Expert Systems with Applications最新文献_第8页

HDFNet:Hybrid-domain fusion network for medical image restoration 用于医学图像恢复的混合域融合网络HDFNet

IF 7.5 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Expert Systems with Applications

Pub Date : 2026-01-23 DOI: 10.1016/j.eswa.2026.131268

Yuqi Liu , Liqun Lin , Shunzhou Wang , Si Chen , Chao Zeng , Nanfeng Jiang , Da-Han Wang

Medical Image Restoration (MIR) represents a classic yet challenging task in computer vision and image processing, aiming to reconstruct High-Quality (HQ) medical images from their Low-Quality (LQ) counterparts degraded by factors such as low-dose acquisition, limited resolution, or modality-specific noise. The diversity of degradation types complicates the design of MIR models that can be applied across different modalities and restoration tasks. However, existing MIR approaches mainly focus on designing specialized network architectures with limited ability to generalize across multiple MIR scenarios. To address this issue, we first analyze diverse medical image degradations from a unified frequency-domain perspective. Building on this insight, we propose the Hybrid-Domain Fusion Network (HDFNet), an efficient hybrid-domain framework that integrates spatial and frequency priors for CT and MRI restoration. Specifically, the proposed HDFNet adopts a dual-domain hybrid structure that performs multi-scale receptive field modeling in both spatial and frequency domains: the spatial domain preserves local anatomical structures, while the frequency domain enhances global consistency and suppresses artifacts via Fast Fourier Transform (FFT)-based operations. In particular, we introduce frequency-aware Harmonic Positional Encoding (HPE) and Frequency Adaptive Convolution (FAC) to extract rich semantic frequency features tailored to different degradation types. Extensive experiments on Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) restoration tasks demonstrate that the proposed HDFNet provides a robust and efficient solution for CT and MRI restoration, and shows promising generalization to broader MIR tasks.

医学图像恢复（MIR）是计算机视觉和图像处理领域的一项经典但具有挑战性的任务，旨在从因低剂量采集、有限分辨率或模态特定噪声等因素而退化的低质量（LQ）医学图像中重建高质量（HQ）医学图像。降解类型的多样性使MIR模型的设计复杂化，这些模型可以应用于不同的模式和恢复任务。然而，现有的MIR方法主要侧重于设计专门的网络架构，在多个MIR场景中泛化的能力有限。为了解决这个问题，我们首先从统一的频域角度分析了不同的医学图像退化。基于这一见解，我们提出了混合域融合网络（HDFNet），这是一种高效的混合域框架，集成了CT和MRI恢复的空间和频率先验。具体来说，提出的HDFNet采用双域混合结构，在空间和频域进行多尺度感受野建模：空间域保留局部解剖结构，而频域增强全局一致性，并通过基于快速傅里叶变换（FFT）的操作抑制伪影。特别是，我们引入了频率感知谐波位置编码（HPE）和频率自适应卷积（FAC）来提取针对不同退化类型的丰富语义频率特征。对磁共振成像（MRI）和计算机断层扫描（CT）恢复任务的大量实验表明，所提出的HDFNet为CT和MRI恢复提供了一种鲁棒和高效的解决方案，并有望推广到更广泛的MIR任务。

{"title":"HDFNet:Hybrid-domain fusion network for medical image restoration","authors":"Yuqi Liu , Liqun Lin , Shunzhou Wang , Si Chen , Chao Zeng , Nanfeng Jiang , Da-Han Wang","doi":"10.1016/j.eswa.2026.131268","DOIUrl":"10.1016/j.eswa.2026.131268","url":null,"abstract":"<div><div>Medical Image Restoration (MIR) represents a classic yet challenging task in computer vision and image processing, aiming to reconstruct High-Quality (HQ) medical images from their Low-Quality (LQ) counterparts degraded by factors such as low-dose acquisition, limited resolution, or modality-specific noise. The diversity of degradation types complicates the design of MIR models that can be applied across different modalities and restoration tasks. However, existing MIR approaches mainly focus on designing specialized network architectures with limited ability to generalize across multiple MIR scenarios. To address this issue, we first analyze diverse medical image degradations from a unified frequency-domain perspective. Building on this insight, we propose the Hybrid-Domain Fusion Network (HDFNet), an efficient hybrid-domain framework that integrates spatial and frequency priors for CT and MRI restoration. Specifically, the proposed HDFNet adopts a dual-domain hybrid structure that performs multi-scale receptive field modeling in both spatial and frequency domains: the spatial domain preserves local anatomical structures, while the frequency domain enhances global consistency and suppresses artifacts via Fast Fourier Transform (FFT)-based operations. In particular, we introduce frequency-aware Harmonic Positional Encoding (HPE) and Frequency Adaptive Convolution (FAC) to extract rich semantic frequency features tailored to different degradation types. Extensive experiments on Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) restoration tasks demonstrate that the proposed HDFNet provides a robust and efficient solution for CT and MRI restoration, and shows promising generalization to broader MIR tasks.</div></div>","PeriodicalId":50461,"journal":{"name":"Expert Systems with Applications","volume":"310 ","pages":"Article 131268"},"PeriodicalIF":7.5,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080787","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Multi-agent role-playing by LLMs and LMMs: An explainable open-world multi-modal crisis tweet classification method llm和lmm的多智能体角色扮演：一种可解释的开放世界多模态危机推文分类方法

IF 7.5 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Expert Systems with Applications

Pub Date : 2026-01-23 DOI: 10.1016/j.eswa.2026.131325

Tong Bie , Yongli Hu , Yu Fu , Linjia Hao , Tengfei Liu , Huajie Jiang , Junbin Gao , Yanfeng Sun , Baocai Yin

