Applied Intelligence最新文献

Industrial knowledge management (KM) remains highly fragmented, with knowledge capture, structuring, and application often treated as isolated activities. Conventional approaches depend on static documentation and experience-based practices, which limit their responsiveness in dynamic, operational environments. In this paper, we design a comprehensive framework for the integration of artificial intelligence (AI) and augmented reality (AR) in industrial maintenance and diagnostics. The framework leverages AI techniques, including natural language processing (NLP) for extracting and structuring domain knowledge and machine learning (ML) models for predictive fault classification. These AI capabilities are seamlessly combined with AR technologies to deliver immersive, context-aware, in-situ task guidance, thereby enhancing decision-making, reducing downtime, and supporting efficient, knowledge-based maintenance processes. A case study is employed to demonstrate the framework’s feasibility by developing a prototype system deployed in a maintenance setting. The AI module clusters maintenance actions and predicts task categories, while the AR component renders step-by-step maintenance instructions anchored in the physical workspace. This integration enables the transition from unstructured, digital maintenance records to step-by-step in-situ repair guidance, reducing reliance on expert memory or static manuals. The results show that combining AI-driven text analytics with AR-based visualisation creates a cohesive knowledge workflow that improves operational efficiency. The proposed framework offers a scalable approach to embedding KM into frontline industrial routines, laying the foundation for more adaptive, technician-centred knowledge systems.

Detecting small infrared targets is a critical challenge in computer vision, focusing on identifying and localizing minimal-pixel-sized targets within infrared imagery. This task faces significant challenges due to the extremely small size of the targets, variability in target dimensions and shapes across diverse scenes, intricate background complexities, and instances of target occlusion. In this study, we introduce an enhanced U-Net model featuring three novel modules: the Residual Coordinate Attention Block (RCA-Block) integrates coordinate attention and residual structures to enhance feature representation; the Dynamic Context-aware Multi-scale Fusion Module (DCMFM) dynamically fuses multi-scale features based on target characteristics; and the Multi-dilation Depthwise Separable Convolutional Module (MDSCM) employs multi-dilation depthwise separable convolutions to capture spatial features across varying receptive fields. Experiments demonstrate that the proposed model synergizes multi-scale feature fusion and spatial information enhancement, achieving superior accuracy in small target detection, robust noise resistance.

Accurate traffic flow forecasting serves as a critical technical foundation for building intelligent transportation systems and enhancing road network operational efficiency. Traffic flow patterns exhibit significant multi-time-scale characteristics, encompassing both long-term macroscopic trends and short-term fluctuations. The dynamic interactions across these scales constitute a core challenge in traffic prediction. Although a few studies have considered multi-scale feature extraction, they often adopt fixed paradigms to process information from different temporal scales, failing to effectively model the dynamic correlation mechanisms among local temporal patterns at multiple scales as traffic conditions evolve. Furthermore, dynamic spatial dependency modeling based on graph convolution is often hampered by the over-smoothing of neighborhood information, which weakens the discriminative power of critical spatio-temporal features and limits the model’s adaptability to complex traffic scenarios. To address these issues, we propose a Spatio-Temporal Dual-Graph Fusion Network (STDGFN). The model first partitions traffic flow sequences based on multi-scale time windows, encoding each segment into a temporal pattern representation that captures local contextual information. It then employs stacked Spatio-Temporal Feature Extraction Blocks (STFEB) to capture complex spatio-temporal dependencies at varying granularities, including dynamic interactions among multi-scale temporal patterns and spatial correlations between nodes. To mitigate the over-smoothing caused by deep graph convolution, we design a Spatial-Aware Gated Pattern Enhancer (SAGPE) that integrates node-level spatial attributes into a gating mechanism, effectively reinforcing the spatial heterogeneity of each node. Experimental results demonstrate that STDGFN outperforms existing state-of-the-art baseline methods on five real-world traffic datasets.

