Applied Intelligence最新文献

Multivariate time-series (MTS) anomaly detection plays a crucial role in ensuring the reliable operation of industrial, transportation, and financial systems. However, the coexistence of long- and short-term temporal dynamics together with complex inter-variable interactions poses significant challenges for existing methods, particularly in feature-scale selection and spatio-temporal dependency modeling. To address these limitations, SAST is proposed as a scale-adaptive spatio-temporal modeling framework that provides a unified representation for temporal evolution modeling, structured variable dependency learning, and data-driven scale adaptation. SAST incorporates a temporal-aware Mixture-of-Experts (MoE) architecture composed of expert subnetworks with heterogeneous receptive fields, where temporal priors are leveraged to dynamically activate the most relevant expert combinations, enabling input-dependent multi-scale feature selection. Along the temporal dimension, SAST employs multi-scale patch partitioning and cross-patch attention to jointly capture short- and long-range temporal dependencies. Along the spatial dimension, a graph neural network guided by shared node and scale embeddings explicitly models multi-scale structural relations among variables. Extensive experiments on four public benchmark datasets show that SAST achieves superior accuracy and robustness over existing state-of-the-art methods, demonstrating its strong capability in multivariate time-series anomaly detection.

Chronic kidney disease (CKD) affects nearly 10% of the global population, where accurate pathological diagnosis and lesion grading are critical for personalized treatment. Although multiple stainings of consecutive sections provides complementary diagnostic cues, cross-staining representation remains constrained by structural misalignment and pronounced heterogeneity.To address these challenges, we propose a matching-driven, spatially enriched multiple instance learning (SEMIL) framework. First, a hybrid feature matching strategy is employed to establish spatial correspondences between pathological entities across differently stained consecutive slices, thereby capturing cross-slice structural associations. Second, a hierarchical entity–context descriptor is introduced to enhance cross-staining pathology feature representation, incorporating spatial enrichment into multiple instance learning. Furthermore, clinical embedding is integrated as prior knowledge to improve case-level feature representation for pathological diagnosis. We constructed two datasets for CKD diagnosis and kidney tissue damage grading. Extensive experiments demonstrate that SEMIL consistently outperforms standard multiple instance learning baselines, achieving gains of up to 4–5%. Visualization further confirms the effectiveness of entity-level spatial modeling. The proposed framework substantially improves MIL-based pathological diagnosis and tissue damage grading in CKD, while offering a generalizable paradigm for case-level pathology AI with potential applicability across other subspecialties.

Loop-closure detection is a fundamental challenge in simultaneous localization and mapping (SLAM). LiDAR sensors offer advantages, such as field-of-view (FOV) and robustness to perceptual variations, making them widely adopted for loop-closure detection tasks. However, the limited horizontal and vertical resolution of LiDAR reduces the robustness of lightweight loop-closure detection methods based on 3D point cloud histograms or bird-eye-view (BEV) representations. Moreover, conventional lightweight methods struggle to accurately recognize environments containing both large- and small-scale structures or those with uniform building heights. To overcome these limitations, this study proposes a line shape descriptor approach. The method constructs descriptors by characterizing the geometric shape and scale of each LiDAR line and evaluates similarity by computing the mean cosine distance between corresponding line descriptors across point clouds. To mitigate line drift caused by LiDAR motion or variations in the ground slope at loop-closure locations, a descriptor alignment technique based on a partitioned line sliding window is introduced. Furthermore, a two-stage search algorithm is incorporated to improve detection efficiency. Experimental evaluations on public and self-collected datasets demonstrate that the proposed method achieves significant improvements in loop-closure detection accuracy and robustness.

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder involving multiple pathological changes. MRI and FDG-PET imaging provide complementary structural and functional information, and their combined use significantly enhances prediction accuracy. However, current multimodal algorithms struggle to capture structural abnormalities across spatial scales in MRI and localized metabolic changes in FDG-PET. Furthermore, these methods often fail to account for the heterogeneity between modalities, leading to feature conflicts or information loss during the fusion process. To address these challenges, this paper proposes a multimodal AD prediction model comprising three key components: (1) Multi-Scale Context-Aware Network (MSCA): effectively captures diverse lesion information from MRI images, enhancing sensitivity to structural abnormalities. (2) Metabolic Abnormality Focus Network (MAFN): focuses on critical metabolic regions in FDG-PET images. (3) Cross-Modal Feature Constraint Fusion Module (CMFC): CMFC integrates intra-modal feature optimization and inter-modal dynamic interactions to adaptively balance and fuse features across both modalities. This design enhances the representation capability of the fused features for lesion regions. Experimental results demonstrate that the proposed model achieves a classification accuracy of 87.02% on the AD vs. MCI vs. NC task, outperforming existing AD prediction algorithms.

