏置RRT、双向RRT时效提升27.1%和12.2%。通过仿真与实车测试验证,该方法在复杂障碍物场景下可稳定输出平滑轨迹,证明该方法具有时效性与可靠性优势。展开更多
航海领域现有的轨迹相似性度量方法多以传统方法为主,计算复杂度较高,尽管已经提出了一些基于深度学习的方法,但存在时空联合建模不足的问题,导致相似性度量的准确性和鲁棒性有待提升。针对上述问题,提出MDU-Net(Marine Density U-Net)...航海领域现有的轨迹相似性度量方法多以传统方法为主,计算复杂度较高,尽管已经提出了一些基于深度学习的方法,但存在时空联合建模不足的问题,导致相似性度量的准确性和鲁棒性有待提升。针对上述问题,提出MDU-Net(Marine Density U-Net)模型。该模型能够自动提取船舶轨迹的低维特征,从而高效可靠地检索与指定目标相似的轨迹。首先,对轨迹数据进行等时间间隔插值,再采用核密度估计(KDE)生成融合空间与速度信息的核密度灰度图,实现轨迹像素化。然后,采用基于U-Net结构的神经网络进行无监督学习,获得轨迹的低维表示。最后,通过计算低维特征向量间的余弦距离构建相似矩阵,量化轨迹间的相似性。实验结果表明,MDU-Net模型在评估指标上显著优于传统模型与主流深度学习模型,与经典动态时间规整(DTW)、Hausdorff距离、卷积自编码器(CAE)模型相比,MDU-Net模型前10条轨迹命中率提升了7.691、14.741、25.191百分点,充分验证了MDU-Net模型在船舶轨迹相似性度量任务中的优越效果。展开更多
基金Supported by the China National Science and Technology Major Project(2016ZX05020005-001)
文摘Wavelet forced de-noising algorithm is suitable for denoising of unsteady drilling fluid pulse signal, including baseline drift rectification and two-stage de-noising processing of frame synchronization signal and instruction signal. Two-stage de-noising processing can reduce the impact of baseline drift and determine automatic peak detection threshold range for signal recognition by distinguishing different features of frame synchronization pulse and instruction pulse. Rising and falling edge relative protruding threshold is defined for peak detection in signal recognition, which can make full use of the degree of the signal peak change and detect peaks flexibly with rising and falling edge relative protruding threshold combination. A synchronous decoding method was designed to reduce position uncertainty of the frame synchronization pulse and eliminate the accumulative error of time base drift, which determines the first instruction pulse position according to position of the frame synchronization pulse and decodes subsequent instruction pulse by taking current instruction pulse as new bit synchronization pulse. Special tool software was developed to tune algorithm parameters, which has a decoding success rate of about 95% for the universal coded signals. For the special coded signals with check byte, decoding success rate using the automatic threshold adjustment algorithm is as high as 99%.
基金supported by the National Natural Science Foundation of China(12273054).
文摘Automatic modulation classification(AMC)is an essential technique in both civil and military applications.While deep learning has surpassed traditional methods in accuracy,distinguishing high-order modulations remain challenging.Current efforts prioritize complex network designs,neglecting the integration of deep features and tailored feature engineering to reslove high-order ambiguities.Therefore,a multi-feature extraction framework is proposed,which directly concatenates the deep feature extracted by a newly designed lightweight neural network and the proposed spectrum secondary features or de-noised high-order statistical features.The proposed features and lightweight network both demonstrate superior overall accuracy than other competing features or networks.Furthermore,the effectiveness of the feature extraction framework is also validated.The average classification accuracy on high-order modulation sets reaches 67.39% on the RadioML2018.01A dataset,increasing more than 2%compared with the other competitive networks under the framework.The results indicate the effectiveness of the proposed feature extraction framework for its representational ability by combing the deep features with the proposed domain features.
文摘为了解决自动文本摘要任务存在的文本语义信息不能充分编码、生成的摘要语义冗余、原始语义信息丢失等语义问题,提出了一种融合知识和文本语义信息的双编码器自动摘要模型(dual-encoder automatic summarization model incorporating knowledge and semantic information,KSDASum)。该方法采用双编码器对原文语义信息进行充分编码,文本编码器获取全文的语义信息,图结构编码器维护全文上下文结构信息。解码器部分采用基于Transformer结构和指针网络,更好地捕捉文本和结构信息进行交互,并利用指针网络的优势提高生成摘要的准确性。同时,训练过程中采用强化学习中自我批判的策略梯度优化模型能力。该方法在CNN/Daily Mail和XSum公开数据集上与GSUM生成式摘要方法相比,在评价指标上均获得最优的结果,证明了所提模型能够有效地利用知识和语义信息,提升了生成文本摘要的能力。
文摘基于采样的快速扩展随机树(Rapidly-exploring Random Tree,RRT)算法在面向智能车复杂障碍环境中进行运动规划时,普遍存在采样冗余、规划效率低及路径质量欠佳等问题。鉴于此,提出一种将深度学习与RRT算法相结合的DFNN-RRT最优运动规划方法。通过自动点云编码器实现复杂障碍物场景的拓扑特征提取与高维数据编码,将预处理后的环境点云数据嵌入到障碍物空间表征;将障碍空间特征与智能车起始位姿、目标位姿等状态空间变量进行多模态融合,构建时空联合输入表征以驱动规划网络训练。在前向传播过程中,网络通过渐进式优化生成具有目标导向性的知情采样样本;以规划点时间间隔为成本函数核心参数,量化路径生长过程的动态代价。同时构建后验预测状态与目标状态间的加权均方误差损失函数,通过梯度优化实现模型高效收敛。试验结果表明:损失率阈值为0.3时,全局最优解概率最高;5 Hz规划频率下,系统耗时最短;采样规模与计算耗时呈显著正相关。在相同环境配置下,DFNN-RRT在保持RRT算法概率完备性的同时,路径质量指标显著优化:相较于基准RRT算法,计算耗时、单位路径长度、路径曲率及采样密度分别优化38.5%,7.8%、42.85%和52.5%;横向对比偏置RRT、双向RRT时效提升27.1%和12.2%。通过仿真与实车测试验证,该方法在复杂障碍物场景下可稳定输出平滑轨迹,证明该方法具有时效性与可靠性优势。
文摘航海领域现有的轨迹相似性度量方法多以传统方法为主,计算复杂度较高,尽管已经提出了一些基于深度学习的方法,但存在时空联合建模不足的问题,导致相似性度量的准确性和鲁棒性有待提升。针对上述问题,提出MDU-Net(Marine Density U-Net)模型。该模型能够自动提取船舶轨迹的低维特征,从而高效可靠地检索与指定目标相似的轨迹。首先,对轨迹数据进行等时间间隔插值,再采用核密度估计(KDE)生成融合空间与速度信息的核密度灰度图,实现轨迹像素化。然后,采用基于U-Net结构的神经网络进行无监督学习,获得轨迹的低维表示。最后,通过计算低维特征向量间的余弦距离构建相似矩阵,量化轨迹间的相似性。实验结果表明,MDU-Net模型在评估指标上显著优于传统模型与主流深度学习模型,与经典动态时间规整(DTW)、Hausdorff距离、卷积自编码器(CAE)模型相比,MDU-Net模型前10条轨迹命中率提升了7.691、14.741、25.191百分点,充分验证了MDU-Net模型在船舶轨迹相似性度量任务中的优越效果。
