Pub Date : 2026-02-05DOI: 10.1016/j.aei.2026.104395
Jian Long , Bin Wang , Haifei Peng , Hengmin Zhang
Chemical process data exhibit both long-term physical dependencies and short-term dynamic variations due to complex interactions among variables. To simultaneously model these heterogeneous dependencies, this paper proposes a Multi-Source Attention Graph Neural Network (MSAGNN) for soft sensing in chemical processes. MSAGNN adopts a dual-path graph recurrent architecture, where a static graph encodes prior physical relationships, and an adaptive graph structure learning module dynamically captures time-varying correlations from data. A multi-source attention mechanism is further introduced to integrate node and neighborhood information and enhance the representation of spatial–temporal dependencies. The proposed MSAGNN is evaluated on three representative industrial processes, including the Debutanizer Column (DC), the Tennessee Eastman (TE) process, and the Fluid Catalytic Cracking (FCC) unit. Experimental results show that MSAGNN consistently achieves lower RMSE, MAE, and MAPE, and higher R than state-of-the-art deep learning and graph-based models, demonstrating its superior prediction accuracy and robustness. Visualization of the learned dynamic graphs and attention scores indicates that MSAGNN can reveal meaningful variable interactions, confirming the effectiveness and interpretability of the proposed approach for complex chemical processes.
{"title":"A Multi-Source Attention Graph Neural Network for modeling long and short-term dependencies in chemical process forecasting","authors":"Jian Long , Bin Wang , Haifei Peng , Hengmin Zhang","doi":"10.1016/j.aei.2026.104395","DOIUrl":"10.1016/j.aei.2026.104395","url":null,"abstract":"<div><div>Chemical process data exhibit both long-term physical dependencies and short-term dynamic variations due to complex interactions among variables. To simultaneously model these heterogeneous dependencies, this paper proposes a Multi-Source Attention Graph Neural Network (MSAGNN) for soft sensing in chemical processes. MSAGNN adopts a dual-path graph recurrent architecture, where a static graph encodes prior physical relationships, and an adaptive graph structure learning module dynamically captures time-varying correlations from data. A multi-source attention mechanism is further introduced to integrate node and neighborhood information and enhance the representation of spatial–temporal dependencies. The proposed MSAGNN is evaluated on three representative industrial processes, including the Debutanizer Column (DC), the Tennessee Eastman (TE) process, and the Fluid Catalytic Cracking (FCC) unit. Experimental results show that MSAGNN consistently achieves lower RMSE, MAE, and MAPE, and higher R<span><math><msup><mrow></mrow><mrow><mn>2</mn></mrow></msup></math></span> than state-of-the-art deep learning and graph-based models, demonstrating its superior prediction accuracy and robustness. Visualization of the learned dynamic graphs and attention scores indicates that MSAGNN can reveal meaningful variable interactions, confirming the effectiveness and interpretability of the proposed approach for complex chemical processes.</div></div>","PeriodicalId":50941,"journal":{"name":"Advanced Engineering Informatics","volume":"71 ","pages":"Article 104395"},"PeriodicalIF":9.9,"publicationDate":"2026-02-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146188619","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-05DOI: 10.1016/j.aei.2026.104398
Yujie Lu , Tao Zhong , Haoyu Deng , Shuo Wang , Chuan Yang , Xianzhong Zhao
High-fidelity 3D modeling of large-scale construction sites requires the fusion of multi-source point clouds to satisfy the diverse requirements of various engineering tasks. A primary challenge in this process is the non-uniform quality of the data, as the relative precision between different sources varies spatially across the site. This variability can cause holistic fusion approaches to erroneously discard high-quality local data. To address this, the presented framework reframes the global fusion challenge into a more tractable problem of local partitioning and optimization. A novel hierarchical partitioning method creates local units that are geometrically adaptive to building components, enhancing both geometric continuity and processing efficiency. The methodology consists of three sequential stages: (1) high-precision registration of all source data; (2) hierarchical partitioning at both architectural and sub-component levels to create basic fusion units; and (3) multi-strategy local filtering based on data quality within each unit. The framework was validated on a challenging case involving the fusion of multi-source, image-based point clouds from a super-tall building site. The results demonstrate a precision improvement of 6.8% for the entire site and up to 31.9% for detailed regions compared to the simply merged data. The methodology enhances the integrity of as-built models and improves the representation of surface textures and structural dimensions, providing a reliable data foundation for downstream construction management tasks such as spatial progress monitoring.
