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.
Abstract To reconcile the inherent accuracy-efficiency trade-off in on-machine measurement of topologically modified involute tooth flanks, this study develops an on-machine adaptive sampling and path planning strategy grounded in curvature-modification feature mapping. Departing from conventional uniform sampling and singlefeature-based adaptive paradigms, the proposed approach synergistically fuses local curvature and topological modification information, employing the resultant mean curvature as a unified geometric indicator. A PSO-GRNN model is established to nonlinearly map this unified indicator to optimal sampling densities and spatial distributions.Leveraging the surface partitioning concept, a two-stage subdivision scheme integrating fuzzy clustering and region growing partitions the tooth flank into sub-surfaces of approximately equivalent density, enabling adaptive samplingpoint allocation; simultaneously, the initial sampling distribution is optimized subject to fitting-error and tooth profile/lead deviation constraints. A hierarchical path-planning strategy is devised: simulated annealing is employed to optimize the global route, minimizing both stylus orientation changes and traversal distance, while an improved ant colony algorithm integrating A*-guided pheromone initialization and goal-oriented heuristic function optimization is employed to locally replan collision-prone path segments. Experimental validation on the L300G gear grinder demonstrates that the proposed methodology exhibits consistency with Gleason 350GMS reference results, while reducing full-flank sampling time by 40% (from 6 min to 3.6 min) without compromising metrological accuracy in actual on-machine measurement scenarios. Experimental results confirm the proposed approach effectively mitigates the inherent accuracy-efficiency imbalance in traditional uniform sampling strategies, furnishing an accuracyefficiency co-optimization framework for high-precision, efficient on-machine gear measurement.
Abstract Abstract：During physical trial assembly (PTA), conventional measurement methods such as total stations have limited capability in capturing continuous geometric deviations and rely heavily on manual operation. In addition, the PTA process requires substantial resources. This study develops a virtual trial assembly (VTA) method for welded box type steel components using high accuracy point cloud data acquired via terrestrial laser scanning (TLS). The method establishes a segmentation framework operating at the component scale, integrating multi-scale filtering with clustering to reliably isolate individual components. A proxy bounding box is further introduced to identify assembly regions and support subsequent feature extraction. To meet the engineering requirement for reserved welding gaps in welded assemblies, a VTA strategy based on a virtual control plane is developed. Case studies based on steel box girder components and consecutive arch rib segments are conducted to validate the feasibility of the proposed TLS based VTA method for steel bridge construction.
Abstract Industrial equipment fault diagnosis remains challenging under variable-speed operating conditions and extreme label scarcity, where vibration signatures are easily affected by speed-induced spectral variations, operational noise, and amplitude outliers. Although Graph Neural Networks (GNNs) can exploit structural relationships among vibration samples, their application to fault diagnosis is still limited by unreliable initial features and scarce labels, which may lead to distorted graph topology, error-prone message passing, and over-smoothed node representations. To address these issues, this paper proposes the Dual- Stream Gated Fusion Network coupled with a Label Propagation System (DSGF-LPS), a semi-supervised graph learning framework integrating robust graph construction, label propagation, and dual-stream gated fusion. Specifically, Hanning-windowed Fast Fourier Transform (FFT) and Spearman-rank-correlation-based graph construction are first used to obtain noise-resistant frequency-domain topology. Then, the label propagation system expands the limited ground-truth labels to high-confidence pseudo-labels, alleviating insufficient supervision. Finally, the Dual-Stream Gated Fusion Network (DSGF-Net) adaptively fuses local spatial attention captured by Graph Attention Network version 2 (GATv2) and multi-scale spectral semantics extracted by Chebyshev graph convolutions, while LayerScale and global residual connections are introduced to mitigate over-smoothing. Experiments on two rotating machinery datasets demonstrate that DSGF-LPS achieves accurate and stable diagnosis under highly non-stationary and label-scarce conditions, attaining over 97% accuracy with only two labeled samples per class.
