During nuclear decommissioning, large structures often need to be cut into smaller parts before being packed into containers. Cutting and packing incur costs, and improving efficiency in one often increases costs in the other, requiring a trade-off. Considering the operational costs of different cutting tools and container types, determining the optimal number of parts to cut a structure into, and how to pack these parts efficiently, is challenging. This paper presents a novel voxel-based cutting and packing optimization algorithm designed to minimize overall costs for decommissioning nuclear structures. We define cost as comprising two factors: cutting cost, which depends on the amount of cutting required and the type of tool used, and packing cost, which is determined by the number of containers needed and the unit cost of a container. We demonstrate that our algorithm can successfully minimize overall cutting and packing costs. Additionally, we show how our algorithm can be used to compare different decommissioning scenarios by cutting a simple structure found in nuclear decommissioning and packing the pieces into two different container types with different unit costs. Our results also show that the algorithm outperforms a minimal cutting approach, where a minimum number of cuts is used to segment the structure.
In existing hand gesture recognition research, single-modal recognition is commonly used. For example, visual hand gesture recognition uses image information, but it is easily affected by the shooting environment. Another example is using surface electromyography (sEMG) for recognition, but it is susceptible to signal noise. To address the above issues, this paper focuses on the fusion of sEMG and vision of the human hand. We propose a novel approach that fuses the two modalities by using convolutional neural networks (CNN) to improve recognition accuracy. Firstly, using an RGB camera and sEMG armband, we jointly collect sEMG signal and skeleton in real-time, creating our own multimodal dataset for training. Secondly, we design a multimodal recognition network with feature fusion of sEMG and skeleton, to achieve an increase in accuracy. Finally, we built a human-computer interaction system that realizes hand gestures to manipulate a dexterous hand and a robot arm. Experimental results demonstrate that the fusion of the two modalities has complementary effects and effectively improves recognition accuracy.
Continuous curvilinear capsulorhexis (CCC) is a delicate operation that may benefit from robot technology. This paper introduces a new hybrid cataract surgery robot that consists of four parallel prismatic pairs connected in series to a "Rotation-Prismatic" unit. Additionally, the paper proposes a master-slave control strategy and a virtual remote center of motion (RCM) algorithm. The virtual RCM is based on the kinematic model of the robot and enables the setting of an arbitrary point on the forceps as the RCM point. Finally, the effectiveness of the proposed robot is verified through experimental evaluations on ex-vivo pig eyeballs.
Knee osteoarthritis (KOA) is a major cause of morbidity, especially in the older and female population. Automatic medical image segmentation is crucial for improving the diagnostic accuracy of KOA and optimizing treatment plans. Convolutional neural networks have evinced potential in enhancing the accuracy of KOA segmentation. However, CNNs face challenges in handling the complex variability in position, shape, and scale of segmentation targets. Additionally, existing CNN models struggle with accurate processing boundary information, which limits their effectiveness in clinical decision-making. To tackle these challenges, we introduce a deep-learning architecture known as STRR-Net, designed for the automated segmentation of knee joints in MRI scans. We also designed a novel parallel encoder structure to achieve complementary and enhanced feature extraction. In particular, we propose a skip-connection structure for multi-scale fusion, which allows us to simultaneously capture different levels of feature information. Our model attains an average Dice of 0.8176 for the knee dataset, with a score of 0.6653 for cartilage segmentation and 0.8320 for meniscus segmentation. Visualizing the feature maps allows for a detailed assessment of the knee joint’s anatomy, including cartilage thickness, thereby facilitating the timely identification of osteoarthritis.
The shield tunneling method is currently the most widely used underground construction technique, where the formed tunnel is composed of many sequentially arranged segments. Before construction, a preliminary layout of the design tunnel axis (DTA) is required to assess the segments' ability to fit the DTA and determine the number of segments needed. Existing preliminary layout methods only consider the axis fitting deviation of the next ring of segments to select the assembly point, which can lead to significantly increased axis deviations for subsequent segments, potentially exceeding allowable limits. This results in incorrect assessments of the segments' fitting ability, particularly for DTAs with small curvature radius. To address this issue, this paper proposes a segment preliminary layout method that considers the maximum fitting deviation of the next multiple rings of segments. The method uses a differential evolution algorithm to optimize the assembly point combinations of the next multiple segment rings. It employs a segment pose transfer algorithm to calculate the poses of the next multiple rings, and then computes the corresponding axis fitting deviations for these rings. The assembly point combinations with the minimum maximum axis fitting deviation is selected as the optimal solution, and the first assembly point is selected to assembly next segment. This method was applied to a 900m radius circular arc DTA using 28-point large diameter segments with a calculation of five segment rings. The maximum axis fitting deviation across the entire DTA was 9.3 mm, significantly lower than the 52.8 mm deviation obtained using existing layout methods.
