Abstract:The thermal output effect of fiber Bragg gratings(FBG) in low-temperature environments is significant and exhibits strong nonlinear characteristics, which creates difficulties for strain and temperature testing of the common-bottom structure of rocket cryogenic storage tanks. Based on fundamental sensing principles, this study analyzes the thermal output sources of bare FBG and strain-sensing FBG. A decoupling method for the thermal output data of FBG was proposed for ground mechanical tests of rocket low-temperature storage tanks. An experimental system for calibrating the thermal output of FBG in extremely low temperature environments was built, and the thermal output characteristic curve of FBG in the liquid helium room temperature range was obtained. The thermal response relationship between strain-sensing FBG bonded to the same material as the common-bottom structure and bare FBG was established. The application of FBG for online monitoring of strain and temperature in the common-bottom structure of rocket cryogenic storage tanks was carried out. By constructing a strain and temperature sensing network based on FBG, real-time temperature and strain data of the common bottom structure have been obtained. The results indicate that this study can effectively decouple the thermal output data of FBG in extremely low-temperature environments, accurately obtain the strain and temperature distribution of the common-bottom structure of rocket cryogenic storage tanks, and achieve good agreement with strain test data from low-temperature resistance strain gauges, with a correlation correlation exceeding 0.997. This method realizes all-fiber online testing in extreme environments, offering high deployment flexibility and sensor survival rate. It provides a case reference for the large-scale application of fiber optic sensing technology in structural cryogenic testing scenarios, and lays a foundation for the engineering implementation of fiber optic onboard testing.

Abstract:Existing fouling monitoring of tubular heat exchangers relies heavily on historical operating conditions. Under fluctuating and variable operating states, significant deviations arise between the current and historical monitoring data, making it difficult to accurately capture fouling thickness and leading to false or missed alarms. This paper presents a study on digital twin-driven fouling monitoring and thickness quantification for tubular heat exchangers. First, a high-precision and high-fidelity finite element simulation model of the heat exchanger is developed. Through geometric modeling, unstructured tetrahedral meshing, and appropriate simulation assumptions, physical data from all nodes during operation are obtained. Second, key low-dimensional modes are extracted using proper orthogonal decomposition (POD), and parameter generalization is enhanced with radial basis function (RBF) interpolation. On this basis, a data-driven POD-RBF model order reduction method is proposed. An adaptive sampling layout optimization approach is also introduced to reduce computational costs while maintaining accuracy, enabling the generation of multiple training datasets and the construction of a digital twin model capable of real-time state prediction. Model errors are dynamically corrected via Kalman filtering to further improve prediction accuracy. Finally, a fouling detection and thickness quantification model is derived from the heat transfer coefficient formula. Fouling is identified by comparing overall heat transfer coefficient, inlet/outlet temperatures , and other parameters between the physical system and its twin. Thickness quantification is then achieved using the correlation between fouling thermal resistance and thickness. For experimental validation, a transparent-shell heat exchanger was used with saturated calcium sulfate solution as the cold fluid. By monitoring inlet/outlet pressures, flow rates, and temperatures, results show that under healthy conditions, the error between the digital twin and actual operation is within 1%. Under fouling conditions, the fouling detection rate reaches 100%, and the thickness sensing error ranges from 5%~25%, thus realizing reliable fouling monitoring and thickness quantification under variable operating conditions.

Abstract:This paper conducts an in-depth study on the problem of shortcut learning in convolutional neural networks for predicting rolling bearing faults, and proposes a physics-guided convolutional neural network model for fault diagnosis and prediction. Using rolling bearing datasets, this study analyzes the shortcut learning problem that occurs during the training of CNN-based rolling bearing fault diagnosis models, and reveals the existence of the shortcut learning phenomenon: even though the convolutional network achieves an accuracy of over 90% on a specific fault dataset, due to the presence of shortcut learning, the model fails to learn the correct fault features that match the fault theory. Instead, it learns incorrect characteristic frequencies or waveform patterns in the spectrogram. The study also analyzes the generation mechanism of the shortcut learning phenomenon in fault diagnosis, and reveals the generation mechanism. Shortcut learning behavior in convolutional neural networks mainly arises from shortcut opportunities in the dataset, caused by factors such as background noise and assembly, the model's tendency to learn simple feature combinations, and data statistical biases caused by comprehensive errors. Since the fault dataset itself cannot sufficiently constrain the learning of deep neural network models, this paper designs sensitive frequency bands based on bearing fault characteristics according to the characteristic frequencies of rolling bearings. It generates physics-guided data through band-pass filters, constructs physics-guided information, and inputs it into the convolutional neural network model to guide the model to learn correct fault features. Experimental verification shows that the physics-guided convolutional neural network can effectively avoid the shortcut learning problem, accurately extract core fault features, improve the accuracy of fault diagnosis and prediction, and enhance the credibility of the convolutional network-based fault diagnosis model. It has application prospects in fault diagnosis of high-end equipment in fields such as aerospace.

Abstract:When axisymmetric guided wave modes propagate to the defect location in a circular pipe, the uneven reflection of energy caused by the defect generates non-axisymmetric guided waves. The propagation characteristics of these waves have a quantitative relationship with the position and size of the defect, which can be used to detect pipeline defects. However, the propagation characteristics of non-axisymmetric guided waves generated by defects in polyethylene (PE) pipes are currently unclear. To address the issue of cracking at heat-fused joints in buried PE pipes, this study investigates the energy distribution patterns of non-axisymmetric longitudinal guided wave L(M,2) in PE pipes and utilizes these patterns to detect and evaluate defects at heat-fused joints. First, the feasibility of exciting non-axisymmetric guided waves in metal pipes using a discretely distributed piezoelectric array was first theoretically investigated, and the number of piezoelectric elements in the array was optimized. Next, a method for defect-induced non-axisymmetric guided waves was proposed, and simulations confirmed that the generated non-axisymmetric guided waves could be used to detect and evaluate defects. The optimized piezoelectric array was then applied to excite non-axisymmetric guided waves in PE pipes, and their axial propagation characteristics were quantitatively analyzed. Experiments verified that the propagation characteristics of defect-induced non-axisymmetric guided waves in PE pipes aligned with predictions. Subsequently, experiments on PE pipes with heat-fused joints demonstrated that intact joints do not generate non-axisymmetric guided waves. Finally, experiments on buried PE pipes showed that by analyzing the non-axisymmetric guided waves generated by defects at heat-fused joints, it is possible to determine the presence of defects and assess their circumferential locations.

