Abstract:With the continuous development of intelligent security systems, pedestrian retrieval for all-day surveillance has become one of the research hotspots. Thus, the research of visible-infrared cross-modality person re-identification has emerged. The main challenge faced in this task is the huge discrepancy between visible and infrared images of the same pedestrian. Existing methods focus on exploring the shared information and reducing the feature variances of the same pedestrian in the two modalities. To further improve the accuracy of the task, this paper proposes multi-granularity shared-disentangling relation network for re-identification. By embedding the shared-disentangling module, the parameter-sharing branch of the backbone is replicated and decomposed, thus breaking limitations of the original benchmark model in multi-granularity feature extraction. By designing the multi-granularity relation feature learning module, the modality-invariant correlation information of the pedestrian body is fully explored, enhancing the learning of the shared features. And through constructing a loss function in multiple levels, effective supervision is available for the training of the model, and the global-local feature alignment scheme is optimized. The proposed algorithm obtains superior performance on both public datasets named SYSU-MM01 and RegDB. The Rank-1 and mAP in All-search mode on the SYSU-MM01 dataset can reach 74.70% and 71.79% respectively. In both retrieval modes of RegDB, Rank-1 and mAP are higher than 90%, and the accuracy is superior to many state-of-the-art methods. Experiments demonstrate the advantages of this network in cross-modality feature alignment and complex scene adaptation.

Abstract:In industrial production, prolonged and high-intensity operations can lead to worker fatigue, increasing the risk of safety incidents. Existing research has shown that contact-based physiological features can effectively represent fatigue status, but using contact-based instruments to monitor fatigue in industrial environments interferes with operations. Therefore, fatigue detection based on surveillance video has become a more practical choice. Current methods mainly focus on mouth and eye features, failing to comprehensively reflect fatigue status. To address this issue, we propose a non-intrusive fatigue detection method that integrates facial appearance and physiological representation, utilizing a video-based dual-branch network model for monitoring worker fatigue. First, we locate the facial areas of interest in the video and segment these areas. By extracting changes in skin reflectance due to variations in capillary blood volume, we construct a physiological spatiotemporal map. Next, we build a dual-branch 3D convolutional network to extract facial appearance and physiological feature representations separately. Finally, we fuse these features and input them into a fully connected layer to map the final fatigue detection results. The proposed method is validated using a fatigue dataset obtained from simulated industrial production tasks. Experimental results demonstrate that the fatigue detection accuracy, based on the integration of facial appearance and physiological features from video, reaches 88%, offering higher accuracy and stronger applicability in industrial settings compared to existing technologies.

Abstract:A portable ultrafine electronic endoscope system based on the STM32F407IGT6 microcontroller has been developed to address the growing demand for high-quality imaging in narrow, confined spaces within the medical field. The system utilizes the OV6946 camera module for image acquisition, and the analog signals are converted to digital form through the OV426 bridge chip. Image data is then transmitted to the host computer via the USB interface. The host computer leverages the LabVIEW platform, integrating advanced image processing techniques, including vertical stripe noise filtering, RAW color restoration, and an improved contrast limited adaptive histogram equalization (CLAHE) algorithm, to enhance image quality significantly. Experimental results demonstrate the effectiveness of the denoising and image enhancement processes. After denoising, the peak signal-to-noise ratio (PSNR) of the processed image reached 37.65 dB, with a structural similarity (SSIM) of 0.970 8, indicating minimal information loss and high structural integrity. Additionally, the image’s local contrast was improved from 3.32 to 13.16, and the average gradient increased from 7.08 to 28.05, which highlights a substantial enhancement in contrast and sharpness, particularly in the vascular regions. The system achieved a real-time processing frame rate of 30 frames per second, with a processing delay of 33 ms, satisfying the stringent requirements for high real-time performance and high-quality imaging in medical diagnostics. In terms of hardware design, the system’s compact size and reduced weight represent a significant improvement compared to traditional endoscopy systems, enhancing both portability and operational flexibility. The experimental results indicate that this system excels in terms of imaging quality, real-time performance, and portability. It offers promising potential to assist clinical diagnoses, improving the accuracy and efficiency of medical procedures. Moreover, the system shows broad applicability in clinical settings, particularly for endoscopic examinations in confined spaces such as the gastrointestinal and respiratory tracts.

