学报信息
出版视野
友情链接
网站访问量
今日访问量: 3
高光谱目标检测：从传统方法到人工智能的演化
Hyperspectral Target Detection:The Evolution from Conventional to Artificial Intelligence
张志利;孙海翔;丁遥;鲁振航;李怀淳;李健;吴俊豪; ZHANG Zhili;SUN Haixiang;DING Yao;LU Zhenhang;LI Huaichun;LI Jian;WU Junhao;为系统剖析高光谱目标检测（hyperspectral target detection,HTD）技术由传统数学建模向深度学习框架演进的发展脉络，定量评估了各类代表性算法的性能表现。在研究方法上，横向对比了约束能量最小化与自适应匹配滤波等经典模型，深入解析了三大前沿智能机制的核心创新：一是将卷积神经网络（convolutional neural network,CNN）作为基于空间-光谱联合表征的非线性关系度量器；二是利用Transformer的多头自注意力机制结合生成式自监督范式，准确捕获全局长程光谱依赖；三是依托自编码器潜在空间中的多维物理约束，实现微弱目标与复杂背景的深度解耦。基于Sandiego1、Sandiego2与Cuprite数据集，以特征曲线下面积（area under curve, AUC）为核心评价指标，对10种不同算法开展了实验验证。结果表明：受制于单像素特征拟合与高维冗余，传统方法的AUC值均普遍难以突破0.81；相反，深度学习模型的AUC则全面跃升至0.86以上。具体而言，CNN架构的SCL-HTD模型凭借卓越的长距离相关性捕获能力，在背景高度混杂的Sandiego2数据集上的AUC取到最大值0.897 4；基于CNN的S2ADet算法在空间目标分布规则的Cuprite数据集中AUC值达到0.894 5；而融入无监督重建机制的自编码类模型NUN-UTD同样在Sandiego1中取得了AUC值高达0.895 2的最佳表现。由于单一网络模型已触及特征表征的瓶颈，未来高光谱探测技术的破局点在于构建自适应混合架构及多粒度特征融合机制，并向无监督与弱监督范式全面倾斜，以此来攻克复杂环境泛化与小样本特征迁移的问题。
To systematically analyze the development trajectory of hyperspectral target detection(HTD)technology from traditional mathematical modeling to deep learning frameworks, the performance of various representative algorithms was quantitatively evaluated. The study performed a horizontal comparison of classic models such as constrained energy minimization and adaptive matched filtering, as well as comprehensively analyzed the core innovations of three cutting-edge intelligent mechanisms: reshaping a convolutional neural network(CNN)into nonlinear relationship measurers based on joint spatial-spectral representation; precisely capturing long-range spectral dependencies across the entire image by combining the multi-head self-attention mechanism of the Transformer with a generative self-supervised paradigm;and achieving deep decoupling of weak targets from complex backgrounds through multi-dimensional physical constraints in the latent space of autoencoders. Based on the Sandiego1, Sandiego2, and Cuprite datasets, a detailed validation of 10 different algorithms using the area under the curve(AUC)as the core evaluation metric was conducted. The results indicate that, constrained by single-pixel feature fitting and high-dimensional redundancy, traditional methods generally struggle to exceed an AUC of 0.81; in contrast, deep learning models achieve a comprehensive AUC exceeding 0.86. The CNN-based SCLHTD model achieves a peak AUC of 0.897 4 on the mixed-background Sandiego2 dataset due to its outstanding long-range correlation capture ability; the CNN-based S2 ADet algorithm achieves an AUC of 0.894 5 on the spatially regular target-distributed Cuprite dataset; and the unsupervised reconstruction-integrated autoencoder model NUN-UTD also achieves the best AUC of 0.8952 on the Sandiego1 dataset. As single network models have reached the bottleneck of feature representation, the future breakthrough in hyperspectral detection technology lies in constructing adaptive hybrid architectures and multi-granularity feature fusion mechanisms, with a shift toward unsupervised and weakly supervised paradigms to address the problems of complex environment generalization and few-shot feature transfer.
