2025年 04期
基于强化学习的改进RRT~*路径规划
Improved RRT~* Path Planning Based on Reinforcement Learning
张艳珠;侯亢钧;陈勇;李婷雪;李妍; ZHANG Yanzhu;HOU Kangjun;CHEN Yong;LI Tingxue;LI Yan;针对RRT~*算法在路径规划中面临搜索效率不高、易于陷入局部最优等问题，提出一种结合强化学习的Q-RRT~*算法。该算法将Q-Learning算法和RRT~*算法相融合，首先引入转角偏向策略增强路径搜索时的导向作用、减少无效节点的生成，提升算法的搜索效率；其次通过动R搜索算法动态地调整搜索半径，进一步优化路径的质量和冗余节点的产生；最后对生成的路径使用三次B样条插值法和冗余节点删除法进一步优化路径质量。在二维和三维环境下的仿真实验结果表明，改进后的Q-RRT~*算法和RRT、RRT~*和RL-RRT算法相比，路径规划时长平均快39.7%,迭代次数平均减低27.9%,路径长度平均缩短16.3%。
In order to solve the problems of low search efficiency and tendency to fall into local optimum in RRT~* path planning, a Q-RRT~* algorithm combined with reinforcement learning was proposed, which fused the Q-Learning algorithm and the RRT~* algorithm.Firstly, the corner bias strategy was introduced to enhance the guiding effect of path search, reduce the generation of invalid nodes, and improve the search efficiency of the algorithm.Secondly, the search radius was dynamically adjusted by the dynamic R search algorithm to further optimize the quality of the path and the generation of redundant nodes.Finally the cubic B-spline interpolation method and the redundant node deletion method were used to further optimize the path quality of the generated path.Simulation results in 2D and 3D environments show that the improved Q-RRT~* algorithm is 39.7%faster on average, 27.9%less iteration and 16.3%shorter in path length than RRT,RRT~* and RL-RRT algorithms.
基于文本图表示学习的人格分类方法
Personality Classification Method Based on Text Graph for the Representation of Learning
刘猛;范摇珊;刘芳;张德育;贡胜男; LIU Meng;FAN Yaoshan;LIU Fang;ZHANG Deyu;GONG Shengnan;针对网络用户的传统人格分类方法提取文本语义特征不充分、分类准确率低的问题，提出一种基于文本图表示学习的人格分类方法。该方法利用自然语言处理技术，并结合深度学习和图网络模型，设计一种自适应图卷积网络(adaptive graph convolutional network, ADGCN),通过自适应调整机制优化节点表示，平衡了节点特征的局部与全局信息。在Kaggle数据集上的测试实验表明，F1分数最高为80%,且平均F1分数达到71.14%,比传统机器学习方法和预训练模型BERT提高近20%,展现了模型计算效率上的优越性。
To solve the problems of insufficient text semantic features and low classification accuracy of traditional personality classification methods for network users, a personality classification method based on text graph representation learning is proposed.This method uses natural language processing technology, combined with deep learning and graph network model, to build a new type of network user personality classification model, and designs an adaptive graph convolutional network(ADGCN).The node representation is optimized by an adaptive adjustment mechanism, which balances the local and global information of node features.Experiments on the Kaggle dataset show that the F1 score is up to 80%,and the average F1 score reaches 71.14%,which is nearly 20% higher than the traditional machine learning method and BERT pre-training model, showing the superiority of the model's computational efficiency.