Abstract:The significant variation in cardiac dimensions across different age groups and the faster heart rate in children result in more blurred cardiac borders compared to adults, impacting the segmentation of echocardiography. To address the above problems, the segmentation model called H2Former is improved, and the model called TPA-H2VSS combining the attention and state space is proposed to segment the left ventricle of pediatric echocardiography. Firstly, this paper replaces the Transformer block with the visual state space (VSS) block to enhance the model’s advantage in long-range modeling. Secondly, the temporal attention (TA) module is built between the encoder and decoder in the model to complements and interacts with the semantic information of the left ventricle in the echocardiography video in the temporal dimension. Finally, the positional attention (PA) module is added in the output head to make pediatric echocardiographic left ventricle segmentation more accurate. The experiments were trained, validated, and tested on the pediatric echocardiographic video dataset EchoNet-Pediatrics on the PSAX dataset and the A4C dataset, respectively. Compared with the base model H2Former, Dice, IoU, and accuracy on the PSAX dataset were improved by 0.86%, 1.41%, and 0.15%, respectively, and HD was reduced by 0.219 5. Dice, IoU, and accuracy on the A4C dataset were improved by 0.93%, 1.53%, and 0.2%, respectively, and HD was reduced by 0.167. By comparing with other models, it was demonstrated that the model could effectively segment the left ventricle in pediatric echocardiography and could provide a new solution for the auxiliary diagnosis of congenital heart disease.

Abstract:Fatigue driving detection is critical for traffic safety. Real-time monitoring and accurate identification of a driver’s fatigue level, coupled with an early warning system, can significantly reduce the risk of accidents caused by fatigue. Addressing the challenges of small micro-expression targets and complex background environments in current driving fatigue detection, this paper proposes an improved driving fatigue detection model—YOLOv10-GMF. The model incorporates an enhanced global grouped coordinate attention (GGCA) module, which improves feature representation by weighting feature maps with global information and generating attention maps, thereby enhancing the model’s ability to capture micro-expression features under fatigue conditions. Additionally, a multi-dimension fusion attention (MDFA) module is integrated, which combines multi-scale dilated convolutions with spatial and channel attention mechanisms in parallel to strengthen the model’s recognition ability for image features in complex driving environments. To further optimize the training process, a feedback-driven loss function (FDL) is introduced, effectively accelerating model convergence and improving prediction accuracy. Ablation experiments demonstrate that the YOLOv10-GMF model achieves a detection accuracy of 98.1%, a 14.5% improvement over YOLOv10, with a detection speed of 64.3 fps. Through real vehicle embedded deployment tests, the average fatigue detection process takes 19.0 ms, and the model fully meets the real-time monitoring needs for fatigue driving.

Abstract:Accurate monitoring of animal vital signs is crucial for health management and disease diagnosis. However, detecting these signals poses several challenges. Breathing and heartbeat signals in animals are weak, with heartbeats easily interfered with by breathing harmonics and noise. Additionally, animal physiology differs from humans, and detection environments can be complex. To address these issues, this study explores millimeter-wave radar-based methods for monitoring vital signs. It proposes an improved adaptive unscented Kalman filter combined with wavelet-based spectral estimation. The approach optimizes the adaptive unscented Kalman filter using a noise weighting factor, maintaining its sensitivity to new observations. It also uses different wavelet bases to extract purer signal features based on the distinct characteristics of heart and breathing rates, employing spectral density estimation for calculating these parameters. The algorithm was validated on 29 cattle and 10 dog datasets, showing accurate measurement. The root mean square errors were 0.030 4 and 0.031 5 for breathing frequency, and 0.057 4 and 0.056 9 for heart rate. Compared to traditional peak - detection algorithms, detection accuracy improved by 3.33% and 7.26% for cattle, and 3.65% and 6.96% for dogs. The algorithm offers high accuracy and strong noise resistance, making it valuable for both theoretical and practical vital-sign detection.

Abstract:Two-person interaction action recognition based on skeleton sequence data has broad application prospects. To address the issues of insufficient interaction feature representation and redundant intra-class features in current recognition models, we propose a multi-scale deformable graph convolutional network (MD-GCN) for recognizing two-person interaction actions. First, we construct a two-person interaction hypergraph, including a person pair hypergraph and an interaction relationship matrix. Unlike traditional graphs, this hypergraph better captures the interaction between the two people, enabling a more comprehensive representation of the interaction features. Next, three input branches perform data preprocessing and feature extraction, and then the extracted features are fused and fed into the main branch, which is based on the multi-scale deformable graph convolutional network for action classification. This network learns deformable sampling positions in a multi-modal manner, effectively capturing key interaction features while avoiding feature redundancy and information loss. The proposed MD-GCN achieves a recognition accuracy of up to 98.41% on the 26 interaction action classes from the NTU RGB+D 60 and NTU RGB+D 120 datasets. This approach effectively addresses the challenges of feature representation in two-person interaction action recognition. Experimental results show that the method not only maintains high recognition accuracy but also significantly reduces the computational cost, achieving a good balance between inference performance and accuracy, making it highly valuable for practical applications.

