Applied Intelligence最新文献

With the deep integration of 5G and IoT in urban rail transit, service traffic grows explosively and accurate classification of heterogeneous service flows is essential for safe and efficient railway operations. Urban rail communication systems are subject not only to operational disturbances such as equipment failures and maintenance interference, but also to complex transmission patterns characterized by the interleaving of multi-service traffic flows. Under these operating conditions, the widespread deployment of proprietary protocols and the high prevalence of encrypted traffic further diminish the applicability of conventional port-based and Deep Packet Inspection (DPI) classification methods. To address these challenges, we propose a priority-aware dual-channel feature fusion framework. Raw traffic bytes and statistical features are mapped into grayscale images and processed by a dual-branch architecture: a transfer learning-enhanced ResNet extracts fine-grained byte-level textures, while a lightweight CNN captures macroscopic statistical patterns. A channel attention mechanism dynamically recalibrates cross-modal features, and a novel Priority-Sensitive Loss (PSL) that integrates business-criticality awareness with class-balance weighting to maximize recall for safety-critical services. Evaluated on a real-world urban rail dataset using priority-weighted metrics, the method achieves 98.74% accuracy and 99.24% weighted recall, with recall of 99.94% and 99.41% on the two safety-critical services: Communication-Based Train Control (CBTC) and Emergency Radio Dispatch (ERD), providing a reliable classification foundation for priority-aware resource scheduling in urban rail communications. With only 0.18M parameters and 0.4–0.7 ms end-to-end latency, offering an excellent balance between high-precision classification, low inference latency, and edge-deployment feasibility.

Accurate patient stratification is critical for personalized radiation therapy in head-and-neck cancer (HNC), yet remains challenging due to patient and disease heterogeneity, high-dimensional feature spaces, and limited samples for certain clinical endpoints. To address these challenges, we developed and evaluated a novel multi-task, attention-based deep learning framework for jointly predicting multiple survival endpoints, including Overall Survival (OS), recurrence-free survivals including local failure-free survival (LFFS), regional failure-free survival (RFFS) and distant failure-free survival (DFFS). The proposed DeepCox_MTL framework consists of three key components: (a) a Multi-Modal Feature Extraction (MMFE) module, which integrates low-level semantic features radiomic features with high-level deep imaging representations extracted via an attention-guided global attention module, together clinical variables; (b) a Multi-Stage Feature Selection (MSFS) module, which employs a novel feature selection strategy that integrates a regularized fitness function into an Enhanced Genetic Algorithm (EGA) to identify a compact and informative predictor set; (c) a Multi-Endpoint Survival Prediction (MESP) module. This module leverages a shared backbone network to capture complex non-linear representations across related survival endpoints, while task-specific output heads are optimized using an adaptive weighting loss function that dynamically balances learning based on individual task performance. The framework was validated on a publicly available dataset of 606 patients with oropharyngeal cancer, and demonstrated strong and consistent performance. Specifically, it achieved superior predictions for well-represented endpoints, with C-indices of 0.7770 for OS, 0.8018 for LFFS, as well as notable performance gains for sparse endpoints, with C-indices of 0.7507 for RFFS, 0.7179 for DFFS. In summary, by leveraging joint low-level semantic radiomic features and high-level deep learning embeddings, and clinical variables within an adaptive multi-task learning framework, DeepCox_MTL effectively captures inter-task relationships to deliver more comprehensive and accurate multi-endpoint survival predictions. These results underscore its strong potential for clinical translation, particularly in complex and imbalanced datasets.

Human Activity Recognition (HAR) has become increasingly relevant due to the proliferation of sensing devices and their applications in healthcare, smart environments, and human–machine interaction, yet building reliable systems remains challenging in real-world conditions, particularly when integrating wearable and ambient data in intelligent environments. Recent approaches have explored multimodal learning to fuse heterogeneous sensor data, federated learning to preserve data privacy, and continual learning to incrementally adapt to new activities and contexts. This work presents a focused mini-review on the intersection of Multimodal Learning, Federated Learning, and Continual Learning applied to HAR, emphasizing how these paradigms can complement each other to enable more realistic, scalable, and privacy-preserving systems. The review synthesizes recent advances across these domains, highlighting fusion strategies, personalization methods, incremental learning techniques, and federated protocols. The study surveys state-of-the-art approaches, analyzes common challenges, and categorizes contributions along architectural, algorithmic, and deployment axes. The works reviewed indicate an increasing trend toward adaptive fusion mechanisms, generative replay, local personalization, and communication-efficient optimization as means to mitigate catastrophic forgetting, data heterogeneity, and modality imbalance. Finally, this review outlines research opportunities for multimodal federated continual HAR, identifies unresolved trade-offs among privacy, adaptability, and scalability, and suggests future directions, such as standardized benchmarks and interdisciplinary collaboration, to bridge laboratory settings and real-world deployment.

