Efficient and reliable monitoring of processes in coal-fired boilers requires advanced combustion diagnostic and optimization methods. This paper introduces an integrated system based on machine learning and deep learning technologies to diagnose and optimize combustion regimes in pulverized coal-fired boilers, aiming to improve efficiency and safety. An artificial neural network accurately simulated thermogravimetric mass loss curves, achieving an average coefficient of determination of 99%. Deep learning methods were employed to detect combustion regimes by monitoring the coal flame and identifying anomalies in flame images. For datasets lacking precise measurements, an unsupervised autoencoder was developed. It achieved an average precision of 77% and a recall of 66%. For datasets with measurements, a supervised convolutional neural network provided a higher average recall of 89%. Various machine learning algorithms were employed to predict deviations from stable combustion modes, and a long short-term memory network with an attention mechanism performed best. It had a mean absolute percentage error of up to 8% and an average coefficient of determination of 91%.
{"title":"System for diagnosis and optimization of combustion in pulverized coal boilers based on artificial intelligence methods","authors":"S.S. Abdurakipov, E.B. Butakov, E.P. Kopyev, D.M. Markovich","doi":"10.1016/j.engappai.2026.114103","DOIUrl":"10.1016/j.engappai.2026.114103","url":null,"abstract":"<div><div>Efficient and reliable monitoring of processes in coal-fired boilers requires advanced combustion diagnostic and optimization methods. This paper introduces an integrated system based on machine learning and deep learning technologies to diagnose and optimize combustion regimes in pulverized coal-fired boilers, aiming to improve efficiency and safety. An artificial neural network accurately simulated thermogravimetric mass loss curves, achieving an average coefficient of determination of 99%. Deep learning methods were employed to detect combustion regimes by monitoring the coal flame and identifying anomalies in flame images. For datasets lacking precise measurements, an unsupervised autoencoder was developed. It achieved an average precision of 77% and a recall of 66%. For datasets with measurements, a supervised convolutional neural network provided a higher average recall of 89%. Various machine learning algorithms were employed to predict deviations from stable combustion modes, and a long short-term memory network with an attention mechanism performed best. It had a mean absolute percentage error of up to 8% and an average coefficient of determination of 91%.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"169 ","pages":"Article 114103"},"PeriodicalIF":8.0,"publicationDate":"2026-02-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146175352","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-09DOI: 10.1016/j.engappai.2026.114062
Wenkai Pang , Zhi Tan
Object detection is a crucial task in computer vision. Despite recent advancements in generalized object detection, accurately detecting small objects remains a significant challenge due to feature degradation in neural networks. Although increasing the input image resolution is widely recognized as the most effective countermeasure against feature degradation effects, this approach inevitably results in substantial computational cost increases. To this end, we propose the Communicating Vessels Detection Network, which presents two major contributions. Initially, the Communicating Vessels Backbone enhances the resolution of deep feature maps and fuses them with shallow-layer features, effectively mitigating feature degradation. In addition, we design the Global Information Coding Module, which adaptively captures contextual dependencies from horizontal, vertical, and global directions, providing critical guidance for small object detection. On the challenging Vision Meets Drone: A Challenge 2019 dataset, the Communicating Vessels Detection Network improves the mean average precision, the average precision for small objects, and the average precision for medium objects by 1.5%, 0.64%, and 1.02%, respectively. Furthermore, our model demonstrates superior performance on the Traffic Signal Detection dataset and two Safety Helmet Wearing Detection datasets compared to state-of-the-art methods.
