Purpose The aim of this paper is to investigate the use of soft flexible fingers with variable stiffness and friction (VSF) as an effective solution for soft-robot in-hand manipulation. Design/methodology/approach The proposed VSF structure comprises a plastic bag, jamming layers, a low-friction surface and a high-friction surface. By controlling the pressure within the plastic bag, the VSF structure facilitates the transition between high friction with low stiffness and low friction with high stiffness. Through the integration of VSF joints, a five-fingered robotic hand was developed to enable in-hand manipulation. This system successfully demonstrated bottle-rotation experiments without requiring grip release. Findings The experimental results reveal that a single joint equipped with the VSF structure exhibits a 22.35% increase in stiffness under 10 kPa pressure and a 200 g load, as well as a maximum 46.51% variation in friction force at 40 kPa positive pressure. These findings demonstrate the VSF structure’s capability to achieve VSF. By integrating this structure into soft robotic hands, dexterous two-finger gripping and in-hand manipulation with a five-finger hand can be realized. Originality/value In this study, the authors developed a VSF structure that is actuated by positive pressure and integrates both VSF capabilities. Furthermore, the authors incorporated this VSF structure into a five-fingered robotic hand, enabling it to manipulate objects by changing contact surfaces, eliminating the need for the soft hand to rotate objects by grasping and placing. This advantage makes it suitable for manipulating objects in confined spaces, with potential applications in mechanical assembly and logistics sorting.

PurposeThis paper aims to address the low-efficiency issues encountered by reinforcement learning (RL) methods in standard automated assembly scenarios for 3C assembly lines, where robots struggle with vast exploration spaces to execute long-horizon sequences of actions.Design/methodology/approachThis work proposes an RL framework formulated as a hierarchical reinforcement learning (HRL) approach guided by human multimodal demonstration knowledge (MDK). The framework integrates task-specific action sequences, derived from the multimodal fusion algorithm within the MDK acquisition method, into the HRL process to guide the learning. These action sequences consist of predefined action primitives, and the hierarchical policy simultaneously learns both the orchestration of these primitives and the parameters for each primitive.FindingsBy leveraging prior knowledge, this RL framework can effectively direct the initial learning process of the robot, significantly expediting convergence. On the other hand, the hierarchical architecture and action primitive design render its transfer to real-world robots straightforward, offering valuable references for applications in 3C assembly scenarios.Originality/valueA robot learning framework that learns through human hand assembly behavior is proposed and verified, which enables faster convergence and is more convenient to be applied to actual robots.

Traditional manufacturing in the computer, communication, and consumer electronics (3C) industries primarily relies on automation but lacks autonomous learning, decision-making, and adaptability. To address this challenge, this study introduces a multi-layer multi-level knowledge representation (MLMLKR) approach aimed at enhancing the adaptability and accuracy of assembly processes in 3C tasks. The MLMLKR framework comprises five layers: task, area, object, action, and agent, as well as three levels: ontology, template, and instance. In addition, we have developed a comprehensive set of templates, including master templates, task subtemplates, and action subtemplates, to improve the transferability of robotic operations in 3C scenarios. To realize knowledge-enhanced task-to-action reasoning (KET2A), we utilize the learnable composite loss function to combine BERT module, graph attention network module and Seq2Seq module into a unified end-to-end model. Finally, we employ multilayer multi-level knowledge graph and KET2A into multiple assembly scenarios, further validate the capabilities of task transfer and action transfer. The experiment results demonstrated that the action sequence reasoning accuracy reached to 98.1% and the perplexity was 1.019. The reasoned sequence execution of subscriber identity module and memory module assembly tasks were performed in real scenarios. Furthermore, the proposed scheme accomplished congeneric object operating transfer from DDR3 8 G 138 x 38.8 mm to double data rate 3 (DDR3) 4 G 138 x 18.8 mm in memory module assembly task, also from Intel Core CPU to advanced micro devices (AMD) CPU in CPU assembly task. Meanwhile, it implemented congeneric action "press" operating transfer between CPU, memory module, and front camera assembly tasks. In a word, the proposed manipulation skill representation and knowledge reasoning method is feasible and can be applied in intelligent 3C assembly.