Abstract:A multi-scale residual neural network (Res2APCNN) model combining data fusion and adaptive attention mechanism is proposed to monitor the health of aero-engine inter-shaft bearing under strong noise. Firstly, the bearing signals are converted into two-dimensional grayscale images by using the Gram angular difference field (GADF), Gram angular sum field (GASF) and Markov transfer field (MTF) methods. These three images are mapped to the RGB channels respectively to construct composite color images, thus enhancing the capture ability of time series information. Secondly, Res2Net module is introduced to extract multi-scale information through parallel convolution operation, filter noise interference and optimize information flow. Thirdly, the adaptive parallel feature fusion module is embedded to assign differentiated weights to feature dimensions, enabling the screening and amplification of key feature signals. Finally, the fault types of inter-shaft bearings are identified through a feature extraction and classification module. The proposed model is verified by using the bearing datasets of Polytechnic University of Turin in Italy and Harbin Institute of Technology, as well as the self-built test bench dataset. The experimental results show that the proposed Res2APCNN model demonstrates excellent fault diagnosis performance in a strong noise environment. Compared with advanced existing methods, the model achieves a 1.52% increase in accuracy over the IDRSN method on the Turin dataset, a 6.65% increase over the MC-CNN method on the HIT dataset, and a 2.35% increase over the Wen-CNN method on the self-built dataset. Furthermore, the diagnostic accuracy rate of this model exhibits the least fluctuation, indicating superior stability. Even under strong noise conditions, the Res2APCNN model can still maintain a high recognition accuracy and show good anti-interference ability.

Abstract:Insulator defect detection is a crucial task in the intelligent inspection of transmission lines. Currently, there is a shortage of image samples, and generating synthetic samples for data augmentation is an effective solution. However, synthetic samples inevitably exhibit domain distribution differences from real samples. To address this issue, a few-shot cross-domain hybrid transfer supervised domain adaptation model is proposed. This approach utilizes a large number of labeled synthetic images as the source domain and a small number of real images as the target domain, enabling effective use of synthetic samples and optimized alignment of cross-domain feature distributions, thereby improving the performance of insulator defect detection under few-shot scenarios. First, source domain images are adapted to match the target-like class distribution and are used to perform foreground-background hybrid augmentation on target domain images, improving the quality and diversity of target samples. Secondly, cross-domain style perturbation is applied to source images to further reduce the domain distribution gap with the target domain. Finally, a domain classifier based on adversarial training is employed to align cross-domain invariant features between the source and target domains, enhancing the model’s generalization ability across domains. Under the condition of using only 8 real insulator defect images for training, the proposed model achieves a 9.0% improvement in AP50 compared to the baseline detection model. Ablation experiments further evaluate the effectiveness of each module. In addition, the proposed model consistently outperforms other supervised domain adaptation approaches across different insulator defect datasets. For instance, it achieves a 3.6% AP50 improvement over the best competing model on a self-constructed insulator defect dataset, and a 2.4% AP50 improvement on the public insulator defect dataset IDID.

Abstract:In the backdrop of smart factory deployment, while similar rotating machinery managed by different enterprises holds potential for collaborative diagnosis, data privacy regulations prevent sharing. Additionally, operating condition differences result in non-independent and identically distributed data, limiting the generalization ability of diagnosis models across varying conditions. To tackle these challenges, this article proposes a personalized federated domain generalization framework. Without sharing local data, it enhances the generalization and robustness of edge-end diagnosis models through alternating adversarial optimization of inter-device collaboration and local personalized updates. The diagnosis model, built on the latent convolutional network, leverages input-driven feature adaptation for dynamic representation. During collaboration, publicly available datasets facilitate knowledge transfer in a shared space, while consistency constraints improve communication efficiency. In the local update phase, performance constraints and self-distillation preserve local knowledge, ensuring stable classification. Experiments on the Huazhong University of Science and Technology bearing dataset and the Mechanical Comprehensive Diagnosis Platform bearing dataset show that the proposed method achieves average accuracies of 88.96% and 92.33% under global operating conditions, respectively, outperforming several advanced approaches. Edge-end models optimized by the proposed approach improve cross-domain generalization while maintaining reliable local diagnosis performance.

Abstract:Carbon fiber reinforced polymer/steel (CFRP/steel) bonded joints are widely used in the reinforcement of bridge and ship structures, and their mechanical performance is significantly affected by overlap length. To address the insufficient understanding of damage evolution mechanisms and the limited recognition accuracy in existing studies, a damage monitoring and overlap length identification method is proposed by integrating acoustic emission (AE) and digital image correlation (DIC) techniques. Four overlap lengths of 25, 50, 75, and 100 mm are tested under quasi-static tensile loading. During the tests, AE signal features including amplitude, energy, centroid frequency, root mean square (RMS), and duration are collected, while full-field strain distributions are obtained using DIC to analyze the damage initiation, propagation, and failure mode evolution of the joint. The results show that the damage process can be divided into three stages, with five main failure modes identified: steel deformation, fiber fracture, matrix cracking, adhesive failure, and cohesive failure. In addition, overlap length has a significant influence on the mechanical performance of the joints, and the joint with an overlap length of 100 mm achieved an ultimate tensile load of 60.70 kN, approximately twice that of the 75 mm joint. Based on the AE features, an extreme gradient boosting (XGBoost) classification model is formulated for identifying the overlap length of CFRP/steel bonded joints, achieving a recognition accuracy of 94%. Furthermore, the shapley additive explanation (SHAP) method is incorporated into the model to quantify the contribution of each feature, revealing centroid frequency, RMS, and duration as the most critical parameters. This study reveals the influence of overlap length on damage behavior of the joint, effectively uncovers the relationship between AE features and damage states, and improves the accuracy and interpretability of damage prediction, providing valuable guidance for failure mode prediction and structural optimization design of CFRP/steel bonded joints.