Abstract:To address the scarcity of online fault diagnosis algorithms caused by strong nonlinear characteristics and multi-parameter coupling in pressurized water reactor (PWR) nuclear power plant steam generator systems, this paper proposes an online parameter identification method based on a self-inspired genetic algorithm (GA). First, a model-driven self-supervised GA framework is constructed based on parameter identification theory, transforming the fault diagnosis problem into the identification of key system performance parameters. By integrating a high-fidelity system mechanism model, the parameter identification task is reformulated as a function optimization problem, effectively overcoming the limitations imposed by nonlinearities and high-order differential terms in the system equations. Subsequently, a parameter identification method is developed by designing a fitness function based on dynamic time warping and optimizing the GA population iteration strategy using a quasi-Newton gradient descent approach. This replaces the global random search strategy with a gradient-directed search strategy, resolving the slow convergence issue of traditional GAs and meeting the requirements for online parameter identification. Finally, the proposed method is validated using both model data and real system simulator data. Compared to conventional GAs, it reduces parameter identification error by approximately 5% and decreases the average number of convergence steps by 47%, demonstrating the effectiveness of the self-inspired GA-based parameter identification method.

Abstract:Accurate prediction of the projectile launch point can quickly locate enemy threat sources, provide critical intelligence support, and optimize counterattack strategies, holding significant strategic importance in the military field. This study addresses the problem of predicting projectile launch points and proposes a deep learning model that combines temporal convolutional network (TCN), bidirectional gated recurrent unit (BiGRU), and attention mechanism. The model aims to improve ballistic trajectory prediction accuracy, especially in complex battlefield environments, by backwardly inferring enemy projectile launch points to support counterattack strategies. Firstly, based on the ballistic model, a detailed projectile trajectory dataset was constructed by solving the six-degree-of-freedom rigid body ballistic equation for different launch angles and initial velocities. Then, the proposed TCN-BiGRU-Attention model captures long-term dependencies in the trajectory data by introducing the TCN structure and optimizes information weighting using the attention mechanism to enhance prediction accuracy. In simulation validation, compared with models like BiGRU, bidirectional long short-term memory (BiLSTM), and their improved variants, the TCN-BiGRU-Attention model demonstrated significantly superior performance in launch point prediction accuracy, particularly in reducing errors in both range and cross-range directions. Through multiple sets of simulation tests, the results indicate that the TCN-BiGRU-Attention model can stably provide accurate launch point predictions at various launch heights. At sea level, the model’s range error is only 8.3 meters, and the cross-range error is minimal, effectively predicting and striking the enemy’s launch point. This study provides theoretical basis and technical support for the implementation of enemy launch point prediction in future battlefield scenarios.

Abstract:Radio frequency identification (RFID) technology has become increasingly important in intelligent warehouse management and logistics tracking systems. However, conventional commercial RFID systems, operating within the Industrial, Scientific, and medical radio band, are constrained by limited bandwidth, which hinders high-precision carrier phase-based ranging. Moreover, in scenarios such as mobile robots and handheld readers, the deployment of multi-antenna arrays is impractical due to space limitations, posing further challenges to angle-based localization techniques. To address these issues, this paper proposes a synthetic aperture-based RFID localization system utilizing a single antenna and a single-frequency point. The system constructs a spatially non-uniform virtual linear array by collecting phase sequences and corresponding timestamps during antenna motion. A fast coarse angle estimation algorithm based on the derivative of the phase sequence is introduced to reduce the search space and improve estimation efficiency. Furthermore, the phase differences between distinct virtual antenna positions are used to compute differential distances to the target tag. These distance differences, combined with angle information, are formulated into a hyperbolic localization model. The final coordinates of the RFID tag are estimated using a particle swarm optimization algorithm with adaptive weighting. Experimental results validate the effectiveness of the proposed system, the median error of differential distance estimation is 4.2 cm, the median angle estimation error is 1°, and the final localization median error reaches 3.45 cm. The proposed method achieves highaccuracy localization on commercial RFID platforms without additional hardware costs, and thus holds promising practical application value.