目标检测是计算机视觉中的一项重要任务。尽管最近在广义目标检测方面取得了进展，但由于神经网络的特征退化，准确检测小目标仍然是一个重大挑战。虽然提高输入图像分辨率被广泛认为是对抗特征退化效应的最有效的对策，但这种方法不可避免地会导致大量的计算成本增加。为此，我们提出了通信船舶检测网络，它有两个主要贡献。首先，通信血管主干增强了深层特征图的分辨率，并将其与浅层特征融合，有效缓解了特征退化。此外，我们还设计了全局信息编码模块，该模块可以自适应地从水平、垂直和全局方向捕获上下文依赖关系，为小目标检测提供关键指导。在具有挑战性的Vision Meets Drone: A Challenge 2019数据集上，通信船只检测网络将平均精度、小物体的平均精度和中等物体的平均精度分别提高了1.5%、0.64%和1.02%。此外，与最先进的方法相比，我们的模型在交通信号检测数据集和两个安全帽佩戴检测数据集上表现出优越的性能。
{"title":"Communicating vessels detection network for small object detection in realistic scenario","authors":"Wenkai Pang , Zhi Tan","doi":"10.1016/j.engappai.2026.114062","DOIUrl":"10.1016/j.engappai.2026.114062","url":null,"abstract":"<div><div>Object detection is a crucial task in computer vision. Despite recent advancements in generalized object detection, accurately detecting small objects remains a significant challenge due to feature degradation in neural networks. Although increasing the input image resolution is widely recognized as the most effective countermeasure against feature degradation effects, this approach inevitably results in substantial computational cost increases. To this end, we propose the Communicating Vessels Detection Network, which presents two major contributions. Initially, the Communicating Vessels Backbone enhances the resolution of deep feature maps and fuses them with shallow-layer features, effectively mitigating feature degradation. In addition, we design the Global Information Coding Module, which adaptively captures contextual dependencies from horizontal, vertical, and global directions, providing critical guidance for small object detection. On the challenging Vision Meets Drone: A Challenge 2019 dataset, the Communicating Vessels Detection Network improves the mean average precision, the average precision for small objects, and the average precision for medium objects by 1.5%, 0.64%, and 1.02%, respectively. Furthermore, our model demonstrates superior performance on the Traffic Signal Detection dataset and two Safety Helmet Wearing Detection datasets compared to state-of-the-art methods.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"169 ","pages":"Article 114062"},"PeriodicalIF":8.0,"publicationDate":"2026-02-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146175343","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-09DOI: 10.1016/j.engappai.2026.114116
Abdulkerim Ari , Metin Katlav , Izzeddin Donmez , Kazim Turk
Flat-slab systems manufactured with self-compacting concrete (SCC) incorporating low hybrid fiber content offer a promising alternative for improving punching shear performance while enhancing constructability in building applications. In this paper, the punching shear behavior of flat-slabs produced with single, binary, and ternary fiber-reinforced SCC was experimentally investigated in terms of load–deflection response, ductility, toughness, cracking behavior, and failure mode. In parallel, a comprehensive database comprising 268 fiber-reinforced concrete flat-slab test results collected from the literature was established, and artificial intelligence (AI)-based predictive models were developed to estimate punching shear capacity (Vpun). Model performance was evaluated using statistical indicators, whereas SHapley Additive exPlanations (SHAP) feature importance and partial dependence plots (PDPs) were employed to enhance interpretability and reveal the governing parameters influencing punching capacity. The outcomes demonstrate that binary hybrid fiber systems provide the most effective enhancement in punching capacity and post-cracking performance, even at low fiber contents, outperforming conventional solutions such as shear studs. Among the developed AI models, the Extra Trees Regressor and Random Forest algorithms exhibited the highest prediction accuracy for the Vpun. Finally, the AI models were integrated into a user-friendly graphical interface to facilitate practical engineering applications. Overall, this research contributes by experimentally validating low-fiber SCC flat-slabs as an efficient punching solution and by proposing an explainable, data-driven decision-support framework for engineering design.