Abstract:To address the issues of high miss rates for small targets, severe interference from complex backgrounds, and the inability of existing models to balance accuracy and efficiency in infrared detection of thermal defects in smart electricity meters and their junction boxes, we propose a lightweight smart electricity meter target detection algorithm, YOLO-MCSL, based on an improved YOLOv8s architecture. This algorithm aims to meet the urgent need for real-time detection in power field inspections. First, the MobileNetV4 lightweight network is adopted as the backbone to significantly reduce the number of model parameters and computational overhead. Second, the CCFF cross-scale feature fusion module from the RT-DETR model is introduced to enhance the detection capability for multi-scale small thermal defects. Subsequently, a lightweight C2f_Star module is designed to replace the original C2f structure, further compressing the model and improving feature extraction efficiency. Additionally, we construct the LSCD lightweight shared convolution detection head, which reduces redundant computation through parameter sharing. Furthermore, we combine the Focaler-SIoU loss function to optimize the bounding box regression process, enhancing the differentiation between easy and hard samples. Finally, we apply a layer-wise adaptive amplitude pruning algorithm to structurally prune the model, achieving a balance between performance and lightweight design. Experiments were conducted on a selfconstructed infrared image dataset of thermal defects in smart electricity meters. The results show that in the detection of three key components—junction boxes, battery modules, and displays—the detection accuracy of YOLO-MCSL reached 91.6%, 99.2%, and 99.5%, respectively, with an overall mAP@0.5 of 97.9%. Compared to the YOLOv8s baseline model, the number of parameters was reduced to 1.749 M (a reduction of 84.3 %), computational complexity was reduced to 5.7 GFLOPs (a reduction of 80.2%), and model memory usage was reduced to 3.8 MB (a reduction of 82.3%). This method provides a high-precision, lightweight, and embeddable solution for smart electricity meter defect detection, showing promising prospects for engineering applications.

Abstract:The adhesive strip of a glass window tends to age over long-term use, undermining the sealing integrity of the structure and posing safety risks. Such damage is often highly concealed, and traditional manual inspections fail to detect it in time, leading to an imbalanced distribution of healthy and faulty data. To address this issue, this study proposes a nonintrusive fault diagnosis method that integrates piezoelectric impedance technology with a Transformerbased deep learning model. To tackle the core challenges of scarce fault samples and imbalanced data distributions in practical applications, we innovatively propose a dataaugmentation model—CBAMVAE, which combines a convolutional block attention mechanism with a variational autoencoder. By learning the distribution of real fault data, the model generates synthetic samples to expand the dataset and enhance the Transformer’s generalization capability. Furthermore, to improve the quality of generated data and increase diagnostic accuracy, we incorporate the PPIR technique, combining it with CBAM-VAE to form the PPIR-CBAM-VAE collaborative optimization method. PPIR retains critical resonance peak features, removes nonpeak points, and applies linear interpolation to restore nonpeak regions, thereby enriching samples diversity while significantly enhancing dataset stability. Experimental results show that under a highly challenging healthytofault ratio of 20∶3, the PPIRCBAMVAE method achieves a diagnostic accuracy of 92.13%. When the imbalance ratio is 4∶1, the accuracy improves markedly from 92.27% (baseline) to 96.45%, greatly boosting the recognition of minorityclass faults samples. This study establishes a systematic fault diagnosis framework that integrates EMI technology, a Transformer model, and the innovative PPIRCBAMVAE data augmentation method, providing a highly sensitive and broadly applicable solution for health monitoring of building sealing systems. The adhesive strip of a glass window tends to age over long-term use, undermining the sealing integrity of the structure and posing safety risks. Such damage is often highly concealed, and traditional manual inspections fail to detect it in time, leading to an imbalanced distribution of healthy and faulty data. To address this issue, this study proposes a nonintrusive fault diagnosis method that integrates piezoelectric impedance technology with a Transformerbased deep learning model. To tackle the core challenges of scarce fault samples and imbalanced data distributions in practical applications, we innovatively propose a dataaugmentation model—CBAMVAE, which combines a convolutional block attention mechanism with a variational autoencoder. By learning the distribution of real fault data, the model generates synthetic samples to expand the dataset and enhance the Transformer’s generalization capability. Furthermore, to improve the quality of generated data and increase diagnostic accuracy, we incorporate the PPIR technique, combining it with CBAM-VAE to form the PPIR-CBAM-VAE collaborative optimization method. PPIR retains critical resonance peak features, removes nonpeak points, and applies linear interpolation to restore nonpeak regions, thereby enriching samples diversity while significantly enhancing dataset stability. Experimental results show that under a highly challenging healthytofault ratio of 20∶3, the PPIRCBAMVAE method achieves a diagnostic accuracy of 92.13%. When the imbalance ratio is 4∶1, the accuracy improves markedly from 92.27% (baseline) to 96.45%, greatly boosting the recognition of minorityclass faults samples. This study establishes a systematic fault diagnosis framework that integrates EMI technology, a Transformer model, and the innovative PPIRCBAMVAE data augmentation method, providing a highly sensitive and broadly applicable solution for health monitoring of building sealing systems.

Abstract:The failure process of localized hydrogen-induced damage (HID) in longdistance oil and gas pipelines involves the coupled effects of material mechanical behavior, magnetic response, and hydrogen diffusion, which makes early detection challenging with conventional non-destructive testing (NDT) techniques, posing a significant threat to the safety of energy transport systems. Weak magnetic inline inspection (WMII) technology, due to its intrinsic sensitivity to early-stage damage in ferromagnetic materials and its capability for online monitoring, exhibits great application potential in the detection and evaluation of HID. Based on this technology and combined with first principles, a multiscale cross-analysis method is proposed to explore the relationship between localized HID and magnetic signal response in pipelines. Furthermore, a Qaverage multi-component magnetic feature fusion parameter is introduced to effectively characterize the failure behavior and hazard level of HID under multi-physical field conditions. The results reveal that under magnetic flux leakage (MFL) testing, the saturation magnetization aligns the magnetic domains in a highly ordered state, rendering it difficult for hydrogeninduced high-pressure stress concentration zones to cause significant perturbations in the overall leakage flux. Consequently, no distinct magnetic response features are observed. In contrast, the characteristic components of the weak magnetic signal exhibit an average increase of approximately doubling compared with the metal magnetic memory method, demonstrating superior applicability and effectiveness for the detection of early-stage HID. In addition, with increasing internal pipeline pressure and external excitation intensity, the stress concentration and magnetic domain reconstruction behavior in hydrogen-enriched zones intensify. Specifically, the Qaverage response curve shows a nonlinear increase with rising internal pressure, with average response growth rates of 137 and 195 A·m-1/MPa for Q235 and Q345 steels, respectively. With increasing excitation intensity, the response increases approximately linearly, and the corresponding rates are 61.24 A·m-1/A for Q235 and 69.06 A·m-1/A for Q345 steels.