Abstract Tropospheric delay modeling accuracy is a critical factor affecting the performance of Precise Point Positioning Real-Time Kinematic (PPP-RTK). However, reliable tropospheric delay estimation remains challenging in regions characterized by sparse reference stations and complex terrain. To address this issue, this study proposes a novel tropospheric delay modeling method that integrates tropospheric delays derived from the Pangu-Weather Model with Global Navigation Satellite System (GNSS) observations to enhance PPP-RTK positioning performance. The proposed method first estimates tropospheric delays from both techniques at different locations and subsequently models the inter-technique bias. The method is validated using a network of 274 GNSS stations across Europe. Results from a one-month experiment demonstrate that the Pangu-Weather Model achieves zenith wet delay (ZWD) estimation accuracy at the level of approximately 2~3cm. In regions with sparse reference stations and complex terrain, the proposed method improves ZWD accuracy, with root mean square (RMS) reductions of 36.1%, 54.7%, and 68.3% compared with three conventional methods. In PPP-RTK positioning, the proposed method enhances convergence speed, accuracy, and stability. The average convergence time is reduced by 9.9%, 12.7%, and 11.3%, while the three-dimensional positioning accuracy is improved by up to 12.6%, 27.8%, and 22.8%. The proportion of vertical positioning errors within 5cm increases from 95.6% to 99.4%. Furthermore, large-scale experiments demonstrate consistent improvements across multiple error intervals, with the proportion of errors within 1cm increased by up to 3.2%, 3.8%, and 16.5% in the east, north, and up components, respectively. These results demonstrate the effectiveness, robustness, and applicability of the proposed method, particularly in challenging environments with sparse reference stations and complex terrain.
Abstract In this paper, a physics-informed neural network (PINN) based on transfer-learning is proposed for maglev planar motor (MPM) with passive Halbach permanent magnet array (HPMA) mover to realize high-precision six degree-of-freedom (6-DoF) pose detection. First, large-scale analytical datasets are generated according to 3D analytical magnetic field model based on the surface current method (SCM). Subsequently, a transfer-learning framework consisting of pretraining and fine-tuning is constructed. In the pretraining stage, analytical datasets are used to train a residual neural network to learn the nonlinear mapping between magnetic flux density features and the 6-DoF pose of passive HPMA mover. In the fine-tuning stage, a small amount of data measured by Hall sensors is used to adapt the model by freezing most of the shallow residual blocks in the pretrained network, updating only deep residual blocks and fully connected layers, and the residual of the 3D analytical magnetic field model is introduced into the loss function as a physical constraint term. Experimental results under dynamic conditions show that the maximum absolute translation errors along the X-, Y-and Z-axes are 103.3μm, 113.2μm and 100.4μm, respectively, while the maximum absolute rotation errors around the RX-, RY-and RZ-axes are 9.02′, 6.83′ and 12.11′, respectively. The results show that the proposed method can achieve 6-DoF pose detection of passive HPMA mover in MPM with high precision.
Abstract A two-stage pendulum system, comprising a lower torsion pendulum and an upper torsion pendulum, is widely employed in gravitational experiment. While prior research has focused primarily on the vibration isolation capability of the magnetic damper, this study investigates the dissipation coupling mechanism between the upper and lower torsion pendulums through theoretical analysis and experimental validation. The dynamic model of the two-stage pendulum system is analyzed, and expressions for the quality factor (hereafter denoted as Q value) of each stage are derived. An experimental apparatus is designed and constructed to enable modulation of the Q value of the upper torsion pendulum and then measure the response of the lower torsion pendulum. Results indicate that when the horizontal magnetic field at the center of the upper torsion pendulum is increased from 1.37 mT to 20.63 mT, the $Q$ value of the upper torsion pendulum decreases from 701 to 7, which leads to a reduction in the Q value of the lower torsion pendulum—from 1.6×10^{5} to 9.0×10^{4}. In gravitational experiment with a silica fiber after using the magnetic damper, the Q value of the lower torsion pendulum can be reduced by approximately half, which results in an increase in thermal noise. This work offers valuable insights into the dissipation coupling mechanism of two-stage pendulum system and aids in assessing noise sources in precision measurement.