In the field of visual tracking, it is a significant challenge to accurately capture the dynamic changes of targets in complex scenes. To address this issue, this paper proposes a novel Hierarchical Feature-Aware Network (HFAN) to improve tracking performance. The design of HFAN mainly includes two pivotal components: Feature-Enhanced Unit (FEU) and Hierarchical Feature-Aware Unit (HFAU). FEU enhances the richness and discriminative power of target representations by reinforcing features from the templates and search regions. HFAU establishes comprehensive dependencies among multi-level features to capture the changing characteristics of targets across different spatial hierarchies. Finally, a Siamese tracker called HFANTrack is proposed to improve tracking accuracy and robustness in complex scenarios. Extensive experimental results show that our method achieves competitive tracking performance with a real-time speed of 49.3fps compared to state-of-the-art methods.
A novel magnetic geared composite motor is proposed based on magnetic field modulation, integrating multiple advantages of magnetic gears. To reduce axial size, a coaxial structure is adopted, embedding a permanent magnet synchronous motor inside the magnetic gear. To minimize torque ripple, a novel structure is designed: a high-speed inner rotor with eccentric magnetic poles is connected to a cup skeleton featuring radial magnetized poles on its inner and outer walls. The outer rotor employs a magnetic geared composite motor structure arranged in a Halbach array. The magnet modulation ring skeleton is 3D printed with PLA materials to reduce weight. To assess its effectiveness, finite element method is employed for electromagnetic simulation. The comparison with traditional magnetic geared composite motors which use various magnetization methods, indicates that the novel magnetic geared composite motor effectively reduces torque ripple and also demonstrates an improvement in output torque.
Visual grounding (VG) is a critical task that seeks to identify and localize a specific visual region within a given image based on a corresponding referring expression. Existing approaches to the visual grounding (VG) task can be categorized into three main types: two-stage methods, one-stage methods, and Transformer-based methods, which have achieved high performance. However, most of the methods do not exploit the visual and linguistic information well, limiting the performance of model. In this work, we propose a language-guided visual attention network for visual grounding, which can utilize language to deeply explore visual information by better processing of the relationship between vision and language. Specifically, we utilize BERT, a pre-trained model, to get the word-level and sentence-level linguistic features, which can understand the linguistic information more comprehensively. Inspired by the Transformer architecture, we design the stacked visual attention module, which leverages language to direct the attention of vision. In addition, we discuss several ways of fusing visual and linguistic features, enabling a better fusion of visual-linguistic information to obtain the correct coordinates. In a series of comprehensive evaluations on the ReferItGame benchmark dataset, our proposed model is shown to establish a new performance standard.
Accurate modeling of a soft robot remains a significant challenge in soft robotics. A precise model enables the more accurate control of the soft robot. Among various simulation methods, the finite element method (FEM) offers a more realistic representation of soft robot behavior. However, during simulations, soft robot bodies may experience collisions, especially with complicated structures, which results in challenges to traditional discrete collision detection. Collision detection is an algorithm that determines the limitations on the surface of the convolutions to avoid penetration. In this paper, a novel continuous collision detection algorithm is designed and proposed to benefit the accuracy performance of the FEM simulation during the contraction of a soft actuator, which contains a complex bellow structure with eight convolutions. The details of the algorithm are presented, along with initial validation results. The preliminary results demonstrate the potential of the proposed algorithm to enhance the accuracy of soft robot simulations. Further development and extensive algorithm validation are ongoing to contribute to the advancement of soft robotics modeling techniques.
Gaze-based interaction technology has been prompted in response to the problem of interaction obstacles of some people caused by the limitations of keyboard and mouse control in the current human-computer interaction. The key step in the vision tracking technology, the pupil, has the problem of low accuracy of pupil center positioning due to the occlusion of the eyelid in the image and the reflection of part of the light, which seriously affects the effect of the human-computer interaction technology based on vision tracking. Therefore, this paper proposes a pupil localization algorithm based on dual edges and occlusion compensation. Firstly, preprocessing such as Gaussian blur is applied to the eye image. Then, histogram-based thresholding is used to binarize the image, obtaining binary image edges and eye image edges. Pupil contour extraction is performed based on dual edges, followed by ellipse fitting using the least squares method to obtain pupil information. Taking into account the problem of poor pupil localization under occlusion, this paper also presents a method to acquire effective edge information and compensate contour points to enhance the accuracy of localization. The experimental results demonstrate that the method proposed in this paper achieves an accuracy of 70.85% on the public dataset LPW, which is 2% higher than the LVCF method. Furthermore, it can effectively locate the pupils under various occlusion conditions.
