Chinese Haipai New Year paintings are an important part of the country’s intangible cultural heritage, and their digital preservation holds great significance. This paper proposes PSE-Net (Pyramid Scale Expansion Network), a deep learning-based segmentation method specifically designed to handle the complex textures and intricate compositions of these artworks. By constructing a dedicated large-scale dataset, we trained PSE-Net to achieve high-precision segmentation by incorporating attention mechanisms and multi-scale feature fusion to better capture detailed features. Experimental results demonstrate that the proposed method outperforms existing approaches (such as ResNet) in terms of segmentation performance, yielding superior results in edge preservation. This work establishes the first automated tool for the pixel-level analysis of Haipai New Year paintings, thereby facilitating museum digitization, art history research, and education. Furthermore, it offers new insights for the image processing and digital preservation of other traditional artworks.

Objective Bridges are critical to transportation infrastructure, but aging and extreme loads make traditional monitoring methods (e. g., accelerometers, video) inadequate for real-time, high-precision structural health monitoring (SHM); they often suffer from narrow coverage, poor environmental adaptability, and high false alarm rates. Phase-sensitive optical time-domain reflectometry (Phi-OTDR), a distributed optical fiber sensing (DOFS) technology, offers high sensitivity, long-range monitoring, and immunity to electromagnetic interference, making it ideal for bridge micro-vibration detection. However, Phi-OTDR signals are weak, spatiotemporally coupled, have a low signal-to-noise ratio (SNR), and are multimodal, posing challenges to traditional demodulation methods. Existing approaches (e.g., STFT, wavelet analysis, SVM, ResNet) struggle with the trade-off between resolution and efficiency, rely on inefficient manual feature extraction, or exhibit poor robustness under low SNR conditions. This study aims to develop a high-precision, lightweight, and noise-robust Phi-OTDR-based method for bridge vibration recognition. Methods A full-process optimization scheme was proposed. For signal acquisition and preprocessing, a scaled bridge model was employed, with approximately 8 m of serpentine-laid single-mode fibers used to collect Rayleigh backscattering (RBS) signals for four events: vehicle load, structural deformation, damage, and human intrusion. The RBS signals were formed into two-dimensional spatiotemporal "waterfall plots", which were then processed via mean filtering, differencing, and normalization. A two-dimensional Fourier transform (2D-FFT) was applied to map the waterfall plots into the wavenumber-frequency domain, thereby decoupling guided waves, separating modes (e.g., L(0, 1)), and capturing dispersion features for "fingerprint" recognition. For classification, a lightweight HF-EfficientNet was designed. This 7-stage model employed HybridFusedMBConv+SE (Stages 1-2) for shallow edge feature extraction, FusedMBConv+CBAM (Stage 3) for efficiency, and MBConv+ECA (Stages 4 -7) for deep semantic learning. The model was trained using cross-entropy loss, SGD, cosine annealing, and data augmentation. For noise suppression, an adaptive noise suppression-generative adversarial network (ANS-GAN) was introduced, which utilized a spatial correlation matrix to guide the generators to focus on cooperative vibration regions; an ANSModule filtered noise through dynamic frequency coefficients and multi-objective loss optimization. Results and Discussions When compared with four mainstream convolutional neural network models (ResNet18, EfficientNetB0, EfficientNetV2, and MobileNetV2) under the same experimental conditions, the proposed HF-EfficientNet model demonstrated significant advantages in both recognition performance and efficiency. In terms of classification performance, the HF-EfficientNet achieved an overall accuracy of 96.7 %, a precision of 97.29 %, a recall of 97.22%, an F1-score of 97.25 %, and a false alarm rate (FAR) as low as 0.76%. Specifically, its F1-score was more than 1.5 percentage points higher than that of ResNet18 (95.55 % ) and EfficientNetV2 (95.46 % ), while its false alarm rate was nearly halved compared to ResNet18 (1.47 % ) and EfficientNetV2 (1.49% ). In terms of model efficiency, the HF-EfficientNet had a parameter size of only 5.93 MB, which was much smaller than ResNet18 (11.18 MB) and EfficientNetV2 (20. 18 MB), making it well-suited for deployment on edge computing devices in on-site bridge monitoring scenarios. From the perspective of confusion matrix analysis, the HF-EfficientNet exhibited the highest diagonal values, indicating that it had the strongest ability to correctly classify each type of bridge vibration event; especially for highly similar events (such as bridge structural deformation and local damage), it showed fewer misclassifications compared to ResNet18 and EfficientNetV2. The results of t-SNE feature space visualization further confirmed the superiority of the HF-EfficientNet: the feature clusters of the four vibration event categories in the HF-EfficientNet showed the clearest separation, with minimal overlap between the green (structural damage) and blue (human intrusion) categories. In contrast, ResNet18 had severe overlap among multiple categories (green, purple, and blue), which significantly affected its classification accuracy. Additionally, in the low SNR (3 dB) test, the recall rate of the HF-EfficientNet only decreased by 2.1 percentage points, while that of EfficientNetV2 decreased by 6.6 percentage points, thus fully verifying the strong noise robustness of the HF-EfficientNet. Conclusions This method addresses the limitations of traditional approaches through 2D-FFT-based feature extraction, a lightweight HF-EfficientNet, and ANS-GAN noise suppression. It delivers high accuracy, a low false alarm rate, and compatibility with edge computing, thereby providing reliable SHM for bridges. Future work will focus on compressed sensing and transfer learning for broader infrastructure applications.

