In this paper, a fuzzy logic-based fault-tolerant attitude control strategy is proposed for the attitude tracking of a quadrotor unmanned aerial vehicle (UAV) subject to actuator faults. The attitude dynamics of the quadrotor are represented using modified Rodrigues parameters. Inspired by the biological trial-and-error mechanism that reinforcement learning (RL) emulates, the proposed method is developed by integrating fuzzy logic systems (FLSs) with RL. To enhance the autonomous learning capability and tracking performance of the UAV system, actor–critic (AC) learning is introduced as an effective RL method. A cost function defined in terms of tracking errors is introduced, and an FLS is incorporated into the critic to approximate the cost function for performance evaluation. The actor is responsible for generating the control input based on the critic signals. Concurrently, another FLS is employed to approximate system uncertainties and actuator bias faults. Furthermore, to meet increasingly stringent control requirements, performance constraints are imposed to guarantee prescribed tracking performance. The system stability and convergence of tracking errors are analyzed using Lyapunov stability theory. Finally, simulations are conducted to verify the effectiveness of the proposed adaptive fault-tolerant attitude control scheme.

Traditional malware classification methods heavily rely on extensive labeled data and single modal features, which limits their adaptability to evolving threats. In this paper, we propose an intelligent multimodal fusion framework that leverages complementary information from static and dynamic analysis for few-shot malware classification. Specifically, we convert malware binaries into grayscale images to capture static characteristics and extract application programming interface (API) call sequences to represent dynamic behaviors. To effectively integrate these heterogeneous modalities under limited data conditions, we introduce a lightweight graph neural network-based intelligent feature fusion module. This module segments modality-specific features, constructs a bipartite graph between segments, and performs cross-modal message passing to learn fine-grained correlations. The fused representations are then used in a prototypical network for few-shot classification. We construct two malware datasets augmented with multimodal features and conduct extensive experiments under few-shot settings. Results demonstrate that our approach significantly outperforms both unimodal baselines and naive fusion methods, achieving up to 95.73% accuracy in 5-way 5-shot classification. Ablation studies and efficiency analysis confirm that our fusion module adds minimal computational overhead while enhancing both accuracy and interpretability. This work highlights the potential of intelligent multimodal integration for robust malware classification with limited labeled data.

The significant deformations and nonlinear characteristics of soft robots pose substantial challenges for accurate modeling. Although various dynamic modeling methods for soft actuators have been explored, existing approaches have excessively long computation times, making them unsuitable for real-time control of soft actuators. To address these issues, this paper proposes an efficient dynamic modeling method for soft actuators. The core idea is to ensure model accuracy by integrating moment-curvature equation with the Lagrangian equation. Additionally, the dynamic model is simplified using Taylor expansion to enhance computational efficiency without compromising control accuracy. The model also accounts for the actuator’s gravity and the buoyancy effects of water on its motion. To validate the effectiveness of our proposed model, we performed dynamic model verification experiments in a laboratory setting. The experimental results indicate that the model achieves an error rate of less than 9.23%, with computation times ranging from 0.0094 to 0.015 s. This approach offers a new solution for real-time control of soft actuators.

Lithium-ion batteries are core components of renewable generation and energy-storage systems and are widely deployed in PV/wind grid scheduling and e-mobility. Accurate remaining useful life (RUL) prediction is essential for operational stability and cost control. This paper proposes a battery life prediction approach that integrates channel-grouping half-convolution (CGHF) and a monotonic decreasing head (MDH) within a Mamba-based sequence modeling framework. CGHF reduces computational redundancy while strengthening multi-scale temporal representations; the selective state-space module of Mamba efficiently captures long-range dependencies; MDH imposes an explicit “non-increasing capacity” constraint at the decoder to enhance robustness and interpretability. Experiments on the National Aeronautics and Space Administration (NASA) Randomized Battery Usage Dataset and the Tongji University (TJU) Commercial Lithium-Ion Battery Cycling Dataset demonstrate superior RUL accuracy, achieving minimum capacity-prediction mean absolute errors (MAEs) of 0.0081 and 0.0009 Ah, respectively, outperforming strong baselines under the same settings. The method improves accuracy while maintaining fast inference, suggesting potential applicability to online health monitoring and maintenance planning, subject to further validation under more diverse operating conditions.