To investigate tensile damage characteristics and mechanisms in depth, this study employed acoustic emission (AE) technology for real-time monitoring of the concrete axial tension process and integrated k-means clustering analysis to achieve self-organized classification of the collected AE signals. By applying correlation analysis of AE feature parameters and the coefficient of variation to characterize data dispersion, the optimal clustering results were selected. These results, combined with temporal analysis dividing the damage stages, collectively achieved the classification of concrete damage severity. Finally, through analyzing the underlying AE feature parameters within the clustering results, an effective interpretation of the concrete axial tension damage mechanisms at different severity levels was accomplished. This research method provides new insights into nondestructive testing and structural health monitoring of concrete structures, offering potential value for engineering applications. In the future, the method can be extended to more complex structural systems, and its applicability in long-term service environments can be explored to advance the development of intelligent monitoring and prediction technologies.

As a prevalent form of multimodal data, video data plays a crucial role in numerous applications, offering various benefits. Meanwhile, video integrity and source issues also pose security risks. Video data is multimodal, containing a container describing video coding and packaging, along with a video data stream featuring visual and audio information. Many works on video integrity and source analysis focus on video containers, and they overlook the fact that a malicious user can readily manipulate these traces within the containers by reconstructing them without transcoding. In our research, we propose a hierarchical fingerprint learning framework through multi-level fusion for video integrity and source analysis. Our approach integrates video encoding attributes, extracting multi-level features from both decoded video key frames and reference frames. We model the dependencies between these features based on encoding characteristics, effectively revealing hidden clues in spatial and temporal domains related to various video processing techniques. Additionally, we introduce a hierarchical framework to fuse multiple clues from different groups of pictures (GOPs), facilitating collaborative feature learning across multiple GOPs. Extensive experiments on publicly available datasets validate the effectiveness of our method in tasks related to video integrity verification and source identification. Our approach provides support for ensuring the credibility and traceability of video content in consumer applications.

Current service robots without learning ability are not qualified for many complex tasks. Therefore, it is very significant to decompose the complex task into repeatable execution unit. In this paper, we propose a complex task representation method based on dynamic motion primitives, and use hierarchical knowledge graph to represent the analytic results of complex tasks. To realize the execution of complex robot manipulation tasks, we decompose the semantic tasks into the minimum motion units that can be executed by the robot and combine the multi-modal information: posture, force and robot joint parameters, obtained by the sensors. We use the knowledge graph to record the end-effector required by the robot to perform different tasks and make appropriate selection of end-effector according to different needs. Finally, Taking the long sequence complex task of service scene as an example, we use UR5 robot to verify the effectiveness and feasibility of this design.

In order to solve the problem of unsatisfactory DeepLabv3+ segmentation of some small target objects and the details loss caused by only one scale coding features in the decoding stage, proposed a semantic segmentation algorithm based on attention mechanism and strip pooling . In the coding stage, a strip of attention refinement module is introduced after the backbone network to increase the receptive field in the horizontal and vertical directions of the coding stage and improve the perception ability of small target information. Refactoring ASPP modules with improved mixed strip pooling modules to enhance contextual information relevance and generate richer high-level semantic features. In the decoding stage, a spatial attention feature fusion module is designed to integrate multi-scale features extracted from the backbone network and enrich low-level semantic features. Experiments in Cityscapes datasets show that the algorithm can effectively enhance the feature extraction capability and small target perception, and alleviate the problem of detail loss.

As the labor cost of manual assembly increases, it is urgent to use industrial robots to carry out assembly tasks. This work presents an industrial robot skill manipulation method based on knowledge graph to realize the intelligent assembly of products. Firstly, the knowledge base based on parameter expression is designed and deployed on the cloud server, so as to facilitate the real-time communication of robots when they perform tasks. In the second stage, the manipulation skills in the assembly scene are expressed in knowledge and stored in the knowledge base, and each action primitive is mapped into the bottom-level code that the robot can execute. Finally, the business system including the execution terminal and turntable is designed, and the memory stick assembly experiment is carried out to verify the effectiveness of this design.