{"title":"A 3D geometry adaptive approach for fusing heterogeneous point clouds in construction sites","authors":"Yujie Lu , Tao Zhong , Haoyu Deng , Shuo Wang , Chuan Yang , Xianzhong Zhao","doi":"10.1016/j.aei.2026.104398","DOIUrl":"10.1016/j.aei.2026.104398","url":null,"abstract":"<div><div>High-fidelity 3D modeling of large-scale construction sites requires the fusion of multi-source point clouds to satisfy the diverse requirements of various engineering tasks. A primary challenge in this process is the non-uniform quality of the data, as the relative precision between different sources varies spatially across the site. This variability can cause holistic fusion approaches to erroneously discard high-quality local data. To address this, the presented framework reframes the global fusion challenge into a more tractable problem of local partitioning and optimization. A novel hierarchical partitioning method creates local units that are geometrically adaptive to building components, enhancing both geometric continuity and processing efficiency. The methodology consists of three sequential stages: (1) high-precision registration of all source data; (2) hierarchical partitioning at both architectural and sub-component levels to create basic fusion units; and (3) multi-strategy local filtering based on data quality within each unit. The framework was validated on a challenging case involving the fusion of multi-source, image-based point clouds from a super-tall building site. The results demonstrate a precision improvement of 6.8% for the entire site and up to 31.9% for detailed regions compared to the simply merged data. The methodology enhances the integrity of as-built models and improves the representation of surface textures and structural dimensions, providing a reliable data foundation for downstream construction management tasks such as spatial progress monitoring.</div></div>","PeriodicalId":50941,"journal":{"name":"Advanced Engineering Informatics","volume":"71 ","pages":"Article 104398"},"PeriodicalIF":9.9,"publicationDate":"2026-02-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146187868","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-05DOI: 10.1016/j.aei.2026.104393
Qiubing Ren , Ruizhe Liu , Mingchao Li , Zhiyong Qi , Xuhuang Du , Jin Yuan
Accurate dam displacement prediction is vital for optimizing maintenance and ensuring structural safety. Nevertheless, current models often struggle to effectively capture the complex relationships between structural responses and environmental variables, alongside the interactions between temporal dynamics and multivariate data, resulting in suboptimal predictive accuracy. Therefore, we propose a dual-branch interactive fusion network (DBIFN) for dam displacement prediction using parallel temporal representation and gated cross-attention. The dual-branch architecture, which parallelly integrates the enhanced Transformer (eTransformer) and long short-term memory (LSTM), is designed to optimize feature extraction and interaction modeling across multiple dimensions. Specifically, eTransformer is dedicated to extracting features from targeted displacement sequences, while LSTM effectively processes auxiliary environmental dynamics, enabling a comprehensive analysis of underlying patterns within monitoring data. To fully fuse the interpreted temporal features from dual-branch outputs, we introduce a new cross-attention module to utilize the multi-dimensional gated attention unit to efficiently encode them into semantic representations, followed by a Kolmogorov-Arnold network mapping for further representation enhancement. The effectiveness of the proposed model is validated using real-world monitoring datasets collected from a concrete dam project, with experiments conducted across multiple monitoring points. Results demonstrate that DBIFN achieves superior prediction accuracy compared to both single-branch and conventional baseline models. Across all monitoring points, the proposed model can effectively capture temporal variations, attaining an average coefficient of determination of over 0.95 on the test set and outperforming comparative models in most metrics. Furthermore, statistical significance testing confirms the reliability and reproducibility of the results, while computational efficiency is maintained within inference time constraints. These findings offer valuable insights into the practical application of DBIFN-based monitoring models and support informed decision-making.