Objective In plant factories, accurate detection of fruit ripeness plays a key role in realizing intelligent supplemental lighting and resource optimization. However, tomato ripeness detection still faces many challenges, such as subtle inter-class differences, the difficulty of detecting small fruit targets, and the computational resource limitations of real-time edge deployment. Existing deep learning-based detectors are insufficient in fine-grained feature extraction, and the efficiency of multi-scale feature fusion is limited, leading to decreased accuracy in complex agricultural environments. To address these problems, this study proposes an improved tomato ripeness detection model based on Mamba-YOLO-T, which enhances Backbone feature representation and Neck multi-scale fusion capability to improve detection accuracy while ensuring real-time performance. Methods This paper builds upon the lightweight Improved-Mamba-YOLO-T framework and introduces targeted architectural improvements, primarily in two aspects. First, for Backbone enhancement, we propose the channel-aware feature enhancement ODSSBlock (CAFE-ODSSBlock). This novel module integrates channel attention mechanisms with state space sequence modeling. It selectively amplifies fine-grained features relevant to ripeness estimation while preserving the model's inherent capacity for long-range contextual modeling, thereby significantly improving feature discriminability. Second, we reconstruct the Neck for semantic fusion enhancement by incorporating the Zoom_cat and semantic detail injection (SDI) modules, along with an adaptive spatio-temporal fusion (ASTF) mechanism. Through multi-scale feature modeling and residual fusion pathways, this enhances Neck structure substantially boosts recognition accuracy and robustness under complex background conditions. Results and Discussions Effectiveness of the proposed method is rigorously validated through comprehensive ablation and comparative experiments. Ablation studies demonstrate that integrating the CAFE-ODSSBlock into the Backbone network significantly enhances feature discriminability, leading to noticeable improvements in both mean average precision when the intersection-over-union (IoU) threshold is 0.50 (mAP@0.50) and 0.50?0.90 (mAP@0.50 & ratio;0.95) metrics (Table 1). Furthermore, the introduction of the multi-scale dynamic semantic interaction neck (MDSI-Neck) effectively compensates for insufficient information flow between high-level and low-level features from the Backbone. In comparative experiments on a common tomato dataset, Improved-Mamba-YOLO-T achieves state-of-the-art performance with 90.4% mAP@0.50 and 76.0% mAP@0.50 & ratio;0.95, surpassing mainstream detectors including YOLOv8, YOLOv11, and RT-DETR-ResNet50 (Table 2). The model also maintains a high recall rate and balanced recognition performance across different ripeness stages on a cherry tomato dataset. For practical deployment, real-world testing confirms that Improved-Mamba-YOLO-T retains its lightweight characteristics while achieving enhanced accuracy, attaining a real-time inference speed of up to 17 frame/s on a Jetson Orin NX platform. This result indicates a superior balance between detection accuracy and computational efficiency compared to the latest models such as YOLOv11 and YOLOv12, confirming its strong suitability for edge computing applications in plant factory environments. Conclusions This paper proposes an Improved-Mamba-YOLO-T model for tomato ripeness detection in plant factory environments. By integrating CAFE-ODSSBlock, ASTF, Zoom_cat, and SDI modules, the model significantly enhances fine-grained feature extraction and multi-scale fusion capability, achieving higher detection accuracy and robustness under complex lighting conditions. Experiments show that Improved-Mamba-YOLO-T outperforms mainstream detection models in both accuracy and real-time performance, and can be deployed on edge computing platforms. Its combination with an intelligent spectral regulation system further demonstrates its application potential in precision agriculture, with possible extensions to growth monitoring, automated lighting adjustment, and related scenarios.