With rapid advances in research on salient object detection in remote sensing images (RSI-SOD), one critical open question for researchers is how to enhance model performance to tackle cross-architecture information interaction challenges, which is vital for effectively extracting information from objects with irregular shapes and complex backgrounds. Thus, this paper accordingly proposes a Sine Function-guided Master–Slave Collaborative Network, which adopts an adaptive sine bridging mechanism and a master–slave architecture to achieve efficient information interaction between the Transformer and CNN branches. First, a master–slave collaborative encoder is designed, where the Transformer encoder serves as the master to model global information, and the CNN encoder with an adaptive weight less than 1 serves as the slave to supplement local information. A sine function modulation term is introduced to guide the outputs of the slave encoder, thereby alleviating feature redundancy between the two encoders. Second, on the master branch, a multi-scale difference–local extraction module (MDLM) is designed to locate salient regions and refine details, and a bidirectional weighted-guided fusion (BWGF) module is developed to guide cross-level feature fusion using fused information as weights. Finally, on the slave branch, a directional frequency attention module (DFAM) is designed to enhance the horizontal and vertical directional information of features, enabling effective modeling of irregular global topologies. Extensive experiments on three benchmark datasets demonstrate that the proposed method significantly outperforms 21 state-of-the-art approaches.

Video anomaly detection aims to identify events that deviate from normal behavior patterns in surveillance videos. Since real anomalies are rare, diverse, and difficult to enumerate, this paper considers the unsupervised one-class setting in which only normal videos are available for training. To improve temporal representation, constrain the reconstruction of unseen abnormal patterns, and reduce background interference in frame-level scoring, we propose a spatiotemporal reverse distillation framework with memory-augmented pseudo-anomaly learning. A pretrained Inflated 3D ConvNet (I3D) is employed as a frozen teacher to extract multi-scale spatiotemporal features. These heterogeneous features are fused and compressed by a three-dimensional one-class bottleneck embedding module, followed by sparse memory retrieval and a three-dimensional student decoder that reconstructs the hierarchical teacher representations. During training, temporally shuffled clips generated exclusively from normal videos are used as pseudo-anomalies, and a margin-based objective encourages them to maintain sufficiently large teacher–student discrepancies. During inference, multi-scale cosine-discrepancy maps and an MSE-based discrepancy map are fused and converted into frame-level anomaly scores through foreground-guided spatial aggregation, overlapping-clip temporal aggregation, and Gaussian smoothing. To avoid test-set-dependent configuration selection, configuration search is performed using only training-derived validation data. After configuration freezing, the final models are retrained from scratch on the complete official normal training sets and evaluated on the benchmark test sets only after training is completed. Across five independent runs, the proposed method achieves micro frame-level AUCs of (varvec{96.12 pm 0.47%}), (varvec{91.06 pm 0.41%}), and (varvec{74.62 pm 0.38%}) on UCSD Ped2, CUHK Avenue, and ShanghaiTech, respectively. Additional experiments with two-dimensional Reverse Distillation, bottleneck removal, teacher adaptation, alternative pseudo-anomaly strategies, foreground-scoring ablations, robustness evaluation, and parameter sensitivity further analyze the behavior of the proposed framework. Code is available at: https://github.com/SVIL2024/ST-RD-VAD.

Mining top-k high-utility itemsets (THUI) is one of the focuses of intense research interest, and the main task is to find k high-value elements within a transactional dataset. Usually, high-utility items are characterized not by one dimension but by the combination effect of multiple dimensions. However, in traditional algorithms, factors are considered separately, making it impossible to mine items with multiple properties. Therefore, in this study, we presented a novel algorithm to mine top-k high-threat-and-frequency event sets simultaneously. Concretely speaking, inspired by the general utility mining model (GUMM), we first introduced a linear function to combine the effects of threat and frequency to evaluate events’ importance, built an enumeration tree to list promising candidates, and proposed pruning algorithms based on THUI to improve mining efficiency. It is worth noting that our TTFE algorithm can be easily extended to mine event sets with multiple external features (more than 2). Experimental results demonstrate that TTFE can achieve sufficient performance on mining the top-k high value items in eight transaction datasets, whilst generally outperforming time consumption by 63.72%, memory consumption by 71.32%, and minimum threshold by 171.56%, compared to baseline algorithm THUFI, the extension version of the state-of-the-art THUI. The TTFE algorithm also outperforms in these three metrics by 85.93%, 89.82%, and 24.24%, respectively, compared to the GUMM algorithm. Additionally, by adjusting the weight parameters (alpha ) and (beta ), our algorithm can flexibly degenerate to known top-k utility mining algorithms. Notably, it can be used to discover the top-k important events among service logs to capture security risks of Industrial Internet of Things (IIoT).