机器学习赋能有限样本数据下锂离子电池健康状态估计方法综述
Review of Machine Learning-Based Methods for Lithium-ion Battery State of Health Estimation with Limited Sample Data
李爱华;王鹏翔;王涛;许晓东;项纪钧; LI Aihua;WANG Pengxiang;WANG Tao;XU Xiaodong;XIANG Jijun;锂离子电池健康状态（state of health,SOH）准确估计是保障电池系统安全运行与健康管理的关键技术。实际应用中，电池全生命周期实验周期长、标注成本高，且真实运行数据常表现为低标注样本、短时充电稀疏样本和工况覆盖不足，有限样本数据已成为制约机器学习SOH估计精度与泛化能力的重要因素。基于此，综述了有限样本数据下机器学习SOH估计方法的研究现状，重点分析了回归方法、浅层学习、深度学习与迁移学习4类方法的性能差异，并从估计精度、计算复杂度、部署条件和适用性等方面进行对比分析。结果表明，回归方法具有概率推理优势但计算成本较高，浅层学习结构简洁但外推能力有限，深度学习特征提取能力强但高度依赖样本规模，迁移学习在有限样本数据下表现更优，但仍面临负迁移和训练稳定性不足等问题。未来应重点突破无标签数据挖掘、物理约束融合、多智能感知及实际工况部署等难题，以实现稳健、可迁移且可解释的锂离子电池SOH估计。
Accurate state of health(SOH)estimation of lithium-ion batteries is a key technology for ensuring safe battery operation and health management. In practical applications, full-life-cycle battery experiments involve long durations and high labeling costs, whereas real-world operating data are often characterized by scarce labeled samples, sparse short-duration charging samples, and insufficient coverage of operating conditions. Therefore, limited sample data has become a major factor restricting the accuracy and generalization of machine learning-based SOH estimation. This study reviewed the current research status of machine learning-based SOH estimation methods with limited sample data, focusing on the performance differences of four categories of methods, including regression, shallow learning, deep learning, and transfer learning. The methods were compared in terms of their estimation accuracy, computational complexity, deployment requirements, and applicability. The results show that regression methods have advantages in probabilistic inference but suffer from high computational cost; shallow learning methods feature simple structures but are limited in their extrapolation capability; deep learning methods excel in feature extraction but rely heavily on the sample size; while transfer learning methods perform better under limited sample data, they still face problems such as negative transfer and training instability. Future research should focus on unlabeled data mining, physical constraint integration, multimodal sensing, and practical deployment to achieve robust, transferable, and interpretable SOH estimations for lithium-ion batteries.
新型发射方式中永磁同步电机控制方法综述
Review of Control Methods for Permanent Magnetic Synchronous Motors in New Launching Modes
陈桂明;许令亮;茶正忠;王保成; CHEN Guiming;XU Lingliang;CHA Zhengzhong;WANG Baocheng;在军事科技快速发展背景下，导弹新型发射方式对驱动系统性能提出了更高要求。永磁同步电机（permanent magnetic synchronous motor, PMSM）因其高功率密度、高效率及优异动态响应特性，在新型发射系统中得到广泛应用，其控制策略直接影响系统整体性能与可靠性。针对不同发射场景对电机控制性能的差异化需求，从转速外环、电流内环、转速-电流一体化单环控制及无传感器控制4个层面，对PMSM控制方法进行了系统综述。在转速外环控制方面，对自抗扰控制、滑模变结构控制及预测控制等方法进行了分析，这些方法在提升系统抗扰能力和动态响应方面具有显著优势；在电流内环控制方面，对滞环控制、反馈线性化控制、无模型及机器学习控制等方法进行了梳理，不同方法在控制精度与实现复杂度之间呈现出不同权衡关系；在转速-电流一体化控制方面，分析了无源控制、模型预测控制及反步控制等方法在简化控制结构与提升动态性能方面的应用潜力；在无传感器控制方面，对基于基波数学模型和高频信号注入方法进行了对比分析，揭示了其在不同转速区间的适用特性。最后，对不同控制方法的优缺点及适用场景进行了梳理总结，并探讨了PMSM控制技术的发展趋势。
Against the backdrop of rapid advancements in military technology, new missile launch modes impose higher demands on drive system performance. Permanent magnetic synchronous motors(PMSMs)are widely employed in novel launch systems owing to their high power density, efficiency,and excellent dynamic response characteristics, with their control strategies directly influencing the overall performance and reliability of the system. To address the differential requirements for motor control performance under various launch