This study explores the application of the Stable Diffusion Models (SD) in generating Haipai New Year paintings and investigates the approaches and methods for constructing an intelligent creative platform. By leveraging digital intelligence technology, this research aims to promote the innovative development of Haipai New Year paintings. The study employs model fine-tuning and technological innovation by integrating the Low-Rank Adaptation (LoRA) and ControlNet models into the SD, optimizing the quality and style of generated images to ensure consistency with the Haipai New Year painting style. Based on this model, an interactive and participatory intelligent painting creation platform is designed and developed. The key innovation of this research lies in the integration of the Stable Diffusion model with LoRA and ControlNet, introducing a novel approach to model fine-tuning through parameter blending. The significance of this study lies in its breakthrough in the digital revitalization of Haipai New Year paintings by leveraging advanced artificial intelligence technologies. The findings highlight the potential of AI in enhancing the visibility and sustainability of traditional cultural heritage. Using Haipai New Year paintings as a case study, this research offers practical insights for future interdisciplinary studies in cultural preservation.

With the growing integration of deep learning and educational informatization, applying artificial intelligence to classroom behavior analysis has garnered significant attention. This article specifies 14 types of classroom behaviors and their classification criteria. By clipping and frame extraction from surveillance videos, target detection, manual annotation, temporal association, and other operations, a multi-label behavior dataset was created. This article also proposes a Weakly supervised fine-grained classification SlowFast SlowFast behavior recognition algorithm, which improves the accuracy of recognizing small difference classroom behaviors from an intra-class classification perspective. By using attention-guided local feature enhancement in the path, weakly supervised fine-grained classification of behavior target local features was achieved. Experimental results showed the algorithm improves behavior recognition accuracy by 4%-11% for specific behaviors and 5.75% overall, contributing to teaching quality evaluation systems.

IntroductionRegional citrate anticoagulation is a preferred option for renal replacement therapy in critically ill patients. However, current implementations ignore individual differences that may exist in the fluctuation of patients' ionized calcium levels. To address this problem, individualized citrate and calcium supplementation models were established based on the pharmacokinetic and clearance characteristics of citrate, and an automated regional citrate anticoagulation system was built with these models as its core to facilitate the treatment of clinical patients. This study was designed to preliminarily evaluate the safety and efficacy of this system, the SuperbMed (R) RCA-SP100 automated regional citrate anticoagulation system, in prolonged intermittent renal replacement therapy.MethodsSeven patients undergoing prolonged intermittent renal replacement therapy completed treatment with the SuperbMed (R) RCA-SP100 system. In vivo and in vitro ionized calcium levels were measured every hour before and after the start of dialysis. The accuracy and alarm sensitivity of the pumps were also monitored.ResultsDuring seven treatments, the average extracorporeal ionized calcium level was 0.34 +/- 0.02 mmol/L, and the mean ionized calcium level in vivo was 1.09 +/- 0.07 mmol/L. No patient required intervention, and there was no filter coagulation. The pumps all had an absolute accuracy less than 5%, and alarms could be triggered precisely.ConclusionsWe reported on an automated system that allows for individualized citrate and calcium supplementation in prolonged intermittent renal replacement therapy and enables the precise and secure implementation of regional citrate anticoagulation.