Abstract:The large-aperture mirror deployment system is primarily designed to address the launch challenges posed by telescopes whose apertures exceed the constraints of the fairing envelope. It is imperative to ascertain that the stowed telescope adheres to the constraints imposed by the fairing envelope. The system must ensure that the stowed telescope conforms to the fairing envelope while minimizing the overall mass of the deployment mechanism and reducing actuator design complexity. To cope with the large number of design parameters and evaluation metrics, an optimization approach combining a multi-objective optimization algorithm with parameter correlation analysis is proposed to improve system performance. First, a multi-objective particle swarm optimization (MOPSO) algorithm incorporating penalty functions is introduced. By exploiting the coupling relationships among evaluation metrics, the number of optimization objectives is reduced. The penalty functions are then used as criteria to evaluate whether parameter schemes satisfy functional requirements of the deployment system, thereby enabling multi-objective optimization while ensuring the reliability of the algorithm outcomes. Subsequently, simulations are conducted by varying initial conditions (including parameter search bounds and the radial envelope of the stowed primary mirror) to analyze the correlations among initial conditions design parameters, and evaluation metrics. Finally, an optimal parameter scheme is selected by grouping evaluation metrics and applying a stepwise scoring and screening strategy. Compared with the original design, the optimal scheme achieves a 20.90% reduction in total mass, a 28.75% reduction in maximum equivalent moment of inertia, and reductions of 64.04% and 67.04% in the stroke lengths of the two rockers, respectively. The large-aperture mirror deployment system is then reconstructed based on the optimal parameters, and a random vibration analysis is performed under the stowed condition. The results demonstrate that, without changing the material, the optimized system continues to meet the required mechanical conditions.

Abstract:Graphene aerogel (GA) has been shown to possess vibration responsiveness and significant potential for developing novel accelerometers. To explore the response characteristics of GA-based sensors under low-frequency and low-amplitude vibrations, a new GA accelerometer was designed and assembled based on the mechanical properties of the prepared GA. In this design, GA serves simultaneously as both the elastic element and the sensing element, achieving dual-function integration that simplifies the overall structure. The response characteristics of the GA accelerometer were tested on a vibration table and further verified by monitoring motor vibration in a vacuum pump. The results demonstrated that the GA accelerometer exhibits high sensitivity and linearity in the time-domain signal, with a sensitivity exceeding 3 mV/g under 1g acceleration—higher than the previously reported 2.6 mV/g. Meanwhile, the frequency-domain signal showed high precision and stable repeatability, with a maximum relative error rate of 0.44%, which is an order of magnitude lower than the 5.46% of commercial sensors. In low-frequency and lowamplitude vibration tests, the frequency-domain signal of the GA accelerometer exhibits a large DC offset, masking relatively small frequency-domain components; however, clear frequency-domain signals could be obtained after removing the DC offset via high-pass filtering. Moreover, during vibration monitoring, the GA accelerometer produced stable time-domain response to both constant-frequency and variable-frequency vibrations, and time-frequency analysis of the output signals aligned well with the preset operating conditions. In practical monitoring applications, the GA accelerometer also showed excellent performance: results from three frequency tests under identical conditions were highly consistent, and the monitoring signals exhibited significant output changes in response to abnormal bodyshaking events. These findings confirm its strong potential for equipment condition monitoring applications.

Abstract:Rotor speed fluctuation causes non-stationary instantaneous phase, while key-phasor signal jitter leads to reference phase drift. The superposition of these interferences significantly increases the error of synchronous phase measurement, severely restricting the dynamic identification and balancing of rotor systems. To address this issue, a tracking regression method based on instantaneous phase resampling is proposed for high-precision synchronous phase measurement under complex operating conditions. First, the instantaneous phase of the vibration signal is calculated via zero-phase shift wideband bandpass filtering and the Hilbert transform to fully extract the rotor′s speed fluctuation information. The instantaneous phase is then tracked and resampled using an interpolated and up-sampled key-phasor sequence, leveraging the key-phasor signal to mitigate the phase non-stationarity caused by speed fluctuation. A first- and second-order cyclostationary model is constructed to accurately describe the resampled phase, which quantifies the key-phasor jitter interference as additive noise. Furthermore, linear regression is applied to the resampled phase to effectively suppress this noise, and an asymptotically unbiased estimate of the intercept yields a precise measurement of the synchronous phase. Finally, the superior antiinterference capability of the method is evaluated through simulations and experiments. Simulations show phase errors were reduced by 70.4% and 40.5% compared to conventional methods. Experiments demonstrate <2° phase fluctuation under strong noise, enabling a 92.2% vibration reduction in a single balancing run. The method provides a robust solution for high-precision phase measurement under complex interferences, supporting applications like rotor fault diagnosis and dynamic balancing.

Abstract:To address the combined requirements for lightweight design and misalignment tolerance in wireless power transfer (WPT) systems for mobile load scenarios, this article proposes an engineering-oriented lightweight design method for a grid-type flat spiral pad (GFSP) magnetic coupler with a central windowed ferrite structure. First, GFSP coils are utilized as the transmitting and receiving units of the WPT system. Based on mutual inductance theory, a mechanism model is formulated to analyze the coupling coefficient under distributed winding, revealing how the winding configuration improves the coupling performance and reduces mutual inductance fluctuation under positional offsets. On this basis, a central opening is introduced in the ferrite plane of the GFSP structure. A corresponding parameter design process is proposed to balance structural weight reduction and coupling performance retention, and a representative parameter design is provided. Then, ANSYS/Maxwell simulation models of the original and lightweight GFSP magnetic couplers are established for comparative analysis of coupling coefficient retention. Results show that the proposed design significantly reduces ferrite weight while maintaining robust coupling performance under lateral misalignment, varying transmission distances, and vertical tilt conditions. Finally, a 200 W experimental prototype based on a single-switch dual-branch P#LCC-S compensation topology is established. Compared with the original design, the magnetic coupler̓s weight is reduced by 25%. Under 60% lateral offset, 90° vertical tilt, and 80~120 mm transmission distance, the system maintains an output voltage above 25 V and transfer efficiency above 80%, evaluating the effectiveness and engineering feasibility of the proposed design.

