Abstract:Embodied intelligent human-following mobile robots are intelligent systems capable of recognizing, tracking, and coordinating their movement with human targets. They have gradually evolved from "functional tools" to "context-aware intelligent agents", marking a paradigm shift from "operating in the environment" to "understanding and acting in the environment through their bodies". Based on the embodied intelligence framework of "environment-body-intelligence", this paper systematically reviews the research progress and development trends of human-following mobile robots. First, it elaborates on the research background and objectives of embodied intelligence, discussing its scientific essence centered on environmental interaction, bodily experience, and emergent intelligence. Second, it outlines the technological evolution, typical application scenarios, and industrial trends. By conducting an in-depth analysis of key technologies such as multimodal sensors, computing and processing units, and motion execution mechanisms, the paper further explores the application requirements in typical scenarios such as warehousing and logistics, public transportation, and elderly care and rehabilitation. Next, it focuses on natural and efficient human-robot interaction mechanisms and collaborative strategies, which include human-robot interaction methods based on explicit commands, implicit states, physical contact, and social-emotional cues, as well as interaction strategies based on reinforcement learning, imitation learning, and transfer learning, all aimed at enhancing the robots′ operational fluency and social acceptance in social environments. Finally, the paper summarizes the main technical challenges in areas such as deep environmental cognition, dynamic bodily adaptation, and continuous intelligent evolution, and provides an outlook on future development directions.

Abstract:Most existing embodied manipulation methods predominantly rely on end-to-end learning strategies. Although performing adequately on specific tasks, these approaches often treat physical interaction as a black box, thereby sacrificing interpretability and suffering from insufficient geometric grounding, large error accumulation in long-horizon planning, and weak adaptability to complex dynamics. To address fine-grained manipulation in unstructured scenarios, this article proposes an embodied manipulation framework based on a 3D geometric world model （3D-GWM） and conditional flow matching, termed 3D-GWM, which comprises three modules: 3D perception and compression, a generative geometric world model, and conditional flow-matching-based action generation. In the perception module, a complementary dual-domain feature enhancement algorithm is proposed to jointly capture spatial high-frequency details and global frequency-domain context, together with an instruction-driven dynamic token compression mechanism to reduce computation while preserving feature fidelity. In deployment evaluations, the number of input visual tokens is reduced from 2 048 to 128, while the peak GPU memory usage is reduced from 58.7 GB to 17.2 GB. In the cognition module, an intermediate-frame prediction strategy is adopted. By predicting stage-wise 3D geometric terminal states of sub-tasks, it provides intermediate goal guidance with causal logic for long-horizon manipulation and mitigates error accumulation during planning. In the execution module, a conditional flow matching strategy is employed. By using geometric anchors to construct target vector fields, it generates smooth trajectories that satisfy physical contact constraints and reduces the risk of singularities commonly encountered in rigid inverse-kinematics solvers. Experimental results show that 3D-GWM achieves an average success rate of 89.4% in simulation and 82.0% on the real robot for complex-logic and high-precision tasks; compared with representative baseline methods, the average success rate increases by 7.0% and 8.75%, respectively.

Abstract:To address the structural complexity, actuation redundancy, and insufficient compliance of fully actuated dexterous hands, an underactuated dexterous hand based on a rigid-flexible coupled structure is proposed. According to the motion characteristics of the human hand, three motors are employed to drive five fingers, achieving compliant coupled bending and multi-finger coordinated control. Each finger consists of a composite structure of a rigid skeleton and flexible joints, with cable used to imitate tendon traction, and torsion spring-pin structures introduced at the joints to emulate the elasticity of soft tissues, enabling passive adaptation to the target shape and compliant rebound during grasping. Tactile sensor arrays are integrated on the fingertips and palm, which, together with motor encoders, form a combined force-position perception system capable of real-time detection of contact state, grasping force and joint posture, thereby effectively enhancing environmental perception and grasp stability. The hand weighs only 482 g and demonstrates strong task adaptability, achieving 82.2% success in power grasps and 46.7% in precision grasps across 16 representative tasks, and it can serve as an end-effector to perform grasping, transportation, and stacking operations, enabling interaction in multiple scenarios. In multi-object tactile recognition experiments, the proposed sensing approach achieves an average accuracy of 96.5% for six target objects, validating its effectiveness in tactile perception. During constant-force grasping and disturbance compensation tests, force fluctuation remains within ±0.3 N and the system rapidly restores stability after sudden load changes, indicating robust and stable control performance. Overall, the proposed dexterous hand achieves simplified structure, compliant grasping, and multimodal sensing capabilities, making it suitable for human-robot interaction and mobile manipulation applications.

Abstract:To address the limitations of existing continuum robots in the in-situ inspection of aero-engines, specifically their poor driving capability, low control accuracy, weak load capacity, and insufficient structural stiffness, this article proposes a variable stiffness continuum robot. This design aims to resolve the trade-off between high flexibility and low stiffness, thereby improving control accuracy. Inspired by the environmental characteristics of engine blades and the adsorption mechanism of octopus tentacles, an adsorbable continuum robot composed of multi-segment mortise-and-tenon flexible joints and pneumatic adsorption units connected in series is designed. By actively adhering to the contact environment, the robot achieves structural locking, enhancing its overall stiffness. First, assuming constant curvature, a geometric kinematic model is established to analyze the mapping between bending deformation and structural parameters. Then, with tip deflection as the evaluation criterion, an equivalent stiffness model is formulated based on the Euler-Bernoulli beam theory and the continuity principle. Furthermore, finite element analysis is utilized to simulate the continuum stiffness, investigating the influence of geometric parameters (e.g., the position, quantity, and strength of the adsorption units) to optimize the structural layout. Finally, an experimental platform is established to experimentally analyze and evaluate the proposed structure and the established model. Results show that the equivalent bending stiffness can be regulated from 31.1 N/m in the initial flexible state to 93.3 N/m in the fully adsorbed state. Under the same load, the tip deformation is reduced by 61.3%, achieving an approximate threefold stiffness gain. It verifies the validity of the proposed variable stiffness scheme and the established model.