An industrial control process multiple-input multiple-output (MIMO) coupled system is analyzed in this study as an example of a Loss of Coolant Accident (LOCA) simulation system. Ordinary control algorithms can complete the steady state of the control system and even reduce the response time to some extent, but the entire system still consumes a large amount of energy after reaching the steady state. So a multivariable decoupled energy-saving control method is proposed, and a novel energy-saving function (economic function, Eco-Function) is specially designed based on the active disturbance rejection control algorithm. Simulations and LOCA simulation system tests show that the Eco-function algorithm can cope with the uncertainty of the multivariable system's internal parameters and external disturbances, and it can save up to 67% of energy consumption in maintaining the parameter steady state.

冷却剂丧失事故(loss of coolant accident,LOCA)试验是模拟核电设备在运行中突然出现冷却剂丧失而导致仓内温度骤然升高的场景,检验核电站用电缆、传感器等设备性能的一种标准化测试流程.为解决流入LOCA试验仓内高温蒸汽的温度控制具有非线性、迟滞大以及时变性的技术难题,使用 Fluent 软件模拟 LOCA试验仓中的温度变化和蒸汽流场,并用MATLAB仿真对比模糊专家PID与经典PID的差异.以设计基准事故(design basis accident,DBA)鉴定曲线CPR1000 和AP1000 为标准,在现场进行了瞬态热冲击试验、喷淋降温试验和温度平衡试验.通过仿真和现场试验结果表明,采用模糊专家PID联合控制的温控系统比经典PID控制响应速度更快、鲁棒性更高、超调量更小,误差控制在±5℃以内,符合DBA曲线的温控要求,能够满足LOCA温控系统的自动化测试需要.

In recent years, Chinese society has become increasingly ageing while the fertility rate continues to decline. This situation has led to a growing public demand for rehabilitation devices such as walking aids. A novel intelligent robotic walker named ReRobo Walker is proposed which can assist groups of elderly people with dysfunctional legs with rehabilitation training and indoor and outdoor walking. We design robust mechanical structures for robotic walker, install special 3D force sensor, 2D LIDAR and other sensors, and design new algorithms to enable intelligent functionality while guaranteeing the safety of robotic walker.(1) Real-time monitoring of the user’s physical status, such as falls, through laser range sensors and 3D force sensor; (2) Modelling of the scene through 2D LIDAR for path planning, obstacle avoidance and navigation functions; (3) Precise control through LADRC-based algorithm for uphill assistance, downhill control and prevention of sharp shifts of the robotic walker. Experiment results demonstrate the solid mechanical structure, stable reliability and the effectiveness of intelligent control algorithms of the intelligent robotic walker.

为帮助下肢受伤和出行不便的患者及老人进行康复训练或日常步行活动,设计了 一款基于FreeRTOS实时操作系统的智能助行器.本设计以STM32作为主控制器,搭配压力采集、超声波测距等外围电路.软件设计上,采用多任务的编程方式,将整个助行器系统的应用程序分为防跌倒任务、导纳控制任务、意图分析任务、压力信号采集任务、电机驱动控制任务这5个任务,并为各任务合理地分配优先级和系统资源以对助行器进行控制.经测试,所设计的智能助行器最终实现助行器柔顺控制、上坡助力、下坡控速以及主动安全防护,达到了设计目的.

Aiming at the problem of object model identification of modern industrial process control systems, a new closed-loop moment parameter identification online method based on the data of normal operation of the running system is proposed. In this method, only one step response data of the system is required, and appropriate convergence factors are introduced into the Laplace formula, the trapezoidal integral method is used to calculate the values of two derivatives of the transfer function, then the four unknown parameters of the second-order model can be solved by fitting the data with the least square method, and the target model can be identified. Finally, the simulation results of building different objects through Matlab show that the identification method has general applicability and good robustness with high recognition, and it is not sensitive to noise signals.