基于改进BiGRU-TCN混合模型的风机轴承温度异常预警方法
Anomaly Warning Method for Wind Turbine Bearing Temperature Based on Improved BiGRU-TCN Hybrid Model
张佳;关启学;姜月秋; ZHANG Jia;GUAN Qixue;JIANG Yueqiu;为解决风力发电机轴承温度预测准确性较低而影响故障预警系统性能的问题，提出一种基于改进双向门控循环单元(bidirectional gated recurrent unit, BiGRU)和时间卷积网络(temporal convolutional network, TCN)的风机轴承温度异常预警方法(BiGRU-TCN)。首先采用bin方法对噪声数据进行清洗，减小其对预测模型准确性的干扰；然后引入TCN捕捉序列依赖性，并结合BiGRU建立融合模型，对清洗后数据进行特征提取，再加入自注意力机制，提高模型在数据波动幅度较大时的预测能力；最后采用滑动窗口算法分析预测值与真实值之间的残差，设置故障预警阈值。实验结果显示：相较于其他常见模型，本文模型预测结果的平均绝对误差(MAE)平均低0.571,均方误差(MSE)平均低3.601;基于本文模型设置的预警方式实现了在异常发生前3天预警，为风电场的运维管理提供了有力支持。
To address the issue of low accuracy in predicting wind turbine bearing temperatures, which negatively affects the performance of fault warning systems, a wind turbine bearing temperature anomaly warning method based on an improved bidirectional gated recurrent unit(BiGRU)and temporal convolutional network(TCN)is proposed(BiGRU-TCN).First, noise data is cleaned using the bin method to reduce their interference with the prediction model's accuracy.Then, TCN is introduced to capture sequence dependencies, which is combined with BiGRU to establish a fusion model for feature extraction from the cleaned data.A self-attention mechanism is further incorporated to enhance the model's prediction capability under significant data fluctuations.Finally, a sliding window algorithm is applied to analyze the residuals between predicted and actual values, and a fault warning threshold is set.Experimental results show that, compared to other common models, the proposed model has an average reduction of 0.571 in mean absolute error(MAE)and an average reduction of 3.601 in mean squared error(MSE).The warning mechanism based on the proposed model successfully provides a warning three days before an anomaly occurs, offering strong support for the operation and maintenance management of wind farms.
一种面向工业物联网的轻量级共识机制
A Lightweight Consensus Mechanism for IIoT
周小明;刘微;张文波; ZHOU Xiaoming;LIU Wei;ZHANG Wenbo;传统区块链共识机制存在资源消耗大、效率低的问题，针对面向工业物联网的分布式网络结构提出一种基于信誉的轻量级共识机制(lightweight reputation-based consensus mechanism, LRBCM)。LRBCM算法的共识流程包括预共识和共识两个阶段：在预共识阶段，每个验证者对接收到的事务进行签名与信息合并，并使用Gossip协议传播，直至大多数验证者就事务的数量和顺序达成共识；在共识阶段，进行节点的局部信誉和全局信誉计算，选取全局信誉值最高的共识节点创建和发布当前区块链的区块，选取结果具有随机性，为认证过程中的数据安全提供了保障。仿真实验结果表明，LRBCM的计算时间复杂度较低，与ReCon机制相比，其吞吐量综合提升约11.79%,平均时延降低约12.87%,能够有效应用于面向工业物联网的分布式网络结构。
To address the issue of large resource consumption and low efficiency of traditional blockchain consensus mechanisms, a reputation-based lightweight consensus mechanism(LRBCM)is proposed for the distributed network structure of industrial internet of things(IIoT).The LRBCM includes a pre-consensus stage and a consensus phase.In the pre-consensus phase, each validator signs the received transactions and information merging, disseminates them using the Gossip protocol, until a majority consensus is achieved on the number and order of transactions.In the consensus stage, the local and global reputation calculation of the node is made and the node with the highest global reputation value is selected via the reputation model to create and publish the block of the current blockchain.This selection is random, providing a guarantee for data security during the authentication process.Simulation results demonstrate that LRBCM has lower complexity of computation time, with an approximately 11.79% increase in throughput compared to the ReCon mechanism and a reduction of about 12.87% in average delay, making itself a suitable solution for the distributed network structure of IIoT.