采用低混合纤维含量的自密实混凝土（SCC）制造的平板系统为改善冲剪性能，同时增强建筑应用中的可施工性提供了一种有希望的替代方案。本文从荷载-挠曲响应、延性、韧性、开裂行为和破坏模式等方面对单、双、三元纤维增强SCC平板的冲剪行为进行了实验研究。同时，建立了一个综合数据库，包括从文献中收集的268个纤维钢筋混凝土平板试验结果，并开发了基于人工智能（AI）的预测模型来估计冲剪能力（Vpun）。采用统计指标评价模型性能，采用SHapley加性解释（SHAP）特征重要性和部分依赖图（pdp）增强可解释性，揭示影响冲压能力的控制参数。结果表明，即使在纤维含量较低的情况下，二元混合纤维体系也能最有效地提高冲孔能力和开裂后性能，优于剪切螺柱等传统解决方案。在已开发的人工智能模型中，Extra Trees Regressor和Random Forest算法对Vpun的预测精度最高。最后，将人工智能模型集成到用户友好的图形界面中，以方便实际工程应用。总的来说，本研究通过实验验证了低纤维SCC平板作为一种有效的冲压解决方案，并为工程设计提出了一个可解释的、数据驱动的决策支持框架。
{"title":"Experimental investigation and explainable artificial intelligence-based modeling of punching shear behavior in self-compacting concrete flat-slabs with low hybrid fiber content","authors":"Abdulkerim Ari , Metin Katlav , Izzeddin Donmez , Kazim Turk","doi":"10.1016/j.engappai.2026.114116","DOIUrl":"10.1016/j.engappai.2026.114116","url":null,"abstract":"<div><div>Flat-slab systems manufactured with self-compacting concrete (SCC) incorporating low hybrid fiber content offer a promising alternative for improving punching shear performance while enhancing constructability in building applications. In this paper, the punching shear behavior of flat-slabs produced with single, binary, and ternary fiber-reinforced SCC was experimentally investigated in terms of load–deflection response, ductility, toughness, cracking behavior, and failure mode. In parallel, a comprehensive database comprising 268 fiber-reinforced concrete flat-slab test results collected from the literature was established, and artificial intelligence (AI)-based predictive models were developed to estimate punching shear capacity (<em>V</em><sub><em>pun</em></sub>). Model performance was evaluated using statistical indicators, whereas SHapley Additive exPlanations (SHAP) feature importance and partial dependence plots (PDPs) were employed to enhance interpretability and reveal the governing parameters influencing punching capacity. The outcomes demonstrate that binary hybrid fiber systems provide the most effective enhancement in punching capacity and post-cracking performance, even at low fiber contents, outperforming conventional solutions such as shear studs. Among the developed AI models, the Extra Trees Regressor and Random Forest algorithms exhibited the highest prediction accuracy for the <em>V</em><sub><em>pun</em></sub>. Finally, the AI models were integrated into a user-friendly graphical interface to facilitate practical engineering applications. Overall, this research contributes by experimentally validating low-fiber SCC flat-slabs as an efficient punching solution and by proposing an explainable, data-driven decision-support framework for engineering design.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"169 ","pages":"Article 114116"},"PeriodicalIF":8.0,"publicationDate":"2026-02-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146175308","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-09DOI: 10.1016/j.engappai.2026.114061
Taoran Song , Hao Pu , Hong Zhang , Paul Schonfeld , Lihui Peng
Vertical alignment design is crucial for a railway project since it largely determines its construction investment, lifecycle costs, and various other impacts. However, among the theoretically-infinite numbers of possible alternatives, it is difficult to optimize a vertical alignment matching the drastically-undulating terrain line along the railway, while also considering large structure tradeoffs (such as bridges and tunnels) and various design constraints. To solve this problem, a two-stage method is proposed for automated vertical alignment optimization. In stage I, the railway terrain line is converted to a spatial wave and modelled through the spectral signature analysis of a topography-driven Fast Fourier Transform (FFT). Afterward, by identifying the key terrain characteristic locations based on the derived Fourier series function, feasible search regions for vertical alignment design are determined by processing specific design constraints. For stage II, an A-Star algorithm is customized for vertical alignment search. First, A-Star nodes are discretized within the above feasible search spaces. Then, a comprehensive constraint-handling operator is devised to guarantee a solution’s feasibility during optimization. Moreover, a deterministic simulation algorithm is integrated to create a weighted directed graph for A-Star path generation. Ultimately, the developed method is applied to a complex mountain railway alignment case. The algorithm performances of the two stages are both discussed in detail.