Abstract:Wheeled mobile manipulators operating in indoor unstructured environments, which are subject to nonholonomic constraints such as a minimum turning radius and the inability to perform instantaneous lateral translation, face significant challenges in whole-body motion planning. Existing decoupled approaches often result in fragmented final paths due to the disconnected planning between the base and the manipulator. Meanwhile, high-dimensional sampling methods suffer from low computational efficiency and poor realtime performance. To address these problems, an innovative hierarchical path planning method is proposed. The method first employs an improved bidirectional dual-tree rapidly-exploring random tree (RRT) algorithm. This algorithm directly incorporates a nonholonomic constraint model and simultaneously explores both forward and backward solution spaces to eliminate redundant computation, thereby generating a collision-free base path embedded with spatiotemporal information. Subsequently, for each temporal node along this base path, a set of feasible manipulator configurations is defined, all satisfying conditions including joint limits and static collision avoidance. Using these sets, a temporal-node-constrained RRT algorithm generates a manipulator path that is strictly synchronized with the base motion. Finally, the complete whole-body motion path is synthesized by aligning the base and manipulator waypoints through index matching. Both ROS simulations and physical experiments show that the algorithm proposed decouples the high-dimensional planning problem, enabling coordinated motion of the mobile base and manipulator in complex constrained environments while generating collision-free trajectories throughout the entire motion in real time. The experimental results indicate that the algorithm has significant advantages in planning time, path length, and smoothness. Compared with the RRT*-Connect algorithm, in complex scenarios, the proposed approach reduces planning time by 29.16% on average, decreases path length by 16.68%, and improves path smoothness by 56.13%.

Abstract:The research of reinforcement learning-based motion control has progressed significantly for the bipedal and quadrupedal robots. However, the deep reinforcement learning control of octopod robots is still in the exploratory stage due to the issues such as high degrees of freedom redundancy, easy interference at the foot end, and complex dynamic modeling. To address this problem, this paper proposes a motion control framework of crab-like octopod robots based on the deep reinforcement learning. Firstly,, the mechanism model and reference gait of the crab-like robot are designed by analyzing the body structure and gait of crabs, which provides the physical prior knowledge for reinforcement learning. Subsequently, an end-to-end deep reinforcement learning training framework is designed to imitate the reference actions of crab gaits, where a behavior cloning initialization strategy is introduced to accelerate the convergence of the policy. At the same time, a phased reinforcement learning method is adopted to gradually increase the terrain difficulty, ensuring that the robot can stably transfer from the flat ground gait to complex terrains. This paper further compares and analyzes different learning rate adjustment strategies as well as the adaptability of the proximal policy optimization (PPO) algorithm with actor-critic (A2C), soft actor critic (SAC), and deep deterministic policy gradient (DDPG) algorithms. Additionally the generalization ability of model under the dynamic disturbances and diverse terrains through domain randomization training and behavior cloning + staged reinforcement learning methods is studied. The simulation experiment results show that the proposed framework outperforms other algorithms in training efficiency, with the gait stability index( i.e., the standard deviation of the five-step center of mass height) of approximately 0.004 m. The robot can achieve the stable movement on complex terrains such as slopes and steps, whose motion performance is superior to that of the trained model. Finally, a preliminary prototype experiment was conducted to verify the feasibility and stability of proposed method in a real physical environment. In conclusion, the proposed deep reinforcement learning framework combing the biological gait imitation, behavior cloning initialization and phased reinforcement learning provides a new technical approach for the motion control of octopod robots.

Abstract:Considering the requirements of high-precision cross-scale, good feature universality, and collaborative measurement with multi-styluses for assembly pose of large components, a contact-type measuring robot was constructed, which used six-axis collaborative robot as motion carrier and elastic styluses as measuring sensors. It features programmable automation, wide operating range, high repeatability in measurement accuracy, and strong product adaptability. Most existing calibration methods of robot tool coordinate system assume the tool to be a rigid body, which makes it difficult to accurately calibrate the tool coordinate system for elastic styluses and fails to unify the relationships among multiple styluses. To address these issues, the calibration method of elastic styluses tool coordinate systems for contact-type measuring robot based on standard sphere was proposed. Firstly, the calibration principle and error analysis of elastic stylus tool coordinate system were derived. Secondly, the iterative calibration methods of multiple elastic styluses tool coordinate systems based on standard sphere were proposed. The calibration methods of stylus diameters, main elastic stylus tool coordinate system, and star-shaped elastic styluses tool coordinate systems were established, respectively. This allowed for the acquisition of tool coordinate systems of multi-styluses and the realization of collaborative measurement between multi-styluses. Finally, measurement experiments were conducted. The results demonstrate that the contact-type six-axis measuring robot system achieved the stylus diameter calibration accuracy of better than 0.007 mm, the tool coordinate system calibration accuracy of better than 0.016 mm, the single-point repeatability measurement accuracy of better than 0.016 mm, the spatial measurement accuracy of better than 0.030 mm. Experiments also show that the product's position measurement accuracy is better than 0.050 mm, and attitude measurement accuracy is better than 0.010°. These results demonstrate that the contact-type measuring robot possesses significant potential for a wide range of engineering applications, such as automated assembly and feature measurement.

Abstract:Backdrivability describes the ability of a robot joint reducer to passively transmit external power during interaction with the environment, and is an important factor affecting the accuracy of robot force feedback. The reverse starting torque is commonly used to characterize a reducer′s backdrivability. However, studies have shown that reverse starting torque represents only a special case of backdrivability, corresponding merely to the specific boundary condition in which the reducer transitions from rest to motion. Therefore, it cannot provide accurate torque information for robot joint feedback control in engineering applications. A comprehensive review of the research and application status of backdrivability in robot joint reducers indicates that most existing studies consider only the simple influences of —moment of inertia and friction, while neglecting the effects of gear ratio and operating speed. Moreover, prior work has focused on static testing and evaluation of reverse starting torque, without addressing the backdrivability of the reducer during motion. To overcome these limitations, this paper conducts an in-depth investigation of the reducer′s bidirectional transmission mechanism and defines the backdrive torque as a parameter to characterize backdrivability. Based on a deconstruction analysis of backdrivability, a backdrive torque model is established that incorporates gear ratio, moment of inertia, friction, and speed. Furthermore, a dynamic characterization and testing method for reducer backdrivability is proposed based on energy transmission, enabling quantitative evaluation of backdrivability under different speeds and operating conditions. The effects of these four parameters on reducer backdrivability are analyzed theoretically, and experimental studies verify the dynamic nature of reducer backdrivability, revealing the inadequacy of using a single reverse starting torque index to evaluate reducer backdrivability. Finally, the theoretical and practical value of reducer backdrivability are highlighted.

