Abstract Short-window IMU preintegration factors are the measurement interface through which high-rate inertial samples enter many graph-based estimators. Most learned inertial correction methods, however, optimize raw samples, bias terms, propagated states, or trajectory-level losses rather than the preintegrated factor residual itself. We present MED-IMU, a factor-interface residual module for short-window IMU preintegration. Although the formulation retains rotation, velocity, and position residuals, the present evidence supports MED-IMU primarily as a rotation-correction framework; velocity and position outputs are treated as secondary diagnostics. MED-IMU predicts residual corrections for 0.25 s factors and uses a training-time chained rotation consistency loss so adjacent factors compose coherently. On EuRoC, MED-IMU improves over AirIMU at 0.25 and 0.50 s, while paired moving-block bootstrap intervals overlap at 0.75 and 1.00 s. On TUM-VI, MED-IMU gives lower rotation-factor error than AirIMU across 0.25–1.00 s, reducing the 1.00 s error from 1.192 to 0.781 degrees. Direct EuRoC-to-TUM-VI transfer degrades performance, but target-domain output calibration with 10–300 s of target-domain calibration data partially recovers held-out target-room accuracy. A controlled SO ( 3 ) factor-insertion analysis evaluates whether corrected rotation factors can be inserted into a shared graph, but it is not a full visual–inertial odometry or PGO benchmark. Finally, an uncertainty evaluation shows that the current covariance head is best interpreted as auxiliary diagnostic output rather than reliable direct backend weighting. These results support rotation-centered factor-residual learning as a practical interface between neural inertial correction and preintegration-based estimation, with explicit limits on full-factor, cross-domain, and backend claims.
Abstract In modern industrial systems, vibration signals acquired by sensors are frequently subject to varying degrees of distortion and degradation. Furthermore, the high cost of annotation in industrial settings leads to a severe scarcity of labeled fault samples. In recent years, self-supervised contrastive learning shows great promise for fault diagnosis under limited labeled data. However, most existing methods rely on instance-level contrastive learning, resulting in insufficient inter-class discriminability in practical applications. Additionally, they fail to effectively integrate time–frequency complementary information with the structured semantics of soft labels, restricting representation capacity and generalization. To address these issues, this paper proposes a self-supervised framework called time–frequency feature-level independence contrastive learning with semantic consistency learning. It combines multi-scale feature extraction with multi-level collaborative optimization to learn discriminative representations from unlabeled vibration data. Specifically, we first design a multi-scale feature extraction module to extract and coordinate local time–frequency and global features. Within the multi-level collaborative optimization module, we introduce a time–frequency feature-level independence contrastive learning component. It effectively refines the feature representations and enhances their discriminability. Furthermore, we integrate a semantic consistency learning module to strengthen semantic constraints, boosting the model’s robustness and generalization under data distribution shifts. Extensive experiments on three public datasets demonstrate that the proposed method can efficiently extract generalized feature representations from unlabeled vibration signals and achieves outstanding diagnostic performance. The source code will be available later at https://github.com/szq0816/TFFICL-SCL .
Abstract High-performance magnetic shielding systems require accurate shielding factor (SF) evaluation, yet testing ultra-high-performance systems (SF > 10 8 ) remains challenging due to the need for impractically large external excitation fields. We developed a testing system specifically designed for this task, based on a portable atomic magnetometer, achieving a sensitivity better than 1 f T/Hz 1/2 . By applying an order of 10 4 nT external field via uniform coils and extracting signals via noise spectrum analysis, we achieved the capability to test SFs up to the order of 10 10 at a signal-to-noise ratio of 2. Testing of 10‐layer magnetic shielding cylinders revealed a pronounced frequency dependence of the SF within the 1–100 Hz band. The radial SF ranges from 1.7 × 10 9 to 5.0 × 10 10 , while the axial SF varies between 4.8 × 10 7 and 1.3 × 10 10 , exhibiting the characteristic of an initial increase followed by a decrease with frequency. Furthermore, the phase-frequency characteristics, which are often overlooked in previous studies, were also tested. The results revealed pronounced hysteresis, which is of great significance for practical applications involving real-time magnetic field closed-loop systems.