How to better leverage Large Language Models (LLMs) and Large Multi-modal Models (LMMs) in downstream tasks has become a prominent research topic. In the field of crisis tweet classification, existing approaches often fail to simultaneously handle open-world scenarios, missing modalities, and explainable classification, while still relying on supervised training data. To address these challenges, we propose a Multi-Agent Role-Playing framework (MARP) in which multiple LLMs and LMMs collaborate through specialized role assignments. MARP includes a LMM-based image analysis expert agent and four LLM-based agents: ordinary social media user, humanitarian organization staff member, content verification expert and long-text summarization expert. The social media user and humanitarian staff member are assigned distinct tasks, gathering tweet information from different perspectives by consulting the image analysis expert. The verification expert analyzes interactions to identify potential issues, while the summarization expert consolidates chat logs into summaries. These summaries are processed by the classification expert to generate predictions while also serving as explanatory rationales. We also propose a Query-aware Dynamic Masking (QDM) that selectively filters irrelevant image regions based on cross-modal similarity, enhancing LMMs’ focus on question-relevant visual content. Experiments on the CrisisMMD dataset under both open-world and missing-modality settings demonstrate that MARP achieves zero-shot accuracy improvements of 6.44% and 2.47% over state-of-the-art baselines, respectively. MARP exhibits strong performance in a training-free manner, while also providing explanatory rationales.

如何更好地利用大型语言模型（Large Language Models, llm）和大型多模态模型（Large Multi-modal Models, lmm）在下游任务中发挥作用已成为一个突出的研究课题。在危机推文分类领域，现有方法往往无法同时处理开放世界场景、缺失模态和可解释分类，而仍然依赖于监督训练数据。为了应对这些挑战，我们提出了一个多代理角色扮演框架（MARP），其中多个llm和lmm通过专门的角色分配进行协作。MARP包括一个基于llm的图像分析专家代理和四个基于llm的代理：普通社交媒体用户、人道主义组织工作人员、内容验证专家和长文本摘要专家。社交媒体用户和人道主义工作人员被分配了不同的任务，通过咨询图像分析专家，从不同的角度收集tweet信息。验证专家分析交互以识别潜在的问题，而总结专家将聊天日志合并为摘要。分类专家对这些摘要进行处理，以生成预测，同时也作为解释的依据。我们还提出了一种基于查询感知的动态掩蔽（QDM），它基于跨模态相似性选择性地过滤不相关的图像区域，增强了lmm对问题相关视觉内容的关注。在CrisisMMD数据集上的实验表明，在开放世界和缺失模态设置下，MARP的零射击精度比最先进的基线分别提高了6.44%和2.47%。MARP以一种无需培训的方式展示了强大的性能，同时也提供了解释性的理由。

{"title":"Multi-agent role-playing by LLMs and LMMs: An explainable open-world multi-modal crisis tweet classification method","authors":"Tong Bie , Yongli Hu , Yu Fu , Linjia Hao , Tengfei Liu , Huajie Jiang , Junbin Gao , Yanfeng Sun , Baocai Yin","doi":"10.1016/j.eswa.2026.131325","DOIUrl":"10.1016/j.eswa.2026.131325","url":null,"abstract":"<div><div>How to better leverage Large Language Models (LLMs) and Large Multi-modal Models (LMMs) in downstream tasks has become a prominent research topic. In the field of crisis tweet classification, existing approaches often fail to simultaneously handle open-world scenarios, missing modalities, and explainable classification, while still relying on supervised training data. To address these challenges, we propose a Multi-Agent Role-Playing framework (MARP) in which multiple LLMs and LMMs collaborate through specialized role assignments. MARP includes a LMM-based image analysis expert agent and four LLM-based agents: ordinary social media user, humanitarian organization staff member, content verification expert and long-text summarization expert. The social media user and humanitarian staff member are assigned distinct tasks, gathering tweet information from different perspectives by consulting the image analysis expert. The verification expert analyzes interactions to identify potential issues, while the summarization expert consolidates chat logs into summaries. These summaries are processed by the classification expert to generate predictions while also serving as explanatory rationales. We also propose a Query-aware Dynamic Masking (QDM) that selectively filters irrelevant image regions based on cross-modal similarity, enhancing LMMs’ focus on question-relevant visual content. Experiments on the CrisisMMD dataset under both open-world and missing-modality settings demonstrate that MARP achieves zero-shot accuracy improvements of 6.44% and 2.47% over state-of-the-art baselines, respectively. MARP exhibits strong performance in a training-free manner, while also providing explanatory rationales.</div></div>","PeriodicalId":50461,"journal":{"name":"Expert Systems with Applications","volume":"311 ","pages":"Article 131325"},"PeriodicalIF":7.5,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146081701","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A knowledge-driven pricing model for supply chain coordination under correlated demand 需求关联下供应链协调的知识驱动定价模型

IF 7.5 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Expert Systems with Applications

Pub Date : 2026-01-23 DOI: 10.1016/j.eswa.2026.131290

Ata Allah Taleizadeh , Madjid Tavana , Razieh Sadeghi , Hamidreza Abedsoltan

The growing competition between national brands and store brands has intensified the need for decision-support systems to guide pricing and coordination strategies in retail supply chains. This study develops an analytical game-theoretic framework that integrates stochastic demand modeling, behavioral customer preferences, and computational optimization for dual-brand supply chains. The system models a two-echelon structure consisting of a national-brand manufacturer and a retailer who also produces a substitutable store brand, acting simultaneously as a collaborator and a competitor. Customer demand for both products is correlated and influenced by relative price and quality, introducing substantial uncertainty. A Stackelberg game framework is employed to capture the hierarchical decision-making process, in which the manufacturer sets the wholesale price, and the retailer optimizes retail prices for both products. Analytical and simulation-based evaluations reveal that conventional revenue-sharing contracts fail to coordinate the supply chain when demand correlation and substitution effects are present. To address this limitation, a two-part tariff contract is designed within the analytical framework to align incentives and achieve coordination. The results demonstrate that the proposed approach enhances profitability, system efficiency, and brand competitiveness, providing actionable insights for decision-makers in retail, apparel, and fast-food industries.