Abstract:To investigate the energy-efficiency collaborative optimization mechanism of excavator robots under typical operating conditions, this paper establishes a co-simulation platform integrating the mechanical system, hydraulic system and trajectory tracking control system of excavator robots for three typical working conditions. Quintic polynomial interpolation is used for trajectory planning, and fuzzy PID (F-PID) is adopted for controller design. Based on the principle of trajectory tracking, a distributed hydraulic power unit drive circuit is proposed. Meanwhile, an integrated synergistic optimized particle swarm optimization (ISO-PSO) algorithm is proposed to optimize the parameters of the F-PID controller, and a multi-index collaborative evaluation model is further established to evaluate the performance of control strategies. Simulation results indicate that the comprehensive multi-index values of the lower boom, upper boom, arm and bucket optimized by the ISO-PSO algorithm are significantly decreased under all three working conditions. Taking WC-I as an example, compared with the conventional proportional-integral-derivative (PID), F-PID, and PSO-optimized F-PID, the average multi-index comprehensive values based on ISO-PSO-optimized F-PID are reduced by 63%, 54%, and 30% for the lower boom, respectively; 63%, 56%, and 38% for the upper boom, respectively; 64%, 56%, and 30% for the stick, respectively; and 41%, 34%, and 12% for the bucket, respectively. Finally, an experimental verification system for trajectory tracking control of excavator robots is constructed based on electromechanical–hydraulic integrated modeling and the co-simulation platform, demonstrating the feasibility of the proposed system.

Abstract:Aiming at the problems of excessive self-weight and poor flexibility of the passive lower extremity exoskeleton assist robot (PLEAR), this paper systematically optimizes the leg structure and hip joint structure of PLEAR. Firstly, a multi-condition topology optimization method based on the variable density method is applied to optimize the leg structure made of short carbon fiber composite with a target weight reduction of 20%. This significantly reduces the structural self-weight and the moment of inertia during the swing phase. Secondly, to simulate the biological motion characteristics of the human hip joint, an abduction/adduction revolute joint is added on the basis of the original two degrees of freedom, upgrading the hip joint to three degrees of freedom. This modification effectively eliminates involuntary shaking and improves human-machine compatibility. A human-machine integration simulation platform is established using OpenSim. The simulation results show that after the overall structural optimization, the root-mean-square reduction rates of the hip, knee, and ankle joint moments of the wearer reach 21.62%, 24.11%, and 12.07% respectively. Meanwhile, the root-mean-square reduction rates of the metabolic values of the main lower limb muscles range from 11.14% to 15.47%, indicating a significant improvement in the assistance effect. To further verify the simulation results, a PLEAR prototype is developed, and 8 subjects are recruited to conduct experiments involving level walking and walking up and down 20° slopes with a 15 kg load. The prototype experimental results show that compared without PLEAR, the optimized PLEAR reduces the root-mean-square values of EMG signals of the main lower-limb muscles by 6.73% to 12.84% during level walking and by 7.57% to 13.79% during slope walking. The experimental results not only verify the effectiveness of structural optimization in reducing the muscle burden and energy consumption of the wearer, but also show that the proposed evaluation method can accurately quantify the assistance efficiency of the exoskeleton, providing a reliable basis for the performance testing of related equipment.

Abstract:To address the significant decline in movement intention recognition accuracy caused by the regular physiological feature drift of surface electromyography (sEMG) signals under varying loads in lower limb exoskeleton applications, a recognition method based on muscle physiological characteristics and a bi-branch cross-scale multi-head attention residual network (BCM-ResNet) is proposed. First, the regular pattern of sEMG signal changes under load, which can be termed "physiological reconstruction", is quantitatively revealed through time-frequency analysis. Agglomerative hierarchical clustering confirms that the sEMG feature space exhibits distinct "low, medium, and high" cluster distributions within the 0~14 kg range. This finding demonstrates that load is a key variable driving the regular displacement of the feature space rather than a random disturbance. Secondly, a differentiated feature extraction strategy is proposed by considering the functional heterogeneity of the thigh and calf muscles. A bidirectional long short-term memory (BiLSTM) network is employed to capture the strong temporal dynamics of the thigh muscles (the primary power source), while a multi-scale convolutional neural network (Multi-scale CNN) is utilized to extract local fine-grained features of the calf muscles (the precision regulators). A bi-branch fusion architecture with a multi-head attention (MHA) mechanism is constructed to achieve dynamic decoupling and weighted fusion of intention and load features, utilizing a residual network (ResNet) to enhance the abstraction of high-level features. Additionally, the sensor configuration is optimized through Pareto analysis, determining a 12-channel unilateral layout on the right leg as the optimal setup. Experiments show that the proposed method achieves an average recognition accuracy of 94.78% on a self-built dataset and 92.85% on an international open-source dataset containing 22 subjects, validating the model′s generalization performance. Finally, migration tests on an edge computing terminal verify the algorithm′s execution stability under sub-millisecond inference latency. This research provides reliable algorithmic support for the precise intention recognition of exoskeletons in complex loading environments.

Abstract:To address the poor real-time performance of asymmetric gait generation and the lack of human-robot physical interaction compliance faced by hemiplegic stroke patients using lower-limb exoskeletons for rehabilitation, this article develops a lightweight quasi-direct drive knee exoskeleton and proposes an adaptive compliant control framework based on online incremental dynamic movement primitives (DMP). In terms of hardware design, the system combines a coaxially integrated quasi-direct drive module with carbon fiber links, which significantly reduces the overall system inertia and joint back-drivability impedance, providing a highly transparent physical foundation for compliant interaction. At the control strategy level, the intention perception layer utilizes an adaptive frequency oscillator to smoothly extract the motion phase of the reference side. The trajectory planning layer proposes an incremental DMP algorithm integrating a time-sharing computation mechanism. By evenly distributing complex weight update tasks across multiple control cycles, this algorithm breaks through the computational bottleneck of embedded hardware, achieving online decoupled adjustment of spatiotemporal features and waveform amplitudes of the reference gait while satisfying 1 kHz high-frequency control demands. The underlying impedance controller then converts the trajectory deviations into compliant assistive torques. Multi-condition bilateral wearing experiments (flat ground, slope climbing, varying frequency, and varying amplitude) involving 7 subjects show that the system can accurately generalize the reference side′s motion rhythm. The bilateral kinematic consistency coefficient (PCC) reaches up to 0.977 on flat ground, with a tracking error controlled at 8.30°. Residual physical interference is restricted within ±10 N·m in the transparent mode, and a synergistic assistive torque of approximately ±20 N·m is smoothly output in the assistive mode. Without interfering with the human′s inherent gait rhythm, the proposed system successfully achieves the online real-time generation and adaptive compliant tracking of asymmetric gaits.