{"title":"Two-stage automated design of railway vertical alignments with topography-driven Fourier transform and constrained A-Star search","authors":"Taoran Song , Hao Pu , Hong Zhang , Paul Schonfeld , Lihui Peng","doi":"10.1016/j.engappai.2026.114061","DOIUrl":"10.1016/j.engappai.2026.114061","url":null,"abstract":"<div><div>Vertical alignment design is crucial for a railway project since it largely determines its construction investment, lifecycle costs, and various other impacts. However, among the theoretically-infinite numbers of possible alternatives, it is difficult to optimize a vertical alignment matching the drastically-undulating terrain line along the railway, while also considering large structure tradeoffs (such as bridges and tunnels) and various design constraints. To solve this problem, a two-stage method is proposed for automated vertical alignment optimization. In stage I, the railway terrain line is converted to a spatial wave and modelled through the spectral signature analysis of a topography-driven Fast Fourier Transform (FFT). Afterward, by identifying the key terrain characteristic locations based on the derived Fourier series function, feasible search regions for vertical alignment design are determined by processing specific design constraints. For stage II, an A-Star algorithm is customized for vertical alignment search. First, A-Star nodes are discretized within the above feasible search spaces. Then, a comprehensive constraint-handling operator is devised to guarantee a solution’s feasibility during optimization. Moreover, a deterministic simulation algorithm is integrated to create a weighted directed graph for A-Star path generation. Ultimately, the developed method is applied to a complex mountain railway alignment case. The algorithm performances of the two stages are both discussed in detail.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"169 ","pages":"Article 114061"},"PeriodicalIF":8.0,"publicationDate":"2026-02-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146175309","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-09DOI: 10.1016/j.engappai.2026.114068
Jinjing Zhu , Ruixiang Bai , Yaoxing Xu , Jiantong Wang , Longchao He , Zhenkun Lei , Cheng Yan
Artificial intelligence (AI), particularly machine learning (ML), has advanced rapidly and is increasingly applied across scientific domains. This paper presents a review of ML applications in predicting mechanical properties and optimizing designs of composite materials, analyzing 165 studies published between 2015 and 2025. Valued in aerospace for lightweight and high-strength properties, composites present modeling challenges due to nonlinear behavior, multi-scale characteristics, and manufacturing sensitivity that limit the accuracy and efficiency of traditional methods. ML techniques have demonstrated remarkable potential in overcoming these limitations through data-driven paradigms, particularly in multi-scale modeling. This study examines the evolutionary trajectory of supervised, unsupervised, reinforcement, and deep learning (DL) within algorithmic frameworks, with a focus on applications in multi-scale analysis of composite materials. At the microscale, ML enables modulus prediction, strength evaluation, and interface analysis via representative volume elements (RVE). For macroscale mechanical analysis, it improves predictions of nonlinear responses, damage progression, and impact behavior in laminates for the rapid optimization design of composite stiffened plates based on buckling and collapse load. Special attention is given to damage evolution, health monitoring, multi-physics field and multi-scale mechanical properties of composite materials. Future research directions, including the development of hybrid data-physics models, digital twin integration, and cross-disciplinary collaboration for enabling breakthroughs in intelligent composite design, are also prospected.
{"title":"Application of machine learning to mechanical properties of composite materials: A ten-year review (2015–2025)","authors":"Jinjing Zhu , Ruixiang Bai , Yaoxing Xu , Jiantong Wang , Longchao He , Zhenkun Lei , Cheng Yan","doi":"10.1016/j.engappai.2026.114068","DOIUrl":"10.1016/j.engappai.2026.114068","url":null,"abstract":"<div><div>Artificial intelligence (AI), particularly machine learning (ML), has advanced rapidly and is increasingly applied across scientific domains. This paper presents a review of ML applications in predicting mechanical properties and optimizing designs of composite materials, analyzing 165 studies published between 2015 and 2025. Valued in aerospace for lightweight and high-strength properties, composites present modeling challenges due to nonlinear behavior, multi-scale characteristics, and manufacturing sensitivity that limit the accuracy and efficiency of traditional methods. ML techniques have demonstrated remarkable potential in overcoming these limitations through data-driven paradigms, particularly in multi-scale modeling. This study examines the evolutionary trajectory of supervised, unsupervised, reinforcement, and deep learning (DL) within algorithmic frameworks, with a focus on applications in multi-scale analysis of composite materials. At the microscale, ML enables modulus prediction, strength evaluation, and interface analysis via representative volume elements (RVE). For macroscale mechanical analysis, it improves predictions of nonlinear responses, damage progression, and impact behavior in laminates for the rapid optimization design of composite stiffened plates based on buckling and collapse load. Special attention is given to damage evolution, health monitoring, multi-physics field and multi-scale mechanical properties of composite materials. Future research directions, including the development of hybrid data-physics models, digital twin integration, and cross-disciplinary collaboration for enabling breakthroughs in intelligent composite design, are also prospected.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"169 ","pages":"Article 114068"},"PeriodicalIF":8.0,"publicationDate":"2026-02-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146175354","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-09DOI: 10.1016/j.engappai.2026.114122