Abstract:Due to the effects of coating, material birefringence， and internal stress, polarization aberration is ubiquitous in high numerical aperture (NA) optical systems. In this case, the imaging quality is highly dependent on the polarization of incident light. Conventional methods to characterize the polarization aberration suffer from complex systems and measurement inconvenience. To address this problem, a back-and-forth optical setup is proposed to measure polarization-resolved wavefront and reconstruct the Jones matrix of the objective in this work, which helps reduce the system complexity. Firstly, a numerical model based on the ray tracing method and scalar diffraction theory is established to simulate intensity distributions of focused light fields at different axial positions. Secondly, the problem of wavefront retrieval is transformed into an optimization problem, and an inverse model is formulated for retrieving the Jones matrix utilizing hybrid particle swarm optimization and gradient descent algorithms, where the deviation between modelled and target intensity distributions is minimized by optimizing the coefficients of Zernike polynomials accounting for the polarization aberration. Thirdly, numerical simulation is implemented to examine the model accuracy under predefined polarization aberrations, showing a good agreement between reconstructed and target Jones matrix elements with retrieval errors less than 10-3. Finally, experimental measurements of polarization aberration are performed on a commercial high-NA microscope objective to retrieve the coefficients of Zernike polynomials, where modelled intensity distributions agree well with the target. Theoretical and experimental results demonstrate that the proposed algorithm can effectively retrieve the objective′s polarization aberration with the advantages of a simple configuration and user-friendly operation. It is anticipated that this work provides a novel technical approach for manufacturing and characterizing high-NA optical systems.

Abstract:Aiming at the problem that high-orbit satellite target-pointing on-orbit autocollimation systems cannot achieve real-time light spot localization under limited computational and storage resources, this paper proposes a storage-free real-time localization method for autocollimator light spots based on a parallel pipeline architecture. By utilizing the decomposition characteristics of centroid calculation, a stepwise centroid computation method is designed: The row or column centroids are calculated first and then combined into a two-dimensional centroid, realizing storage-free computation of the light spot centroid and avoiding the global storage requirements for the original image data. A sliding correlation filtering method and its FPGA implementation scheme are developed based on the parallel pipeline architecture. A Gaussian negative second derivative template was used to effectively suppress background gradient noise and random noise through real-time correlation calculation between pixels in the sliding window and the template. Hardware optimization designs such as pipeline multipliers and additive tree accumulators were utilized to ensure synchronous filtering operations during data streaming input, reducing noise impact while ensuring real-time computation. The method was verified through simulations and actual hardware deployment. Results demonstrate that, under the premise of ensuring computational accuracy, the proposed method completes filtering and centroid calculation within 246 clock cycles after reading the light spot information, utilizing only a small amount of FPGA on-chip RAM. At a 25 MHz clock frequency, the computation time is only 9.84 μs, with an average deviation of 0.032 pixels, achieving both high precision and real-time capability. This method can significantly improve the real-time feedback of on-orbit monitoring data for optical payload line-of-sight pointing in high-orbit satellites, providing real-time data support for tracking and locating high-speed maneuvering targets, with important application value and prospects in aerospace remote sensing.

Abstract:Ultrasound bone evaluation is of great significance for the diagnosis and prevention of metabolic bone disease in neonates. Dual-energy X-ray absorptiometry, a routine clinical technique, is difficult to meet the clinical needs of long-term monitoring of neonatal bone status due to ionizing radiation. Biochemical index testing, as an invasive detection method, is likewise unsuitable for evaluating neonatal bone status. Ultrasonic guided wave detection technology, with its advantages of non-invasiveness and absence of ionizing radiation, has been applied to the diagnosis of osteoporosis and fracture evaluation. However, clinical research on applying ultrasonic guided waves to evaluate neonatal cortical bone quality still needs further development. To address this issue, we propose an ultrasonic guided wave wavenumber dispersion extraction method based on the Matrix Pencil spectral estimation, aiming to achieve accurate evaluation of the sound velocity and cortical thickness of the neonatal tibia. First, a single-layer bone plate model and a multi-channel ultrasonic phased array guided wave field were constructed through numerical simulation. Experimental results showed that the relative errors of thickness assessment for isotropic plates with thicknesses of 1, 2, and 4 mm were all less than 3%. A horizontal comparison of cortical thickness assessment results under signal-to-noise ratios of 10 and 5 dB verified that the evaluation accuracy was less affected by noise. Furthermore, this wavenumber dispersion extraction method was applied to evaluate the longitudinal wave velocity in neonatal tibial cortical bone. The results indicated that the relative error between the estimated cortical sound velocity and the clinical ultrasonic measurement results was less than 10%, and the detected cortical thickness was significantly correlated with various neonatal physical measurement values (p-value < 0.001). This study demonstrates the clinical application value of the ultrasonic guided wave method in assessing neonatal tibia bone quality and effectively expands its scope of application in clinical practice.

Abstract:LiDAR-based 3D object detection achieves superior performance. However, the unevenly distributed point clouds on foreground objects can weaken their geometric representation. Meanwhile, far-away objects typically have very few points, which further impairs detection performance. In this article, a novel framework PUDet is presented, which integrates generative models into discriminative detectors. A point cloud upsampling network is leveraged with prior knowledge to enhance the geometric details of foreground objects, aiding the detector in achieving more accurate prediction. PUDet incorporates two key modules: LDEM for nearby objects, which optimizes point distribution while minimizing computational costs, and DDAM for distant objects, which increases point density to better delineate object contours. To evaluate the optimization of geometric contours, the uniform loss of close and long-distance targets before and after enhancement is experimentally compared, showing the efficacy of LDEM and DDAM. This article also displays the attention maps on object point clouds, explaining the observed accuracy gains. Experimental results on the KITTI testing set show that the proposed framework improves the baseline CT3D by 1.84 mAP, confirming the effectiveness of PUDet. This work introduces a novel approach to 3D object detection, enhancing precision and reliability in object recognition for applications like autonomous driving.