Abstract:Accurate real-time perception of human muscle force is a core prerequisite for achieving "assistance-on-demand" control in lower-limb exoskeletons. To address the vulnerability of traditional surface electromyography (sEMG)based motion intention recognition to failure under challenging real-world conditions such as muscle fatigue, electrode displacement, and skin perspiration, this paper proposes a non-invasive lower-limb muscle force prediction and control method based on the multimodal fusion of depth vision and inertial measurement unit (IMU). First, a multimodal visual and kinematic dataset was constructed involving five healthy subjects across five typical daily terrains, including level ground and sloped walking at various gradients. The internal muscle forces of the quadriceps and hamstrings were calculated offline using an sEMG-driven Hill-type musculoskeletal model to serve as the supervised ground truth. Second, a dual-branch spatiotemporal feature fusion deep network was designed, which extracts spatial-geometric features from depth image sequences and temporal dynamics from inertial motion data through parallel channels. By incorporating an adaptive attention mechanism for the intelligent, heterogeneous weighted fusion of these features, the proposed framework completely eliminates the reliance on sEMG sensors during the practical application phase, thereby bypassing signal degradation issues. To bridge the gap from perception to closed-loop control, an adaptive impedance control framework driven by real-time predicted muscle force was established, continuously mapping variations in muscle force to the assistance stiffness of the exoskeleton. Experimental results demonstrate that the proposed method significantly outperforms unimodal approaches, yielding a robust cross-subject average coefficient of determination (R2) exceeding 0.81 and a root-mean-square error (RMSE) of less than 6 N. Furthermore, the predicted force curves maintain highly consistent profiles under simulated muscle fatigue interference, exhibiting superior robustness against physiological disturbances. Finally, platform validation on a commercial lower-limb exoskeleton system achieved high temporal phase synchronization and low lateral drift, providing a promising and viable new paradigm for the highly robust, assistance-on-demand control of exoskeletons in complex, dynamic environments.

Abstract:Self-balancing lower-limb exoskeletons can achieve autonomous and stable walking without the aid of crutches or external support frames in rehabilitation training and assisted walking scenarios. Their dynamic stability critically depends on accurate perception of the center-of-mass (CoM) state of the human-exoskeleton coupled system. Due to factors such as inertial measurement unit (IMU) drift, uncertainties in kinematic models, and the compliant characteristics of human-robot coupling, existing CoM state estimation methods still suffer from insufficient accuracy and limited robustness during complex dynamic walking tasks. To address these challenges, this paper proposes a high-precision CoM state estimation method for self-balancing lower-limb exoskeletons based on projectionconstrained multi-sensor fusion. Within a Kalman filtering framework, the proposed method integrates IMU measurements from the CoM and both feet, plantar force/torque sensor data, and lower-limb kinematic constraints to accurately estimate the CoM position, orientation, and velocity. Furthermore, a projection constraint strategy incorporating both equality and inequality constraints is introduced to effectively suppress estimation divergence caused by noise accumulation, model inaccuracies, and abnormal measurements, thereby significantly enhancing robustness. Straight-line walking experiments conducted on the self-balancing lower-limb exoskeleton AutoLEE-G3 demonstrate that the proposed method achieves high consistency with motion capture measurements for CoM state estimation in both translational and rotational directions, with no noticeable drift during long-duration dynamic walking. Additional ablation experiments further verify the critical role of projection constraints in improving estimation accuracy and stability. These results indicate that the proposed method provides a high-precision and reliable state perception foundation for balance control and safe walking of self-balancing lower-limb exoskeletons.

Abstract:Lower-limb exoskeleton robots play an important role in rehabilitation training and walking assistance, where accurate lower-limb joint angle prediction is a key technology for achieving natural and coordinated gait. However, since surface electromyography (sEMG) signals are susceptible to individual differences and variations in movement patterns, achieving cross-task generalization for joint angle prediction remains a major challenge. To address the above issues, this paper proposes a cross-task generalization-oriented multi-joint angle prediction framework for the lower limb, which includes a non-redundant muscle synergy feature extraction algorithm and a joint angle prediction network with cross-speed and cross-subject generalization capabilities. First, the feature extraction algorithm employs non-negative matrix factorization to extract activation coefficient matrices from sEMG signals, and further uses the Uniform Manifold Approximation and Projection algorithm for nonlinear dimensionality reduction to obtain low-dimensional features with higher discriminative power as model inputs. Second, the joint angle prediction network extracts cross-task shared features through pre-training and leverages a small number of samples from new tasks for fine-tuning, enabling rapid adaptation to new tasks. Based on a self-constructed dataset and a public dataset, this paper conducts feature extraction method comparison experiments, model ablation experiments, and cross-speed and cross-subject generalization experiments. Experimental results show that in cross-speed experiments on the self-constructed dataset, the prediction errors of hip and knee joint angles range from 1.703 1° to 2.790 4°, with correlation coefficients all exceeding 0.98. Compared with baseline methods, the proposed method achieves significant improvements in both prediction accuracy and stability, demonstrating its effectiveness and generalizability. Finally, physical experiments on exoskeleton-assisted walking are conducted, further validating the feasibility and application potential of the proposed method in practical lower-limb exoskeleton applications.

Abstract:The existing discrete architecture graph neural networks (GNNs) are unable to extract deep spatiotemporal representations, the single variable dependencies in existing GNNs tend to learn a single type of time scale pattern, and fixed graph structures in existing GNNs cannot accurately represent the entire degradation process of aeroengine performance, resulting in poor prediction accuracy in predicting remaining useful life (RUL) of aeroengines. To address these issues, a novel GNN named multi scale dynamic graph neural ordinary differential equations (MDGODE) is proposed for RUL prediction of aeroengines. In MDGODE, a multi-scale framework is first established to capture the laten temporal dependence of aeroengine sensor signals at different time scales; secondly, a multi-scale dynamic graph structure learning module is designed to automatically adjust the spatial interaction of different sensor signals and capture the dynamic correlation between sensor signals at different time scales; finally, a new type of spatiotemporal coupled graph neural ordinary differential equations is constructed to capture the deep spatiotemporal representation in the performance degradation data of aeroengines. Due to the above characteristics, the proposed MDGODE has strong spatiotemporal representation capability and robustness to complex non-stationary performance degradation processes of aeroengines. Therefore, the prediction accuracy of MDGODE for aeroengine RUL is significantly improved compared to existing mainstream aeroengine RUL prediction methods. The proposed RUL prediction method based on MDGODE achieves a root mean square error (RMSE) of 14.06 and a Score of 879.36 for the RUL prediction values corresponding to the test set of FD004 subset in the aeroengine performance degradation dataset C-MAPSS. This experimental result verifies the effectiveness and superiority of the proposed method.