Traditional walkers are commonly used for the elderly in social life, which solves the basic problem of walking, but it is difficult to ensure safety when a fall occurs, and the human-computer interaction is poor. The image recognition method or the IMU sensor method fixed on the user, such as a wearable watch, is used by most of the current fall detection methods. Wearable sensors require the user's wearing operation, which is a little troublesome, and the detection accuracy is related to the way of wearing. The image recognition method requires a high-priced camera and a fixed installation position, which is unable to adapt to outdoor activities. We investigated a low-cost method of mounting the sensor on the body of the walker. We propose in this paper an improved fall detection method, namely Precondition and Limit Threshold SPRT (PLT-SPRT), and a novel fall detection system on the smart walker based on PLT-SPRT. The signals of the upper and lower limb sensors are fused based on the Kalman filter algorithm, and the admittance control parameters are obtained through the system identification method. In this study, the improved sequential probability ratio test algorithm is used to set the null hypothesis and the alternative hypothesis, construct the likelihood ratio and optimize the decision function, which is used to judge whether falls occur. The system is simulated by Matlab software, the user intention after fusion is more accurate, and the optimized decision function is judged accurately. Verified by the embedded system based on STM32 of the smart walker equipment in the real world, it can accurately identify the fallen state, with low detection delay, and the fallen state is detected about 160ms earlier than the traditional threshold-based detection algorithm, at the same time, the accuracy is higher than 94.9%, which meets the high real-time requirements of fall detection and is the ideal solution for smart walkers.

针对叉车式自动导引车(AGV)的车体结构和工作特点,提出了基于模糊控制的单目视觉和激光雷达导引AGV的运动控制方案.在单目视觉识别目标的基础上,远距离通过YOLOv5算法进行图像识别,导引AGV运动到目标前方的最佳工作位置;近距离采用模糊控制算法控制AGV,通过激光雷达精确限位,对AGV运动轨迹进行纠偏,完成对目标的插取与运送.测试结果表明:采用单目视觉远距离导引、激光雷达近距离纠偏与模糊策略控制相结合的方式可以很好地完成工作要求,所设计的控制系统工作横向距离误差为±3 cm,角度误差值为±5°,单次工作成功率大于95％,应用效果良好.

BACKGROUND:Regional citrate anticoagulation (RCA), a complex and effective technique, is recommended as the anticoagulation of choice for continuous renal replacement therapy. One of its key objectives is to keep the ionized calcium in the targeted range. In this study, we aimed to develop an automated RCA based on online monitoring of the ionized calcium concentration and closed-loop feedback.METHODS:We constructed calcium-selective electrodes with liquid inner contact, which measured a potentiometric signal as the output. We tested the responses, stability, and selectivity of the electrodes in flowing fluid containing calcium chloride. We compared the measurement accuracy between the electrodes and an i-STAT system in vivo. Moreover, we established closed-loop feedback using a proportional-integral-derivative controller model. We performed simulated automated RCA both in vivo and in vitro.RESULTS:The electrode gave a Nernstian response to the variation of ionized calcium concentration. It showed high stability and a relatively short response time. Changes in the fluid flow rate, solution PH, and addition of metal ions including Mg2+ and K+ did not interfere with the measurements of ionized calcium. These measurements in whole blood by the electrode were very close to those assessed by the i-STAT system. The feedback control system responded quickly to an abnormal ionized calcium concentration and regulated the infusion rates of calcium or citrate to maintain the concentration of ionized calcium within the targeted range.CONCLUSIONS:We successfully trialed automated RCA, which may help simplify the complexities of RCA in the future.

Deep neural networks (DNNs) have achieved excellent performance in image classification research, part of which is due to the large-scale training data with accurate annotations. However, it is expensive and time-consuming to collect such clean data. In contrast, when collecting a dataset by crawling from websites, noisy labels are ubiquitous, which makes it easy for deep neural networks (DNNs) to overfit noisy labels and cause performance degradation. Most recent efforts have been focused on defending noisy labels by roughly ignoring some samples with high losses, which are treated as noise or reweight the training data in the loss function. Both strategies inevitably have priori conditions, such as a clean validation set or a ground-truth noise transition matrix, which are impractical in real-world datasets. In this paper, we propose a novel end-to-end framework for noise correction, called End-to-end Correction with Mixup and Balance terms (ECMB). ECMB can completely correct noisy labels to true labels and keep the number of each class more balanced. This framework uses a backbone network that is pre-trained by using an improved Mixup entropy instead of the traditional cross entropy, and does not need any extra conditions. In addition, we introduce a new balance term that can update noisy labels more accurately. Compared with other state-of-the-art methods, the experimental results on publicly available CIFAR-10, CIFAR-100 and Clothing1M datasets demonstrate that our method has superior performance. (C) 2021 Elsevier B.V. All rights reserved.