民族品牌和商店品牌之间日益激烈的竞争，加强了对决策支持系统的需求，以指导零售供应链中的定价和协调策略。本研究发展了一个分析博弈论框架，整合了随机需求模型、客户行为偏好和双品牌供应链的计算优化。该系统模拟了一个两级结构，包括一个全国性品牌制造商和一个零售商，后者也生产可替代的商店品牌，同时充当合作者和竞争者。顾客对这两种产品的需求是相互关联的，并受相对价格和质量的影响，从而产生了很大的不确定性。采用Stackelberg博弈框架捕捉分层决策过程，其中制造商设定批发价格，零售商优化两种产品的零售价格。基于分析和仿真的评估表明，当存在需求相关性和替代效应时，传统的收益共享契约无法协调供应链。为了解决这一限制，在分析框架内设计了一个两部分的关税合同，以调整激励并实现协调。结果表明，该方法可提高企业的盈利能力、系统效率和品牌竞争力，为零售、服装和快餐行业的决策者提供可操作的见解。

{"title":"A knowledge-driven pricing model for supply chain coordination under correlated demand","authors":"Ata Allah Taleizadeh , Madjid Tavana , Razieh Sadeghi , Hamidreza Abedsoltan","doi":"10.1016/j.eswa.2026.131290","DOIUrl":"10.1016/j.eswa.2026.131290","url":null,"abstract":"<div><div>The growing competition between national brands and store brands has intensified the need for decision-support systems to guide pricing and coordination strategies in retail supply chains. This study develops an analytical game-theoretic framework that integrates stochastic demand modeling, behavioral customer preferences, and computational optimization for dual-brand supply chains. The system models a two-echelon structure consisting of a national-brand manufacturer and a retailer who also produces a substitutable store brand, acting simultaneously as a collaborator and a competitor. Customer demand for both products is correlated and influenced by relative price and quality, introducing substantial uncertainty. A Stackelberg game framework is employed to capture the hierarchical decision-making process, in which the manufacturer sets the wholesale price, and the retailer optimizes retail prices for both products. Analytical and simulation-based evaluations reveal that conventional revenue-sharing contracts fail to coordinate the supply chain when demand correlation and substitution effects are present. To address this limitation, a two-part tariff contract is designed within the analytical framework to align incentives and achieve coordination. The results demonstrate that the proposed approach enhances profitability, system efficiency, and brand competitiveness, providing actionable insights for decision-makers in retail, apparel, and fast-food industries.</div></div>","PeriodicalId":50461,"journal":{"name":"Expert Systems with Applications","volume":"310 ","pages":"Article 131290"},"PeriodicalIF":7.5,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080970","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A modular retrieval-generation-feedback framework for large language model recommendation 用于大型语言模型推荐的模块化检索-生成-反馈框架

IF 7.5 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Expert Systems with Applications

Pub Date : 2026-01-23 DOI: 10.1016/j.eswa.2026.131282

Zhisheng Yang, Li Li

Large Language Models (LLMs) have shown strong generative capabilities in recommendation systems, giving rise to the LLM-as-Recommender (LLM-as-RS) paradigm. However, two key challenges remain: (1) LLMs are static and costly to update with new item information, and (2) directly inputting extensive user data is constrained by maximum input length, which may hinder recommendation quality. To address these issues, we propose two retrieval-augmented generative recommendation frameworks: RAG-LLM-RS and RPRAG-LLM-RS. RAG-LLM-RS enhances personalization by retrieving multiple behaviorally similar users with an untrained retriever and constructing distributed prompts. Building on this design, RPRAG-LLM-RS introduces a pretrained recommendation-oriented retriever and a lightweight evaluation module (R-Critic), which dynamically optimizes retrieval by minimizing the KL divergence between the retrieval distribution and the generation-based scoring distribution. This yields a closed-loop retrieval-generation-feedback mechanism that continuously refines which users are selected for conditioning, without retraining the LLM itself. Experiments on real-world movie and book datasets show that both frameworks significantly improve recommendation accuracy, personalization, and adaptability compared with non-tuning baselines. On the movie dataset, compared with the strongest non-tuning baseline, RAG-LLM-RS improves Precision@3, HR@3, and NDCG@3 by up to +4.77 / +5.70 / +5.83 absolute points, respectively, while RPRAG-LLM-RS further increases these gains to as much as +5.60 / +7.30 / +7.69 absolute points; consistent and stable improvements are also observed at cutoffs 5 and 10. The book dataset exhibits the same trend, where both frameworks outperform the strongest non-tuning baseline and maintain leading performance across different cutoffs (top-3/5/10). Moreover, both frameworks support real-time adaptation without retraining and operate under LLM input-length constraints, highlighting their scalability and practicality for real-world recommendation scenarios.

大型语言模型（llm）在推荐系统中显示出强大的生成能力，从而产生了LLM-as-RS （LLM-as-RS）范式。然而，仍然存在两个关键挑战：(1)llm是静态的，并且使用新项目信息进行更新的成本很高；(2)直接输入大量用户数据受到最大输入长度的限制，这可能会影响推荐质量。为了解决这些问题，我们提出了两个检索增强生成推荐框架：ragg - llm - rs和rprg - llm - rs。RAG-LLM-RS通过使用未经训练的检索器检索多个行为相似的用户并构建分布式提示来增强个性化。基于这一设计，rprg - llm - rs引入了一个预训练的推荐导向检索器和一个轻量级评估模块（R-Critic），该模块通过最小化检索分布和基于生成的评分分布之间的KL差异来动态优化检索。这产生了一个闭环检索-生成-反馈机制，该机制可以不断地改进选择哪些用户进行条件反射，而无需重新训练LLM本身。在真实世界的电影和书籍数据集上的实验表明，与非调优基线相比，这两种框架都显著提高了推荐的准确性、个性化和适应性。在电影数据集上，与最强的非调优基线相比，ragg - llm - rs分别将Precision@3， HR@3和NDCG@3提高了+4.77 / +5.70 / +5.83绝对点，而rprg - llm - rs进一步将这些增益提高到+5.60 / +7.30 / +7.69绝对点；在截止值5和10处也观察到一致和稳定的改进。book数据集显示了相同的趋势，其中两个框架都优于最强的非调优基线，并在不同的截止点（top 3/5/10）保持领先的性能。此外，这两个框架都支持实时自适应，无需再训练，并在LLM输入长度约束下运行，突出了它们在现实世界推荐场景中的可扩展性和实用性。