Abstract:Aiming at the problem that online detection of weld defects during laser welding of carbon steel thin plates is susceptible to vibration interference, and traditional non-destructive testing methods fail to achieve highly reliable real-time monitoring in confined detection spaces, a novel online weld detection method is proposed based on laser-excited ultrasonic waves and ultrasonic reception by dual electromagnetic acoustic transducer (EMAT). A finite element model for the 'one-transmit and two-receive' weld detection process of laser electromagnetic ultrasound is established. Taking 4 mm-thick carbon steel butt welds with 2 mm-diameter circular hole defects as the detection objects, the effects of parameters such as laser spot shape and size, number of turns and spacing of EMAT coils, and distance between the excitation point and receiving points on the ultrasonic transmission coefficient are analyzed, followed by experimental verification. Univariate analysis indicates that the excitation efficiency of a sheet-shaped spot is superior to that of a circular spot, and higher defect identification sensitivity is achieved when the coil has 4 turns with a turn spacing of 2.0 mm, and the excitation point is positioned at distances of 11 and 36.4 mm from the two receiving EMATs, respectively. On this basis, the optimal parameter combination is obtained through orthogonal experimental design, and a mobile laser-electromagnetic ultrasonic testing system is established, and experiments are conducted on butt welds containing circular holes of different diameters. Experimental results show that for the optimized laser electromagnetic acoustic transducer (Laser-EMAT) system, the attenuation amplitudes of the transmission coefficient for Φ3 mm and Φ2 mm circular holes are 1.53 times and 3.96 times those before optimization, respectively, indicating higher defect detection sensitivity. For example, the deviation coefficient caused by defects is 1.91 times and 4.98 times that before optimization, realizing more reliable detection. For instance, the signal variation coefficient in the defect-free area decreases from 6.77% to 5.41%. The optimized Laser-EMAT system yields a deviation degree of 0.62 and a 3.36% reduction in the transmission coefficient for a Φ1 mm-diameter circular hole. Although these results do not meet the requirements for reliable detection, the system eliminates the negative fluctuations corresponding to the worst parameter combination and significantly improves the detection sensitivity and reliability.

Abstract:The designed service life of cross-linked polyethylene (XLPE) cables is typically 30~40 years. The service life of cables in some early urbanized areas is gradually approaching this limit, leading to frequent faults that seriously impair the reliability of power systems. Cable joints are the weak link in cable lines, and their faults account for approximately 31% of all cable faults. In response to the above characteristics, this paper proposes a multi-strategy fusion cable joint fault diagnosis model (AKF-MTF-GADF-ViT-BiLSTM) based on the improved cuckoo-catfish optimization algorithm (ICCOA). First, the model uses adaptive Kalman filtering (AKF) to perform denoising and restoration processing on the original data, which effectively suppresses the nonlinear interference caused by environmental disturbances at the acquisition site and sensor jitter. Then, Markov transition field (MTF) and gramian angular difference field (GADF) are utilized to capture the dynamic and static fusion features of the data. A dual-path Vision Transformer (ViT) is adopted to extract the multi-dimensional correlation information of the features. Finally, the processed features are input into a bidirectional long shortterm memory network (BiLSTM) to learn the spatial features and temporal dependencies, yielding the final fault diagnosis results. In this article, ICCOA is introduced to tune parameters such as the number of feedforward network hidden units, the number of attention heads, and the decay factor of the ViT-BiLSTM model. This optimization addresses the problems of slow convergence and overfitting of the original model. The proposed model is verified for multi-fault diagnosis using measured data from an experimental platform and simulation data from SolidWorks Simulation. The overall diagnostic accuracy reaches 95.1%, which is 24.8% higher than that of the ablation model without filtering, 31.9% and 17.0% higher than those of the two ablation models with single-mode feature extraction respectively, and up to 10.9% higher than the three mainstream industry models of 1D-CNN-BiLSTM, CNN-LSTM, and VGAF-CLT. These results demonstrate that the proposed method effectively overcomes the limited generalizability of single models and improves the accurate identification of multiple cable joint fault types.

Abstract:Flip-chip microbump vibration signals are characterized by microscale features, strong interference in acquired signals, and severe confusion of multi-modal components, which give rise to complicated mapping relationships of defect features and a remarkable drop in defect recognition accuracy. It severely restricts the accuracy of defect identification and industrial applicability, becoming a key bottleneck that urgently needs to be addressed in the field of micro-nano electronic manufacturing defect detection. To tackle the deficiencies of existing knowledge distillation, including insufficient exploitation of logical knowledge, weak utilization of inter-class relationships, and limited capacity to achieve accurate defect identification, this article proposes a feature logic space grouping distillation (FLSGD) algorithm. This algorithm analyzes explicit and implicit knowledge around the logic of knowledge transfer, and projects interactive information into three probability spaces, namely target space t, intra-class space o＼t and outer-class space c＼o. It decouples the traditional Kullback-Leibler (KL) divergence to realize directed learning of logical features. The subjective and objective weighting method is adopted to quantify the weights of the three spaces and determine the parameter guidance values, so as to flexibly regulate the focus of knowledge transfer. To verify the effectiveness of the proposed algorithm, experiments are conducted on two ultrasonic vibration datasets derived from PCB substrate chip and silicon-based dummy chips. A single/dual-source ultrasonic excitation detection platform is established for validation, which verifies that the target space plays a dominant role in FLSGD, while the intra-class and outer-class spaces serve auxiliary functions. Experimental results demonstrate that FLSGD outperforms mainstream feature distillation methods in terms of recognition accuracy across various teacher-student network combinations, achieving a maximum accuracy of 96.47% and 97.69% on the two datasets, respectively. FLSGD can effectively exploit the modal characteristics of vibration signals, enhance the perception of key features, improve defect recognition accuracy and accelerate model convergence speed without introducing additional computational modules or increasing computing power consumption, thus providing an effective solution for lightweight defect detection in industrial applications.

Abstract:The excessive reliance on minutiae features in partial fingerprint recognition, leads to degraded recognition performance under conditions of small-area capture and low-quality images. To address the issue, the article propose a partial fingerprint recognition method based on multi-level feature fusion. First, the input fingerprint image undergoes normalization, estimation of the orientation and frequency fields, region segmentation, and Gabor enhancement to improve the stability of subsequent feature extraction. Secondly, during the search phase, a 256-dimensional normalized search vector is constructed using local curvature information, and K-means clustering is employed to filter candidate samples, thereby reducing the computational overhead of large-scale matching. In the fine-matching phase, a local structural model is formulated by combining minutiae information, neighborhood ridge line directions, and local texture information. For each minutiae, four neighborhood descriptor points are sampled, and the number of descriptors per image is controlled between 50 and 200 to balance matching efficiency and feature discrimination capability. Furthermore, dynamic programming, breadth-first search, and linear transformation estimation are integrated to achieve minutiae matching and image alignment. Subsequently, convex and concave ridge shape features are extracted based on ridge width variations, and three-level feature supplementary matching is performed within the overlapping regions. Finally, adaptive weighted fusion of level-2 and level-3 features is performed based on the number of available minutiae, forming a recognition framework characterized by level-1 retrieval, level-2 dominant matching, and level-3 compensation enhancement. Experiments are implemented on the FVC2002, FVC2004, and a self-built fingerprint database. The results show that the proposed method achieves a lower equal error rate (EER) than the comparison methods across different databases and under varying capture conditions. In the smallest capture area scenario, the EER is reduced to 3.35%~6.21%. This method effectively utilizes ridge structure to supplement discriminative information when minutiae are insufficient, exhibiting high recognition accuracy and robustness.