Abstract:To address the issue that single-frame perspective data struggle to depict the complete contour of large-scale targets, thereby limiting size measurement, this article proposed a large-target size measurement method of LiDAR based on visual semantic constraints, with automobiles as the research object for large-target measurement. Firstly, this method achieves spatio-temporal synchronization of LiDAR, camera, and inertial measurement unit data through joint calibration and timestamp nearest-neighbor matching. Subsequently, a mobile cart is used to acquire information about the entire target in three dimensions. The simultaneous localization and mapping technology is employed to reconstruct the contour of the measured target. In this module, the algorithm′s accuracy is enhanced through a ground-based residual optimization and loop closure detection framework. After denoising the point cloud, a ground segmentation algorithm is used to separate ground points from non-ground points, and a pass-through filter is applied to ensure the segmentation effect. Meanwhile, a target detection algorithm is utilized to obtain the category and position information of the target in the image. Next, through an adaptive threshold point cloud clustering method, the centers of different point cloud clusters are visually projected, and the point cloud corresponding to the target is located according to the visual target detection results. Finally, a contour fitting algorithm is designed to complete the contour fitting of the target point cloud. Then, a three-dimensional box fitting algorithm is used to calculate the target′s size. Experimental results show that for large-sized objects such as automobiles. In a parking lot with a large number of vehicles, the proposed method yields a maximum error and an average error of less than 1.97% and 0.82% respectively for vehicle length, less than 3.26% and 2.08% respectively for width, and less than 3.99% and 1.99% respectively for height, demonstrating promising prospects for engineering applications.

Abstract:Human pose estimation is one of the important supporting technologies for Industrial Manufacturing 5.0, which has already been applied in various scenarios such as action recognition, human-computer interaction, and digital twin. However, in complex industrial scenes, objects such as notice boards, pipes, and columns can easily cause local or global occlusions for workers, leading to errors in joint points localization by human pose estimation models and a decrease in the performance of the human pose estimation model. To address this problem, this article proposes a human pose estimation performance enhancement method for complex industrial scenes, which firstly structurally models the key points of the human body based on VQ-VAE model, mapping joint features to a quantized latent space to improve the accuracy of human pose estimation when occlusion occurred. Then, to address the problem of insufficient worker occlusion dataset, a dynamic data augmentation and training method is innovatively proposed. In the process of model training, industrial scene-specific worker occlusion images are generated dynamically using real industrial scene occlusion objects by evaluating the human pose estimation results of the model for the next model training, further enhancing the model′s robustness in human pose estimation tasks. The experimental results show that the method proposed in this article achieves an average precision (AP) improvement of 3.8% and an average recall (AR) improvement of 2.7% over the advanced method PCT on the self-constructed dataset and is able to effectively cope with the human occlusion problem in complex industrial scenes.

Abstract:With the development of electronic products in the direction of high performance and miniaturization, printed circuit boards (PCBs), as the core carrier of electronic systems, are becoming more and more complex in design and manufacturing, with more closely arranged components and a finer structure, which puts forward higher requirements for component detection and defect detection. Although the target detection models based on convolutional neural networks represented by YOLO have received a lot of research, these models are only designed for a single defect or component detection scenario, and have a limited effect on the detection of small targets and dense scenarios. The emergence of RT-DETR has enabled Transformer-based end-to-end detection models to perform excellently in real-time detection. Therefore, based on the RT-DETR model, this article proposes an end-to-end real-time target detection model MFSF-DETR, based on a Transformer for PCB scenarios. Firstly, the Faster-CGLU Block is used to replace the Block layer in the backbone network, the channel attention mechanism is refined, and the entanglement transformer block (ETB) is introduced to integrate the frequency domain with the spatial domain to enrich the deep semantics. Then, the rep adaptive weighted cross-scale feature fusion (RAWCFF) is designed to replace the CNN-based cross-scale feature fusion and form a new feature fusion encoder with the cross-scale feature shift fusion (CFSF)to realize the deep interaction between neighboring and non-neighboring features. Finally, the proposed model is evaluated using the PCB defect dataset DsPCBSD+, the PKU-Market-PCB dataset, the PCB component dataset PCB_WACV, and the PCB and drone dataset VisDrone2019 to assess its detection performance and generalization ability in PCB scenarios. The experimental results show that the MFSF-DETR model achieved the highest accuracy of 85.6%, 98.1%, and 89.9% in defect and component detection, respectively, which is 3.1%, 1.0%, and 3.8% higher than the baseline model. Meanwhile, the FPS indicators also reached 120.2, 57.1, and 71.8, respectively, achieving efficient and high-precision detection in the PCB background.

Abstract:In the field of collaborative classification between hyperspectral images and LiDAR data, although CNN and Transformer have shown keen insight into local features and global dependencies in image processing and data analysis, their collaborative mechanisms have not been fully explored, and the potential for cross-modal feature complementarity has not been effectively unleashed. Therefore, this article proposes a multimodal collaborative land-cover classification method for remote sensing data that combines CNN with Transformer for hyperspectral images and LiDAR data. Firstly, the model performs dimensionality reduction on hyperspectral images through principal component analysis to remove redundant spectral information. Then, it uses CNN layers to capture local texture features, and constructs a global spectral-spatial representation using the Transformer self-attention mechanism. Then, through a bidirectional feature interaction mechanism, the global contextual information from the Transformer is injected into the CNN feature channels, while the local details extracted by the CNN are fed back into the Transformer branch. Cross-scale feature alignment is achieved through the feature coupling unit, enhancing the joint extraction ability of the model for the global structure and local details of hyperspectral images. For LiDAR data, a dynamic convolution cascade module is used to effectively capture elevation information and contextual relationships. Finally, a cross-modal feature fusion module is used to achieve deep interaction and fusion of dual source data features, improving the classification accuracy of complex land features in the complementary semantics of dual modalities. Experiments on three publicly available datasets—Houston 2013, Trento, and Augsburg—showed that the overall classification accuracy of our proposed method reached 99.85%, 99.68%, and 97.34%, respectively, with average accuracies of 99.87%, 99.34%, and 90.60%. This improvement in classification accuracy compared to mainstream methods such as GLT and HCT fully demonstrates the advantages and effectiveness of our proposed method for multimodal data collaborative classification.