Abstract Maritime over-the-horizon (OTH) two-station angle-of-arrival localization is affected by both Earth-curvature modeling errors and systematic bearing biases introduced by tropospheric scattering. To improve localization accuracy while retaining the existing two-station direction-finding architecture, this paper proposes a WGS-84-constrained weighted least-squares (WLS) method with tropospheric equivalent bearing-bias compensation. The method formulates the bearing geometry in the WGS-84 ellipsoidal reference frame and represents the angular deviation between the dominant scattered arrival direction and the true geometric bearing as a station-specific equivalent bias estimated from calibration data or prior information. Random direction-finding errors and residual compensation uncertainty are incorporated into the observation covariance matrix of the WGS-84-constrained WLS formulation. Monte Carlo simulations under a representative maritime OTH configuration show that the proposed method reduces the localization root-mean-square error (RMSE) from 50.71 km with conventional planar intersection and 11.07 km with uncompensated WGS-84-constrained WLS to 1.49 km. When the distance from the fixed station to the target varies from 400 km to 800 km, the RMSE remains below 2 km. With imperfect compensation, the RMSE remains approximately 4 km even when the standard deviation of the residual compensation error increases to 0.30 ∘ . Tests on an independent validation dataset generated from realistic maritime scenario parameters yield an RMSE of 1.48 km, compared with 50.87 km and 11.85 km for the two baseline methods. These results demonstrate that the proposed method can jointly mitigate long-range curvature errors and tropospheric-scattering-induced bearing biases, offering a practical means of improving maritime OTH localization accuracy.
Abstract In fields such as robotics and autonomous driving, multi-lidar data fusion is often employed to obtain a comprehensive view of the environment, thereby enhancing environmental perception capabilities. Achieving robust multi-lidar fusion relies on precise calibration between sensors. This paper proposes a dual-lidar extrinsic calibration method based on semantic segmentation and tetrahedral structures. The method is realized by a calibration board placed within the common field of view of the lidars while simultaneously capturing static point cloud data from each lidar, without the need for additional sensors or well-initialized extrinsic parameters. Ground removal is then performed using point cloud processing algorithms. Subsequently, the MA-PointMLP semantic segmentation network is introduced to extract the segmented calibration plate, forming a three-sided point cloud together with the ground. Finally, the lidar extrinsic parameters are solved through nonlinear optimization based on feature point pairs from the calibration board. A tetrahedral structure composed of a triangular plate and the ground was constructed for experimental validation. Experimental results indicate that the average absolute error of the proposed method is 0.029 m, with a root mean square error of 0.031 m. The proposed method outperforms the existing calibration methods, providing a research foundation for subsequent multi-lidar data fusion.
Abstract Accurate standard deviation estimation is a prerequisite for reliable control-chart limits, process-capability indices, and ISO GUM Type A evaluation of the experimental standard deviation in automated measurement and process-control systems. The classical corrections (Bessel, n − 1.5 , Gurland–Tripathi) remain systematically biased at every finite sample size. For the symmetrically truncated normal distribution arising from range-limited sensors, saturating analog-to-digital converters, and bounded manufacturing processes, no compact closed-form correction exists. Compact, deployable bias-correction formulas for both distributional settings are obtained here by coupling large-scale Monte Carlo simulation ( 2 × 10 8 iterations per configuration) with symbolic regression. For the normal distribution, tabulated correction factors are validated against the exact chi-distribution c 4 ( n ) benchmark, achieving | E − σ | < 1.5 × 10 − 6 ; a compact linear fractional function (LFF) provides a deployable closed form with residual estimator bias at most 1.5 × 10 − 3 (at n = 2 ), falling below 10 − 4 for n ⩾ 9 . For the symmetrically truncated normal distribution, a second-degree rational correction factor b ( n , r ) is identified for the truncated region ( r < 4 , n = 2 to 100 ), with the correction reverting to the unrestricted normal form for r ⩾ 4 , reducing uncorrected biases of 10 to 25 % in constrained measurement scenarios where no analytical alternative exists. The bias-variance trade-off is quantified explicitly: the mean squared error overhead of the unbiased estimator falls below 5 % for n ⩾ 9 and below 1 % for n ⩾ 39 . Both corrections are validated on two independent open datasets of real measurements with different measurands and instrument classes; they reduce the systematic bias by one to two orders of magnitude relative to classical estimators. The residual that remains after correction is bounded by | γ 2 | / ( 8 n ) , where γ 2 is the excess kurtosis of the measured process, so the user can estimate this bound in advance from a measurable property of their own data. Both formulas require only elementary arithmetic and execute on the same controller or instrument that performs the measurement.