Abstract Online remaining useful life prediction for aero-engines is challenged by evolving operating conditions and degradation patterns, which may cause static offline models to lose reliability after deployment. Full retraining and unrestricted historical replay are also impractical for resource-constrained prognostic systems. This study proposes a continual-adaptation framework based on a function-aware heterogeneous constraint strategy. Representation-related parameters are protected through orthogonal gradient descent, whereas dynamics-related parameters are regularized using dynamically updated Fisher information. Residual-driven replay is further introduced to prioritize difficult samples under limited buffer capacity. The framework is evaluated on the four C-MAPSS subsets using an engine-level streaming protocol and five independent random seeds. After five adaptation stages, the proposed method reduces root mean square error (RMSE) by 9.3%, 5.0%, 6.5%, and 2.7% on FD001–FD004, respectively, relative to the pretrained model. It also achieves the lowest final RMSE among representative continual-learning baselines and reduces the mean RMSE-based forgetting from 0.13 cycles for the strongest external baseline to 0.07 cycles. The average update time is 11.86 s per incremental batch, with a peak GPU memory usage of approximately 300.5 MB. These results demonstrate an effective balance among predictive accuracy, knowledge retention, and bounded resource overhead under non-stationary degradation conditions.
Abstract Existing intelligent fault diagnosis methods based on multimodal fusion face the problem of significant differences in the representation capabilities of different modal data for machine faults, making it difficult to achieve optimal cross-modal data fusion and accurate fault identification. This study proposes a prior-enhanced cross-modal vibration and acoustic data fusion network based on directed attention mechanisms to address the aforementioned issue. First, through preliminary experiments in fault diagnosis, the differences in fault sensitivity between vibration and acoustic data are measured to determine the prior dominant data modality. Then, based on the directed cross-attention mechanism, a prior-dominant modality-weighted fusion of vibration and acoustic data features is realized. This process allows for unidirectional feature information transfer from the dominant data modality to the weaker one, avoiding reverse information contamination. Thus, cross-modal fusion features that are more sensitive to machine faults can be extracted. Finally, the extracted cross-modal fusion features are used to achieve fault diagnosis. The results of two machine fault experiments demonstrate that, compared with the state-of-the-art methods, the proposed method can achieve a significant leading advantage in the same diagnostic tasks.
Abstract In challenging urban scenarios with low illumination and changing fields of view, the accuracy and stability significantly degrade for the existing multi-sensor fusion positioning methods. To address this degradation, a multi-modal tightly coupled positioning framework based on the error-state iterated Kalman filter is established, integrating thermal camera, LiDAR, and IMU. In addition, an adaptive enhancement method based on field of view perception is proposed within this framework. After data preprocessing, targeting the measurement characteristics of point cloud spatial distribution in open and narrow field of view scenarios, an adaptive factor to capture field of view features is designed. This factor dynamically adjusts the current-frame point cloud density, root voxel map resolution, and maximum iteration number in layer-by-layer updates, establishing a quantitative mapping from field of view characteristics to front-end and back-end system parameters. The proposed method is validated on both open-source and self-collected datasets in urban scenarios, exhibiting visibly better positioning accuracy and stability, with a 18.78% average reduction in absolute trajectory error over the second-best method. And the effectiveness of each module is verified through ablation experiments. The open-source code is available at: https://github.com/GNSSer-zzh/A–I-LITO .