Traffic accident prediction is a crucial component of Intelligent Transportation Systems (ITS), playing a vital role in reducing accident rates and enhancing urban traffic safety. However, accurately forecasting accident risk remains challenging due to the complex spatial–temporal dependencies and multiple external influencing factors. To address this issue, this study proposes a Region Feature Enhancement and Multi-view Diffusion Graph Convolutional Network (RFE-MDGCN) for traffic accident risk prediction. The model introduces a Region Feature Enhancement (RFE) module that employs channel-level and spatial-level attention mechanisms to capture spatial dependencies and highlight region-specific risk features. A Multi-view Diffusion Graph Convolutional Network is further designed to model similarity and semantic relationships among urban regions through a graph diffusion process, thereby enriching node representations with high-order neighbor information. To capture temporal dependencies, a Spatial–Temporal Fusion module integrating a Long Short-Term Memory–Transformer Encoder (LT Encoder) is developed to dynamically learn both short-term continuity and long-term periodicity of accident risk. Experimental results on two real-world datasets, New York City (NYC) and Chicago (ZJG), show that the proposed method achieves improvements of 11.45% in Recall and 16.10% in Mean Average Precision (MAP) over state-of-the-art baselines, demonstrating its robustness and effectiveness for traffic accident risk prediction.

With the acceleration of the global energy transition, photovoltaic (PV) power generation, as an essential component of clean energy, has become increasingly important in promoting green and low-carbon development and achieving sustainable energy goals. However, the time-series data generated by PV power generation systems often contains outliers caused by equipment failures or environmental changes. Previous studies have shown that reconstruction-based anomaly detection methods are deficient in simultaneously modeling the global dependencies and local dynamics of time-series data, and are susceptible to model overfitting. Against the background of rapid expansion in PV installed capacity and sustained growth in monitoring data, coupled with a scarcity of precise annotations, unsupervised anomaly detection in multivariate time series has become a critical issue for ensuring the safe and economical operation of large-scale PV plants. For this reason, an unsupervised anomaly detection model based on GTBAD (GVSAO-Transformer-BiLSTM for anomaly detection) is proposed in this article. The model introduces an improved Snow Ablation Optimization algorithm (GVSAO) during the offline training phase to optimize model parameters, thereby enhancing global search capability and improving the efficiency of parameter optimization. The encoder employs a Transformer module with positional encoding to capture global dependencies in the time series, and regularization techniques are incorporated at each layer to mitigate overfitting. During decoding, a BiLSTM module is used to further exploit local temporal patterns via bidirectional recurrent modeling, enabling comprehensive utilization of contextual information in PV data. Finally, the reconstructed sequence is generated through a fully connected output layer, and anomaly detection is performed by comparing the reconstruction error against a predefined threshold. In this paper, we conduct comparison experiments between the GTBAD model and existing anomaly detection methods on four public datasets and verify the effectiveness of each module through ablation experiments. In addition, this study incorporates attention visualization and variable-level attribution analysis to enhance model interpretability and improve its operational usability. The experimental results show the effectiveness and superiority of the GTBAD model in PV anomaly detection.

This study presents a deployment-oriented fish-counting pipeline for multivariate aquaculture environments, where illumination, turbidity, depth, and fish density jointly affect detection reliability. Using a dataset of 6,500 labeled 1080p images collected across Vietnamese aquaculture sites, we benchmark two unmodified Ultralytics detectors (YOLO11 and YOLOv8), quantify the contribution of scenario-informed preprocessing and augmentation, and export the selected model to ONNX for execution on heterogeneous hardware (GPU, CPU, and embedded edge devices). On the test set, YOLO11 achieves (:mA{P}_{50-95}=:0.993) and YOLOv8 achieves 0.995; under the operational constraint of (:mA{P}_{50-95}ge::0.99), YOLO11 is selected due to consistently lower latency. ONNX export accelerates YOLO11 inference by ~ 25% while preserving accuracy within (:varDelta:mA{P}_{50-95}le::0.001), and a Wilcoxon signed-rank test confirms that the latency advantage is statistically significant (p < 0.05). Field deployment in three fish ponds further demonstrates reliable counting and proof-of-concept alerting via short-horizon aggregation of frame-wise counts. Importantly, the contribution of this work is system-level: an environment-stratified evaluation protocol, a data-centric robustness analysis, and an ONNX-based “train once, deploy across devices” monitoring pipeline—rather than a new detection architecture.