scenarios, this study provided a systematic review of PMSM control methods from four perspectives: speed outer-loop, current inner-loop, integrated single loop control speed-current,and sensorless control. In the realm of speed loop control, methods such as active disturbance rejection,sliding mode variable structure, and model predictive controls were analyzed, highlighting their significant advantages in enhancing disturbance rejection capability and dynamic response. For current inner-loop control, methods including hysteresis, feedback linearization, model-free, and machine learning-based controls were reviewed, revealing varying trade-offs between control accuracy and implementation complexity.In terms of integrated speed-current control, the application potential of passivity-based, model predictive,and backstepping controls in simplifying control structures and improving dynamic performance was discussed. For sensorless control, a comparative analysis of fundamental wave model-based and high-frequency signal injection methods was conducted, demonstrating their suitability across different speed ranges. Finally, the advantages, disadvantages, and applicable scenarios of different control methods were summarized, and future development trends in PMSM control technology were explored.
基于散射功率驱动自适应邻域基三维类高斯核的多时相PolSAR图像边缘检测
Edge Detection for Multi-Temporal Polarimetric Synthetic Aperture Radar Images Based on Span-Driven Adaptive Neighborhood-Based Three-Dimensional Gaussian-Like Kernel
郑晓龙;王钰清;赵维恒;肖政举;康帅直; ZHENG Xiaolong;WANG Yuqing;ZHAO Weiheng;XIAO Zhengju;KANG Shuaizhi;针对多时相极化合成孔径雷达（polarimetric synthetic aperture radar,PolSAR）图像边缘检测存在的时变散射表征不足、协方差估计偏差、时空信息融合低效等问题，提出一种基于散射功率驱动自适应邻域（span-driven adaptive neighborhood,SDAN）基三维类高斯核的多时相PolSAR图像边缘检测方法。首先，采用球不变随机向量（spherically invariant random vector,SIRV）模型对PolSAR数据进行统计表征，实现极化分量协方差矩阵与纹理分量的联合建模；其次，在二维高斯核的基础上，以SDAN为空间支撑构建基于SDAN的二维高斯核，有效提升异质区域协方差矩阵估计精度；最后，将SDAN基二维高斯核与时间维度的一维自适应卷积核加权融合，形成SDAN基三维类高斯核，并结合自适应滞后阈值完成边缘图优化。实验结果表明：所提方法在两组数据集上的精确率分别达到0.84和0.94，召回率分别达到0.79和0.82，整体效率优于现有主流边缘检测方法，在时变散射区域边缘检测中具有显著优势。
To address the issues of insufficient temporal scattering characterization, covariance estimation bias, and inefficient spatiotemporal information fusion in edge detection for multi-temporal polarimetric synthetic aperture radar(PolSAR)images, an edge detection method based on span-driven adaptive neighborhood(SDAN)-based three-dimensional Gaussian kernel was proposed. First, the spherically invariant random vector model was employed to statistically characterize the PolSAR data, achieving joint modeling of polarimetric covariance matrices and texture components. Subsequently, building upon the two-dimensional Gaussian kernel, an SDAN-based two-dimensional Gaussian kernel was constructed with the SDAN as the spatial support, effectively improving the covariance matrix estimation accuracy in heterogeneous regions. Finally, the SDAN-based two-dimensional Gaussian kernel was weighted and fused with a one-dimensional adaptive convolutional kernel along the temporal dimension to form an SDAN-based three-dimensional Gaussian kernel, which was combined with an adaptive lag threshold to optimize the edge map.The results demonstrate that the proposed method achieves precision rates of 0.84 and 0.94, recall rates of 0.79 and 0.82, respectively. The overall efficiency is superior to that of existing mainstream edge detection methods, exhibiting significant advantages in edge detection for temporally varying scattering regions.