{"title":"A modular retrieval-generation-feedback framework for large language model recommendation","authors":"Zhisheng Yang, Li Li","doi":"10.1016/j.eswa.2026.131282","DOIUrl":"10.1016/j.eswa.2026.131282","url":null,"abstract":"<div><div>Large Language Models (LLMs) have shown strong generative capabilities in recommendation systems, giving rise to the LLM-as-Recommender (LLM-as-RS) paradigm. However, two key challenges remain: (1) LLMs are static and costly to update with new item information, and (2) directly inputting extensive user data is constrained by maximum input length, which may hinder recommendation quality. To address these issues, we propose two retrieval-augmented generative recommendation frameworks: RAG-LLM-RS and RPRAG-LLM-RS. RAG-LLM-RS enhances personalization by retrieving multiple behaviorally similar users with an untrained retriever and constructing distributed prompts. Building on this design, RPRAG-LLM-RS introduces a pretrained recommendation-oriented retriever and a lightweight evaluation module (R-Critic), which dynamically optimizes retrieval by minimizing the KL divergence between the retrieval distribution and the generation-based scoring distribution. This yields a closed-loop retrieval-generation-feedback mechanism that continuously refines which users are selected for conditioning, without retraining the LLM itself. Experiments on real-world movie and book datasets show that both frameworks significantly improve recommendation accuracy, personalization, and adaptability compared with non-tuning baselines. On the movie dataset, compared with the strongest non-tuning baseline, RAG-LLM-RS improves Precision@3, HR@3, and NDCG@3 by up to +4.77 / +5.70 / +5.83 absolute points, respectively, while RPRAG-LLM-RS further increases these gains to as much as +5.60 / +7.30 / +7.69 absolute points; consistent and stable improvements are also observed at cutoffs 5 and 10. The book dataset exhibits the same trend, where both frameworks outperform the strongest non-tuning baseline and maintain leading performance across different cutoffs (top-3/5/10). Moreover, both frameworks support real-time adaptation without retraining and operate under LLM input-length constraints, highlighting their scalability and practicality for real-world recommendation scenarios.</div></div>","PeriodicalId":50461,"journal":{"name":"Expert Systems with Applications","volume":"310 ","pages":"Article 131282"},"PeriodicalIF":7.5,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146081054","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

CellMixer: Pathological image classification using dual-branch VMamba with randomly mixing gradient features data augmentation CellMixer：病理图像分类使用双分支vamba随机混合梯度特征数据增强

IF 7.5 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Expert Systems with Applications

Pub Date : 2026-01-23 DOI: 10.1016/j.eswa.2026.131298

Enhui Chai , Xinyu Gu , Zheng Lu , Tianxiang Cui

Pathological diagnosis is crucial for patient care, and Region of Interest (ROI) analysis serves as a key pathological method for extracting local cellular details to guide precise clinical decision-making. While most of the current foundation models have shown promise in ROI pathological image classification, existing approaches often fall short in addressing the unique characteristics of pathology ROI data from three aspects simultaneously: (1) inter-class similarity, (2) complex global patterns, and (3) multi-scale granularity. To address them, we propose CellMixer, a novel framework designed to extract and integrate local-global ROI pathological image representations. The key innovation lies in the synergistic integration of three corresponding domain-aware components: (1) To amplify subtle morphological distinctions, we designed a data augmentation (GradMix), which selectively fuses gradient maps and pixel-level features to enhances low-level feature sensitivity, directly improving discrimination of visually similar classes; (2) To capture both localized patterns and global tissue structures across ROI regions, we proposed Dual-branch VMamba Block (DVB), which enhances long-range dependency modeling and simultaneously extracts cell-level fine-grained features; (3) To fuse local and global features to concurrently represent intra-class homogeneity and inter-class heterogeneity across scales, a novel feature fusion strategy (Insert-Merge (InM)). Extensive experiments on 8 public pathology ROI datasets demonstrate that CellMixer consistently outperforms existing methods, proving task-specific model, even with limited data, yields superior visual representations to generic foundation models.

病理诊断对患者护理至关重要，感兴趣区域（ROI）分析是提取局部细胞细节以指导精确临床决策的关键病理方法。虽然目前大多数基础模型在ROI病理图像分类中显示出前景，但现有方法往往无法同时从三个方面解决病理ROI数据的独特特征：(1)类间相似性，(2)复杂的全局模式，(3)多尺度粒度。为了解决这些问题，我们提出了CellMixer，这是一个新的框架，旨在提取和整合局部-全局ROI病理图像表示。关键创新在于三个相应的领域感知组件的协同集成：(1)为了放大细微的形态差异，我们设计了一个数据增强（GradMix），它有选择地融合梯度图和像素级特征，以增强低级特征的灵敏度，直接提高视觉相似类的识别；(2)为了捕获ROI区域的局部模式和全局组织结构，我们提出了双分支vamba块（Dual-branch vamba Block， DVB），该方法增强了远程依赖建模，同时提取细胞级细粒度特征；(3)为了融合局部和全局特征，在不同尺度上同时表现类内同质性和类间异质性，提出了一种新的特征融合策略（Insert-Merge, InM）。在8个公共病理ROI数据集上进行的大量实验表明，CellMixer始终优于现有方法，证明了特定任务模型，即使数据有限，也能产生优于通用基础模型的视觉表现。