Gas pipeline accidents can cause severe consequences and are influenced not only by explicit causes that directly trigger accidents but also by latent contributing factors. Although such risk factors may not directly trigger accidents and are difficult to identify, they may be closely associated with explicit causes, accident development, and consequence severity. However, existing studies have mainly focused on directly identifiable explicit causes, with limited attention to latent contributing factors and their associations with explicit causes. Based on 1624 gas accident records from the U.S. Pipeline and Hazardous Materials Safety Administration, a BERT-based conditional probability model is developed for risk-factor identification. With the assistance of large language models, domain knowledge is incorporated into the network to identify explicit causes and latent contributing factors and estimate the probability of latent contributing factors conditional on specific explicit causes. By integrating these probabilities with accident causes, temporal characteristics, and consequence severity, a Gas Leak Accident Knowledge Graph (GLAKG) is constructed. Probability-weighted topological indicators are then derived from the GLAKG to quantify associations among accident factors and their risk implications. Experimental results show that the method achieves precisions of 0.9933 and 0.7034 and F1 scores of 0.9861 and 0.7245 for explicit cause and latent contributing factor classification, respectively. The framework integrates factor probabilities and consequence information to identify key explicit causes and latent contributing factors. The resulting risk associations help safety officers prioritize maintenance, hazard inspection, and prevention, while supporting accident investigation by identifying explicit and latent contributing factors.

Rapid prediction of scattered pressure fields is essential for acoustic metasurface design. Although the finite element method produces high-fidelity simulations, its high computational cost renders the exploration of large design spaces impractical. Deep learning models offer an efficient alternative, but existing approaches often struggle to reconstruct localized high-gradient structures. To overcome this challenge, we propose an Attention-Driven Gradient Balancing Operator Network (ADGB-ONet). ADGB-ONet adopts a progressive representation strategy comprising spectral enrichment, condition-dependent feature interactions, and spatially adaptive reconstruction. The Fourier Embedding layer first maps the raw spatial coordinates into multi-scale sine and cosine bases, thereby enriching the coordinate representation and enhancing the ability of the trunk network to resolve rapid spatial variations in scattered pressure fields. The Multi-Branch Attention Mechanism (MBAM) then couples the enriched spatial features with frequency and geometric descriptors and replaces fixed fusion rules with condition-guided feature interactions. The Adaptive Attention Reconstruction Module (AARM) subsequently reconstructs the modulated features through multiple branches with different transformation scales. A Gate Neural Network (Gate NN) generates coordinate-dependent weights to adaptively aggregate the branch outputs, enabling spatially adaptive reconstruction across regions with heterogeneous variations. This learned coordinate-dependent weighting allows the model to adjust reconstruction sensitivity without requiring explicit gradient information. Linear attention further captures spatial dependencies among neighbouring sampling points. Consequently, these components collectively enhance the reconstruction of localized high-gradient structures in scattered pressure fields. Experimental results show that ADGB-ONet achieves the lowest prediction error among the compared models, with a mean relative error (MRE) of 2.88%. This corresponds to a 23.4% reduction compared with OFormer, the strongest baseline in terms of MRE, and 58.0% reduction compared with the baseline MIONet. Ablation studies further verify the effectiveness of the proposed architectural design, with AARM providing the largest individual performance gain. These results establish ADGB-ONet as a reliable surrogate model to accelerate acoustic metasurface design.

Prediction of drug resistance in Mycobacterium tuberculosis is often hindered by complex variant interactions and variable sequencing quality. We present the Interpretable Variant-Aware Multi-Path Network (VAMP-Net), a novel architecture addressing these challenges through a dual-pathway approach. Path-1 utilizes a Set Attention Transformer to model permutation-invariant variant sets and capture epistatic dependencies, while Path-2 employs a 1D-CNN to analyze VCF quality metrics for adaptive confidence scoring. Evaluated on four critical anti-TB drugs, VAMP-Net significantly outperforms baseline CNN and MLP models, achieving accuracies > 95% and AUCs (sim )0.97 for Rifampicin and Rifabutin. Feature attribution analysis via Integrated Gradients successfully recovered canonical targets (rpoB, embB, katG) and discovered high-impact novel loci. Functional enrichment confirmed these novel variants constitute non-random metabolic modules (p=0.00239) centered on cell-wall remodeling. Furthermore, systematic ablation of the Quality-Aware pathway demonstrates that the model performs a learned “integrated audit,” prioritizing the Fraction of Supporting Reads and relative confidence over raw depth to mitigate technical noise. This dual-layer interpretability, bridging genomic pathogenicity with technical reliability, establishes a new paradigm for robust, auditable, and clinically actionable resistance prediction, positioning VAMP-Net as an important tool for both diagnostic classification and mechanistic discovery in clinical genomics.