Le Kang , Mengyao Dai , Hui Li , Xinhuai Wang
Vortex beams featuring low-sidelobe, null-steering, co-divergence and multi-mode capabilities hold significant promise for advanced radio-frequency systems. To achieve this, a hybrid neural network (NN)-based framework is proposed for phased arrays. It employs a multilayer perceptron (MLP) followed by a physics-informed network, integrating surrogate modeling for the array with optimization for element excitations. The established model learns the non-linear mapping from input orbital angular momentum (OAM) features to output excitations, and facilitates an optimization process that operates simultaneously with training. The physics-informed network further shifts the learning paradigm from mere data fitting to finding physically guided solutions. This mitigates the reliance on densely layered architectures, and empowers the model with enhanced functionality for beamforming. For verification, numerical experiments involving various scenarios have been conducted. For a 16 × 16 uniform rectangular array without failure, the generated vortex beams achieve sidelobe levels ≤ −19.5 decibel (dB), null depths ≤ −20.2 dB, mode purities ≥85 %, and null steering up to (30°, 30°). Co-divergence and coaxial multi-mode capabilities are also supported. The implementation, involving 512 variables and 5 constraints, requires 1.7 × 107 floating-point operations (FLOPs) and an average time cost of 20 s. Meanwhile, it enables excitation quantization as well as compatibility with diverse array configurations and element patterns. Compared with the state-of-the-art beamforming methods, this work demonstrates concurrent multifunctionality, broader applicability to complex and higher-dimensional problems, and superior computational efficiency.
{"title":"Versatile electromagnetic vortex beamforming for phased array using hybrid neural network-based modeling and optimization","authors":"Le Kang , Mengyao Dai , Hui Li , Xinhuai Wang","doi":"10.1016/j.engappai.2026.114122","DOIUrl":"10.1016/j.engappai.2026.114122","url":null,"abstract":"<div><div>Vortex beams featuring low-sidelobe, null-steering, co-divergence and multi-mode capabilities hold significant promise for advanced radio-frequency systems. To achieve this, a hybrid neural network (NN)-based framework is proposed for phased arrays. It employs a multilayer perceptron (MLP) followed by a physics-informed network, integrating surrogate modeling for the array with optimization for element excitations. The established model learns the non-linear mapping from input orbital angular momentum (OAM) features to output excitations, and facilitates an optimization process that operates simultaneously with training. The physics-informed network further shifts the learning paradigm from mere data fitting to finding physically guided solutions. This mitigates the reliance on densely layered architectures, and empowers the model with enhanced functionality for beamforming. For verification, numerical experiments involving various scenarios have been conducted. For a 16 × 16 uniform rectangular array without failure, the generated vortex beams achieve sidelobe levels ≤ −19.5 decibel (dB), null depths ≤ −20.2 dB, mode purities ≥85 %, and null steering up to (30°, 30°). Co-divergence and coaxial multi-mode capabilities are also supported. The implementation, involving 512 variables and 5 constraints, requires 1.7 × 10<sup>7</sup> floating-point operations (FLOPs) and an average time cost of 20 s. Meanwhile, it enables excitation quantization as well as compatibility with diverse array configurations and element patterns. Compared with the state-of-the-art beamforming methods, this work demonstrates concurrent multifunctionality, broader applicability to complex and higher-dimensional problems, and superior computational efficiency.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"169 ","pages":"Article 114122"},"PeriodicalIF":8.0,"publicationDate":"2026-02-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146175353","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-09DOI: 10.1016/j.engappai.2026.114105
Jing Li , Haidong Zhang , Zhuoma Dawa , Yanping He
In highly uncertain real-world environments, making robust decisions amid incomplete information and the cognitive biases of decision-makers remain critical challenges in project management and other complex system decision-making-related domains. Although the three-way decision-making (3WD) method based on hesitant fuzzy (HF) environments provides an effective approach for managing uncertainty, the current research still has shortcomings in several key areas. On the one hand, the distance formulas that are commonly used for computing hesitant fuzzy elements (HFEs) generally suffer from insufficient sensitivity to the information captured by the score function and weak discriminative power. On the other hand, the process of determining loss functions is subjective and fails to consider the behavioral psychological factors of decision-makers, making it difficult to reflect the cognitive characteristics of humans during actual decision-making processes. These issues collectively limit the adaptability and practicality of the existing decision-making methods in complex real-world environments. To address the aforementioned issues, this study is aimed at constructing an HF 3WD framework that possesses both cognitive rationality and computational robustness. To this end, the core contributions of this work are as follows. First, a