MSMA自感知执行器的信号解耦及实验验证
Signal Decoupling and Experimental Verification of MSMA Self-sensing Actuator
鲁军;刘文龙;魏凡博;宋金泽;李昱昊; LU Jun;LIU Wenlong;WEI Fanbo;SONG Jinze;LI Yuhao;基于磁控形状记忆合金(magnetically controlled shape memory alloy, MSMA)可逆特性研制的自感知执行器可用于振动的主动控制。然而，自感知执行器的传感线圈中混合有执行信号，需要对混合信号进行解耦才能提取出传感信号，从而实现振动的有效控制。为此，设计一种新的MSMA自感知执行器解耦方法，其采用变步长最小均方算法设计自适应噪声抵消器，该抵消器可实现步长因子的动态调整，从而提高算法收敛速度。搭建实验平台进行仿真测试，实验结果表明：自适应噪声抵消算法可有效实现混合信号的解耦，配合前馈-PID控制算法可在MSMA自感知执行器主动振动控制中获得较好的消振效果。
A self-sensing actuator developed based on the reversible properties of magnetically controlled shape memory alloy(MSMA)can be used for active vibration control.However, the sensing coil of the self-sensing actuator is mixed with execution signals, and the mixed signals need to be decoupled in order to extract the sensing signals and achieve effective vibration control.To this end, a new decoupling method for MSMA self-sensing actuators is designed, which uses the variable step size minimum mean square algorithm to design an adaptive noise cancellation device.This eliminator can dynamically adjust the step size factor, thereby improving the convergence speed of the algorithm.A test platform for simulation testing is built and the experimental results show that the adaptive noise cancellation algorithm can effectively achieve decoupling of mixed signals.Combined with the feedforward PID control algorithm, it can achieve good vibration damping effect in the active vibration control of MSMA self-sensing actuators.
基于ACFM的厚涂层下裂纹检测技术优化
Optimization of Thick Coating Crack Detection Technology Based on ACFM
隋涛;焦权明;李昀恒;张双楠; SUI Tao;JIAO Quanming;LI Yunheng;ZHANG Shuangnan;Shenyang Ligong University;针对现有交流电磁场检测(ACFM)技术对10 mm以上涂层的金属工件裂纹缺陷检测灵敏度低的问题，提出一种优化磁路的ACFM方法。通过有限元仿真软件建立厚涂层下ACFM仿真模型，分析不同检测频率和磁轭跨度在10 mm提离高度下缺陷信号的变化规律，选取最优探头参数组合以改变磁路。搭建检测系统进行实验验证，结果表明：采用仿真得到的探头参数组合，检测系统可实现对10 mm厚涂层下工件表面缺陷的有效检出；当涂层厚度大于4 mm时，选用30 mm跨度磁轭比选用10 mm跨度磁轭显示出更高的检测灵敏度，表明研究结果对10 mm以上厚涂层ACFM检测的开展具有参考价值。
An optimized magnetic circuit alternating current field measurement(ACFM)detection method is proposed for inspecting metal components with thick coatings, to improve the low sensitivity of current ACFM technology in detecting crack defects larger than 10 mm.A simulation model for ACFM under thick coatings is established using finite element simulation software, analyzing variations in defect signals at a 10 mm lift-off height with different detection frequencies and yoke spans.Optimal parameter combinations are selected to modify the magnetic circuit and construct a detection system for experimental validation.The results show that the probe parameters deriving from simulations enable effective detection of surface defects under 10 mm thick coatings with high sensitivity.For coating thicknesses exceeding 4 mm, a 30 mm yoke span exhibits higher detection sensitivity compared to a 10 mm span.These findings suggest that the research is valuable for advancing ACFM technology in detecting coatings thicker than 10 mm.
