Visual autoregressive (VAR) models have recently emerged as a promising family of generative models, enabling a wide range of downstream vision tasks such as text-guided image editing. By shifting the editing paradigm from noise manipulation in diffusion-based methods to token-level operations, VAR-based approaches achieve better background preservation and significantly faster inference. However, existing VAR-based editing methods still face two key challenges: accurately localizing editable tokens and maintaining structural consistency in the edited results. In this work, we propose a novel text-guided image editing framework rooted in an analysis of intermediate feature distributions within VAR models. First, we introduce a coarse-to-fine token localization strategy that can refine editable regions, balancing editing fidelity and background preservation. Second, we analyze the intermediate representations of VAR models and identify structure-related features, by which we design a simple yet effective feature injection mechanism to enhance structural consistency between the edited and source images. Third, we develop a reinforcement learning-based adaptive feature injection scheme that automatically learns scale- and layer-specific injection ratios to jointly optimize editing fidelity and structure preservation. Extensive experiments demonstrate that our method achieves superior structural consistency and editing quality compared with state-of-the-art approaches, across both local and global editing scenarios.
Previously, we proposed a power-free isokinetic training robot designed to provide resistive isokinetic training for knee injury patients during advanced-stage rehabilitation. However, patients in the early stage often lack sufficient muscle strength and necessitate assistive support. To address this limitation, this study introduces a hybrid assistive-resistive isokinetic training robot that integrates active assistance for early-stage knee rehabilitation and power-free resistive training for advanced stages. The system features a compact mechanical design and a reconfigurable control circuit capable of dynamically switching among three modes: active, passive (regeneration), and passive (consumption). Ten healthy subjects and ten knee-injury patients participated in the experimental validation. The results confirmed the adaptability of the system across multiple rehabilitation stages. These findings demonstrate the feasibility of the hybrid assistive-resistive isokinetic training robot and highlight the potential of the system for both clinical application and home-based rehabilitation. Future work will focus on extending the system to multi-joint training and enhancing control algorithms for broader patient populations.
Differentiating benign and malignant pulmonary nodules on CT suffers from frequent missed tiny lesions and severe background interference. To balance detection accuracy and lightweight deployment, this paper proposes a novel algorithm named FDC-DETR by improving RT-DETR. The backbone adopts the Fourier convolution-enhanced C2f module to model long-range global dependencies. The dynamic entropy-aware attention module eliminates redundant computation and boosts feature learning of micro nodules. The context-refined attention fusion module suppresses disturbances from blood vessels and bronchi. On LUNA16 dataset, FDC-DETR reaches 88.9% mAP@0.5 with merely 16.57M parameters and raises mAP@0.5 by 3.2% against baseline RT-DETR. It achieves outstanding generalization on Tianchi dataset and meets the demands of lightweight real-time clinical screening.
Traditional online monitoring systems often rely on high-quality data samples or external expert knowledge, and their model accuracy degrades significantly when equipment states deviate from expected conditions over time. In the preresearch stage of mechanical equipment, researchers face not only data scarcity and severe class imbalance but also performance degradation caused by nonstationary processes and varying operational conditions. To address these challenges, this paper proposes a novel fusion-enhanced model based on local neighborhood standardization and recursive quadratic kernel entropy component analysis. The proposed LNS-RQKECA framework constructs a data-driven vibration signal analysis model for early fault detection in mechanical equipment under multiple load conditions, relying solely on historical data from normal operation without requiring external experience or prior knowledge. The local neighborhood standardization technique transforms the multiload problem into a single-modal problem, enabling effective load-case fusion. To counteract the feature dilution effect induced by sample fusion, the proposed quadratic kernel entropy component analysis employs advanced kernel-based entropy optimization to enhance feature discriminability through deeper data mining. Additionally, a recursive updating mechanism ensures adaptive model adjustment, allowing the system to handle nonstationary load conditions dynamically. This proposed model is validated using both public datasets and real-world operational data. Experimental results demonstrate that under multiload conditions, the proposed model achieves strong generalization capability and nonstationary process adaptability with minimal training data. Compared with the traditional method, the false missing rate and false alarm rate are reduced by 12% and 15%, respectively.
Electrochemical impedance spectroscopy (EIS) is widely applied for corrosion characterization of pipeline steel under complex environmental conditions. However, conventional EIS analysis relies heavily on expert experience for equivalent circuit identification and initial parameter estimation, which limits its scalability and repeatability in long-term industrial monitoring. To address this challenge, this study proposes an engineering-oriented intelligent analysis framework based on random forest models to enable automated equivalent circuit identification and parameter initialization. A complete workflow is established, including data preprocessing, model construction, and post-processing. In the preprocessing stage, a physics-informed multi-strategy normalization scheme is designed to handle the heterogeneous and multi-scale characteristics of EIS features while preserving the intrinsic physical meanings of different impedance components. Within the proposed framework, a random forest classifier is employed to automatically identify equivalent circuit types, followed by circuit-specific multi-output random forest regressors to predict initial values of circuit parameters, thereby enabling stable and efficient nonlinear fitting. The proposed method is validated using atmospheric EIS data of pipeline steel and demonstrates robust performance in both circuit classification and parameter regression tasks, with consistent improvements over conventional machine learning baseline methods. Experimental results show that the proposed framework effectively reduces reliance on expert intervention while maintaining high accuracy, robustness, and interpretability. By integrating data-driven modeling with electrochemical physical insights, this work provides a practical and reliable solution for automated EIS analysis in industrial corrosion monitoring applications.