传统基于稀疏表示的人脸识别方法因未充分利用样本包含的信息,而存在较低的鲁棒性和识别率等问题.提出基于小波变换的多字典人脸识别方法.提取人脸图像的多尺度纹理特征,构建多字典,每个字典分别对应样本的原始图像和不同尺度的纹理信息.为提高人脸识别的准确性,在训练阶段利用相同的系数表示将不同字典相关联,深入挖掘了人脸图像与其不同尺度纹理间的共性,探索人脸不同特征的内在联系.较仅聚焦于样本单一特征的传统方法,克服了忽视样本不同特征间联系的不足,更深刻地挖掘训练样本人脸特征且突出了不同类训练样本的个性特征.在多个人脸数据库上的实验结果表明,识别性能有明显的提升.

随着智能技术的不断发展,数字孪生技术应用领域得到逐步推广.将数字孪生技术引入到图书馆馆藏资源共享模式中,梳理了数字孪生技术的概念与特征,研究了数字孪生技术在图书馆应用的必要性,数字孪生技术的可视化、评估工具属性为其在图书馆领域的广泛应用奠定基础.

The increasing popularity of electric vehicles (EVs) has posed significant challenges to the existing charging facilities. The existing research works on battery swapping station (BSS) are usually based on the same type of EV battery. This paper introduces a BSS model with multiple battery types for handling the EVs’ requests with diverse battery types. To maximize the battery diversity at BSS, we propose a novel objective function with weighted summarized variables and the average charging time of each battery type. We specially designed a dynamic charging schedule strategy to optimize the output power of the charging units at a BSS. Experimental results have been illustrated to verify the effectiveness of the proposed dynamic charging schedule model with multiple battery types.

Answer selection task is an important task in question answering systems. In this work, we propose several deep learning methods to address answer selection task. Current answer selection tasks use LSTM networks to learn the contextual information of query and candidate answer sequences, but the LSTM network suffers from the problem of gradient instability and fail to extract local information. Aiming to solve these problems, we first introduce fusion layer with residual ideas to alleviate gradient instability. Then we further introduce CNN networks to capture local n-gram information. In addition, we introduce one-way and two-way attention mechanism respectively, in order to capture the interaction between query and candidate answer, and further improve model performance. Experimental results of two public datasets InsuranceQA and WikiQA show that our methods outperform baseline methods, which conclude the effectiveness of our methods proposed.

As we know, image colorization is widely used in computer graphics and has become a research hotspot in the field of image processing. Current image colorization technology has the phenomenon of single coloring effect and unreal color, which is too complicated to be implemented and struggled to gain popularity. In this paper, a new method based on a convolution neural network is proposed to study the reasonable coloring of human images and ensures the realism of the coloring effect and the diversity of coloring at the same time. First, this paper selects about 5000 pictures of people and plants from the Imagenet dataset and makes a small dataset containing only people and backgrounds. Secondly, in order to obtain the image segmentation results, this paper improves the U-net network and carries out three times of down sampling and three times of up-sampling. Finally, we add the expanded convolution, and use the sigmoid activation function to replace the ReLU (The Rectified Linear Unit) activation function and put the BN (Batch Normalization) before the activation function. Experimental results show that our proposed image colorization algorithm based on the deep learning scheme can reduce the training time of the network and achieve higher quality segmentation results.

Objectives: The aim of this study was to investigate the detection efficacy of deep learning (DL) for automatic breast ultrasound (ABUS) and factors affecting its efficacy. Methods: Females who underwent ABUS and handheld ultrasound from May 2016 to June 2017 (N = 397) were enrolled and divided into training (n = 163 patients with breast cancer and 33 with benign lesions), test (n = 57) and control (n = 144) groups. A convolutional neural network was optimized to detect lesions in ABUS. The sensitivity and false positives (FPs) were evaluated and compared for different breast tissue compositions, lesion sizes, morphologies and echo patterns. Results: In the training set, with 688 lesion regions (LRs), the network achieved sensitivities of 93.8%, 97.2% and 100%, based on volume, lesion and patient, respectively, with 1.9 FPs per volume. In the test group with 247 LRs, the sensitivities were 92.7%, 94.5% and 96.5%, respectively, with 2.4 FPs per volume. The control group, with 900 volumes, showed 0.24 FPs per volume. The sensitivity was 98% for lesions > 1 cm3, but 87% for those ≤1 cm3 (p < 0.05). Similar sensitivities and FPs were observed for different breast tissue compositions (homogeneous, 97.5%, 2.1; heterogeneous, 93.6%, 2.1), lesion morphologies (mass, 96.3%, 2.1; non-mass, 95.8%, 2.0) and echo patterns (homogeneous, 96.1%, 2.1; heterogeneous 96.8%, 2.1). Conclusions: DL had high detection sensitivity with a low FP but was affected by lesion size. Advances in knowledge: DL is technically feasible for the automatic detection of lesions in ABUS.