novel HF distance measure is developed, significantly improving the ability to distinguish fuzzy information differences. Second, a novel -dominance relation is introduced, and the conditional probability is calculated using a data-driven approach, eliminating the reliance on expert scoring and thereby improving the objectivity and accuracy of the conditional probability. Finally, an objective loss function is established, effectively capturing the decision-maker’s nonlinear value perceptions and comparative psychology in gain and loss scenarios. Furthermore, comparative experiments and parameter analyses are conducted in big data scenarios to validate the fact that the proposed method outperforms the existing methods in terms of classification accuracy and decision stability, demonstrating superior effectiveness and robustness. We believe that by simulating human judgments made under uncertainty, this method opens up new avenues for implementing artificial intelligence-based decision-making systems in high-risk scenarios.
{"title":"Hesitant fuzzy three-way decision-making for large-scale data based on a new distance measure and behavioral theory","authors":"Jing Li , Haidong Zhang , Zhuoma Dawa , Yanping He","doi":"10.1016/j.engappai.2026.114105","DOIUrl":"10.1016/j.engappai.2026.114105","url":null,"abstract":"<div><div>In highly uncertain real-world environments, making robust decisions amid incomplete information and the cognitive biases of decision-makers remain critical challenges in project management and other complex system decision-making-related domains. Although the three-way decision-making (3WD) method based on hesitant fuzzy (HF) environments provides an effective approach for managing uncertainty, the current research still has shortcomings in several key areas. On the one hand, the distance formulas that are commonly used for computing hesitant fuzzy elements (HFEs) generally suffer from insufficient sensitivity to the information captured by the score function and weak discriminative power. On the other hand, the process of determining loss functions is subjective and fails to consider the behavioral psychological factors of decision-makers, making it difficult to reflect the cognitive characteristics of humans during actual decision-making processes. These issues collectively limit the adaptability and practicality of the existing decision-making methods in complex real-world environments. To address the aforementioned issues, this study is aimed at constructing an HF 3WD framework that possesses both cognitive rationality and computational robustness. To this end, the core contributions of this work are as follows. First, a novel HF distance measure is developed, significantly improving the ability to distinguish fuzzy information differences. Second, a novel <span><math><mi>o</mi></math></span>-dominance relation is introduced, and the conditional probability is calculated using a data-driven approach, eliminating the reliance on expert scoring and thereby improving the objectivity and accuracy of the conditional probability. Finally, an objective loss function is established, effectively capturing the decision-maker’s nonlinear value perceptions and comparative psychology in gain and loss scenarios. Furthermore, comparative experiments and parameter analyses are conducted in big data scenarios to validate the fact that the proposed method outperforms the existing methods in terms of classification accuracy and decision stability, demonstrating superior effectiveness and robustness. We believe that by simulating human judgments made under uncertainty, this method opens up new avenues for implementing artificial intelligence-based decision-making systems in high-risk scenarios.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"169 ","pages":"Article 114105"},"PeriodicalIF":8.0,"publicationDate":"2026-02-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146174976","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-09DOI: 10.1016/j.engappai.2026.114107
Muhammad Arsal, Ubaid ur Rehman, Abaid ur Rehman Virk
E-commerce websites are increasingly depending on machine learning (ML) algorithms to improve their recommendation systems to deliver personalized user experiences and enhance customer satisfaction. ML models provide useful solutions, but choosing the best ML model is still difficult, especially for e-commerce platforms' recommendation systems. This selection challenge involves dual aspects (bipolarity), which means that both the positive and negative aspects of the evaluation criteria for ML models must be taken into consideration. It is a multi-criteria decision-making (MCDM) problem with uncertainty. However, previous research has failed to consider the bipolar character of these model selection criteria. Finding the right weights for the evaluation criteria also becomes important in MCDM, particularly when handling the bipolarity and uncertainty present in actual e-commerce situations. To overcome this limitation, we introduce a novel bipolar fuzzy (BF)-method based on the removal effects of criteria (MEREC)-elimination and choice translating reality (ELECTRE-I) methodology that combines the method based on the removal effects of criteria (MEREC) method's objective weighting capability with the elimination and choice translating reality (ELECTRE-I) technique's decision-ranking power. The methodology is used in a real-world case study that focuses on choosing the best ML model for an e-commerce platform's recommendation system. The proposed structure provides a strong decision-support tool that tackles the difficulties associated with choosing an ML algorithm and captures the dual nature of evaluation criteria. Comparative findings show how well the BF-MEREC-ELECTRE-I method works to provide data-driven, interpretable, and useful recommendations for e-commerce applications.