Abstract:Conventional control methods struggle to satisfy the high-performance speed regulation requirements of switched reluctance motors (SRM) under variable-speed and variable-load operating conditions. This paper proposes a super-twisting control algorithm (STCA)-based speed regulation method for SRMs. A mathematical model of SRM was established, followed by an in-depth investigation into the systematic design methodology of STCA implementation under variable operating conditions. To optimize control parameters, a multi-objective optimization fitness function was constructed and addressed through the fruit fly optimization algorithm (FOA), with detailed analysis of its implementation process. Comparative simulations and experiments with conventional gradually steady signal error control (GSSEC) algorithm demonstrate significant improvements: Steady-state control accuracy increased by 1.213%; For abrupt speed changes, transient process durations reduced by 22.3% (sudden increase) and 26.6% (sudden decrease); Under load disturbances, speed overshoots decreased by 61.09% (sudden increase) and 62.28% (sudden decrease), with corresponding transient durations reduced by 36.36% and 38.88%. The proposed method exhibits superior steady-state and dynamic performance metrics compared with GSSEC, demonstrating substantial practical application value.

Abstract:Prefabricated steel structures are widely used in construction projects due to their advantages such as light weight, high strength, and convenient construction. However, their fire resistance and anti-corrosion performance are insufficient in complex service environments, and they need to rely on multiple protective coatings (such as fireproof layers and anti-corrosion layers) to ensure long-term durability. The precise measurement of coating thickness is a key link in quality control. The existing ultrasonic testing technology is confronted with problems such as echo aliasing at multi-layer interfaces and noise interference, resulting in insufficient measurement accuracy. To this end, this paper proposes a coating thickness detection method based on the combination of high-frequency ultrasonic pulse reflection method and sparse decomposition matching pursuit (MP) algorithm, and improves the detection accuracy through signal separation and reconstruction. Firstly, a finite element model of the protective coating for prefabricated steel structures was established based on COMSOL to simulate the ultrasonic detection process of water immersion, analyze the propagation characteristics of ultrasonic waves in multi-layer media, and reveal the physical mechanism of interface echo. Aiming at the problem of aliasing signal separation, a sparse decomposition MP algorithm that optimizes the over-complete atomic dictionary of Chirp is proposed. Signal denoising and feature reconstruction are achieved by iterative matching of the optimal atoms, and the separation effects of the traditional wavelet transform modulus maximum method are compared. The simulation results show that under the same signal-to-noise ratio condition, the root mean square error of the reconstructed signal by the MP algorithm is significantly lower than that of the wavelet transform method. The relative errors in the thickness detection of its fireproof layer and anti-corrosion layer are also better than those of the wavelet transform method. To verify the practicability of the method, a 20 MHz water-immersed probe was used to test the anti-corrosion layer and fireproof coating test blocks, and the true thickness was calibrated in combination with the metallographic method. The echo of the superimposed interface was successfully separated through the MP algorithm. The relative errors of the thickness of the fireproof layer and the anti-corrosion layer were calculated to be -4.64% and -4.65% respectively, which were significantly better than -7.15% and -7.28% of the wavelet transform. Therefore, the method proposed in this paper has better practical detection application scenarios.

Abstract:Aiming at the trouble which the tractive performance of one or more motors in intercity train tractive system decreases due to uncertainties such as parameter perturbations and unknown disturbances, resulting in the reduction of total traction, a multi-motor predefined time total collaborative control method based on virtual spindle is proposed. Firstly, considering parameter perturbations and unknown disturbances, the state equations of the multi-motor of the intercity train are established based on the tracking error of the output torque of multiple motors versus a given torque. Secondly, a novel predefined time sliding mode surface is used to design a predefined time total cooperative controller based on the total cooperative consistent control algorithm; meanwhile, the virtual spindle control strategy is used to feedback the sum of multi-motor torque to the total cooperative controller to ensure that the total train traction force is constant, and at the same time, the synchronization control performance of multi-motor is improved. Finally, the proposed control algorithm is verified by comparing simulation and hardware-in-the-loop experiments with PI control and integral sliding mode control. The total output torque of the multi-motor is kept stable within a predefined time, and the output torque of the system is able to track the given torque within 0.005 s, with a tracking error of no more than 0.03%. The results show that the method enhances the synchronization control performance of a multi-motor traction system when there are parameter perturbations and unknown disturbances in the motor.