{"title":"CellMixer: Pathological image classification using dual-branch VMamba with randomly mixing gradient features data augmentation","authors":"Enhui Chai , Xinyu Gu , Zheng Lu , Tianxiang Cui","doi":"10.1016/j.eswa.2026.131298","DOIUrl":"10.1016/j.eswa.2026.131298","url":null,"abstract":"<div><div>Pathological diagnosis is crucial for patient care, and Region of Interest (ROI) analysis serves as a key pathological method for extracting local cellular details to guide precise clinical decision-making. While most of the current foundation models have shown promise in ROI pathological image classification, existing approaches often fall short in addressing the unique characteristics of pathology ROI data from three aspects simultaneously: (1) inter-class similarity, (2) complex global patterns, and (3) multi-scale granularity. To address them, we propose CellMixer, a novel framework designed to extract and integrate local-global ROI pathological image representations. The key innovation lies in the synergistic integration of three corresponding domain-aware components: (1) To amplify subtle morphological distinctions, we designed a data augmentation (GradMix), which selectively fuses gradient maps and pixel-level features to enhances low-level feature sensitivity, directly improving discrimination of visually similar classes; (2) To capture both localized patterns and global tissue structures across ROI regions, we proposed Dual-branch VMamba Block (DVB), which enhances long-range dependency modeling and simultaneously extracts cell-level fine-grained features; (3) To fuse local and global features to concurrently represent intra-class homogeneity and inter-class heterogeneity across scales, a novel feature fusion strategy (Insert-Merge (InM)). Extensive experiments on 8 public pathology ROI datasets demonstrate that CellMixer consistently outperforms existing methods, proving task-specific model, even with limited data, yields superior visual representations to generic foundation models.</div></div>","PeriodicalId":50461,"journal":{"name":"Expert Systems with Applications","volume":"310 ","pages":"Article 131298"},"PeriodicalIF":7.5,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146081056","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A sliding window feature extraction and weight adaptive random forest-based method for TBM tunnel rock mass grade identification 基于滑动窗特征提取和权值自适应随机森林的TBM隧道岩体等级识别方法

IF 7.5 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Expert Systems with Applications

Pub Date : 2026-01-23 DOI: 10.1016/j.eswa.2026.131336

Honggan Yu , Shuzhan Xu , Xin Yin , Mahdi Hasanipanah

Identifying the rock mass grade of the tunnel face is the basis for analyzing the stability of the surrounding rock and predicting the performance of the tunnel boring machine (TBM). Current research on using TBM’s tunneling parameters to identify rock mass grade usually overlooks the geological information contained in the changes of tunneling parameters, and the machine learning methods rarely consider the impact of class weights. Therefore, this paper proposes a sliding window feature extraction and weight adaptive random forest-based method for rock mass grade identification. Firstly, the raw data is preprocessed, including parameter screening based on experience and random forest, outlier detection based on isolated forest etc. After that, a novel sliding window-based feature extraction method is proposed, which can extract geological-related features from the changes in tunneling parameters. Finally, a weight adaptive random forest algorithm is proposed, and the particle swarm optimization is used to obtain the optimal class weights. On-site data from a water conveyance project was used to validate the effectiveness of the proposed method. The results show that the proposed sliding window feature extraction method can significantly improve the model’s performance compared with directly using tunneling parameters as the model’s input. Moreover, the proposed weight adaptive random forest algorithm can effectively suppress misclassification caused by high similarity among classes, and its performance is better than random forest, adaptive boosting, extreme gradient boosting, and light gradient boosting machine. Therefore, the proposed method can accurately identify the rock mass grade, which has essential engineering value.

确定巷道工作面岩体等级是分析围岩稳定性和预测隧道掘进机性能的基础。目前利用掘进机掘进参数识别岩体品位的研究往往忽略了掘进参数变化所包含的地质信息，机器学习方法也很少考虑类权值的影响。为此，本文提出了一种基于滑动窗口特征提取和权值自适应随机森林的岩体品位识别方法。首先对原始数据进行预处理，包括基于经验和随机森林的参数筛选、基于孤立森林的离群点检测等；在此基础上，提出了一种基于滑动窗的特征提取方法，该方法可以从掘进参数的变化中提取地质相关特征。最后，提出了一种权值自适应随机森林算法，并利用粒子群算法获得最优的类权值。利用某输水工程的现场数据验证了该方法的有效性。结果表明，与直接使用隧道参数作为模型输入相比，所提出的滑动窗口特征提取方法能显著提高模型的性能。此外，所提出的权重自适应随机森林算法能有效抑制由于类间高度相似而导致的误分类，其性能优于随机森林、自适应增强、极端梯度增强和轻梯度增强机。因此，该方法能准确识别岩体品位，具有重要的工程价值。

{"title":"A sliding window feature extraction and weight adaptive random forest-based method for TBM tunnel rock mass grade identification","authors":"Honggan Yu , Shuzhan Xu , Xin Yin , Mahdi Hasanipanah","doi":"10.1016/j.eswa.2026.131336","DOIUrl":"10.1016/j.eswa.2026.131336","url":null,"abstract":"<div><div>Identifying the rock mass grade of the tunnel face is the basis for analyzing the stability of the surrounding rock and predicting the performance of the tunnel boring machine (TBM). Current research on using TBM’s tunneling parameters to identify rock mass grade usually overlooks the geological information contained in the changes of tunneling parameters, and the machine learning methods rarely consider the impact of class weights. Therefore, this paper proposes a sliding window feature extraction and weight adaptive random forest-based method for rock mass grade identification. Firstly, the raw data is preprocessed, including parameter screening based on experience and random forest, outlier detection based on isolated forest etc. After that, a novel sliding window-based feature extraction method is proposed, which can extract geological-related features from the changes in tunneling parameters. Finally, a weight adaptive random forest algorithm is proposed, and the particle swarm optimization is used to obtain the optimal class weights. On-site data from a water conveyance project was used to validate the effectiveness of the proposed method. The results show that the proposed sliding window feature extraction method can significantly improve the model’s performance compared with directly using tunneling parameters as the model’s input. Moreover, the proposed weight adaptive random forest algorithm can effectively suppress misclassification caused by high similarity among classes, and its performance is better than random forest, adaptive boosting, extreme gradient boosting, and light gradient boosting machine. Therefore, the proposed method can accurately identify the rock mass grade, which has essential engineering value.</div></div>","PeriodicalId":50461,"journal":{"name":"Expert Systems with Applications","volume":"311 ","pages":"Article 131336"},"PeriodicalIF":7.5,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146081692","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A lightweight block stochastic configuration network with fisher information for soft sensor applications 用于软传感器应用的具有fisher信息的轻量级块随机配置网络