Abstract:Because of the challenges faced by GNSS/INS high-precision integrated navigation models in complex urban canyon environments, which are prone to occlusion, multipath effects, and fault interference, this article proposes a hierarchical resilient tightly coupled RTK/INS integrated navigation method that ensures system accuracy and robustness by implementing a multi-stage framework of detection, fault exclusion and multi-source enhancement. Initially, during the satellite RTK positioning stage, the method achieves rapid fault identification in the observation data of each epoch by introducing a Chi-square test-based fault detection method. If a fault is detected, a solution separation method is applied to accurately identify and isolate the faulty satellite, thereby enhancing the reliability of the satellite navigation system. However, due to the inherent limitations of threshold detection methods, while a loose threshold setting helps ensure the timely detection of large faults, it may also trigger false alarms, leading to the incomplete exclusion of some small faults. To further improve the resilience and reliability of the integrated navigation system, this article employs the IGG-III robust estimation method. By dynamically adjusting the observation weights within the integrated navigation system, this approach effectively enhances the system's capability to suppress small faults that fall below the threshold, thereby boosting its overall performance in complex environments. The experimental results indicate that the proposed algorithm reduces the eastward positioning error by 34.29% and the northward error by 13.22%. Notably, it achieves a significant 55.87% reduction in the upward positioning error. The overall performance evaluation results show that the proposed algorithm improves 3D positioning performance by 46% compared to conventional methods, which strongly validates its effectiveness and robustness in urban canyon environments.

Abstract:This study addresses challenges in conventional threeaxis force sensors—namely the coupling between shear and normal forces, insufficient flexibility, and complex decoupling algorithms—which limit their application in precision teleoperation of robots. To overcome these issues, we propose a three-dimensional magnetic tactile sensor based on a single-layer annular sinusoidal magnetic film that enables real-time, self-decoupled measurement of tri-axial forces. Through the single-layer magnetic film structure, the sensor achieves self-decoupling of three-dimensional force, thereby avoiding multi-dimensional force coupling and the dependence on complex algorithms. Calibration experiments demonstrate that the sensor exhibits excellent performance with high sensitivity and high accuracy. Along the Z-axis, the measurable range is 0～15 N with a sensitivity of 0.014 7 kPa-1 and a root-mean-square error of 0.001 5 mT. Along the X- and Y-axes, the measurement range is -5 to 5 N, with sensitivities of 0.020 3 and 0.020 9 kPa-1, and RMSEs of 0.002 and 0.001 9, respectively. The sensor body adopts a flexible structural design, featuring low power consumption and a high output frequency. In slip-control experiments, a slip-control strategy was implemented on the end effector of the robotic arm. By mounting the designed tactile sensor on an augmented-intelligence gripper, the system monitors the normal and tangential contact forces in real time and, together with an incremental PID controller, rapidly adjusts the gripping force upon slip detection to stably grasp a container with a dynamically increasing load. The average response time is 113.3 ms, demonstrating the stability and reliability of the proposed slip-control strategy in practical applications.

Abstract:In view of the problems such as low efficiency of path planning, high risk of collision, and excessive map exploration caused by fixed parameters in the scene of varying complexity of real water area maps, such as rivers and coasts, a path planning algorithm for unmanned ships based on dynamic potential field guidance of local map complexity is proposed. The algorithm can significantly improve the planning performance and safety in complex waters by dynamically adjusting parameters through real-time perception of the environment. Firstly, according to each extended node, the obstacle collision detection is implemented in real time, and the local map complexity of the node is calculated. Based on this complexity, the gravitational strength and repulsive strength of the potential field are dynamically adjusted, the hybrid weights of random exploration and potential field guidance in the node generation strategy are adjusted adaptively, and the switching of the three-level safety step size is used to solve the adaptation problem of the traditional fixed parameters in the complex overlapping waters in the actual environment. Meanwhile, a compound navigation potential field integrating target gravity, obstacle repulsion, and boundary repulsion is constructed to solve the problem of obstacle avoidance and welt safety balance in narrow waterways and improve the safety of path planning. Then, the redundant nodes are removed by pruning, and the key turning points are optimized by stretching. The cubic B-spline smooth path is used to ensure the curvature continuity, improve the path feasibility, and meet the requirements of unmanned ship mission execution. Through electronic chart simulation and unmanned ship experiment, it shows that the proposed algorithm maintains advantages in planning time, path length, sampling points, and path feasibility. Compared with the improved algorithm DVSA-RRT, in a complex environment, the planning time is shortened by 87.27%, the path length is shortened by 21.6%, and the sampling points are reduced by 78.53%, which improves the efficiency of path planning, reduces the path planning space, and meets the requirements of the unmanned ship mission.

Abstract:In recent years, with the continuous development of the industrial robot technology, the application of robots equipped with laser rangefinders for multipose measurement and the demand for handeye calibration have been increasing, placing higher requirements on calibration accuracy. However, traditional methods often rely on dedicated calibration objects or sensors, which are complex to operate and costly. To address this issue, this paper proposes a robot handeye calibration method based on the deep Qnetwork(DQN) algorithm and circular contour fitting. The DQN algorithm controls the two end joints of the robot to drive the laser rangefinder such that its return value is minimized. On this basis, a kinematic model of the manipulator is established to compute the theoretical coordinates of the light spot. By setting multiple angle values of joint one, the light spot forms a circular trajectory on the target plane. Circular fitting is then applied to the collected light spot coordinates, and an optimization model with equality constraints is constructed to solve for the calibration parameters. MATLAB-based simulations verified the feasibility of the method, analyzing the influence of initial values of angular and displacement parameters on calibration results, as well as robustness against laser ranging noise. Comparative experiments demonstrate that the proposed method achieves higher accuracy than other calibration approaches. An experimental system was also built, and calibration parameters were obtained using the proposed method. Experimental results show that the scanning error of the calibrated system does not exceed 05 mm, meeting the accuracy requirements of industrial applications. The method requires no additional expensive calibration objects, relying only on singlepoint measurements from the laser rangefinder and geometric constraints. It significantly reduces calibration cost and operational complexity while maintaining good noise resistance, making it well suited for highprecision industrial on-site calibration.