Abstract:Motor-cognitive integrated rehabilitation has proved effectiveness in patients with central nervous system disorders, including stroke. However, the dynamic interactions among brain regions that underlie these benefits are not well understood, and quantitative characterization of network-level interaction effects related to motor-cognitive function is essential for understanding and optimizing rehabilitation interventions. This paper presents a brain functional connectivity analysis method based on optimal-transport subnetworks to extract explainable motor-cognitive neural features from EEG signals. First, the experimental paradigm of grip force control with and without visual feedback was designed, during which EEG and grip force data were recorded simultaneously. Then, the functional connectivity matrix was constructed based on phase locking values and the most critical node connections were selected by the Kruskal algorithm, producing the optimal transport subnetwork with a refined network representation. Further, intra-and inter-ROI connectivity strength features were extracted by combining the optimal subnetwork with sliding task windows, enabling analysis of the grip force control mechanism and quantification of the dynamic changes of brain functional connectivity across different task stages. Fifteen subjects were recruited in the experiment, and the results demonstrated that visual feedback significantly enhanced specific intra-and inter-regional functional connectivity, with distinct enhancement patterns at different stages. In the preparation stage, high-frequency signals enhanced the connectivity within the prefrontal cognitive area. In the initiation stage, low-frequency oscillations enhanced coupling among the occipital visual cortex, central motor area and somatosensory association cortex. In the stable stage, low and high frequency signals jointly orchestrated connectivity between cognitive and motor areas to sustain continuous and stable grip control. Thus, this paper has established an explainable method to characterize brain region interaction and extract interpretable features, providing a technical tool and quantitative evidence for motor-cognitive integrated rehabilitation.

Abstract:To improve the subdivision accuracy of the capacitive angular displacement encoder, and address the problem of nonlinear distortion caused by parasitic capacitance, processing and installation error interference of the sensing signal, this article proposes an encoder sine and cosine signal subdivision error compensation method based on the time domain signal, the state duty cycle. Based on the timespace periodic sine and cosine signal sensing model of the capacitive encoder, the partial derivative model of the corresponding relationship between the gain error, the zero position offset error, and the time point of the time-domain signal is designed. Different from the traditional complex fitting algorithm, this method uses the high-frequency clock to collect and record the time information of the voltage threshold intersection of the pre-set periodic signal in real time, divides the sine and cosine single periodic signal into four time-domain state intervals, and defines the proportion of the duration of each interval in the whole cycle as "state duty cycle". The time point and zero offset of the actual signal are calculated based on the recorded time information and voltage threshold information, and the real-time correction of subdivision error caused by gain error and zero offset error is further realized. In this paper, a 16-bit small capacitive encoder with an outer diameter of d=50 mm manufactured by PCB process is used as the experimental object, and the subdivision error before and after compensation is tested and analyzed by using this method. The results show that the measurement error of the encoder can be reduced from ±30″ to ±10″, and the standard deviation of error can be reduced from 16.48″ to 6.85″. This method breaks through the traditional matrix operation compensation framework and is easy to implement at the hardware level. It has lower time and space complexity, faster response time, and can effectively improve the environmental adaptability, dynamic response characteristics of the encoder, and the reliability and stability of angle measurement.

Abstract:In power electronic systems, the variation in on-state resistance of the power switching device metal-oxide-semiconductor field-effect transistor (MOSFET) serves as a key indicator reflecting the device′s health condition, power loss level, system operating efficiency, and reliability. To address the problems of low accuracy, insufficient utilization of topological structure information, and weak model interpretability in existing MOSFET on-state resistance parameter identification methods under complex coupling operating conditions, this article proposes a two-stage hybrid model for multi-parameter identification by fusing Kolmogorov-Arnold networks (KAN) and heterogeneous graph Transformer (HGT). Firstly, time-domain statistical features and frequency-domain wavelet packet local energy values are extracted from the output voltage signal. A combined feature optimization strategy based on the Pearson correlation coefficient and mutual information is utilized to select features with high sensitivity and low redundancy as the key feature subset. Secondly, multiple parallel KAN sub-networks are established to realize the preliminary nonlinear mapping of on-state resistance for each MOSFET. Finally, a heterogeneous graph reflecting the physical semantics of the circuit is constructed, and topologyaware structured compensation is performed on the preliminary identification residuals through HGT to capture the nonlinear characteristics of devices and the coupling relationship of the system, so as to obtain the final identification results. Experimental results show that the average relative error of the proposed method for MOSFET on-state resistance identification in the three-phase space vector pulse width modulation (SVPWM) rectifier system is 0.962%, which is superior to traditional neural networks and homogeneous graph models. The average relative error of identification in hardware experiments is 1.992%, showing favorable generalization and robustness. Finally, this article verifies the effectiveness of feature selection and the physical consistency of model decision-making through Shapley additive explanations (SHAP) analysis and attention visualization, which provides a new idea for the high-precision and interpretable intelligent diagnosis of power electronic devices.

Abstract:Microwave technology serves as a pivotal technique for achieving high-precision measurement of rotating blade tip clearance (BTC). However, in 120 GHz microwave measurement systems, amplitude modulation, DC offsets, and amplitude-phase imbalance introduced by hardware imperfections and clutter reflections cause distortion of echo signals; Moreover, conventional static calibration methods are difficult to adapt to dynamic operating conditions, making integrated calibration and measurement challenging; Furthermore, phase-difference-based ranging methods are limited by the half-wavelength constraint of radio-frequency signals, resulting in a restricted measurement range. To address these issues, a signal calibration and beyond-half-wavelength BTC measurement method for rotating conditions is proposed. A dynamic calibration method based on robust segmented fitting and K-nearest-neighbor (K-NN) parameter estimation is developed to compensate signal distortions online during blade rotation. Second, to overcome the problem of phase wrapping and jumping in beyond-half-wavelength measurements, a phase accumulation correction strategy based on De Moivre′s formula is proposed, enabling continuous displacement demodulation beyond the half-wavelength limit through phase alignment and fine compensation of BTC interval signals. Experimental results show that, with laser sensor measurements as references, the average absolute errors of microwave BTC measurement are 1.611 and 1.563 μm for rectangular and curved blades, respectively, and the maximum repeatability errors are 0.623 and 0.481 μm. In a 5 mm beyond-half-wavelength continuous measurement experiment, the average demodulation linearity of 12 blades reaches 99.932%. The results demonstrate the accuracy and robustness of the proposed method for dynamic BTC measurement.