Abstract For bearing fault diagnosis, vibration analysis is one of the most primary methods in engineering practice. Many traditional methods are ineffective in engineering applications with variable operating conditions. Tacholess envelope order analysis (TLEOA) technology shows promising application prospects for bearing fault diagnosis in conditions where speed varies, since it does not require additional speed measurement equipment to obtain phase information. Nevertheless, the widespread application of this method requires optimizing the instantaneous phase (IP) extraction algorithm of the reference shaft (RS) under the condition of no rotational speed first, including the requirement for prior knowledge to determine harmonic order and the starting point for ridge tracking, artificial initialization of certain important parameters, and inaccurate phase estimation. To address these issues, this article proposes an adaptive TLEOA approach based on surrogate data test. This approach can adaptively and accurately extract IP of RS from vibratory signal without the requirement for manual parameter initialization. Based on IP of RS, the initial vibratory signal is resampled with equal angular increments to convert it from the time domain to the angular domain, overcoming the adverse impact of variable speed on diagnosis. Finally, resonance demodulation is performed on the resampled signal, and envelope order analysis is performed on the demodulated signal to achieve fault identification. Experimental results demonstrate that the adaptively estimated IP using this approach closely aligns with measurements obtained by the tachometer, and the accuracy of fault diagnosis is superior to many existing methods. It provides a better diagnostic approach in various industrial settings.
Abstract With the global rise of the low-altitude economy, unmanned aerial vehicle (UAV) applications are experiencing explosive growth across various sectors. However, small-object detection in UAV imagery remains a critical bottleneck and urgent challenge. This is primarily due to the extreme representational fragility of targets (typically smaller than 32 × 32 pixels) and complex background interference. To address these challenges, this paper proposes Directional Multi-axis Structural DETR (DMS-DETR), a robust small-object detection method tailored for UAV scenarios, built upon the Real-Time Detection Transformer (RT-DETR) framework. Specifically, an SDNet backbone is designed to mitigate the irreversible loss of fine-grained details during downsampling. Within this network, the Space-to-Depth (S2D) rearrangement module achieves lossless spatial transformation to preserve critical geometric cues. Simultaneously, Star Operation is introduced at deep layers to explicitly encode high-order channel interactions, significantly enhancing semantic expressiveness without increasing channel overhead. Furthermore, the Directional Region Attention Module is proposed to combine directional modeling with regional statistics, thereby alleviating background interference and highlighting orientation-specific texture contours. To bridge the pronounced semantic gap between low-level spatial details and high-level semantic information, the multi-directional feature enhancement module is designed to expand receptive fields and suppress environmental noise, ensuring robust multi-scale feature alignment. Extensive experiments on two typical UAV datasets verify the superiority of the proposed method. On the VisDrone dataset, DMS-DETR achieves a 3.6% increase in mAP while reducing the parameter count by 17%. On the AI-TOD dataset, mAP increases by 4.8% while maintaining excellent inference performance. These results demonstrate that DMS-DETR provides an effective accuracy–efficiency trade-off for UAV small-object detection, offering a promising dataset-validated algorithmic approach for UAV small-object detection.
Abstract Rotating blades are critical components in aero-engines and fan systems for lighter-than-air vehicles. Reliable condition monitoring is essential, yet speed fluctuations often compromise measurement accuracy. To address the significant errors of traditional blade tip timing under such fluctuations, this paper proposes a Convolution-based BTT (C-BTT) method. By decoupling the Theoretical Time of Arrival from instantaneous speed calculations, the method effectively eliminates speed-induced errors. Numerical results show that C-BTT maintains a dimensionless error below 4 × 10 −4 rad across the full speed range. Additionally, convolution smoothing improves computational efficiency for real-time applications. Simulations and experiments demonstrate the effectiveness of C-BTT in various fluctuation scenarios, providing a robust basis for non-contact dynamic stress measurement.
Abstract To address central pixel dependency and degraded matching reliability in low-texture or occluded regions, this paper proposes a robust binocular ranging framework integrating an improved Census-BT cost with a multi-resolution ROI pyramid strategy. The approach employs an annular Gaussian-weighted Census transform combined with weighted BT costs to alleviate the limitations of traditional Census transforms, significantly enhancing robustness against noise and complex environments. To optimize computational efficiency for real-time applications, an enhanced YOLO11 model is integrated to generate region of interest (ROI) layers for cross-scale disparity refinement, achieving a 2.8% improvement in precision compared to the baseline. This localization facilitates progressive disparity refinement across pyramid layers, enabling reliable depth estimation without the extensive supervised retraining characteristic of learning-based models. Evaluated on benchmarks, the algorithm demonstrates a favorable accuracy and efficiency trade-off compared to traditional non-learning baselines, achieving an 11.33% D1-all error rate on KITTI 2015 and a minimum mismatch rate of 9.52% on Middlebury 2021 under challenging conditions. System-level validation using a ZED stereo camera substantiates the framework’s practical efficacy. In real-world ranging experiments across diverse scenarios, the system maintains a maximum relative ranging error of 3.19% with practical measurement accuracy consistently exceeding 96.81%. With an execution time below 0.75 seconds, the proposed framework achieves an effective balance between accuracy and processing speed, making it suitable for robotic perception and industrial object measurement on resource constrained edge devices.