As global warming intensifies, extreme weather events are becoming increasingly frequent. Humanity is compelled to confront unprecedented challenges. There is a growing public clamor for transitioning to low-carbon living and sustainable production. This comprehensive survey focuses on traditional machine learning, advanced deep learning and latest federated learning methods for carbon emission, carbon footprint and carbon pricing prediction tasks. Firstly, a systematic overview of carbon emission prediction and management with some traditional machine learning methods are provided, including carbon emission prediction model, carbon market and carbon price prediction and industrial and energy optimization. Then, we conduct a comprehensive summary of some advanced deep learning methods for above tasks. In recent years, federated learning, as a distributed learning mode with security and privacy protection, has been widely used in some sensitive domains of data security. Hence, we further discuss about federated learning pattern for carbon emission, including green federated learning, carbon-efficient federated learning, energy-efficient federated learning methods. Finally, we propose some potential research problems as future directions from various aspects, including model interpretability, robustness and generalization across different regions and sectors, real-time prediction and adaptive learning systems, multi-source heterogeneous data fusion and optimization, data security and privacy protection.

While the current preference optimization methods for aligning large language models have demonstrated promising performance, extensive reliance on large amounts of manually annotated preference data presents significant barriers for broader application. The acquisition of annotated data is costly and inherently subjective, bringing challenges to the model’s fairness, robustness and trustworthiness. We propose a Self-Iterative Preference Alignment (SIPA) method, integrating both On-policy and Off-policy optimization strategies to reduce this reliance. We investigate the crucial role of self-optimization in model alignment by conducting experiments on a biased dataset and the evaluations are designed using the LLM-as-Judge framework. The experimental results indicate that our method improves the preference alignment performance of the baseline model effectively. Specifically, it achieves better performance with only ten percent human-annotated preference data, demonstrating its strong potential for application in low-resource scenarios. The data distribution shifts during iterative training are analyzed by visualizations, highlighting SIPA’s effectiveness in enhancing model alignment with human preferences.

To address the dual challenges of high parameter complexity and lack of interpretability in deep neural networks, this study proposes KAN-RCNN—a novel object detection framework based on the mathematical formulation of Kolmogorov-Arnold Networks (KANs). By integrating KANs with conventional CNN architectures, comparative experiments on the PASCAL VOC 2012 benchmark dataset demonstrate that KAN-RCNN achieves: 1) 13.6% parameter reduction compared to the original Faster R-CNN baseline; 2) 1.3% improvement in detection accuracy; 3) enhanced model interpretability. Through systematic validation with 1D synthetic signals, MNIST grayscale images, and multimodal data from PASCAL VOC 2012, the experimental results confirm that KAN-RCNN maintains competitive detection performance while attaining superior computational efficiency. This research provides new methodological insights for developing efficient and interpretable computer vision models.

Liquid sloshing in moving open containers poses significant risks in various industrial and engineering applications, often leading to spillage, contamination, and reduced operational safety. Effective control of sloshing is therefore critical for ensuring product integrity and preventing losses during transportation. This paper presents three novel Deep Reinforcement Learning (DRL)-based feedback control frameworks for automatic motion planning of an open cylindrical liquid container moving along a straight-line trajectory. The sloshing dynamics are modeled as a nonlinear underactuated system—specifically, a simple pendulum mounted on a moving cart—to capture the essential fluid-structure interaction while enabling control design in a simulation environment. Each proposed framework employs a DRL agent trained using the Deep Deterministic Policy Gradient (DDPG) algorithm to generate optimal control actions that minimize sloshing and reduce overall travel time. The agents are trained in a closed-loop feedback setting using the pendulum-cart model to ensure robustness and adaptability to dynamic disturbances induced by the sloshing liquid. The performance of the proposed DRL-based frameworks is rigorously evaluated and benchmarked against several conventional control strategies, including Super Twisting Control (STC), Linear Quadratic Regulator (LQR) and adaptive Sliding Mode Control (ASMC), under disturbance condition. Furthermore, to validate the practical applicability of the learned policies, the DRL-generated trajectories are tested in open-loop simulations using FLOW-3D computational fluid dynamics (CFD) software. This dual-layered validation approach demonstrates the effectiveness and robustness of the proposed methods in achieving efficient, spill-free transport in liquid handling systems.

Convolutional neural networks (CNNs) are a powerful tool for image-related applications due to their ability to learn features of images hierarchically. However, even after more than a decade, CNNs still present many challenges. Among these challenges, there is the arbitrary choice of parameters that makes the design of CNNs a difficult task. This work presents a new CNN model -SevenNet- for classifying tomato leaf diseases from the PlantVillage dataset. SevenNet’s Architecture has been built from scratch using a formulation extracted through extensive experimentation. SevenNet’s main advantages are the large number of extracted feature maps, fast convergence, and an overall reduced number of learnable parameters. A detailed study explored training the network on different data partitions, ranging from standard partition to cross-validation split in addition to other non-standard partition. Validation of SevenNet has been conducted against several state-of-the-art models, with all networks being trained from scratch. Obtained results were not only found to be outstanding and comparable to leading models, but SevenNet’s architecture demonstrated distinctive advantages, matching the performance of these established models. Notably, SevenNet’s convergence has been achieved more rapidly in terms of accuracy and loss. Additionally, the highest overall accuracy has been achieved when tested with an unusual partition (10% training, 10% validation, 80% test). The proposed CNNs were also found to be superior in terms of execution speed and convergence, solidifying SevenNet’s advantages over existing approaches.