Abstract In industrial visual inspection, unsupervised anomaly detection has significant application value due to the elimination of anomaly labeling requirements. However, existing methods often rely on independent modeling by category, leading to high storage and maintenance costs; unified multi-category modeling is susceptible to the diversity of normal patterns, resulting in approximate identity mappings and weakening anomaly representation capabilities. To address these issues, we propose a hierarchically conditioned denoising and guidance framework (HCDG), which combines adaptive hierarchical feature fusion with error-aware conditional denoising. HCDG integrates shallow texture and deep semantic features and uses noise prediction errors to guide adaptive denoising in the feature bottleneck. A feature-guided decoder reconstructs normal features, and reconstruction and noise prediction errors are jointly used for image-level and pixel-level anomaly scoring. HCDG achieves competitive overall performance on MVTec AD, reaching 99.7% I-AUROC, 99.8% I-AP, and 99.4% I-F1-max at the image level. At the pixel level, HCDG attains 98.4% P-AUROC, 70.2% P-AP, 69.9% P-F1-max, and 95.0% P-AUPRO. These results suggest that the proposed denoising and guidance strategy not only preserves image-level discrimination but also yields more stable pixel-level localization under unified multi-class training.
Abstract To address the limitations of traditional contact-based vibration measurement, including restricted sensor installation, sensor-induced loading, and the difficulty of separating multiple periodic vibration components, this paper proposes a high-speed video-based non-contact vibration analysis method using adaptive fusion micro-vibration measurement and periodic-subspace-based adaptive mode decomposition (PS-AMD). Rolling bearing fault diagnosis is used as a representative case study to validate the effectiveness of the proposed method in periodic impact vibration extraction and component separation. The method consists of two stages, namely visual micro-vibration extraction and periodic mode decomposition. In the displacement extraction stage, color video frames are first transformed into the luminance-inphase-quadrature (YIQ) color space, and dynamic fusion weights are constructed according to channel variances within the region of interest to obtain a more robust single-channel texture field. Multiscale and multi-orientation complex Gabor filter banks are then employed, and the local phase differences between adjacent frames are used to estimate subtle displacements, thereby generating a visual displacement time series and realizing visual micro-vibration extraction. In the periodic component separation stage, to address spectral crowding and mode mixing caused by the superposition of multiple periodic impulsive components under compound fault conditions, a PS-AMD is introduced. Significant periods are selected using a criterion that jointly considers energy contribution, impulsiveness, and periodic reliability, and sequential residual reconstruction is further adopted to achieve multi-periodic mode separation. Experimental results under both single-fault and compound-fault conditions show that the extracted visual displacement signals are in good agreement with synchronously acquired acceleration signals in terms of dominant envelope spectral peaks, harmonic structures, and modulation characteristics. The proposed method can effectively identify inner-race and outer-race fault characteristic frequencies and achieve clear periodic mode separation under compound fault conditions. These results indicate that the proposed method can extract weak vibration responses and separate periodic components under non-contact measurement conditions. It provides an interpretable visual measurement and periodic decomposition approach for vibration analysis of rotating machinery, while the rolling bearing experiments further verify its effectiveness in fault feature identification.