IF 7.5 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Expert Systems with Applications

Pub Date : 2026-01-23 DOI: 10.1016/j.eswa.2026.131285

Xinyu Zhou , Jun Lu , Jinliang Ding , Yongchao Zhang

Block stochastic configuration network improves the training efficiency of stochastic configuration network by adding multiple nodes per iteration. However, this acceleration comes at the cost of significantly increased model complexity. To address this limitation, we propose a novel lightweight block stochastic configuration network with fisher information, termed Fisher Information Block Stochastic Configuration Network (FIBSCN). During training process, FIBSCN evaluates the fisher information of each newly generated node, enabling the algorithm to discard redundant nodes with weaker contribution to the model. This selective node retention mechanism effectively suppresses unnecessary model growth and substantially reduces structural complexity while preserving predictive accuracy. Building on the FIBSCN, we further develop an Online FIBSCN (OnFIBSCN), which incorporates regularization constraint term based on fisher information. OnFIBSCN can maintain model stability under new streaming data and retain essential historical knowledge. The effectiveness is demonstrated by comprehensive experiments, including three benchmark regression tasks and a real-world mineral process for hematite application.

块随机配置网络通过每次迭代增加多个节点来提高随机配置网络的训练效率。然而，这种加速是以显著增加模型复杂性为代价的。为了解决这一限制，我们提出了一种具有fisher信息的新型轻量级块随机配置网络，称为fisher信息块随机配置网络（FIBSCN）。在训练过程中，FIBSCN对每个新生成节点的fisher信息进行评估，使算法能够丢弃对模型贡献较弱的冗余节点。这种选择性节点保留机制有效地抑制了不必要的模型增长，在保持预测准确性的同时大大降低了结构复杂性。在FIBSCN的基础上，我们进一步开发了基于fisher信息的正则化约束项的在线FIBSCN （OnFIBSCN）。OnFIBSCN可以在新的流数据下保持模型的稳定性，并保留必要的历史知识。通过综合实验，包括三个基准回归任务和一个实际的赤铁矿选矿过程，验证了该方法的有效性。

{"title":"A lightweight block stochastic configuration network with fisher information for soft sensor applications","authors":"Xinyu Zhou , Jun Lu , Jinliang Ding , Yongchao Zhang","doi":"10.1016/j.eswa.2026.131285","DOIUrl":"10.1016/j.eswa.2026.131285","url":null,"abstract":"<div><div>Block stochastic configuration network improves the training efficiency of stochastic configuration network by adding multiple nodes per iteration. However, this acceleration comes at the cost of significantly increased model complexity. To address this limitation, we propose a novel lightweight block stochastic configuration network with fisher information, termed Fisher Information Block Stochastic Configuration Network (FIBSCN). During training process, FIBSCN evaluates the fisher information of each newly generated node, enabling the algorithm to discard redundant nodes with weaker contribution to the model. This selective node retention mechanism effectively suppresses unnecessary model growth and substantially reduces structural complexity while preserving predictive accuracy. Building on the FIBSCN, we further develop an Online FIBSCN (OnFIBSCN), which incorporates regularization constraint term based on fisher information. OnFIBSCN can maintain model stability under new streaming data and retain essential historical knowledge. The effectiveness is demonstrated by comprehensive experiments, including three benchmark regression tasks and a real-world mineral process for hematite application.</div></div>","PeriodicalId":50461,"journal":{"name":"Expert Systems with Applications","volume":"310 ","pages":"Article 131285"},"PeriodicalIF":7.5,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146015658","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Beyond clicks: Measuring attractiveness and satisfaction in e-commerce using Bayesian models with conversion signals 超越点击：使用带有转换信号的贝叶斯模型测量电子商务的吸引力和满意度

IF 7.5 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Expert Systems with Applications

Pub Date : 2026-01-23 DOI: 10.1016/j.eswa.2026.131299

Hacer Turgut , Afra Arslan , Ömür Bali , Mehmet Yasin Ulukuş

Understanding user behavior is essential for improving user experience and maximizing conversion rates on e-commerce platforms. To more accurately capture user satisfaction, the iLab Click and Conversion Dynamic Bayesian Network (iCCDBN) is introduced, a novel click model that jointly incorporates click and post-click conversion signals. iCCDBN employs separate satisfaction parameters for clicks and conversions, enhancing interpretability while maintaining computational efficiency. The probabilistic formulation of the model is derived, and parameter estimation is carried out using the Expectation-Maximization (EM) algorithm. For evaluation, iCCDBN is compared with established click models on large-scale interaction logs from a real estate marketplace. Results show that iCCDBN, together with strong baselines, achieves the lowest click-through rate prediction errors, with optimal performance observed when (query, item) pairs have at least 60 historical sessions. In satisfaction prediction, iCCDBN surpasses the Dynamic Bayesian Network (DBN) with a lower mean squared error (0.1927 vs. 0.2313). KL divergence analysis further demonstrates that iCCDBN achieves an 8.6% reduction in KL divergence when evaluated on raw prediction scores. When score ranges are normalized via min-max scaling, thereby emphasizing distributional shape rather than scale, the relative improvement increases to 17.4%. These findings highlight the benefits of integrating conversion data and refined behavioral structures into click models, offering a more faithful representation of user satisfaction.