The recent success of deep convolutional neural networks (CNNs) is mostly due to the availability of large-scale datasets with accurate annotations. However, the collection of such large datasets with clean annotations is time-consuming and not always feasible. In this paper, we propose a novel framework for learning with noisy labels, called the Cyclical training method with Joint loss and Co-teaching strategy net (CJC-net), where the net means our method is insensitive to the structures of the CNN. CJC-net pretrains two networks simultaneously and then performs the cyclical training strategy under an improved co-teaching method based on the two pretrained networks. During the training process, we adjust the learning rates of the two networks to make the network states peri-odically transfer from overfitting to underfitting. The cumulative loss of each sample under two networks is recorded; and the higher the cumulative loss of a sample is, the higher the probability of identifying it as a noisy label or a hard label. Then, we remove those samples with a high loss and fine-tune the two networks using the remaining data. The experimen-tal results on several datasets demonstrate that CJC-net is superior to many state-of-the-art methods. (c) 2021 Elsevier Inc. All rights reserved.

以移动机器人为研究对象,针对机器人系统模型的不确定性和非线性,提出一种反步运动学控制与径向基神经网络(RBFNN)最小参数学习法(MPLM)结合的滑模控制(SMC)算法.使用径向基神经网络逼近系统中的不确定项,消除系统中不确定因素的影响,增强系统的鲁棒性.针对径向基神经网络中权值难以快速调整的缺陷,采用最小参数学习法设计单个参数以代替神经网络中的权值,以此来优化算法复杂度,增强控制的实时性,通过Lyapunov稳定性判别方法,证明了系统在有限时间内的收敛性与稳定性.实验表明:该方法可使控制器的结构简化,鲁棒性、实时性显著提高,易于在实际应用中实现.

针对目前人体动作同步机器人造价高,使用过程繁琐,需要穿戴多种传感器设备等缺点,设计并研制了一种不需要穿戴传感器的人体手臂动作同步机器人.介绍了机器人的结构和系统组成,给出了Kinect读取和共享数据以及转换为机械臂可用数据的方法,设计PID结合PWM脉宽调速的控制驱动模块,完成了机械臂控制系统的软件.结果表明该人体动作同步机器人在未经培训的操作者指挥下,能够从不同初始位置完成人类肢体动作发出的指令,并且精度较高.这一研究可为无传感器肢体交互智能机器人研制提供参考.

为了提高外科手术机器人的工作空间和手术效率,缩小机器人的操作空间和结构尺寸,以适应针对不同病人的工作特点和手术环境.研制了一种新型的7自由度外科手术机器人(AOBO),设计了每个关节的动力机构以及专用的手术器械.肩关节机构处通过1个谐波减速器使得整体运动更加平稳,负载能力更大;大臂和小臂采用高强度低质量的碳纤维材质制造减少了由于臂长给电机增加的负载,同时建立了其运动学方程,仿真了机器人的工作空间.实物测试实验表明,该外科手术机器人具有合适的工作空间和很好的动作灵活性,能够满足外科手术对于机器人的准确定位和精确操作的要求.

This paper built a multi-joint manipulators model with a set of nonlinear location equations on a 6-DOF surgical robot. Further,it presented a calculation method of these nonlinear location equations based on Lyapunov stability theorem. With this method the operation of nonlinear location equations was simplified and simulated easily,which was benefit for high-speed manipulator. Experimental results show that the algorithm is suitable for solving multijoint manipulators nonlinear location equations,and it has a higher accuracy and convergence speed.