Expert Systems最新文献

Large-scale multi-objective optimization problems (LSMOPs) are characterised by concurrent optimization of multiple conflicting objectives and no fewer than 100 decision variables. They widely exist in the fields of practical engineering and scientific research. Over the past decade, many large-scale multi-objective evolutionary algorithms (LSMOEAs) have emerged to address LSMOPs. This paper systematically reviews and comprehensively analyzes the ideas, advantages, disadvantages, and latest developments of these LSMOEAs. Firstly, it introduces the relevant concepts of LSMOEAs. Then classify them into four categories: decision variable grouping-based LSMOEAs, non-grouping dimensionality reduction-based LSMOEAs, effective offspring generation-based LSMOEAs, and learning models-based LSMOEAs. It analyzes representative algorithms in each category, elaborating on their core strategies, advantages, and disadvantages. Finally, it explores the applications of LSMOEAs in computer vision, like tackling pixel-level correlation, high-resolution feature redundancy, dynamic target tracking, and complex visual modelling. This paper provides readers with a comprehensive and systematic overview of LSMOEAs, serving as a valuable reference for both researchers entering this field and practitioners seeking to select appropriate algorithms for practical problems.

Traditional imaging- and audio-based sensing systems often face issues with environmental interference, privacy, and security. Wi-Fi Channel State Information offers a non-invasive alternative for human behaviour sensing but lacks precision in full-body activity recognition. This study presents a sustainable edge computing system that integrates a 3D Convolutional Neural Network and a Bidirectional Gated Recurrent Unit (Bi-GRU) with attention for real-time human skeleton pose reconstruction. By aligning Wi-Fi CSI with Kinect-captured posture data, the system extracts spatial–temporal features to generate accurate 3D skeletal models. It accurately identifies trunk and posture by combining the precision of vision-based recognition with the non-invasive advantages of CSI-based sensing. Leveraging edge computing enhances energy efficiency and reduces cloud transmission needs, making it suitable for sustainable healthcare, smart homes, and remote monitoring in resource-limited settings.

In recent years, deep learning-based vision technologies have gradually become a research hotspot for concrete defect classification. However, existing concrete defect classification datasets generally suffer from issues such as blurry boundaries, labeling errors, limited data sources and a lack of diversity in defect types, which are insufficient to meet practical application requirements. At the same time, current methods still face challenges such as data scarcity, single-modal information and high computational resource demands. To address these issues, this paper proposes a novel concrete defect classification dataset: ConDef, and introduces an innovative multi-view feature enhancement method: MVFE-MC. This method combines large language models (LLMs) with deep learning techniques to achieve high-precision concrete defect classification in low-resource and few-shot environments. The ConDef dataset is constructed through multiple channels to ensure data diversity and high quality, and a triple-validation mechanism is adopted to ensure labeling accuracy. The MVFE-MC method introduces a Multi-View Feature Module and a Multimodal Classifier Module, effectively integrating visual and textual information to enhance the model's ability and robustness in fine-grained defect recognition. Experimental results show that MVFE-MC achieves state-of-the-art performance on two concrete defect classification datasets, Concrete Crack Image and ConDef, validating the effectiveness and superiority of the proposed method.