Abstract In urban canyons, multipath and non-line-of-sight reception introduce substantial pseudorange biases that degrade smartphone Global Navigation Satellite System (GNSS) positioning. To improve practical training efficiency beyond the multilayer perceptron (MLP)-based PrNet, we develop a hybrid model combining a convolutional neural network (CNN) and a long short-term memory (LSTM) network. The CNN extracts local features from GNSS observations, while the LSTM models temporal dependencies in pseudorange errors. The CNN-LSTM model was evaluated on the Google Smartphone Decimeter Challenge 2021 (GSDC 2021) dataset against weighted least squares (WLS), PrNet, CNN, and LSTM under consistent conditions. Autocorrelation and ablation analyses indicated temporal dependence in pseudorange residuals and complementary contributions from the CNN and LSTM components. Compared with PrNet, CNN-LSTM reduced the median epoch time by 30.1%, increased training throughput by 43.1%, and reduced graphics processing unit (GPU) training-step latency by 33.9%. For the overall trajectory, CNN-LSTM achieved a horizontal positioning root mean square error (RMSE) of 4.638 m, 5.40% lower than PrNet, while reducing the mean absolute error (MAE) and GSDC score by 9.39% and 5.43%, respectively. It also achieved the lowest values for all three metrics in urban interchange, building-obstructed, and open-sky environments, with improvements of 53.13%–73.96% over WLS. These results show that CNN-LSTM provides a better balance between positioning accuracy and practical GPU training efficiency than PrNet under the evaluated conditions.
Abstract Reliable condition monitoring of permanent magnet synchronous motor (PMSM) encoders under extreme industrial noise remains a persistent challenge due to severe feature corruption. To address this, this paper proposes a Physics-Informed Multimodal Front-end Enhancement (PI-MFE) framework. The framework systematically internalizes physical priors as inductive biases into the feature extraction hierarchy to actively disentangle nonlinear noise distortions. For the electromagnetic modality, a Physics-Constrained Dual Attention Network (PCDAN) is orchestrated by a novel spectral consistency constraint. This mechanism explicitly enforces the latent manifold topology to align with authentic phase-modulation harmonics, fundamentally breaking the inherent spectral bias of deep networks. Concurrently, for the mechanical vibration modality, a Hybrid PSO Physics-Informed Neural Network (HPSO-PINN) is formulated as an uncertainty-aware probabilistic surrogate. By adaptively allocating computational queries driven by cognitive uncertainty, this strategy elegantly reformulates the intractable non-convex fitness landscape into an active learning paradigm, decisively circumventing the combinatorial explosion of heuristic evaluations during mode decomposition. A minimalist Linear SVM decodes the enhanced features, with the proposed PI-MFE(Linear-SVM) attaining peak accuracies of 98.1 ± 0.1% in simulation and 95.8 ± 0.3% on the physical platform, while retaining 84.8 ± 0.2% and 82.3 ± 0.5% fusion accuracy under the extreme -5 dB condition, respectively. The small accuracy gap Δ between Linear SVM and RBF-SVM further indicates effective linear separability, confirming that physics-driven front-end enhancement enables robust edge diagnostics.