Consumers increasingly rely on online reviews to make informed decisions, with negative reviews playing a crucial role due to their perceived credibility and detailed insights. Despite their importance, research on negative reviews remains limited. This study explores the application of machine learning models to predict the usefulness of negative online reviews. By conducting a series of experiments, we identified the most impactful features for predicting review usefulness. Our findings reveal that feature combination experiments demonstrated that combining semantic and structural features improves model performance to 98% of accuracy, highlighting the importance of feature quality over quantity. The proposed XGBoost model, leveraging these combined features, consistently performed well, achieving high accuracy and robustness. This model can assist consumers in identifying useful reviews even when vote counts are insufficient, enhancing their decision-making process. The study underscores the critical role of negative reviews and offers practical solutions for improving online review sorting mechanisms.
Assisted exoskeleton robots are gradually being used in industrial production in order to reduce the risk of workers suffering from work-related musculoskeletal disorders (WMSDs). To lower the weight of the exoskeleton and improve its applicability, the active degrees of freedom of exoskeletons are often reduced in prototype designs. However, this may decrease the exoskeleton's output power and working capacity. To address this, we propose a tendon-sheath-based multi-joint actuation method that uses a single motor to sequentially drive multiple degrees of freedom, and applies it to the design of the shoulder and elbow joints of the upper limb exoskeleton. To do this, we develop a clamper remotely controlled by a steering gear, which has three working modes and is used to switch the position of the end anchor point of the tendon sheath artificial muscle. The clamper can clamp wire ropes with diameters of 1-2 mm and provide a maximum clamping force of 800 N. The performance of the multi-joint exoskeleton and the clamper was tested by experiments. The results show that the clamper can quickly control the tendon sheath artificial muscle to achieve the sequential movement of the multiple joints, and it only takes about 100ms to 150ms to switch between the three working modes.
Depression poses a major threat to an individual's overall physical and psychological health, as well as their social functioning. Consequently, it is urgent to design an effective automated diagnostic technique for depression. Recent studies have revealed that multimodal approaches produce better detection performance than unimodal methods. To this end, this paper introduces a multimodal depression detection framework using visual and textual data. Initially, separate networks are employed to extract their shallow features, aiming to capture the diversity across various modalities. The visual features are then integrated into the textual features through a newly designed Textual-dominated Cross-modal Fusion module to highlight the information transmitted by the textual features. Finally, the Multilayer Perceptron-based Hybrid Unit is designed to facilitate deep feature interaction, improving the overall detection performance. Our method was assessed on the NCUDID dataset and achieved promising experimental results with a value of 3.41 for the MAE and 4.59 for the RMSE.
Traditional Simultaneous Localization and Mapping (SLAM) often fails when operating within narrow and repetitive metro train inspection pit environments. A novel SLAM framework is presented in this paper to address this problem. It integrates multi-sensor data and leverages prior knowledge of train structures, ensures consistent trajectory estimation in robot’s unidirectional straight runs up to 192m. By employing feature-based Laser Odometry (LO) and Time-of-Flight (ToF) ranging sensors for train wheel center detection, the SLAM frontend generates two types of keyframes. LO keyframes record timestamps, velocities, and pose estimations. ToF keyframes record timestamps when the ToF laser ranging point reaches the centers of the first and last wheels of each carriage. These heterogeneous keyframes are merged to serve as nodes in the factor graph. In addition to standard pose-to-pose constraints, the framework uses true distances between each carriage's first and last wheels, as well as estimated actual distances between carriages and from the robot's reached boundaries in the inspection pit to the centers of the first and last train wheels, as distance constraints between nodes. Finally, the backend optimization yields a globally consistent trajectory, an optimized feature map, and accurate center positions of each carriage’s first wheel in the inspection pit coordinate system. The proposed framework solves the issues of scale inconsistency in maps generated during real-time SLAM in each inspection session, as well as scale drift encountered during long-distance straight-line movements, ensures accuracy and safety in under-train inspections. The effectiveness of the proposed framework is demonstrated using real-world datasets.