Abstract:Compared with voltage source inverters (VSI), current source inverter (CSI) has several advantages, including voltage boosting capability, shortcircuit and overcurrent protection, lower output current harmonic distortion, and higher operational reliability. However, the direct-current (DC) bus current in a CSI is generated by charging and discharging an inductor from a DC voltage source, and therefore does not exhibit the characteristics of an ideal constant current source. Furthermore, the inability of the CSI to operate in a step-down mode and to facilitate bidirectional power flow significantly limits its applicability in motor drive systems. To overcome the inherent limitations of the conventional CSI, this paper focuses on a two-stage CSI induction motor drive system incorporating a bidirectional chopper and investigates its maximum torque per ampere (MTPA) control strategy. Firstly, by establishing the steady-state mathematical model of the motor and the alternating-current (AC) filter capacitor, expressions are derived for the minimum DC bus current reference and the optimal stator excitation current reference under the MTPA criterion. Subsequently, based on the operating mode of the bidirectional chopper, a hysteresis control strategy for the DC bus is proposed to enable rapid and precise tracking of the DC bus current to its reference under varying operating conditions. Finally, a MATLAB/Simulink simulation model and a digitally controlled experimental platform are established to validate the proposed MTPA control strategy for the CSI driven motor system. The sensitivity of the MTPA calculation accuracy to variations in the motor and capacitor parameters is also analyzed. The results demonstrate that the implementation of a DC bus current hysteresis control based on a bidirectional chopper, the DC side of the CSI exhibits characteristics of a controlled current source. Furthermore, the proposed MTPA control strategy effectively reduces the torque ripple and current harmonic distortion of the induction motor, leading to improved dynamic and steady-state performance.

Abstract:To address the energy-efficient path planning problem of omnidirectional mobile robots in uncertain environments with multiple obstacles, this paper proposes an energy-aware path planning algorithm named energy-aware quadratic programming parameterization soft actor-critic(E-QPSAC), which integrates model predictive control parameterization with an improved soft actor-critic (SAC) framework. First, local path planning is formulated as a Quadratic Programming problem and efficiently solved by an improved primal-dual hybrid gradient algorithm. Second, a 12-dimensional state space (including robot pose, target information, and obstacle information) for SAC is constructed, along with a cascaded network architecture and a dynamic weight fusion mechanism to effectively balance obstacle avoidance and goal orientation. Meanwhile, considering the high energy consumption characteristics of omnidirectional mobile robots caused by environmental uncertainty, system nonlinearity, and motion characteristics, a five-dimensional energy consumption model is established, and an energy-saving reward function is designed based on this model. Finally, comparative experiments are conducted on a simulation platform with SAC, deep deterministic policy gradient (DDPG), and the prioritized replay and LSTM-based twin delayed deep deterministic policy gradient (PL-TD3) algorithm. The experimental results show that the proposed algorithm achieves a task completion rate of 98.7% (85.0% for SAC, 98.5% for DDPG, and 98.2% for PL-TD3), a total energy consumption as low as 18.274 kJ (25.765 kJ for SAC, 22.418 kJ for DDPG, and 21.191 kJ for PL-TD3), and an average path length of 6.45 m (6.56 m for SAC, 6.97 m for DDPG, and 7.13 m for PL-TD3). The proposed algorithm outperforms the others in task completion rate, energy consumption, path length, and robustness. The effectiveness of the proposed method is finally verified through practical experiments.

Abstract:During high-speed and high-precision motion of a gantry platform driven by dual permanent magnet linear synchronous motor (PMLSM), large dynamic synchronization errors are prone to occur due to the difference in dynamic responses between the two motors and the mechanical coupling effect of the crossbeam. Under the constraints of the linear guideways, these errors generate excessive mechanical stress and additional friction losses, severely affecting the motion accuracy and service life of the platform. Moreover, the motion state of the beam mover directly influences the dynamic response of the dual-PMLSM displacement tracking control, ultimately affecting the dynamic synchronization error of the dual PMLSM movers. To address these issues, this article first establishes the dynamic equations of the dual-PMLSM-driven gantry system and reveals the principle by which the beam mover position and its dynamic motion along the crossbeam affect the beam deflection motion. Then, based on the composite displacement error, a cross-coupled sliding mode synchronous control for the dual PMLSM movers is constructed. To mitigate the adverse effect caused by the continuously changing center of mass of the beam during its motion on synchronization suppression, the dynamic model is combined with the real-time position of the beam mover to perform current reconstruction, thereby reducing the dynamic synchronous displacement error. To overcome the control lag problem inherent in traditional fuzzy control that uses error as the feedback core, the net q-axis current of the beam mover, which is directly related to the beam torsion torque, is adopted in the fuzzy control. Taking the net q-axis current of the beam mover as the dominant factor, a fuzzy variable-gain control strategy is constructed by also incorporating the synchronous displacement error and synchronous velocity error of the dual PMLSM movers, so as to further reduce the influence of the beam mover′s dynamic motion along the beam on the dynamic synchronous displacement error. Experimental results show that, compared with the cross-coupled sliding mode synchronous control with constant gains, the proposed control strategy achieves a substantial reduction in the peak synchronization error during both the start-stop processes of the dual PMLSMs and the crossbeam mover.

Abstract:To address the issues of insufficient state observation accuracy and limited energy scheduling efficiency during the active equalization of lithium-ion battery packs, a hierarchical modular equalization control strategy based on an improved extended Kalman filter (EKF) is proposed. The system consists of an intra-module equalization network constructed from multiple improved Buck-Boost multi-path equalization circuits at the bottom layer, enabling direct energy transfer between adjacent and non-adjacent cells, and an inter-module equalization network formed by a bidirectional Flyback equalization circuit and a switch matrix at the top layer, facilitating energy exchange between modules and the entire battery pack. Taking the state of charge (SOC) estimated by the EKF as the equalization criterion, a fuzzy proportional-integral-derivative regulator is introduced to achieve hierarchical parallel coordinated control within and between modules, thereby improving the overall equalization efficiency of the system. To address the problem of increased observation errors caused by decreased sensitivity of the open-circuit voltage (OCV)-SOC curve in the plateau region, a plateau correction coefficient G is introduced into the EKF to adaptively adjust the observation gain, thereby enhancing SOC estimation accuracy. Experimental results show that the proposed equalization strategy achieves a maximum SOC estimation error of 2.9% and a root mean square error of 0.86%, which are 35.7% and 36.8% lower than those of the traditional EKF, respectively. Compared with the proportional-integral-derivative control strategy based on Buck-Boost and Flyback equalization circuits, the equalization time under static conditions is reduced by 37.1%, and under 1C charging and discharging conditions, it is reduced by 58.66% and 60%, respectively, with significantly improved dynamic response performance. To balance real-time performance and engineering feasibility, the Fata Morgana algorithm is employed to optimize the correction coefficient G offline and construct a lookup table. Online computation is achieved through table lookup on the TMS320F28335 platform, with a single control cycle time of 0.32 ms, accounting for only 3.2% of the 10 ms control cycle. The proposed strategy significantly enhances equalization efficiency while enhancing SOC estimation accuracy, providing a viable implementation path with engineering application value for the integrated measurement and control design of lithium-ion battery management systems.