Abstract:Real-time and high-precision detection of smoke and fire is of great significance for fire monitoring and rapid early warning. Addressing the challenge that the current detection methods have difficulty balancing accuracy and real-time performance, as well as the problem of high computational complexity, this paper proposes a refined YOLOv5s smoke and fire detection method. Firstly, the Neck structure was optimized. On the basis of the original FPN-PAN architecture, it adds an additional P6 feature detection layer targeting smaller scales. Then, it enhances the network’s multi-scale feature fusion capability and improve the recognition and localization accuracy for small objects. Secondly, a lightweight modification was applied to the C3 module within the backbone network. C3 modules were replaced with C3RepGhost modules based on structural re-parameterization, effectively reducing the computational load and accelerating the inference process. Furthermore, a large-scale smoke and fire dataset is conducted and it consists of approximately 18 000 images from diverse scenes (including urban streets, forests, and individual flames) for model training and validation. Experimental results demonstrate that the proposed method achieves a mean average precision (mAP) of 0.89 on the above dataset, with an improvement of approximately 29% compared to the original YOLOv5s model. The detection speed reaches 66 fps. The proposed method realizes high-accuracy and real-time smoke and fire detection. Compared to the latest YOLOv11s model, the computational complexity of the refined YOLOv5s method is reduced by 46%, making it more suitable for deployment on edge computing devices.

Abstract:In order to improve the efficiency of the double-sided LCC compensation wireless charging system, a parameter optimization method is proposed to realize the zero-voltage switching of the inverter in both constant current and constant voltage modes by optimizing the parameters of the compensation element. Firstly, the double-sided LCC compensation topology is modeled by Kirchhoff’s theorem, and the constant current and constant voltage output conditions independent of the load are analyzed under the condition of zero phase angle. Secondly, a dual-mode parameter collaborative optimization strategy is proposed: the disturbance coefficient is introduced to quantitatively analyze the influence trend of the parameter disturbance of the compensation element on the imaginary part of the equivalent input impedance of the system in the dual-mode, and the components that can make the equivalent impedance show weak inductance in the dual-mode are screened. Based on this screening result, the parameter optimization method of the system to achieve zero voltage switching in both constant current and constant voltage dual modes is given, so as to realize dual-mode ZVS. Finally, a simulation model of the double-sided LCC compensation wireless charging system and an experimental prototype with an input voltage of 15 V are built for verification. The results show that the system has a constant current and constant voltage output capability independent of the load. The optimized system inverter can achieve zero voltage switching in both modes. The maximum efficiency of the system can reach 91.31%, which is 1% ~ 1.1% higher than the efficiency of the traditional single-mode parameter optimization method, which verifies the effectiveness of the proposed parameter optimization method.

Abstract:A lightweight YOLO-SGLS model is proposed based on the YOLO11 algorithm to address the shortcomings of existing elevator wire rope surface damage detection methods, such as insufficient accuracy and excessive computational complexity. Firstly, StarNet is used to replace the backbone network of YOLO11, and the star operation is used to improve feature extraction and computational performance. Meanwhile, the LSKA module is integrated with SPPF to enhance the feature expression and perception of the model through deep convolution. In addition, the Ghost module is improved using DynamicConv to obtain the Ghost Dynamic Conv （GDC) module, which is combined with C3K2 to reduce computational burden. Finally, an LSCD detection head is designed to improve inference speed. The experiment uses the Cable Damage dataset, which is divided into training, validation, and testing sets. In a specific experimental environment, ablation experiments, generalization experiments, and comparative experiments are conducted. The experiment shows that the YOLO-SGLS model reduces GFLOPs and parameter count by 40% and 36% respectively compared to the original base network YOLO11, improves accuracy by 5.5%, and only decreases average accuracy and recall by 0.3% and 1.9%. In the generalization ability test, the YOLO-SGLS model correctly recognizes 77 images out of 100 new datasets. It has been proven that the lightweight, accuracy, and robustness of the algorithm meet the requirements of elevator wire rope damage detection in practical application scenarios, especially for embedded devices with limited resources.