交通灯配时与车辆诱导协同控制技术研究
Research on Coordinated Control of Traffic Signal Timing and Vehicle Guidance
刘思萌;文峰; LIU Simeng;WEN Feng;智能交通系统(intelligent transportation systems, ITS)中城市交通控制由交通信号灯系统和车辆诱导系统各自独立并相互配合实现，交通信号灯系统在时间上完成交通灯的动态配时，车辆诱导系统在空间上完成交通车辆分流。为更好解决城市道路交通拥堵问题，提高车辆行驶效率，提出一种基于深度强化学习的交通灯配时与车辆诱导协同控制算法。通过交通灯配时系统和车辆诱导系统在信息生成、数据处理和策略执行等多个方面的数据互通，实现对交通灯智能体和车辆智能体的协同控制，从而提升整个路网的综合性能。在SUMO仿真器上的实验结果表明，该算法有效地提高了路网通行效率，且具有显著的稳定性。
In intelligent transportation systems, urban traffic control is achieved through the independent yet coordinated functioning of the traffic signal system and the vehicle guidance system.The traffic signal system dynamically adjusts signal timing over time, while the vehicle guidance system spatially manages traffic flow distribution.To more effectively address urban road traffic congestion and enhance vehicle travel efficiency, a cooperative control algorithm for traffic light timing and vehicle guidance based on deep reinforcement learning is proposed.This algorithm facilitates data exchange between the traffic light timing system and the vehicle guidance system in terms of information generation, data processing, and strategy execution, thereby achieving collaborative control of traffic light agents and vehicle agents, and improving the overall performance of the road network.Experimental results in the SUMO simulator indicate that the proposed algorithm effectively increases traffic efficiency and demonstrates significant stability.
融合局部-全局特征的加密流量分类模型
Encrypted Traffic Classification Model Fusing Local-global Features
田伟祥;张文波; TIAN Weixiang;ZHANG Wenbo;随着互联网技术的发展，对加密流量进行高效分类成为网络管理的重要手段，但现有的分类技术对加密流量的特征提取不充分，且仅关注单一的局部特征或全局特征而忽略了两者之间的有效融合，导致分类准确率较低。针对上述问题，提出了一种融合局部-全局特征的加密流量分类模型(local-global fusion model for encrypted traffic classification, LGF-ETC)。为解决特征提取不充分的问题，设计了特征增强模块(feature enhancement module, FEM),用于增强加密流量特征，以便在后续模型中对特征进行充分捕获；针对局部与全局特征的融合问题，从Swin Transformer网络中提取核心模块用于捕获全局特征，并设计了多尺度局部感知模块(multi-scale local perception module, MSLPM),将其嵌入Swin Transformer核心模块中，以捕获多尺度局部特征，进一步将两特征进行充分融合。实验结果表明，本文LGF-ETC模型的分类准确率达到98.87%,显著改善了现有模型在特征提取和特征融合方面的不足。
With the development of internet technology, efficient encrypted traffic classification has become an important means of network management, but the existing classification techniques do not fully extract the features of encrypted traffic and only pay attention to either local or global features, ignoring the effective fusion of the two, resulting in low classification accuracy.To address this issue, a local-global fusion model for encrypted traffic classification(LGF-ETC)is proposed.To solve the problem of insufficient feature extraction, a feature enhancement module(FEM)is designed to enhance the features of encrypted traffic, so that they can be fully captured in subsequent models.In addition, to address the problem of integrating local and global features, the core module of Swin Transformer network is extracted for capturing global features, and a multi-scale local perception module(MSLPM)is designed to be embedded in the core module of Swin Transformer to capture multi-scale local features, further fully integrating the two features.Experimental results show that the classification accuracy of the LGF-ETC model proposed in this paper is 98.87%,significantly improving the feature extraction and feature fusion of existing models.