Abstract:Due to the sparsity and limited vertical resolution of 3D light detection and ranging (LiDAR) point clouds, LiDAR-inertial measurement unit (IMU) simultaneous localization and mapping (SLAM) systems often lack the effective elevation constraints and suffer from the accumulated vertical drift. To address this issue, this paper proposes a 3D LiDAR elevation-constrained SLAM method for autonomous vehicles by integrating ground curvature fitting with motion prediction. Ground points are first segmented with the spatial pre-filtering and reflectivity-based secondary filtering. Then the local ground points are fitted by a quadratic surface to estimate the ground normal vectors and average curvature, which are combined with predicted vehicle motion to construct a novel elevation constraint. The proposed elevation constraint factor together with LiDAR odometry and IMU pre-integration factors is incorporated into a factor graph optimization framework to achieve multi-source collaborative optimization while maintaining the real-time performance. Experiments on the KITTI dataset and a self-collected campus dataset demonstrate that the proposed method provides strong elevation constraints, reducing the average localization error by up to 46.54% compared to the tightly coupled LiDAR-inertial odometry algorithm via smoothing and mapping (LIO-SAM), which significantly improves the map consistency and trajectory stability.

Abstract:As a critical transition zone between land and sea, tidal flat areas hold immense ecological research value. However, environmental characteristics such as high humidity, strong reflectivity, and complex topography make it difficult for traditional remote sensing methods to achieve high-frequency, high-precision, and non-invasive detailed monitoring. The rapid development of unmanned aerial vehicle (UAV) multispectral remote sensing technology has provided a new technical approach for real-time monitoring of tidal flat areas. Nonetheless, existing methods still have shortcomings in terms of radiometric calibration accuracy, channel consistency, and spectral feature stability, which limit the accuracy and efficiency of tidal flat land cover classification. To address these issues, this study developed an integrated UAV multispectral detection system tailored to the complex tidal flat environment. A combined data processing method was designed, incorporating indoor spectral calibration, outdoor graycard correction, and global affine registration. This scheme effectively mitigates radiometric distortions caused by high humidity and strong reflectivity, eliminates geometric and spectral discrepancies across channels, and significantly enhances the stability and distinguishability of spectral features. Based on this, rapid classification rules were established using a combination of the normalized difference vegetation index (NDVI), normalized difference water index (NDWI), and normalized difference water index normalized red-edge near-infrared index (NRNI). Precise identification of intertidal land features was achieved through multi-index collaborative threshold decision-making. Field validation was conducted in a typical tidal flat area. The experimental results indicate that this method can effectively highlight spectral differences among intertidal land features. By using the thresholds NDVI>0.4, NDVI∈［0, 0.2］∩NDWI<-0.1, NDWI>-0.1∩NRNI> -0.1, enabling precise classification of three types of coastal features: the water′s edge, beach forests, and foreshore gravel. The overall accuracy reached 96.7%, with a Kappa coefficient of 94.5%. With a processing time of 0.001 seconds per sample, the method enables real-time monitoring of intertidal zones at sub-meter resolution, meeting the detailed requirements of ecological research.

Abstract:The assembly of large aerospace structures is a critical step in the manufacturing process, with hole-shaft connections being a common configuration, and the accuracy of their pose measurement directly affects assembly quality. Traditional visual measurement methods are limited by image resolution, especially under long-distance working conditions, making it difficult to meet high accuracy requirements. Furthermore, existing super-resolution technologies are rarely used in industrial measurement and lack optimization methods specifically for the characteristics of hole-shaft images. This article proposes a pose measurement method for aerospace hole-shaft structures based on geometry-aware and pixel-accurate adjustable super-resolution. First, an image degradation strategy is optimized to address the characteristics of hole-shaft images, and a dedicated super-resolution dataset is constructed. Second, a geometry-aware and pixel-accurate adjustable super-resolution network is designed, achieving a balance between pixel-level fidelity and geometric structure quality by introducing low-rank adapters (LoRA) and a loss function based on context loss and edge enhancement gradient variance loss. For pose calculation, a spline feature extraction algorithm is proposed, and a pose optimization model based on forward projection geometric distance and hybrid product is established to address geometric deviations caused by manufacturing errors. Experimental results show that this method improves the position measurement accuracy from 0.008 mm to 0.003 mm and the attitude measurement accuracy from 0.14° to 0.07°, while significantly enhancing measurement stability. In practical applications, this method has been successfully used in helicopter lift system assembly, achieving an attitude adjustment accuracy better than 0.05 mm and an assembly time of less than 30 minutes, verifying its effectiveness and practicality.

Abstract:A visual measurement benchmark algorithm for thread crest line regression, correction, and evaluation is proposed to address the challenges of rapid benchmark positioning for small thread and large-diameter measurement in industry applications. Gaussian filtering is used to eliminate image noise, and the edge transition zone of the thread image is extracted based on a dual threshold binary method. An image coordinate system is then established. Feature points are extracted to preliminarily determine the thread center axis, which is used as the measurement reference. An image rotation regression model is constructed to perform the first correction of the thread measurement benchmark. Using the left thread crest line as the measurement benchmark, a statistical measure of the goodness of fit of a quantitative regression model with a coefficient of determination is used to adjust the measurement benchmark for the second time. Finally, the corrected coefficient of determination is used to evaluate the positioning accuracy of the measurement benchmark. On this basis, a major-diameter measurement algorithm is proposed by statistically analyzing edge transition information in thread crest images. The expected edge of the thread crest is projected onto the x-axis, and large-sample statistical data are used to determine the thread crest-edge position. The thread diameter is then calculated at the expected positions of the left and right thread crests, and the algorithm is calibrated using standard gauge blocks. The algorithm and measurement accuracy were experimentally verified through high-precision edge measurement of measuring blocks. The experimental results show that when the center axis of the equivalent block has a certain inclination angle with the x-axis, the maximum deviation obtained using the proposed algorithm is 0.001 7mm. Measurement experiments were conducted on the same metric thread, M2.5, using the MV-TOSEDP high-precision automatic image measuring instrument and the 103341 thread measuring machine, respectively. When the central axis of the thread has a certain inclination angle with the x-axis, the difference between the two test results is 3.1 μm, and the maximum measurement deviation is 8.8 μm. The average measurement time of the proposed algorithm is 2.480 1 seconds. In summary, the regression correction measurement method for the thread crest line satisfies the fast measurement requirements of metric small threads based on the visual measurement of the large diameter benchmark.