了解用户行为对于改善用户体验和最大化电子商务平台的转化率至关重要。为了更准确地捕捉用户满意度，引入了iLab点击和转换动态贝叶斯网络（iCCDBN），这是一种结合点击和点击后转换信号的新颖点击模型。iCCDBN为点击和转换使用单独的满意度参数，在保持计算效率的同时增强了可解释性。推导了模型的概率表达式，并采用期望最大化（EM）算法进行参数估计。为了进行评估，将iCCDBN与基于房地产市场大规模交互日志的已建立的点击模型进行了比较。结果表明，iCCDBN与强基线一起实现了最低的点击率预测误差，当（查询，项目）对具有至少60个历史会话时观察到最佳性能。在满意度预测方面，iCCDBN以较低的均方误差（0.1927 vs. DBN）优于动态贝叶斯网络（DBN）。0.2313)。KL散度分析进一步表明，当对原始预测分数进行评估时，iCCDBN的KL散度降低了8.6%。当分数范围通过最小-最大尺度归一化，从而强调分布形状而不是尺度时，相对改善增加到17.4%。这些发现突出了将转换数据和精炼的行为结构整合到点击模型中的好处，提供了更忠实的用户满意度表示。

{"title":"Beyond clicks: Measuring attractiveness and satisfaction in e-commerce using Bayesian models with conversion signals","authors":"Hacer Turgut , Afra Arslan , Ömür Bali , Mehmet Yasin Ulukuş","doi":"10.1016/j.eswa.2026.131299","DOIUrl":"10.1016/j.eswa.2026.131299","url":null,"abstract":"<div><div>Understanding user behavior is essential for improving user experience and maximizing conversion rates on e-commerce platforms. To more accurately capture user satisfaction, the iLab Click and Conversion Dynamic Bayesian Network (iCCDBN) is introduced, a novel click model that jointly incorporates click and post-click conversion signals. iCCDBN employs separate satisfaction parameters for clicks and conversions, enhancing interpretability while maintaining computational efficiency. The probabilistic formulation of the model is derived, and parameter estimation is carried out using the Expectation-Maximization (EM) algorithm. For evaluation, iCCDBN is compared with established click models on large-scale interaction logs from a real estate marketplace. Results show that iCCDBN, together with strong baselines, achieves the lowest click-through rate prediction errors, with optimal performance observed when (query, item) pairs have at least 60 historical sessions. In satisfaction prediction, iCCDBN surpasses the Dynamic Bayesian Network (DBN) with a lower mean squared error (0.1927 vs. 0.2313). KL divergence analysis further demonstrates that iCCDBN achieves an 8.6% reduction in KL divergence when evaluated on raw prediction scores. When score ranges are normalized via min-max scaling, thereby emphasizing distributional shape rather than scale, the relative improvement increases to 17.4%. These findings highlight the benefits of integrating conversion data and refined behavioral structures into click models, offering a more faithful representation of user satisfaction.</div></div>","PeriodicalId":50461,"journal":{"name":"Expert Systems with Applications","volume":"310 ","pages":"Article 131299"},"PeriodicalIF":7.5,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080878","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Design of Bayesian acceptance test for multi-state reliability growth of equipment with multinomial distribution 多项分布设备多状态可靠性增长的贝叶斯验收试验设计

IF 7.5 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Expert Systems with Applications

Pub Date : 2026-01-23 DOI: 10.1016/j.eswa.2026.131306

Haobang Liu, Tong Chen , Peng Di, Tao Hu, Haolin Wen, Lisha Zheng, Minggui Li

Traditional equipment reliability acceptance test predominantly relies on binary outcomes (normal/failure), which inadequately capture the multi-state nature of performance degradation in operation. To address this, this paper proposes a knowledge-integrated Bayesian acceptance test framework for multi-state reliability growth, modeled via a multinomial distribution. The framework systematically incorporates domain expert knowledge and simulation test information through Dirichlet prior, and embeds sequential constraints of reliability growth into the Bayesian model, effectively serving as an expert-augmented decision support system. To solve this model, the existing Markov chain-Monte Carlo (MCMC) and Gibbs sampling algorithms are improved to ensure parameter samples extracted in each iteration meet the sequential constraints. Experimental analysis demonstrates that the standard deviation of the reliability estimation results obtained by this research method is reduced from 0.079 to 0.058 compared with the existing Bayesian method, and the operating characteristic (OC) curve is steeper, indicating stronger discrimination and sharper decision-making. This work provides a scalable and knowledge-integrated Bayesian framework that aligns with the development of expert systems for intelligent reliability estimation and acceptance test.

传统的设备可靠性验收测试主要依赖于二元结果（正常/故障），这不能充分反映运行中性能退化的多状态性质。为了解决这个问题，本文提出了一个知识集成的贝叶斯验收测试框架，用于通过多项分布建模的多状态可靠性增长。该框架通过Dirichlet先验算法系统地融合了领域专家知识和仿真试验信息，并将可靠性增长的顺序约束嵌入到贝叶斯模型中，有效地作为专家增强决策支持系统。为了求解该模型，改进了现有的Markov chain-Monte Carlo （MCMC）和Gibbs采样算法，保证每次迭代提取的参数样本满足序列约束。实验分析表明，与现有贝叶斯方法相比，该方法得到的可靠性估计结果的标准差从0.079降低到0.058，OC曲线更陡，具有更强的判别能力和更敏锐的决策能力。这项工作提供了一个可扩展和知识集成的贝叶斯框架，与智能可靠性评估和验收测试专家系统的发展保持一致。