Image inpainting represents a fundamental task in computer vision, focusing primarily on the generation of missing content within an image to restore its integrity and aesthetics. Existing GAN-based approaches often produce content with ambiguity and require a high training difficulties. Moreover, they tend to focus narrowly on damaged regions, leading to edge distortions that hinder generalisation. To address these challenges, we propose an algorithm that consist of two distinct networks. The first network, called Edge-e4e, is designed for initial image restoration and integrates a pre-trained StyleGAN2 as the generator to mitigate edge distortions. This network employs an encoder-StyleGAN2 architecture, where only the encoder part is trained, thereby reducing training costs compared to traditional GAN methods. To resolve ambiguities in the restored content, we incorporate edge information into the damaged regions, guiding the network to generate content that is consistent with the original image. The second network, called Appending network, includes two style-based encoders and a generator to improve the similarity between the images restored by Edge-e4e and the original images. Specifically, we subtract the restored images from the input images in the channel dimension to obtain distortion maps, which serve as a prior to refine the restored images from Edge-e4e. To further enhance the quality of refined images, we propose incorporating plugin and modulate plugin modules for style extraction and fusion. These modules utilise information from the input images and seamlessly integrate it into the style-based generator. Experimental results demonstrate that our algorithm achieves high-fidelity restoration and excellent generalisation, with optimal FID and Lpips metrics of 0.0631 and 0.875, respectively. The code is publicly available at: https://github.com/MengZhen-Chi/Edge-Pries-Image-Inpainting-with-StyleGAN2.

Deep learning models have achieved significant success in time series forecasting, substantially improving predictive accuracy across several applications. The interpretability of these intricate models continues to pose a significant problem, especially in high-stakes fields where transparency and trust are essential. Current explanation strategies are often limited to static assessments, lacking consistency and semantic significance into model behaviour. This article introduces SUE-TS (Surrogate-based Universal Explanation for Time Series), a versatile interpretability framework that develops a simple and interpretable surrogate to replicate the predictive behaviour of opaque forecasting models, therefore addressing existing restrictions. SUE-TS incorporates a SHAP-based closed-loop feedback mechanism that enhances prediction accuracy and semantic consistency within the explanation domain. The methodology guarantees fidelity via dual-consistency evaluation: predictive consistency, which evaluates numerical concordance between surrogate and black-box models, and explanation-space consistency, which confirms semantic coherence through high-dimensional representation analysis. Comprehensive studies on seven benchmark time series datasets and six advanced forecasting systems reveal that SUE-TS consistently attains high approximation fidelity, interpretability, and robustness. The surrogate models reproduce output behaviours and maintain the reasoning logic of the original models, even in complex temporal dynamics. SUE-TS offers a scalable, model-agnostic solution for reliable time series forecasting by reconciling performance with explainability. It provides a significant resource for enhancing interpretability in actual AI applications, facilitating downstream activities such as model audits, knowledge distillation, and decision-support system integration.

Flight trajectory prediction is crucial for enabling the air traffic management system to anticipate future flight conflicts, allowing for the timely adoption of precautionary traffic control measures. To improve the prediction accuracy, some leading-edge approaches leverage multi-agent social neural networks or spatio-temporal graph neural networks to mine the interactions among flights. However, these methods typically focus on learning the temporal states of individual flights driven by multi-flight interactions, while underexploring the dynamic evolution of correlations among different or same flights at varying positions or times, resulting in suboptimal prediction performance. To address this gap, we propose a dynamic spatio-temporal graph attention neural network (DST-GAT) with an encoder-decoder structure. DST-GAT first draws on trajectory pattern mining to outline the multi-flight interaction range, then respectively employs two types of spatio-temporal graph edge encoders to capture the spatio-temporal dynamics of multi-flight interactions as well as the hyper-temporal dynamics of individual flight flying trends. Extensive experiments on Beijing Terminal Manoeuvring Area operation data demonstrate that DST-GAT outperforms state-of-the-art methods in terms of prediction accuracy, highlighting its promising ability to exploit the intricate spatio-temporal correlations among air traffic.

More and more companies have realised that implementing a joint replenishment and delivery (JRD) strategy can lead to significant cost savings. This paper presents a practical JRD problem for heterogeneous products, taking into account resource constraints. An enhanced hiking optimization technique called BHHOA is proposed. BHHOA incorporates a bound heuristic to refine the delivery frequency boundaries and introduces two new population generation methods based on the differential evolution algorithm and attention mechanism. Experimental results demonstrate that BHHOA outperforms six other algorithms with a lower average total cost. The JRD model presented in this study is effectively solved using the BHHOA algorithm. This study provides a practical technical method for companies to implement the JRD strategy.