{"title":"Dual-branch interactive fusion network for dam displacement prediction based on parallel temporal representation and gated cross-attention","authors":"Qiubing Ren , Ruizhe Liu , Mingchao Li , Zhiyong Qi , Xuhuang Du , Jin Yuan","doi":"10.1016/j.aei.2026.104393","DOIUrl":"10.1016/j.aei.2026.104393","url":null,"abstract":"<div><div>Accurate dam displacement prediction is vital for optimizing maintenance and ensuring structural safety. Nevertheless, current models often struggle to effectively capture the complex relationships between structural responses and environmental variables, alongside the interactions between temporal dynamics and multivariate data, resulting in suboptimal predictive accuracy. Therefore, we propose a dual-branch interactive fusion network (DBIFN) for dam displacement prediction using parallel temporal representation and gated cross-attention. The dual-branch architecture, which parallelly integrates the enhanced Transformer (eTransformer) and long short-term memory (LSTM), is designed to optimize feature extraction and interaction modeling across multiple dimensions. Specifically, eTransformer is dedicated to extracting features from targeted displacement sequences, while LSTM effectively processes auxiliary environmental dynamics, enabling a comprehensive analysis of underlying patterns within monitoring data. To fully fuse the interpreted temporal features from dual-branch outputs, we introduce a new cross-attention module to utilize the multi-dimensional gated attention unit to efficiently encode them into semantic representations, followed by a Kolmogorov-Arnold network mapping for further representation enhancement. The effectiveness of the proposed model is validated using real-world monitoring datasets collected from a concrete dam project, with experiments conducted across multiple monitoring points. Results demonstrate that DBIFN achieves superior prediction accuracy compared to both single-branch and conventional baseline models. Across all monitoring points, the proposed model can effectively capture temporal variations, attaining an average coefficient of determination of over 0.95 on the test set and outperforming comparative models in most metrics. Furthermore, statistical significance testing confirms the reliability and reproducibility of the results, while computational efficiency is maintained within inference time constraints. These findings offer valuable insights into the practical application of DBIFN-based monitoring models and support informed decision-making.</div></div>","PeriodicalId":50941,"journal":{"name":"Advanced Engineering Informatics","volume":"71 ","pages":"Article 104393"},"PeriodicalIF":9.9,"publicationDate":"2026-02-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146188615","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-05DOI: 10.1016/j.aei.2026.104417
Jiangting Zhao , Xiaoyu Zhang , Zhanfeng Qi , Kunpeng He , Yandong Yang
Visual-inertial navigation, as a widely adopted integrated navigation approach, serves as a key enabling technology for unmanned systems to operate in unknown environments. To address the challenges of degraded adaptability and localization accuracy in practical scenarios, which often result from diverse platform motions and complex environmental conditions, this paper proposes a robust visual simultaneous localization and mapping (VSLAM) algorithm named MSHK-SLAM. The method is built upon a multi-source information composite constraint model and a heterogeneous kernel function optimization. First, an adaptive keyframe (KF) selection strategy is introduced to prevent KF omission caused by abrupt viewpoint changes during motion, while providing more reliable information for back-end processing. Second, an optimization method based on a heterogeneous kernel function using Gaussian–Cauchy mixture correntropy (GCM) is designed to effectively suppress interference from mixed environmental noise, further enhancing system robustness. Experimental results on public datasets demonstrate that MSHK-SLAM outperforms other state-of-the-art algorithms in both adaptability and accuracy under complex environmental conditions.
{"title":"Robust VSLAM algorithm based on multi-source information composite constraint model and heterogeneous kernel function optimization","authors":"Jiangting Zhao , Xiaoyu Zhang , Zhanfeng Qi , Kunpeng He , Yandong Yang","doi":"10.1016/j.aei.2026.104417","DOIUrl":"10.1016/j.aei.2026.104417","url":null,"abstract":"<div><div>Visual-inertial navigation, as a widely adopted integrated navigation approach, serves as a key enabling technology for unmanned systems to operate in unknown environments. To address the challenges of degraded adaptability and localization accuracy in practical scenarios, which often result from diverse platform motions and complex environmental conditions, this paper proposes a robust visual simultaneous localization and mapping (VSLAM) algorithm named MSHK-SLAM. The method is built upon a multi-source information composite constraint model and a heterogeneous kernel function optimization. First, an adaptive keyframe (KF) selection strategy is introduced to prevent KF omission caused by abrupt viewpoint changes during motion, while providing more reliable information for back-end processing. Second, an optimization method based on a heterogeneous kernel function using Gaussian–Cauchy mixture correntropy (GCM) is designed to effectively suppress interference from mixed environmental noise, further enhancing system robustness. Experimental results on public datasets demonstrate that MSHK-SLAM outperforms other state-of-the-art algorithms in both adaptability and accuracy under complex environmental conditions.</div></div>","PeriodicalId":50941,"journal":{"name":"Advanced Engineering Informatics","volume":"71 ","pages":"Article 104417"},"PeriodicalIF":9.9,"publicationDate":"2026-02-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146187867","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-05DOI: 10.1016/j.aei.2026.104418
Yang Tian , Lei Pan , Yicai Liu , Yahui Zhang , Yushu Li , Yihong Fan , Xiangyu Wang , Liang Li , Bingxin Ma
Path tracking control (PTC) as a core function of autonomous vehicles, has been widely studied, and many prescribed performance control (PPC) methods have been applied. However, PTC faces challenges such as model nonlinearity, parameter variations, and external disturbances. The initial constraint problem (ICP) also limits PPC, as the tracking error often exceeds the performance constraint boundary at the initial stage and must be regulated back within the permissible set within a reasonable time. This article proposes a model-free PPC scheme that considers ICP, achieving the appointed time performance and adaptive adjustment of envelope curvature. First, the path tracking error is abstracted as a preview error to avoid the complexities of time-varying nonlinearities. An appointed time prescribed performance function is proposed, where the envelope becomes concave to accelerate error convergence for small initial errors and convex to mitigate ICP for large initial errors. Subsequently, this function is integrated with a nonlinear manifold control method to ensure the appointed time stability of the preview error and its derivative. Simulations and vehicle experiment results verify the effectiveness, robustness, and efficiency of the proposed controller.