Social robots (SRs) are increasingly expected to assist in healthcare, education, and companionship, thereby addressing the growing need for personalized and affordable health and social care. However, sustaining long-term user engagement remains a major challenge for SRs, largely due to their limited understanding of human mental states. Accordingly, we leverage a recently introduced mathematical dynamic model of human perception, cognition, and decision-making for behavioral control of SRs. By identifying the parameters of this model and deploying it within a model-based behavioral steering system, SRs can autonomously adapt their actions to evolving user mental states, enhancing long-term engagement and personalization. To achieve this, we introduce the first integration of a systems-theoretic cognitive model into a closed-loop predictive behavioral control framework for SRs, formulated as a constrained multi-objective optimization problem that enables transparent, cognition-aware adaptation. In experiments with (varvec{10}) participants interacting with a Nao robot across three chess puzzle sessions ((varvec{45}) to (varvec{90}) minutes each), the identified model achieved a mean squared error (MSE) of (varvec{0.067}) (i.e., (varvec{1.675%}) of the maximum possible MSE) in tracking beliefs, goals, and emotions of participants, and increased engagement by (varvec{16%}) ((varvec{p = 0.009})) compared to a model-free baseline. Post-interaction participant questionnaires further confirmed the perceived engagement and awareness of the model-based controller. Overall, the framework provides a practical pathway toward SRs that autonomously adapt to users in real time, sustain long-term engagement, and ultimately deliver more effective and personalized assistance in domains such as healthcare, education, and companionship.

Recommender systems assist users in navigating information-rich environments by delivering personalized content. While model-based collaborative filtering approaches, such as matrix factorization (MF) and graph neural networks (GNN), are widely adopted, the inherent uncertainty in user preferences and the sparsity of data can lead to unreliable predictions. Confidence estimation has emerged as a strategy to quantify prediction reliability, yet its integration remains unexplored in GNN-based models, and prior methods often degrade accuracy or suffer from convergence issues. This study benchmarks four prominent confidence-aware models—OrdRec, Confidence-aware Probabilistic Matrix Factorization, Confidence-aware Bayesian Probabilistic Matrix Factorization, and Lightweight Beta Distribution across three public datasets: Amazon Movies and TVs, Jester Joke, and Movie Lens. We evaluate these models in terms of rating accuracy, ranking quality, and confidence estimate quality. In addition, we propose a novel confidence-integrated model based on a deep graph attention network architecture. Experimental results reveal that while distribution-based confidence methods are highly sensitive to dataset characteristics and harm accuracy, the proposed method demonstrates consistent performance across all datasets and metrics, outperforming prior distribution-based models. Nevertheless, challenges remain in aligning confidence estimates with prediction error.

The malware classification task involves systematically categorizing malware based on its distinctive characteristics, behavior patterns, and functional attributes. Existing classification methods typically rely on machine learning techniques based on single features or unimodal image classification techniques, which are insufficient to address the current complex and diverse malware threats, making it difficult to fully capture and integrate the multidimensional characteristics exhibited by malware at different levels and their potential interrelationships. To this end, we propose a malware classification method based on multi-modal feature interaction, namely the Multi-modal Interaction Transformer with Cross-Attention (MIT-CA). This model leverages both BYTES files, which represent malware in hexadecimal format, and ASM files obtained from disassembly, converting them into comprehensible textual languages and grayscale images, respectively. By employing multi-modal learning, it aggregates information from multiple data sources, enabling the model to learn more comprehensive representations. During the multi-modal interaction, a transformer encoding structure based on Cross-Attention is used to facilitate information exchange between different modal features. This trick guides the model to learn malware features more comprehensively and allows for independent training of each modal feature. Extensive experiments conducted on malware classification datasets verify the effectiveness of the proposed model, and experimental results demonstrate the outstanding performance of our model in the malware classification task.