{"title":"Assessment and evaluation of machine learning algorithms for recommendation system of E-commerce based on bipolar fuzzy- method based on the removal effects of criteria-elimination and choice translating Reality-I approach","authors":"Muhammad Arsal, Ubaid ur Rehman, Abaid ur Rehman Virk","doi":"10.1016/j.engappai.2026.114107","DOIUrl":"10.1016/j.engappai.2026.114107","url":null,"abstract":"<div><div>E-commerce websites are increasingly depending on machine learning (ML) algorithms to improve their recommendation systems to deliver personalized user experiences and enhance customer satisfaction. ML models provide useful solutions, but choosing the best ML model is still difficult, especially for e-commerce platforms' recommendation systems. This selection challenge involves dual aspects (bipolarity), which means that both the positive and negative aspects of the evaluation criteria for ML models must be taken into consideration. It is a multi-criteria decision-making (MCDM) problem with uncertainty. However, previous research has failed to consider the bipolar character of these model selection criteria. Finding the right weights for the evaluation criteria also becomes important in MCDM, particularly when handling the bipolarity and uncertainty present in actual e-commerce situations. To overcome this limitation, we introduce a novel bipolar fuzzy (BF)-method based on the removal effects of criteria (MEREC)-elimination and choice translating reality (ELECTRE-I) methodology that combines the method based on the removal effects of criteria (MEREC) method's objective weighting capability with the elimination and choice translating reality (ELECTRE-I) technique's decision-ranking power. The methodology is used in a real-world case study that focuses on choosing the best ML model for an e-commerce platform's recommendation system. The proposed structure provides a strong decision-support tool that tackles the difficulties associated with choosing an ML algorithm and captures the dual nature of evaluation criteria. Comparative findings show how well the BF-MEREC-ELECTRE-I method works to provide data-driven, interpretable, and useful recommendations for e-commerce applications.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"169 ","pages":"Article 114107"},"PeriodicalIF":8.0,"publicationDate":"2026-02-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146174977","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-08DOI: 10.1016/j.engappai.2026.114113
Mahdi Movahedian Moghaddam , Hassan Dana Mazraeh , Kourosh Parand
This study introduces a reinforcement learning framework using Double Deep Q-Networks (DDQN) to discover interpretable, symbolic solutions to integral and integro-differential equations. The method leverages Context-Free Grammars to guide expression generation and Physics-Informed Neural Networks (PINN) to optimize coefficients. Evaluated on diverse equations (Fredholm, Volterra, and fractional types), DDQN outperforms standard DQN in stability, convergence, and interpretability, and succeeds in complex cases like population models where DQN fails. This approach is highly valuable for engineering domains such as control theory, electromagnetics, and fluid mechanics, where such equations are prevalent. It provides engineers with compact, analytical expressions that offer immediate physical insight, unlike opaque numerical solutions. These symbolic models enable faster system analysis, real-time control, and robust design optimization, bridging the gap between high-fidelity simulation and practical, interpretable models.