Abstract Rolling bearing fault detection under strong background noise remains a challenging task because fault-induced repetitive transients are usually weakened by the transmission path and masked by random interference. To address this issue, this paper proposes a periodic sparse joint deconvolution (PSJD) method for extracting weak bearing fault impulses from vibration measurements. The proposed method formulates bearing fault detection as an inverse filtering problem, in which the deconvolved signal is expected to exhibit both strong periodicity and high sparsity. Specifically, a multi-period correlation term is constructed to enhance repetitive impulses occurring at the fault characteristic period, while a logarithmic sparsity term is introduced to promote impulsive structures and suppress noise-related components. In addition, a regularization term is imposed on the inverse filter to improve numerical stability and avoid excessive oscillation of filter coefficients. The inverse filter is updated using a gradient-ascent scheme with normalization to remove scale ambiguity. The effectiveness of the proposed method is verified using simulated signals and experimental bearing vibration signals. The results demonstrate that the proposed method can effectively recover weak periodic transients and highlight fault characteristic frequencies in the envelope spectrum, even when the original signal is contaminated by strong noise. Compared with conventional deconvolution methods, the proposed method provides better capability in enhancing repetitive impulsive features and improving the reliability of bearing fault identification.
Abstract Accurately assessing the passive radon adsorption performance of porous materials under natural conditions is of vital importance for environmental radiation protection. Currently, the existing testing methods have obvious limitations: the active adsorption method is affected by forced airflow and it is difficult to extract the true physical adsorption parameters; while the traditional static chamber method faces the dual problems of gas mixture delay and mechanical operation interference. To address the aforementioned bottlenecks, this paper proposes a dual-cycle closed radon monitoring system based on a magnetic triggering mechanism. By placing the sample box with a magnetic control cover in the chamber in advance, the passive adsorption process of the dried coconut shell activated carbon can be triggered instantly without damaging the system’s sealing boundary. The experiment was conducted in two stages: Firstly, the baseline leakage rate of the system was calibrated by monitoring the decay of natural radon. Then, the adsorption process was remotely initiated by a magnetic trigger, and the entire process did not disrupt the sealed state. The full-time concentration data were nonlinearly fitted using a first-order kinetic model that included leakage and adsorption terms, resulting in an average natural leakage rate of (1.19 ± 0.09) × 10 −5 s −1 and a passive radon adsorption rate constant of (3.61 ± 0.20) × 10 −4 s −1 (dry coconut shell activated carbon). The entire adsorption process was driven solely by the natural concentration gradient, and the samples could be pre-stored in sealed containers for a long time before the test. Crucially, the measured adsorption rate is approximately 30 times greater than the baseline leakage, demonstrating the system’s ability to effectively decouple genuine adsorption kinetics from background errors. This low-disturbance paradigm enables reliable evaluation of porous materials for practical indoor radon mitigation.
Abstract Vacuum system finds increasingly widespread application across diverse industries. Precise measurement of vacuum pressure is essential for ensuring the safe and reliable operation of technologies and equipment. For fully sealed or online-operating vacuum chambers, detecting and assessing the internal vacuum state without compromising the structure or functionality of the chamber holds significant importance. This paper proposes a novel contactless online vacuum detection method that combines the temperature characteristics of permanent magnets, electromagnetic induction and the Pirani principle. The AC coil outside the vacuum chamber heats the internal permanent magnet by generating eddy currents. By detecting variations in the electromagnetic force between the permanent magnet and the external DC coil during the cooling process, vacuum pressure is determined. Slower changes in electromagnetic force indicate lower vacuum pressure. This approach offers simpler operation, lower cost, and broader applicability without requiring complex instrumentation. Based on theoretical research and COMSOL simulation design, the proposed contactless vacuum detection method is expected to detect vacuum pressures as low as 10 −2 Pa–10 Pa.