Abstract:In response to the problems of decreased robustness and insufficient feature expression ability caused by noise and outlier interference in traffic anomaly detection models, and low minority class detection rates when dealing with imbalanced high-dimensional massive data, a traffic anomaly detection method based on generative adversarial networks was proposed. Firstly, the clustering based on SCiForest algorithm is used to detect outliers and reduce their impact on the subsequent training of the generative adversarial network. Secondly, a denoising autoencoder-based generative adversarial network (DGAN) is designed to generate reliable synthetic minority class samples. The network defines its training target based on the Wasserstein distance between reconstructed error distributions, effectively alleviating the problem of data imbalance. Again, using a denoising autoencoder (DAE) with the same architecture as the generative adversarial network discriminator, real and synthetic samples are input for reconstruction training, and the optimized encoder part is extracted as the feature extraction and dimensionality reduction module to enhance feature expression ability. Finally, the processed data is input into the feature fusion model of CNN and BiGRU (CNN-BiGRU-FFusion) model, which completes classification and detection based on capturing spatial and temporal features. The accuracy and F1 score on the NSL-KDD dataset reached 92.06% and 92.25%, respectively, verifying the superior performance of the proposed method in network traffic anomaly detection tasks. The feasibility of the method was further validated through experiments on the CICIDS2017 dataset.

Abstract:As the volume of hyperspectral data utilized in planetary exploration continues to grow, the development of efficient and accurate algorithms for data analysis becomes increasingly critical. This study explores the potential of spectral unmixing techniques to analyze hyperspectral images of Mars. Observations from the compact reconnaissance imaging spectrometer for mars (CRISM) serve as the primary dataset. Preprocessing steps, including atmospheric correction and mitigation of the “Smile” spectral effect, are performed to minimize noise and provide a robust foundation for subsequent spectral profile analysis. The number of endmembers in hyperspectral images of the Jezero impact crater is estimated using an eigenvalue-based method. Specifically, the eigenvalue maximum likelihood method is employed to define a likelihood function that determines the optimal number of endmembers by identifying the global maximum without the need for threshold adjustments. This approach achieves reliable results even under low signal-to-noise ratio conditions. Subsequently, the vertex component analysis (VCA) algorithm is applied to decompose and extract the mixed endmembers in the images. The extracted results are compared with the CRISM spectral library, and key absorption features in the spectral curves are analyzed to identify specific minerals. This methodology enables precise identification of mineral components within the Jezero impact crater, including water-bearing silicate and carbonate minerals. These findings suggest that Mars may have once sustained a liquid water environment conducive to life and experienced a warmer, wetter climate during its ancient history.

Abstract:Wafer map defect pattern classification is a critical step in semiconductor manufacturing, significantly impacting product yield and production efficiency. To address the limitations of existing deep learning-based wafer map defect pattern classification methods, such as poor interpretability and high computational resource consumption, this study proposes an improved feature extraction method based on topological data analysis (TDA). By leveraging persistent homology theory, the method constructs Alpha complexes to characterize topological structures in wafer maps and quantifies them into discriminative features. Experimental results on a synthetic wafer map dataset, generated by emulating the geometric distribution characteristics of the WM-811K dataset, demonstrate that replacing the conventional vietoris-rips (VR) complex with the Alpha complex reduces the average complex construction time by approximately 82% and decreases memory usage by 10.09%. Compared to state-of-the-art models including DenseNet121, Swin Transformer, and ConvNeXt, the TDA-based method achieves superior clustering performance, as evidenced by t-SNE visualizations, with a 17.24% improvement in Silhouette Coefficient over the suboptimal ConvNeXt model, along with a 75% reduction in feature extraction time and a 95% reduction in peak memory consumption. When integrated with a support vector machine (SVM) classifier, the TDA-based framework attains an overall classification accuracy of 0.992, outperforming DenseNet (0.989 3) and Swin Transformer (0.982 0).

Abstract:Addressing the challenges in position detection and visualization monitoring for roadheaders, this study introduces a digital twin-driven method for roadheader pose monitoring. Initially, a three-dimensional positioning model and pose solution model for the roadheader, leveraging ultra-wideband technology, are developed. To mitigate the impact of underground non-line-of-sight environments on UWB positioning accuracy, a C-T fusion algorithm based on maximum correlation entropy Kalman filtering is proposed. This algorithm significantly enhances the precision of positioning tags under NLOS conditions. Subsequently, employing digital twin technology, a digital twin system is constructed using Unity3D software, with a MySQL database serving as the medium for data exchange, thereby facilitating real-time three-dimensional visualization monitoring of the roadheader’s position in virtual space. Ultimately, an experimental platform is assembled to conduct position detection experiments for the roadheader. The experimental outcomes demonstrate that the positional error in the roadheader positioning experiment does not exceed 3.44 cm, and the angular error is within 0.34°. The system’s monitoring capabilities are real-time and effective, ensuring the synchronization and consistency of the system, which aligns with the demands for roadheader pose detection and visualization monitoring during the operation at the mining face. This approach offers a novel perspective on the application of digital twin technology in underground tunneling operations within coal mines.