Abstract Coherence scanning interferometry is a non-contact optical measurement technique used for high-resolution, three-dimensional surface topography characterization. The physical principle of this technique allows to measure the shape of an object, whether its surface is smooth or rough. The measurement system is basically an interferometer with an area image sensor at the output and the object to be measured used in place of one of the mirrors. The captured interferograms represent a large amount of data, which results in long measurement times. To reduce the measurement time, we introduce a measurement method based on the use of light polarization. A λ/8 wave plate causes two interference patterns to be generated at the output of the interferometer for each pixel. Two image sensors at the output of the interferometer capture pairs of interferograms. The interferograms in the pair are mutually phase shifted by π/2. The interferogam envelope is easily calculated from the pair of interferograms. Interferograms can be acquired with high undersampling, which allows for short measurement times.
Abstract Precise instance segmentation of surface defects plays a critical role in industrial quality control. However, conventional methods face difficulties in detecting tiny, low contrast, and irregular defects due to insufficient context modeling, limited attention mechanisms, and excessive computation. To address these issues, this paper proposes an efficient and lightweight defect instance segmentation network named DHC-Net. The innovations of DHC-Net lie in three specially designed core modules for micro defect segmentation. Firstly, we constructed an industrial inspection system and built a pixel-level defect dataset USB-SEG. Based on YOLOv8, the proposed Dual-Conv module combines group convolution and point wise convolution to reduce parameters while maintaining feature effectiveness for tiny defects.The HATA hybrid attention mechanism integrates channel attention and window based self-attention to enhance the perception of weak and low contrast defect regions. The C2f_Context module models local, neighbouring, and global context simultaneously to improve the understanding of defect background relationships. Experimental results show that DHC-Net achieves 75.1% mask mAP50 on USB-SEG and 91.6% mask mAP50 on NEU-DET, with competitive performance over mainstream methods and a real time speed of 189 FPS. This research provides an effective solution for high-precision and efficient surface defect inspection in industrial manufacturing.
Abstract A key challenge in federated remaining useful life (RUL) prediction arises from the entanglement between degradation dynamics and client-specific temporal scaling under heterogeneous operating conditions. In such settings, standard federated aggregation mixes trajectories with different degradation rates, leading to systematic bias in the learned latent evolution and consequently distorted RUL estimates. We observe that this issue is fundamentally caused by the lack of identifiability between shared degradation evolution and client-dependent temporal rate variations in conventional federated optimization. To address this limitation, we propose FedMDR, a federated latent dynamics framework that explicitly separates the shared degradation field, client-specific temporal scaling, and condition-dependent drift within a unified continuous-time formulation. Its identifiability-aware design keeps degradation-rate information largely out of the shared representation during aggregation, while a rate-drift diagnostic down-weights unstable client updates. Across five seeds (mean ± 90% CI), FedMDR attained a C-index of 0.776 ± 0.011 and RMSE of 16.3 ± 0.8 cycles on C-MAPSS; on CALCE, it achieved a mean C-index of 0.789 ± 0.014 and Brier score of 0.133 ± 0.010. Together, the results indicate strong predictive discrimination and reliable uncertainty estimates across mechanical and electrochemical degradation domains.
Abstract Advanced semiconductor packaging and heterogeneous integration demand 3D surface metrology that combines submicron precision with industrial-scale throughput. However, traditional optical methods are inherently limited by the slow speed of mechanical scanning. To overcome this, we propose a chromatic differential arrayed-confocal three-dimensional microscopic imaging (CDAC-3DMI) method. By integrating a digital micromirror device for high-speed programmable pinhole generation with a chromatic differential detection principle, the proposed system enables scanless online 3D topography reconstruction without lateral or axial mechanical scanning. Experiments on packaging substrates validate the measurement fidelity, yielding step-height and Sa relative deviations of less than 3% and 1%, respectively, compared with a commercial laser scanning confocal microscope (Olympus OLS5100), within the measurement uncertainties reported in the validation table. CDAC-3DMI reduces the end-to-end full-resolution 3D acquisition time from 20 s to 65 ms under equivalent measurement conditions, corresponding to a 308-fold improvement in throughput. CDAC-3DMI thus provides a robust, high-throughput, and non-destructive solution for submicron characterization of complex microstructures in high-volume microelectronics packaging production lines.