长串联多体AUVs平稳直航的动力分配方法
Thrust Allocation Method for Stable Straight-line Navigation of Long Serially Connected Multi-body AUVs
丁睿;张进;高溪钠;伍泽江; DING Rui;ZHANG Jin;GAO Xina;WU Zejiang;State Grid Anshan Electric Power Supply Company;采用传统在线优化方法进行长串联多体自主式水下航行器(AUVs)动力分配时存在计算周期长、效果不佳等问题，为此提出一种离线模型训练与在线动力优化相结合的新方法。首先采用哈里斯鹰优化算法生成动力分配数据集，以其训练神经网络模型，减少在线处理时间；然后基于当前直航状态调用离线模型和粒子群优化算法，进行在线动力分配。为验证本文方法的有效性，在Matlab环境中对30个单元组成的长串联多体AUVs进行动力分配仿真，结果表明：长串联多体AUVs的单元间距和整体直航速度均快速收敛到目标值；与直接使用遗传算法相比，采用本文方法进行动力分配得到的最大单元间距与理想间距相对偏差降低了37.67%、最小单元间距与理想间距相对偏差降低了6.50%、整体直航速度与目标直航速度相对偏差降低了56.00%。离线训练与在线优化相结合的动力分配方法有效提升了长串联多体AUVs的航行稳定性，可为其在复杂水下环境中的应用提供理论基础和实践指导。
To address the limitations of traditional online optimization methods, such as long computation cycles and suboptimal performance in thrust allocation for long serially connected multi-body autonomous underwater vehicles(AUVs),a novel approach combining offline model training with online thrust optimization is proposed.First, the Harris hawks optimization is employed to generate a dataset for thrust allocation, which is utilized to train a neural network model, thereby reducing online processing time.Second, based on the current direct flight status, the offline model and the particle swarm optimization are invoked for online thrust allocation.To verify the effectiveness of the proposed method, Matlab simulations are carried out for thrust allocation in a long serially connected multi-body AUVs consisting of 30 units.Results show that the inter-unit spacing and overall straight-line navigation speed both rapidly converge to the target values.Compared with directly using the genetic algorithm, the relative deviation between the maximum unit spacing achieved by the thrust allocation method proposed in this study and the ideal spacing is reduced by 37.67%.The relative deviation between the minimum unit spacing and the ideal spacing is reduced by 6.50%.The relative deviation between the overall straight-line speed and the target straight-line speed is reduced by 56.00%.The combination of offline training and online optimization significantly improves the navigation stability of long serially connected multi-body AUVs, providing a theoretical foundation and practical guidance for their application in complex underwater environments.
基于改进DeepLabV3+的息肉分割方法
A Method for Polyp Segmentation Based on Improved DeepLabV3+
王文成;尹哲;李学桐;刘欣然; WANG Wencheng;YIN Zhe;LI Xuetong;LIU Xinran;为解决当前结肠镜检查在分割复杂生长环境中的息肉时面临的精度差、效率低的难题，提出一种基于改进DeepLabV3+的息肉分割方法。首先通过将DeepLabV3+的主干网络Xception替换为参数量更小的MobileNetV2以降低模型的复杂度，并引入注意力机制(CA),以提高模型对息肉的定位精度；其次从MobileNetV2引出一条中级特征，并与经过CA处理后带有位置信息的高级特征共同构建特征金字塔(FPN)进行多尺度融合，并作为解码器特征融合的分支，将融合时的4倍上采样分解为逐层上采样以增加图像细节信息；最后引入多尺度卷积注意力模块(MSCA)聚合上下文信息并输出分割结果。实验结果表明，改进后的网络有效提高了结肠息肉分割的准确率和效率。
To solve the problem of low accuracy and efficiency in colon polyp segmentation in complex growth environments, a polyp segmentation method based on the improved DeepLabV3+is proposed.First, the main trunk network of DeepLabV3+is replaced by MobileNetV2 with fewer parameters to reduce the complexity of the model.At the same time, an attention mechanism(CA)is introduced to improve the precision of the model in locating polyps.Second, a middle-level feature is extracted from MobileNetV2,and a feature pyramid(FPN)is constructed by combining the middle-level feature with the high-level feature after CA attention with position information.The multi-scale fusion is used as a branch of the decoder feature fusion, and the four-fold upsampling during fusion is decomposed into layer-by-layer upsampling to increase image detail information.Finally, a multi-scale convolutional attention module(MSCA)is introduced to aggregate context information and output segmentation results.The experimental results show that the improved network effectively improves the accuracy and efficiency of colon polyp segmentation.