Abstract:To address the issues of high current stress and poor dynamic response in hybrid three-level dual active bridge (DAB) converters, this study proposes an improved extended phase-shift (EPS) modulation scheme and a minimum current stress control strategy based on a super-twisting sliding mode active disturbance rejection controller (STSMC-ADRC). First, the EPS modulation is enhanced by redefining the internal and external phase-shift ratios, ensuring a positive correlation between the phase-shift ratios and transmitted power while reducing coupling between the ratios. Second, mathematical models of transmitted power and current stress under different operating modes are analyzed for the improved EPS modulation. The karush-kuhn-tucker (KKT) conditions are applied to solve for the optimal phase-shift ratio combination that minimizes current stress while satisfying soft-switching constraints. Third, to simplify calculations, a reduced-order model of the converter is established, and the super-twisting sliding mode control algorithm is integrated with active disturbance rejection technology to enhance dynamic performance. Finally, experimental validation is conducted using a prototype. Results demonstrate that compared to traditional ADRC, the proposed STSMC-ADRC reduces the regulation time by 72.4% and voltage fluctuation by 51.7% during sudden load resistance reduction. For input voltage step changes, the regulation time is shortened by 73.7% and voltage fluctuation decreased by 60%. Additionally, the strategy effectively reduces current stress and achieves soft switching. Compared to single phase-shift modulation, efficiency improves by 15% at low power and 9% at high power.

Abstract:Outdoor scenes are rich in feature points with diverse geometric shapes and scales; however, significant illumination variations and high texture repetitiveness often lead to low feature extraction and matching accuracy in conventional visual simultaneous localization and mapping (SLAM) algorithms during 3D reconstruction. To improve mapping accuracy and robustness in complex environments, this paper proposes a visual SLAM algorithm integrated with an attention mechanism, aiming to enhance the feature extraction and matching strategies within SLAM systems. Specifically, a channel-spatial convolutional attention module is embedded into the convolutional layers of the SuperPoint encoder to strengthen the model’s feature detection and matching capabilities. The improved SuperPoint network is then integrated with the backend of the ORB-SLAM2 algorithm, enabling more accurate pose estimation and map construction in complex scenarios. The proposed approach is validated on the KITTI dataset. Experimental results demonstrate that the SuperPoint network integrated with the channel-spatial convolutional attention module significantly improves feature matching accuracy between images while maintaining the stability of keypoints and the discriminability of descriptors. Compared with the original ORB-SLAM2 algorithm, the proposed method achieves a 30.05% reduction in absolute trajectory error (ATE) and a 14.49% reduction in relative pose error (RPE). These results confirm that the proposed SLAM algorithm exhibits stronger robustness and stability in outdoor environments characterized by significant illumination changes and repetitive textures, effectively enhancing the mapping accuracy of SLAM systems in complex outdoor scenes.

Abstract:As a core component of power electronic systems, insulated gate bipolar transistors （IGBT） are susceptible to performance degradation and failure due to electro-thermal stress during practical applications, making accurate remaining useful life prediction crucial. To address the insufficient prediction accuracy of single models in insulated gate bipolar transistors lifetime prediction, this paper investigates a multi-model fusion approach for remaining useful life prediction. The method first employs variational mode decomposition （VMD） to decompose the collector-emitter transient peak voltage, a key characteristic parameter for insulated gate bipolar transistors lifetime prediction, into multiple intrinsic mode components. The low-frequency trend component is predicted using a Gaussian process regression model, while the high-frequency fluctuation components are modeled using neural basis expansion analysis for time series (NBEATS) network. The final prediction is obtained by reconstructing and fusing the predictions of all components. Validation using NASA’s IGBT accelerated aging experimental data shows that the proposed fusion model achieves a 70% reduction in root mean square Error, a 23.2% decrease in mean absolute error, and an improvement in the coefficient of determination to above 0.97 compared to the best single VMD-NBEATS model. By varying the ratio between training and testing sets, the fusion model consistently demonstrates superior performance across different proportions, validating the stability and generalizability of the multi-scale decomposition and differentiated modeling approach. This work provides a novel solution for health monitoring and preventive maintenance of power electronic devices.