Abstract With the rapid advancement of intelligent driving, LiDAR and cameras provide complementary geometric and semantic information. Therefore, fusing these two sensors is a mainstream approach for 3D object detection. However, existing methods still face two key challenges: effectively exploiting complementary information among candidate boxes during post-processing and balancing detection accuracy with computational efficiency. To address these challenges, we first propose an aggregated Euclidean distance weighted box fusion (AED-WBF) method, which aggregates complementary information from multiple candidate boxes during post-processing to improve bounding-box selection and localization accuracy. We further develop a hybrid deformable half-conv (HDHC) module that jointly enhances global and local feature representations through hierarchical offset prediction and local neighborhood attention. By integrating half-conv with a separable self-attention mechanism, HDHC reduces computational complexity while maintaining detection accuracy. Based on AED-WBF and HDHC, we construct EAEPNet, an efficient multilevel LiDAR–camera fusion network for 3D object detection. Extensive experiments are conducted on the KITTI and nuScenes datasets. On the KITTI test set, EAEPNet improves the mean average precision (mAP) by 2.73% over the baseline network. On nuScenes, EAEPNet achieves a mAP of 72.5% and an nuScenes detection score (NDS) of 74.4% on the validation set, as well as a mAP of 73.2% and an NDS of 75.3% on the test set. These results validate the effectiveness of EAEPNet in multi-sensor 3D object detection and spatial measurement. Its strong performance across multiple datasets further demonstrates its potential for intelligent driving and real-time high-precision spatial measurement. The code is available at: https://github.com/juanmao73/EAEPNet .
Abstract Real-time mechanical monitoring of liquid-to-solid phase transitions is essential in materials science and biomedical diagnostics, yet existing instruments face challenges in continuously and nondestructively tracking the complete phase transition process of small-volume samples. Here we present an instrument for monitoring small-volume liquid-to-solid phase transitions based on magnetically driven droplet resonance. By combining phase-sensitive optical coherence tomography (OCT) detection with magnetic excitation using embedded superparamagnetic nanoparticles, the instrument enables continuous measurement of the resonance frequency throughout the entire transition process of a single 30 μ l droplet. We designed a sidewall-free sample base to confine the sample geometry via surface tension, providing consistent vibrational boundary conditions for cross-phase-state measurements. To mitigate electromagnetic interference-induced mechanical noise in the optical detection path under alternating magnetic fields, a non-metallic optomechanical structure was designed, reducing the displacement noise spectral density to below 3 nm Hz − 1 . The instrument achieves frequency measurement stability of ± 0.4 Hz and ± 0.6 Hz (15 min, standard deviation) for liquid and solid standard samples, respectively. Monitoring of agar gelation and porcine whole blood coagulation demonstrates the instrument’s ability to distinguish distinct phase transition mechanisms. This instrument provides a continuous, non-destructive mechanical monitoring tool for small-volume liquid-to-solid transition research and coagulation function assessment.
Abstract Urban planning, disaster assessment, and related sectors have extensively applied deep learning–based techniques for building extraction from remote sensing images as a result of the perpetual growth of computing power and satellite sensors. Although numerous deep learning–based techniques for building extraction are being proposed forth, there are still many obstacles to overcome in the extraction of buildings from remote sensing images. For example, background pixels often occupy a large proportion of remote sensing images, making it challenging to accurately learn complete building edge information. Therefore, this paper proposes a novel building extraction network based on edge decoupling refinement and pixel shuffle reconstruction (EDRPS-Net). EDRPS-Net adopts a dual-branch architecture that uses a convolution-based branch and a transformer-based branch as the core encoder. Furthermore, EDRPS-Net integrates the foreground enhancement module (FEM), the edge decoupling refinement module (EDRM), and the improved efficient subpixel convolutional neural network (ISPCN). First, FEM strengthens the weighting and semantic representation of foreground regions by integrating wavelet transform and attention gate. Second, EDRM decouples decoder features into body and edge features through learnable flow fields, and employs independent stems to model building edges, thereby refining edge information. Finally, this paper designs ISPCN, which integrates efficient subpixel convolutional neural networks with densely connected structures. As a learnable component within the decoder, ISPCN effectively reconstructs high-resolution feature maps, thereby optimizing the recovery of building details and spatial information. Experimental results on the WHU, Massachusetts, and Inria datasets demonstrate that EDRPS-Net achieves Intersection over Union values of 89.92%, 75.30%, and 77.77%, respectively.