Object Detection Algorithms is widely used in transportation. With YOLOv3 however, it is impossible to achieve real-time detection. This paper made some adjustments to YOLOv3, and proposed a new light-scale model named MobileNetv1_yolov3lite. In our MobileNetv1_yolov3lite, we use MobileNetv1 instead of Darknet53 as our backbone network, and we use a newly proposed module yolov3lite for feature fusion. These adjustments achieve significant increases in detecting speed, and can achieve real-time detection. However, it suffers from accuracy loss. In order to improve detecting accuracy, we further modify loss function as well as training methods, which contributes to a higher accuracy.

Effectively improving the competitiveness of ISP has always been a hot topic in the field of communication research. Video as the main body of network traffic, providing video users with high satisfaction network services can enhance the user experience, and thus effectively improve the competitiveness of ISP. Some user behaviours will affect the detection of the stalling event. Therefore, this paper proposes a video stalling decision model based on the video stalling detection method based on HTTPS. The model mainly models the IP source, packet length, packet type, and basic information of the packet data. In the case of the unsolved HTTPS protocol, it can use the limited information to identify the user behaviour and perform the stalling detection, thereby reducing the misjudgement rate of the stalling detection.

文章对服务接触理论的概念进行了阐释,分析了新一代信息技术驱动下的图书馆交互服务模式的发展和演变,阐述了图书馆用户交互服务存在的问题,论述了服务接触理论视阈下图书馆用户交互模式的构建策略,以期为图书馆开展业务工作提供参考和借鉴.

本文针对在线视频服务网站的个性化推荐问题,提出了一种基于用户播放行为序列的个性化推荐策略.该策略通过深度神经网络词向量模型分析用户播放视频行为数据,将视频映射成等维度的特征向量,提取视频的语义特征.聚类用户播放历史视频的特征向量,建模用户兴趣分布矩阵.结合用户兴趣偏好和用户观看历史序列生成推荐列表.在大规模的视频服务系统中进行了离线实验,相比随机算法、基于物品的协同过滤和基于用户的协同过滤传统推荐策略,本方法在用户观看视频的Top-N推荐精确率方面平均分别获得22.3％、30.7％和934％的相对提升,在召回率指标上分别获得52.8％、41％和1065％的相对提升.进一步地与矩阵分解算法SVD++、基于双向LSTM模型和注意力机制的Bi-LSTM+ Attention算法和基于用户行为序列的深度兴趣网络DIN比较,Top-N推荐精确率和召回率也得到了明显提升.该推荐策略不仅获得了较高的精确率和召回率,还尝试解决传统推荐面临大规模工业数据集时的数据要求严苛、数据稀疏和数据噪声等问题.

Many-objective optimization problems take challenges to multi-objective evolutionary algorithms.A number of nondominated solutions in population cause a difficult selection towards the Pareto front.To tackle this issue, a series of indicatorbased multi-objective evolutionary algorithms (MOEAs) have been proposed to guide the evolution progress and shown promising performance.This paper proposes an indicator-based manyobjective evolutionary algorithm called ε-indicator-based shuffled frog leaping algorithm (ε-MaOSFLA), which adopts the shuffled frog leaping algorithm as an evolutionary strategy and a simple and effective ε-indicator as a fitness assignment scheme to press the population towards the Pareto front.Compared with four stateof-the-art MOEAs on several standard test problems with up to 50 objectives, the experimental results show that ε-MaOSFLA outperforms the competitors.