Abstract:To address the issue of inaccurate sediment thickness identification in the existing ultrasonic echo detection method for grading electrodes sediments, an intelligent identification method of grading electrodes sediments based on SVMD-CPO-RF-AdaBoost combining multiscale entropies is proposed. The successive variational mode decomposition method is used to decompose the ultrasonic echo signals, and effective signal components are selected and reconstructed based on the modal selection method to achieve signal denoising. Multiscale permutation entropy and multiscale dispersion entropy are introduced as feature extraction methods to extract the multiscale features of ultrasonic echo signals from grading electrodes with different sediments thicknesses. A dual ensemble learning model of RF-AdaBoost, with random forest as the weak classifier of AdaBoost, is constructed and optimized by the crested porcupine optimizer to achieve intelligent identification of sediment thicknesses. A comparative study was conducted on different multiscale entropies, as well as six models including CPO-RF, CPO-AdaBoost, and CPO-SVM. Experimental results show that the recognition accuracy of sediment thicknesses using a combination of MPE and MDE features is superior to that using a single feature. The proposed method can accurately identify the sediment thickness of the grading electrodes sediments in the range of 0 to 0.8 mm, with a recognition accuracy of 94.50%. Moreover, its precision, recall, and F1 score outperform those of other methods, providing an intelligent identification solution for grading electrodes sediments in ultra-high voltage converter stations.

Abstract:In order to improve the efficiency of circuit breaker life prediction and formulate a reasonable maintenance plan, an IFMD-BiTCN-BiGRU-AT prediction model based on crown porcupine optimization algorithm (CPO) is proposed based on the characteristics that the non-periodic vibration signal of the circuit breaker can fully characterize the residual life. Firstly, the feature mode decomposition method is improved by integrating the fitness function and the new period estimation method to make up for its poor ability to deal with non-periodic signals, and the IFMD adaptive decomposition is realized by using CPO. Secondly, a two-way parallel structure and attention mechanism are introduced. The BiTCN-BiGRU-AT prediction model is constructed to fully extract the important features of time-space, and the CPO is used to search the optimal hyperparameter combination. Finally, the experimental platform of circuit breaker signal acquisition and processing is built for experimental verification. The method is used to predict and design ablation experiments and multi-model comparison experiments. Finally, the fitting degree, MAE and RMSE indexes obtained by this method are 99.28%, 80.33 and 98.17 respectively. Compared with the other three signal processing methods, the prediction fitting degree is increased by 19.7% on average after IFMD processing, and the prediction efficiency is the highest. Compared with other models, the prediction fitting degree of the model is increased by 18.3% on average, and the MAE and RMSE are reduced by 60.9% and 61.6% on average. Experimental results show the effectiveness and performance advantages of the proposed method.

Abstract:In the fields of intelligent transportation systems and security monitoring, the accuracy of target detection technology is of great significance. However, in addition to the normal traffic environment, adverse weather conditions such as rain and snow severely restrict the accuracy of target detection. Rain and snow weather make images blurry, greatly increasing the difficulty of feature extraction for targets such as pedestrians and vehicles, resulting in large errors in detection results and affecting the effective operation of related systems. To address this challenging problem, this paper proposes an optimized target detection method for special weather conditions like rain and snow based on the YOLOv7 algorithm. Firstly, a widely-used dark channel de-fogging algorithm and a rain and snow removal algorithm based on guided filtering are introduced to preprocess images affected by rain, snow, and fog. This effectively eliminates the image degradation caused by weather factors and restores clear details of the images. Secondly, the Deep Image Prior (DIP) module is combined with the convolutional neural network-post-processing (CNN-PP) module. Through weakly supervised learning, the method further excavates the target features in the images, enhancing the algorithm’s recognition ability for targets in complex weather conditions. Extensive experimental results demonstrate that the improved algorithm performs excellently in terms of detection accuracy. Compared with the YOLOv5 algorithm, its detection accuracy has increased by 23.7%, and a significant growth of 11.9% has been achieved compared with the original YOLOv7 algorithm. These results fully prove the effectiveness and superiority of the proposed method in target detection scenarios under special weather conditions. It provides reliable technical support for the stable operation of intelligent transportation, security monitoring, and other fields in adverse weather, showing important practical application value and broad development prospects.