Abstract Early fault prediction in chemical processes remains challenging because incipient signatures are weak, evolve gradually, and are often obscured by dominant process dynamics and measurement noise. This paper proposes a temporal–spectral early-warning network (TS-EWNet), which directly maps historical multivariate measurements to fault-related probabilities without intermediate signal forecasting. A coarse-to-fine architecture first captures the global process context and then enhances weak fault-related variations by integrating temporal responses with low-frequency spectral evidence. TS-EWNet is evaluated on the Tennessee Eastman Process (TEP) benchmark, a first-principles continuous stirred-tank reactor (CSTR) simulation, and a laboratory-scale thermo-hydraulic platform. On the TEP dataset, TS-EWNet achieves an accuracy of 93.67 ± 0.15 % and a Macro-F1 of 93.73 ± 0.19 % at H = 10 , while retaining an accuracy of 89.93 ± 0.22 % at H = 20 . At H = 10 , its Early-F1, stable correct warning rate, and penalized normalized stable detection delay are 68.77%, 69.62%, and 0.3286, respectively. After Holm correction applied separately to each metric, 13 of 14 Macro-F1 comparisons and all 14 Early-F1 comparisons remain statistically significant. On the CSTR dataset, the model achieves 96.62% accuracy and 97.89% Macro-F1 at H = 10 . Under continuous streaming operation on the laboratory platform, it obtains a macro-average accuracy of 77.71% and a Macro-F1 of 70.21%. For a representative F5 trajectory, a stable warning is confirmed 45 s before the reference fault manifestation, while the inference time remains below the 1 s sampling interval. These results demonstrate the effectiveness of TS-EWNet for multi-horizon fault prediction, initial-stage warning, and real-time process monitoring.
Abstract Microcracks and other invisible damage often initiate beneath the protective coating on carbon fiber reinforced polymer (CFRP) surfaces, posing a serious threat to structural integrity. To achieve effective non-destructive testing, a microwave sensor based on a complementary spiral resonator (CSR) is proposed. The sensor adopts a double-layer microstrip structure, with a microstrip transmission line on the top layer and CSR units etched into the ground plane on the bottom layer, which excites deep notches to achieve resonance. When the CSR near-field interacts with a damaged area on the CFRP, local variations in the material’s electromagnetic parameters cause a shift in the sensor’s resonant frequency, enabling damage identification. Two sensor configurations, a single-resonator unit and an array-resonator unit, are designed for crack detection and spatial localization, respectively. Simulation and experimental results demonstrate that the single-resonator sensor can detect microcracks with a width of 0.1 mm beneath the CFRP coating, and evaluate crack size within the 0.5–2 mm range; by identifying the frequency shift patterns of different resonant units, the array-resonator sensor can effectively determine the crack location while maintaining detection sensitivity. This study provides an effective solution for the non-destructive testing of CFRP structures.
Abstract Cable-driven manipulators rely on multiple driving cables to achieve motion, but cable wear during long-term service can reduce motion accuracy. Existing methods have difficulty exploiting physically coupled tension measurements for joint-level fault diagnosis and subsequent cable-level localization. Therefore, this paper proposes a physics-informed two-stage fault diagnosis and localization method. First, based on the mechanical mapping relationship between driving-cable tension and joint angle, a physics-informed graph neural network model is constructed by integrating a joint-state graph structure with physics-constrained loss terms. This model represents relationships between adjacent angle–tension states within each joint and inter-joint tension coupling through a joint-state graph structure and enhances physical consistency through tension consistency, cable-length consistency, and angle-boundary constraints, thereby enabling joint-level fault diagnosis. Subsequently, active equal-distance contraction excitation is designed for the faulty joint to transform cable faults into joint-angle responses. Fiber Bragg grating angle sensing is used to collect joint-angle data, and Gramian angular summation field encoding is combined with a residual network (ResNet) model to localize the faulty cable. Experimental results show that the proposed method achieves an overall accuracy of 96.67% in the fault diagnosis task and an accuracy of at least 93.71% in the fault localization task, demonstrating its good fault-recognition capability under different load and pretension conditions.