{"title":"Appointed-time prescribed performance path tracking control for autonomous vehicles considering initial constraint problem","authors":"Yang Tian , Lei Pan , Yicai Liu , Yahui Zhang , Yushu Li , Yihong Fan , Xiangyu Wang , Liang Li , Bingxin Ma","doi":"10.1016/j.aei.2026.104418","DOIUrl":"10.1016/j.aei.2026.104418","url":null,"abstract":"<div><div>Path tracking control (PTC) as a core function of autonomous vehicles, has been widely studied, and many prescribed performance control (PPC) methods have been applied. However, PTC faces challenges such as model nonlinearity, parameter variations, and external disturbances. The initial constraint problem (ICP) also limits PPC, as the tracking error often exceeds the performance constraint boundary at the initial stage and must be regulated back within the permissible set within a reasonable time. This article proposes a model-free PPC scheme that considers ICP, achieving the appointed time performance and adaptive adjustment of envelope curvature. First, the path tracking error is abstracted as a preview error to avoid the complexities of time-varying nonlinearities. An appointed time prescribed performance function is proposed, where the envelope becomes concave to accelerate error convergence for small initial errors and convex to mitigate ICP for large initial errors. Subsequently, this function is integrated with a nonlinear manifold control method to ensure the appointed time stability of the preview error and its derivative. Simulations and vehicle experiment results verify the effectiveness, robustness, and efficiency of the proposed controller.</div></div>","PeriodicalId":50941,"journal":{"name":"Advanced Engineering Informatics","volume":"71 ","pages":"Article 104418"},"PeriodicalIF":9.9,"publicationDate":"2026-02-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146188121","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-05DOI: 10.1016/j.aei.2026.104399
Yibang Zhou , Xiangkai Li , Yue Yin , Liwei Chen , Haiming Xu , Jiajie Fu , Aoyu Zhou , Jianjun Yi
Robot assembly is a core task in industrial manufacturing systems, and its autonomy and stability have been a research focus. Traditional methods are mostly limited to specific assembly tasks, struggling to robustly accomplish the entire assembly process in unstructured scenarios. This paper proposes a reinforcement learning method based on a hierarchical diffusion policy for robot assembly tasks. The proposed method decomposes the assembly policy space into a high-level task planning controller and a low-level policy controller. The low-level policy controller employs a multimodal conditional diffusion policy, leveraging the environmental perception and task generalization capabilities of visual observations, as well as the contact distribution fitting capability of force observations, to efficiently and robustly accomplish assembly subtasks. The high-level task planning controller learns a task scheduling policy, supporting multi-timescale decision-making in complex dynamic assembly environments. Experimental results demonstrate that the proposed method can stably and efficiently complete assembly tasks on components of different shapes, even in the presence of significant initial pose errors, exhibiting higher learning efficiency and generalization performance compared to traditional methods.