Abstract:Geometric parameter calibration is an effective method to improve the end-effector positioning accuracy of industrial robots, which directly affects operational precision, product quality, and production safety. It is significant to analysis and study on the positioning accuracy reliability of calibrated robots. Firstly, an MDH model is established and the axis measurement method is developed to calibrate the robot geometric parameters in this paper. Secondly, the positioning accuracy reliability is analyzed and formulated, and robot positioning accuracy reliability analysis method based on AMCS is proposed. Finally, calibration experiments are conducted on Staubli TX60 robot using Leica AT960 laser tracker under uncertain factors affected by measurement repeatability, joint motion ranges, joint motion step sizes, joint motion velocities, etc. Experimental results demonstrate that the proposed AMM improves the robot′s positioning accuracy by approximately 22.9%, verifying its effectiveness for geometric parameter calibration. In the meantime, AMCS and MCS are used to calculate the positioning accuracy reliability of calibrated robot under the influence of different measurement factors. The results show that joint motion range, joint motion speed, and measurement repeatability significantly impact the reliability of robot positioning accuracy. When the numerical tolerance values are set to 0.01 and 0.02, he probability distribution function (PDF) characteristics of positioning accuracy reliability obtained by AMCS exhibit maximum relative errors of only 1.1% and 1.9%, respectively, compared with MCS, while computation times are reduced to about 1/4 and 1/9 of MCS. It has been confirmed that the proposed AMCS can control the convergence speed and accuracy of the algorithm by setting different numerical tolerances, providing an efficient and reliable tool for analyzing the positioning accuracy reliability of calibrated robots. It is suitable for practical engineering applications in robot calibration and reliability evaluation.

Abstract:In this article, an adaptive fuzzy fixed-time fault-tolerant control is developed for a steer-by-wire system with actuator fault, input hysteresis, and full-state constraints. Firstly, a dynamic model of the steer-by-wire system considering actuator fault input hysteresis is formulated. The input hysteresis, caused by factors such as the electromagnetic characteristics of the steering motor, mechanical transmission clearance and the delay in sensor signal processing, is characterized using a backlash model. Actuator faults are modeled by incorporating effectiveness factors and bias faults, which reflect the performance degradation and deviation of the steering motor. Then, based on backstepping control theory, fuzzy logic system, and adaptive technology, the compensation method for actuator fault and input hysteresis is designed. In this method, the fuzzy logic system is employed to approximate the unknown nonlinearities in the system, while the adaptive law is designed to update only a single global parameter in real time, thereby effectively reducing computational complexity. To ensure that the system states always remain within predefined constraint boundaries, the barrier Lyapunov functions are introduced to incorporate the constraint conditions of the front wheel angle and its rate of change into the control law design. The method is then analyzed from the perspectives of safety, actuator feasibility, and driving comfort. On this basis, a fixed-time controller is constructed to ensure that the system tracking error converges to a bounded compact set within a fixed time, thereby effectively improving the control accuracy and reliability of the closed-loop system under the influence of complex factors. The experimental results show that the states of the proposed method do not exceed the bounds under both classic scenarios, including double lane change and sharp turn, as well as extreme conditions on a low adhesion road surface. Furthermore, the average maximum error and average root mean square error are 0.038 and 0.006 rad, respectively, which are obviously superior to other methods in existing literature.

Abstract:To address the issue of degraded current tracking accuracy in NPC grid-connected inverters caused by parameter mismatches under conventional deadbeat predictive current control, this paper proposes a model reference adaptive deadbeat predictive control strategy based on Popov′s hyperstability theory. First, a two-step prediction scheme is employed to forecast and compensate the output current, achieving high dynamic response and low harmonic distortion under ideal conditions. Second, to mitigate the influence of inductance parameter variations, a model reference adaptive structure is introduced. By comparing the outputs error of a reference model and an adjustable model, an adaptation law is designed according to Popov′s hyperstability theory, enabling real-time parameter identification and dynamic compensation. This enhances parameter robustness and current prediction accuracy without compromising dynamic performance. Furthermore, a dynamic factor-based SVPWM strategy is incorporated into the control framework. By redistributing the action time of voltage vectors, this method effectively suppresses neutral-point potential fluctuations while synthesizing the desired output voltage vector, thereby improving control quality and operational stability. Finally, both simulation and experimental results demonstrate that, compared with conventional deadbeat predictive current control,the proposed strategy reduces output current THD by 9% under matched parameters and by 28% under parameter mismatch, while improving dynamic response speed by 34%. These results verify the of the proposed strategy in enhancing system robustness, improving waveform quality, and accelerating dynamic response.

Abstract:Aiming at the problem of model mismatch caused by time-varying parameters of permanent magnet synchronous motor (PMSM) in wind turbine pitch system in complex operating environment, a model-free predictive current control (HBF-MFPCC) scheme for PMSMs, integrating ultra-local modeling, an HBF neural network observer, and an improved dual-vector modulation strategy, is proposed. A first-order ultra-local model is employed to construct the predictive model for the proposed model-free current control, enabling future current prediction based solely on historical current and voltage data.. The current value in the future can be predicted only by using the historical information such as current and voltage of the motor, and the dependence on the parameters such as resistance, inductance and flux linkage of the motor is eliminating the dependence entirely, which solves the problem that the traditional model predictive current control (MPCC) depends on accurate motor parameters. A HBF neural network observer is designed to quickly identify the lumped error of the prediction model. The decision tree is used to optimize the center and width of the basis function. The observer has high identification speed and adaptability, which significantly enhances the accuracy of the prediction model. An improved dual-vector optimal duty cycle modulation strategy is adopted. The optimal vector is selected from 19 possible voltage vector combinations to drive the inverter. Adaptive time allocation is then applied to suppress current ripple, thereby improving current tracking performance. The simulation and experimental results show that the proposed HBF-MFPCC strategy can reduce the current tracking error by 50 % and the harmonic distortion rate by 28 % compared with the MPCC strategy under the condition of simulating extreme parameter mismatch. The designed HBF neural network observer can reduce the current tracking error by 53 % and the harmonic distortion rate by 55 %. The improved double vector modulation method can reduce the current tracking error by 24 % and the harmonic distortion rate by 11 %. This scheme can significantly improve the robustness of the system and ensure good current tracking performance.