A novel Additive Friction Stir Deposition Robot is proposed, which is a rigid-flexible coupled redundantly actuated parallel robot designed for Stirring friction stir based additive manufacturing. The compliance of the robot is improved by designing a bidirectional cable-driven rod as the flexible branch chain of the mechanism. The degrees of freedom of the AFSDR are calculated, and the inverse kinematics model is established. Taking into account the geometric constraints in actual working conditions, the workspace performance of the moving platform is evaluated. The force/motion transmission performance indicators LTI and GTI based on the screw theory are computed. Finally, multi-objective optimization of the mechanism is accomplished using the NSGA-II algorithm, resulting in enhanced performance through comparative analysis.
Knee osteoarthritis is a chronic degenerative disease commonly seen in middle-aged and elderly people. At present, the number of patients with knee osteoarthritis is increasing year by year, accounting for 21.51% of the total patients. Today, knee osteoarthritis rehabilitation robots have made breakthrough progress, but safety and portability are still the key issues to be solved, and further research and improvement are needed to achieve a wider application. By studying the functions of common lower limb rehabilitation robots and the rehabilitation needs of patients with knee osteoarthritis, this paper proposes design requirements and schemes. It conducts dynamic and kinematic analysis, and establishes three-dimensional models. The goal is to complete the mechanism design of wearable knee osteoarthritis rehabilitation robots, aiming to enhance the effectiveness and overall performance of the rehabilitation robots. The main feature of this design is the use of cable driver module design and varus and valgus knee orthoses mechanism design, which can help the knee dysfunction patients to provide stable support and efficiently complete rehabilitation training.
In advancing brain-computer interface (BCI), force feedback has demonstrated potential in enhancing neurophysiological interactions during motor tasks. This study investigated the impact of force feedback on brain activity and the accuracy of decoding movement direction. We developed an electroencephalogram (EEG)-based BCI paradigm with four levels of force feedback (i.e., 0 N, 8 N, 16 N, 24 N) applied during right-hand movements to left and right directions. Six participants were involved to ensure robust results. Three deep learning models—DeepConvNet, ShallowConvNet, and EEGNet—were used to decode movement directions. Findings from event-related desynchronization/event-related synchronization (ERD/ERS) and movement-related cortical potentials (MRCPs) indicated that increased force feedback significantly enhanced the brain’s response to motor stimuli. The decoding results revealed that force feedback notably improved decoding accuracy of DeepConvNet and EEGNet, particularly under medium and high-intensity conditions. Specifically, three models demonstrated accuracy improvements of 11%, 4%, and 12% under high-intensity force feedback, respectively. These results suggest that specific force feedback enhances motor area responsiveness, improving movement intention decoding in BCI. Our study confirms the positive impact of force feedback on BCI performance, highlighting the potential of force feedback-based BCI systems.
This paper presents a study on the evaluation of moisture content in in vitro chewed food boluses using image processing techniques. This integration offers a novel, non-destructive method that enhances the accuracy and efficiency of moisture assessments compared to traditional gravimetric methods and external sensors, which are typically more time-consuming and potentially alter the sample's properties. Invitro food samples were taken for moisture analysis from a biomimetic robot equipped with automatic saliva injections. Image skewness, variance, lightness, and Gray-Level Co-occurrence Matrix (GLCM) analysis were found to quantitatively indicate moisture content for two types of in vitro chewed food samples, at different chewing cycles and trajectories. This pilot study was conducted with two food samples, but the approach can be applied to a broader range. The study's findings highlight the effectiveness of image processing in analyzing moisture dynamics within chewed food boluses, providing valuable insights that could enhance the design and functionality of biomimetic chewing robots to resemble human oral processing more closely.
The rapid development of artificial intelligence in the context of Industry 4.0. This study focuses on cable trenches and explores the application of monocular vision technology in cable recognition and localization in detection robots. The underground cable trench environment is complex and poses safety hazards. This study developed a cable recognition system using the YOLOv5 algorithm, which captures video images in real-time using a monocular camera. Combined with the camera's internal parameters, the system identifies and locates target objects. The experimental results show that the proposed algorithm achieves an efficient processing speed of 53.2 frames per second while achieving an average accuracy of 98.3%. Compared with actual measurement results, the effectiveness and accuracy of the algorithm in identifying target size and distance measurement have been verified.
The study of vehicle suspension systems is a critical component in ensuring enhanced comfort and stability for vehicles. In this paper, we propose an event-triggered control strategy for networked suspension systems to withstand denial-of-service (DoS) attacks. This strategy ensures vehicle driving stability and safety, while also minimizing the consumption of network resources. An attack-based Lyapunov function is used to obtain sufficient conditions with exponential stabilization and H∞ properties. Finally, the effectiveness of the controller is validated through system simulations conducted on a bumpy road.