自适应多跳图卷积与Transformer融合的高光谱图像分类方法
Adaptive Multi-Hop Graph Convolution and Transformer Fusion Network for Hyperspectral Image Classification
周正阳;李文浩;王念;朱良玉;崔智高; ZHOU Zhengyang;LI Wenhao;WANG Nian;ZHU Liangyu;CUI Zhigao;针对高光谱图像分类中传统图卷积网络（graph convolutional network,GCN）感受野受限、深层堆叠易导致过平滑以及空谱交互存在信息冗余的问题，提出一种自适应多跳图卷积与Transformer融合的高光谱图像分类方法。该方法设计了自适应多跳图注意力层，通过可学习的注意力向量对0至K跳邻域特征进行自适应加权融合，同时构建了增强型GCN-Transformer融合接口，引入通道注意力机制和一维邻域卷积模块，从而有效压缩了节点嵌入的通道冗余并增强了光谱判别性。在Indian Pines、Salinas和Pavia University数据集上的实验结果表明，所提方法的分类精度分别达到96.05%、96.92%和96.28%，相比经典GTFN方法分别提升了2.05、0.11和1.14个百分点，表明多跳空间拓扑建模与增强光谱序列融合能有效克服高光谱“维度灾难”，显著提升模型对不规则边界的刻画精度；当引入空间图建模时，Transformer模型的总体精度提升了27.91个百分点，且当跳数K=3时模型的自适应加权分类性能达到最优，表明自适应多跳图注意力机制能在较低算力开销下拓宽空间视野并利用权重衰减规律有效降低过平滑风险；在通道注意力模块中引入特征接口后，模型的综合性能达到最优，表明增强型接口能有效压缩交互阶段的信息冗余，大幅增强模型对高度相似地物光谱的判别能力。
To address the issues of limited receptive fields in conventional graph convolutional networks(GCNs), over-smoothing caused by deep network stacking, and information redundancy in spatial-spectral interactions for hyperspectral image classification, an adaptive multi-hop graph convolution and Transformer fusion network was proposed. An adaptive multi-hop graph attention layer was designed to achieve the adaptive weighted fusion of 0-to K-hop neighborhood features via learnable attention vectors. Concurrently, an enhanced GCN-Transformer fusion interface was constructed by integrating a channel attention mechanism and a one-dimensional neighborhood convolution module, thereby effectively compressing the channel redundancy of the node embeddings and elevating the spectral discriminability. The results on the Indian Pines, Salinas, and Pavia University datasets demonstrate that the proposed method achieves classification accuracies of 96.05%, 96.92%, and 96.28%, outperforming the classic GTFN method by 2.05,0.11, and 1.14 percentage points, respectively. These results confirm that the synergistic effect of multihop spatial topology modeling and enhanced spectral sequence fusion can effectively overcome the “curse of dimensionality” in hyperspectral and significantly improve the delineation accuracy of irregular boundaries.The introduction of spatial graph modeling boosts the overall accuracy by 27.91 percentage points, with adaptive weighting achieving optimal performance at a hop count of K=3, indicating that the adaptive multihop graph attention mechanism can broaden the spatial perspectives with lower computational overhead and reduce over-smoothing risks by utilizing learned weight decay regularities. By incorporating a feature interface into the channel attention module, the model yields the best comprehensive performance, demonstrating that the enhanced interface effectively compresses information redundancy during the interaction phase,which strengthens the discriminative capability for highly similar spectral land covers.