{"title":"A double deep Q network-powered approach to discovering symbolic solutions for nonlinear integral equations","authors":"Mahdi Movahedian Moghaddam , Hassan Dana Mazraeh , Kourosh Parand","doi":"10.1016/j.engappai.2026.114113","DOIUrl":"10.1016/j.engappai.2026.114113","url":null,"abstract":"<div><div>This study introduces a reinforcement learning framework using Double Deep Q-Networks (DDQN) to discover interpretable, symbolic solutions to integral and integro-differential equations. The method leverages Context-Free Grammars to guide expression generation and Physics-Informed Neural Networks (PINN) to optimize coefficients. Evaluated on diverse equations (Fredholm, Volterra, and fractional types), DDQN outperforms standard DQN in stability, convergence, and interpretability, and succeeds in complex cases like population models where DQN fails. This approach is highly valuable for engineering domains such as control theory, electromagnetics, and fluid mechanics, where such equations are prevalent. It provides engineers with compact, analytical expressions that offer immediate physical insight, unlike opaque numerical solutions. These symbolic models enable faster system analysis, real-time control, and robust design optimization, bridging the gap between high-fidelity simulation and practical, interpretable models.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"169 ","pages":"Article 114113"},"PeriodicalIF":8.0,"publicationDate":"2026-02-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146174973","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-08DOI: 10.1016/j.engappai.2026.114115
Pengfei Cui , Haowei Wu , Jun Yin
The goal of deep image clustering is to assign images to their corresponding clusters in an unsupervised manner. Contrastive Clustering with Effective Sample pairs construction (CCES) (Yin et al., 2023) has demonstrated the effectiveness of combining ContrastiveCrop (Peng et al., 2022) with nearest-neighbor mining for clustering. However, a key limitation of CCES is that it treats augmentations of different samples from the same class as negative pairs, which introduces false negatives and conflicts with the goal of clustering. To address this issue, we propose a novel contrastive clustering framework termed Reinforced Semantic Information Acquiring for Contrastive Clustering (RASCC). We extend CCES by introducing a pseudo-label-guided dual-level fine-tuning strategy that mitigates the adverse impact of false negatives and improves clustering quality. Specifically, RASCC consists of two stages. In Stage 1, we follow CCES to learn discriminative features and generate high-confidence pseudo labels based on clustering predictions. In Stage 2, we leverage these pseudo labels to perform dual-level fine-tuning. At the instance level, we mitigate false negatives by excluding same-pseudo-label pairs from the negative set. At the cluster level, we improve assignment quality and balance by refining cluster predictions with a class-balanced self-labeling loss that each sample is weighted inversely proportional to the frequency of its high-confidence pseudo-label. We conduct experiments on four challenging datasets. The proposed method surpasses many state-of-the-art deep clustering algorithms, demonstrating the effectiveness of sample pair construction and fine-tuning strategy.
{"title":"Reinforced semantic information acquiring for contrastive clustering","authors":"Pengfei Cui , Haowei Wu , Jun Yin","doi":"10.1016/j.engappai.2026.114115","DOIUrl":"10.1016/j.engappai.2026.114115","url":null,"abstract":"<div><div>The goal of deep image clustering is to assign images to their corresponding clusters in an unsupervised manner. Contrastive Clustering with Effective Sample pairs construction (CCES) (Yin et al., 2023) has demonstrated the effectiveness of combining ContrastiveCrop (Peng et al., 2022) with nearest-neighbor mining for clustering. However, a key limitation of CCES is that it treats augmentations of different samples from the same class as negative pairs, which introduces false negatives and conflicts with the goal of clustering. To address this issue, we propose a novel contrastive clustering framework termed Reinforced Semantic Information Acquiring for Contrastive Clustering (RASCC). We extend CCES by introducing a pseudo-label-guided dual-level fine-tuning strategy that mitigates the adverse impact of false negatives and improves clustering quality. Specifically, RASCC consists of two stages. In Stage 1, we follow CCES to learn discriminative features and generate high-confidence pseudo labels based on clustering predictions. In Stage 2, we leverage these pseudo labels to perform dual-level fine-tuning. At the instance level, we mitigate false negatives by excluding same-pseudo-label pairs from the negative set. At the cluster level, we improve assignment quality and balance by refining cluster predictions with a class-balanced self-labeling loss that each sample is weighted inversely proportional to the frequency of its high-confidence pseudo-label. We conduct experiments on four challenging datasets. The proposed method surpasses many state-of-the-art deep clustering algorithms, demonstrating the effectiveness of sample pair construction and fine-tuning strategy.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"169 ","pages":"Article 114115"},"PeriodicalIF":8.0,"publicationDate":"2026-02-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146175021","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号