{"title":"Design of Bayesian acceptance test for multi-state reliability growth of equipment with multinomial distribution","authors":"Haobang Liu, Tong Chen , Peng Di, Tao Hu, Haolin Wen, Lisha Zheng, Minggui Li","doi":"10.1016/j.eswa.2026.131306","DOIUrl":"10.1016/j.eswa.2026.131306","url":null,"abstract":"<div><div>Traditional equipment reliability acceptance test predominantly relies on binary outcomes (normal/failure), which inadequately capture the multi-state nature of performance degradation in operation. To address this, this paper proposes a knowledge-integrated Bayesian acceptance test framework for multi-state reliability growth, modeled via a multinomial distribution. The framework systematically incorporates domain expert knowledge and simulation test information through Dirichlet prior, and embeds sequential constraints of reliability growth into the Bayesian model, effectively serving as an expert-augmented decision support system. To solve this model, the existing Markov chain-Monte Carlo (MCMC) and Gibbs sampling algorithms are improved to ensure parameter samples extracted in each iteration meet the sequential constraints. Experimental analysis demonstrates that the standard deviation of the reliability estimation results obtained by this research method is reduced from 0.079 to 0.058 compared with the existing Bayesian method, and the operating characteristic (OC) curve is steeper, indicating stronger discrimination and sharper decision-making. This work provides a scalable and knowledge-integrated Bayesian framework that aligns with the development of expert systems for intelligent reliability estimation and acceptance test.</div></div>","PeriodicalId":50461,"journal":{"name":"Expert Systems with Applications","volume":"310 ","pages":"Article 131306"},"PeriodicalIF":7.5,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146081112","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

AAF-Bi-LSTM-NARX: A bidirectional network with adaptive attention fusion for sensor data imputation AAF-Bi-LSTM-NARX：基于自适应注意力融合的传感器数据输入双向网络

IF 7.5 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Expert Systems with Applications

Pub Date : 2026-01-22 DOI: 10.1016/j.eswa.2026.131266

Zhao Zhao, Xian Yi, Jianjun Xiong, Lin Ran, Jieyi Zhao, Yalan Zhu

Sensors in wind tunnel tests occasionally experience data missing due to harsh environmental conditions, which affects data quality and system safety. Traditional interpolation and machine learning methods are difficult to effectively capture the nonlinear and temporal features in the data, while most existing deep learning models rely on unidirectional information flow and fail to make full use of contextual information. To address this issue, a bidirectional LSTM-NARX model based on adaptive attention fusion (AAF-Bi-LSTM-NARX) is proposed for imputing the missing data of wind tunnel sensors.The model extracts forward and backward temporal features respectively through a bidirectional LSTM-NARX network. Furthermore, an adaptive attention fusion module is designed to dynamically fuse the bidirectional prediction results, thereby improving the accuracy of data imputation. Experiments on a real wind tunnel dataset show that the proposed model significantly outperforms comparative models such as LSTM-NARX, Bi-LSTM, LSTM, Informer, GRU-D, and ARIMA under the missing rate of 5%, 10%, and 25%. Ablation studies further verify the effectiveness of the bidirectional structure and the global error verification mechanism, and explore the applicable conditions of local attention under different missing rates. This study provides an efficient and reliable solution for sensor data imputation, and is of great significance for improving the data quality of wind tunnel tests and supporting the development of data-driven wind tunnel systems.

风洞试验中的传感器由于环境条件恶劣，有时会出现数据丢失的情况，影响数据质量和系统安全性。传统的插值和机器学习方法难以有效捕获数据中的非线性和时间特征，而现有的深度学习模型大多依赖于单向信息流，未能充分利用上下文信息。针对这一问题，提出了一种基于自适应注意力融合的双向LSTM-NARX模型（AAF-Bi-LSTM-NARX），用于风洞传感器缺失数据的输入。该模型通过双向LSTM-NARX网络分别提取前向和后向时间特征。设计了自适应注意力融合模块，对双向预测结果进行动态融合，提高了数据输入的准确性。在真实风洞数据集上的实验表明，在缺失率分别为5%、10%和25%的情况下，该模型显著优于LSTM- narx、Bi-LSTM、LSTM、Informer、GRU-D和ARIMA等模型。消融研究进一步验证了双向结构和全局误差验证机制的有效性，探索了不同缺失率下局部关注的适用条件。该研究为传感器数据的输入提供了一种高效、可靠的解决方案，对于提高风洞试验数据质量，支持数据驱动风洞系统的发展具有重要意义。

{"title":"AAF-Bi-LSTM-NARX: A bidirectional network with adaptive attention fusion for sensor data imputation","authors":"Zhao Zhao, Xian Yi, Jianjun Xiong, Lin Ran, Jieyi Zhao, Yalan Zhu","doi":"10.1016/j.eswa.2026.131266","DOIUrl":"10.1016/j.eswa.2026.131266","url":null,"abstract":"<div><div>Sensors in wind tunnel tests occasionally experience data missing due to harsh environmental conditions, which affects data quality and system safety. Traditional interpolation and machine learning methods are difficult to effectively capture the nonlinear and temporal features in the data, while most existing deep learning models rely on unidirectional information flow and fail to make full use of contextual information. To address this issue, a bidirectional LSTM-NARX model based on adaptive attention fusion (AAF-Bi-LSTM-NARX) is proposed for imputing the missing data of wind tunnel sensors.The model extracts forward and backward temporal features respectively through a bidirectional LSTM-NARX network. Furthermore, an adaptive attention fusion module is designed to dynamically fuse the bidirectional prediction results, thereby improving the accuracy of data imputation. Experiments on a real wind tunnel dataset show that the proposed model significantly outperforms comparative models such as LSTM-NARX, Bi-LSTM, LSTM, Informer, GRU-D, and ARIMA under the missing rate of 5%, 10%, and 25%. Ablation studies further verify the effectiveness of the bidirectional structure and the global error verification mechanism, and explore the applicable conditions of local attention under different missing rates. This study provides an efficient and reliable solution for sensor data imputation, and is of great significance for improving the data quality of wind tunnel tests and supporting the development of data-driven wind tunnel systems.</div></div>","PeriodicalId":50461,"journal":{"name":"Expert Systems with Applications","volume":"310 ","pages":"Article 131266"},"PeriodicalIF":7.5,"publicationDate":"2026-01-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146025742","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0