Currently, there are many studies focusing on keypoint extraction, weight estimation, and pose assessment using point clouds from pig backs. But extracting point clouds from complex environments remains challenging, especially when the data collection environment, location and height are changing. This study addresses the challenge of extracting 3D point clouds of pig in complex environments with variable heights and severe interference from a bird’s-eye perspective by proposing a target extraction method based on a single Time-of-Flight (TOF) depth camera. A custom-designed pushcart-based data acquisition equipment was utilized to collect 987 point cloud datasets under diverse conditions, encompassing three height levels to accommodate various pig body sizes and postures. A dynamic algorithm, dynamic point-cloud feature focusing and segmentation (DPFFS), was developed, which integrates a point counting peak statistical filtering module and a multi-dimensional perceptual spatial filtering module to remove ground point clouds and other interfering noise. This way of extracting target does not require pigs to move through specific channels and is also transferable to the segmentation of multiple targets, and other animals, and varied scenes. Experimental results show that the pig body point clouds were highly consistent with the ground truth, as represented by the manual segmentation results, with an average intersection over union (IoU) of 0.984, considering the erroneous segmentation caused by clustering, the IoU is 0.836, mis-segmentation rate is 0.17. After voxel grid downsampling, the DPFFS algorithm achieved an average running time of 0.928 s. It can serve as a pre-processing module for point cloud target extraction in various application scenario, providing accurate preliminary results for tasks such as individual identification, body size measurement, and pig weight estimation, etc.

Data acquisition for river-lake avian species suffers from interference by long-distance imaging, water surface reflections and occlusions, producing low-quality images with motion blur, low contrast and annotation noise. Such defects severely degrade the accuracy of classification models. To address this problem, this paper proposes a processing algorithm for low-quality river-lake bird image data, termed the quality-driven unsupervised data curation and robust learning (QUC-RL) Method. The algorithm is divided into four key modules: First, a multi-dimensional quality-aware representation space for bird image data is constructed by fusing deep semantic features, texture features and four interpretable quality metrics, based on which a quality score is derived. Second, multi-strategy unsupervised auditing and curation for outlier and mislabelled sample localization is implemented with the constructed multi-strategy cluster selection mechanism. Third, the dataset is preliminarily reconstructed via an adaptive distribution preservation strategy based on the obtained quality scores and localization results. Finally, hard subsets within the reconstructed dataset are enhanced by a quality-conditional robust learning framework. Experimental results demonstrate that the overall performance is preserved and the stability of macro-average performance on low-quality subsets is improved, while the dataset scale is reduced by the proposed QUC-RL method. Classification accuracies of 96.580%, a Macro-F1 of 95.807% and a Grade C F1 of 96.362% are achieved by models trained on the reconstructed dataset.

Partial multi-label learning (PML) addresses challenges where each instance is associated with a set of candidate labels that includes both relevant and irrelevant ones. Traditional label disambiguation strategies often overlook the importance of nonlinear subspace structures. The assumption that data points closely adhere to multiple linear subspaces is restrictive and may not hold in certain applications. Linear subspace clustering algorithms frequently struggle with data that lie on multiple nonlinear manifolds, because they focus only on global linear relationships between data points. To address this gap, we propose a novel approach called label co-occurrence guided nonlinear disambiguation for partial multi-label learning (LCND). Specifically, we introduce a label weight-guided kernel low-rank representation to learn an instance affinity matrix in a nonlinear feature space, enabling effective identification of instances with complex nonlinear structures. Meanwhile, we design a weighted Jaccard distance to quantify label relevance by exploiting label frequency and co-occurrence information. By jointly optimizing instance-level and label-level affinity matrices, the proposed method effectively denoises labels under weak supervision. Extensive experimental results demonstrate that LCND significantly outperforms most state-of-the-art PML methods on the vast majority of benchmark datasets and evaluation metrics.