Abstract The power industry is a crucial backbone of global economic and social development. However, external force damage to transmission lines severely threatens the secure and stable operation of smart grids. Existing detection methods often struggle with feature extraction for slender targets, class imbalance, and geometric misclassifications, while lacking automated warning mechanisms. This paper proposes an intelligent detection framework based on 3D point clouds and knowledge graphs (KGs). First, a multi-semantic grouping convolutional block attention module is developed to enhance feature representation for slender and sparse targets. Furthermore, a weighted PolyLoss function is introduced to address class imbalance and geometric similarities between pylons and construction machinery. Finally, a KG for transmission line external force damage detection is designed to fuse unstructured safety regulations with structured 3D point cloud information, enabling automated safety distance warnings. Experimental results demonstrate that the proposed method significantly outperforms baseline models, achieving improvements of 20.04% in mean intersection over union and 12.42% in overall accuracy, while surpassing other methods.
Abstract Reliable visual measurement of insulator defects is essential for safe power-transmission inspection. In unmanned aerial vehicle (UAV)-based inspection, extreme aspect ratios, complex backgrounds, weather-induced degradation, and tiny low-contrast defects introduce substantial uncertainty into defect localization and classification. To improve detection reliability under such uncontrolled conditions, we propose SGMP-DEIM, an end-to-end real-time insulator defect detector based on Semantic-Grouped Mamba and pyramidal multi-granularity feature enhancement. In the backbone, a semantic grouped mamba stage is designed to guide the state-space modeling process with semantic grouping, enabling long-range structural modeling of insulator strings with linear complexity. In the neck, a multi-granularity diffusion pyramid network with a high-frequency dual-view feature aggregation module is developed to preserve high-resolution details and improve the localization of tiny defects. In addition, a clustering-enhanced meta-encoder is introduced to learn semantic prototypes and suppress background-induced false alarms. Experiments on two public insulator datasets show that SGMP-DEIM improves the baseline by 7.2% in AP 50 and 6.9% in AP 50 : 95 , while maintaining a favorable accuracy–efficiency trade-off. A controllable adverse-weather robustness benchmark is further constructed and used only for testing, demonstrating the potential of SGMP-DEIM for reliable visual measurement in complex UAV inspection scenarios.
Abstract Multi-motor synchronous control (MMSC) is essential for the integrated robotic joint, where synchronization performance directly impacts accuracy and stability in space-constrained applications. Traditional control methods often overlook this coupling and fail to balance disturbance rejection with energy optimization-a critical challenge for computationally constrained embedded measurement and control platforms. To address this, this paper proposes an event-triggered model predictive iterative learning (EMPIL) strategy to achieve high-precision angle tracking and coordinated control. Specifically, the scheme achieves dynamic decoupling through lumped uncertainty modeling and active compensation of differential dynamics. Simultaneously, a complementary mechanism utilizing iterative learning control and terminal integral sliding mode control suppresses diverse disturbances to ensure tracking precision without strict reliance on precise measurements. Furthermore, an event-triggered mechanism embedded in the receding horizon framework breaks the traditional periodic model predictive control paradigm. This eliminates computational redundancy and redundant communication overhead to realize an on-demand match between energy efficiency and tracking precision. Simulation and experimental results, incorporating external measurements from a motion-capture system, confirm that the proposed method achieves superior performance in angle tracking accuracy, coordination consistency, and energy efficiency.