{"title":"Robust robotic assembly via hierarchical diffusion policy-guided reinforcement learning","authors":"Yibang Zhou , Xiangkai Li , Yue Yin , Liwei Chen , Haiming Xu , Jiajie Fu , Aoyu Zhou , Jianjun Yi","doi":"10.1016/j.aei.2026.104399","DOIUrl":"10.1016/j.aei.2026.104399","url":null,"abstract":"<div><div>Robot assembly is a core task in industrial manufacturing systems, and its autonomy and stability have been a research focus. Traditional methods are mostly limited to specific assembly tasks, struggling to robustly accomplish the entire assembly process in unstructured scenarios. This paper proposes a reinforcement learning method based on a hierarchical diffusion policy for robot assembly tasks. The proposed method decomposes the assembly policy space into a high-level task planning controller and a low-level policy controller. The low-level policy controller employs a multimodal conditional diffusion policy, leveraging the environmental perception and task generalization capabilities of visual observations, as well as the contact distribution fitting capability of force observations, to efficiently and robustly accomplish assembly subtasks. The high-level task planning controller learns a task scheduling policy, supporting multi-timescale decision-making in complex dynamic assembly environments. Experimental results demonstrate that the proposed method can stably and efficiently complete assembly tasks on components of different shapes, even in the presence of significant initial pose errors, exhibiting higher learning efficiency and generalization performance compared to traditional methods.</div></div>","PeriodicalId":50941,"journal":{"name":"Advanced Engineering Informatics","volume":"71 ","pages":"Article 104399"},"PeriodicalIF":9.9,"publicationDate":"2026-02-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146188826","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-04DOI: 10.1016/j.aei.2026.104368
Guang Yu , Jia Liu , Lai Hu , Anbang Jiang , Liping Wang
Aiming to improve the machining efficiency and accuracy of five-axis hybrid machine tool (FaHMT), the kinematics system of FaHMT was modeled. According to the kinematics model, the velocity mapping model and error transfer model of the FaHMT were obtained by derivation and perturbation differentiation. The three-loop control model of permanent magnet synchronous motor used in machine tools was focused on modeling. Meanwhile, the mapping characteristics of the movement speed of the tool and the feed speed of each driving shaft under different postures of the parallel mechanism were analyzed. In the acceleration experiment, the position where the inflection point of error change appeared at a = 4.6 mm/s2. When the acceleration was greater than this value, the XW axis follow-up error increased significantly. For Z1, Z2 and Z3 axes, their movement positions and errors were consistent basically. The modeling method and practical conclusions analyzed provide effective theoretical support and experimental verification for studying kinematics and acceleration analysis of FaHMT.
{"title":"Motion system modeling and acceleration analysis of five-axis hybrid machine tool","authors":"Guang Yu , Jia Liu , Lai Hu , Anbang Jiang , Liping Wang","doi":"10.1016/j.aei.2026.104368","DOIUrl":"10.1016/j.aei.2026.104368","url":null,"abstract":"<div><div>Aiming to improve the machining efficiency and accuracy of five-axis hybrid machine tool (FaHMT), the kinematics system of FaHMT was modeled. According to the kinematics model, the velocity mapping model and error transfer model of the FaHMT were obtained by derivation and perturbation differentiation. The three-loop control model of permanent magnet synchronous motor used in machine tools was focused on modeling. Meanwhile, the mapping characteristics of the movement speed of the tool and the feed speed of each driving shaft under different postures of the parallel mechanism were analyzed. In the acceleration experiment, the position where the inflection point of error change appeared at <em>a</em> = 4.6 mm/s<sup>2</sup>. When the acceleration was greater than this value, the <em>X<sub>W</sub></em> axis follow-up error increased significantly. For <em>Z</em><sub>1</sub>, <em>Z</em><sub>2</sub> and <em>Z</em><sub>3</sub> axes, their movement positions and errors were consistent basically. The modeling method and practical conclusions analyzed provide effective theoretical support and experimental verification for studying kinematics and acceleration analysis of FaHMT.</div></div>","PeriodicalId":50941,"journal":{"name":"Advanced Engineering Informatics","volume":"71 ","pages":"Article 104368"},"PeriodicalIF":9.9,"publicationDate":"2026-02-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146188618","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-04DOI: 10.1016/j.aei.2026.104400
Jaeyoung Lee , Hyuna Jeon , Uiin Kim , Misuk Kim
Signal data are essential for condition monitoring, fault diagnosis, and decision-making across industrial domains, and research leveraging signal data has been actively pursued in areas such as healthcare and manufacturing. However, acquiring such data is costly and difficult due to factors such as the risk of equipment damage, the need for expert labeling, and the scarcity of fault data. Moreover, collected data often contain sensitive operational information, making sharing difficult, and enterprises are restricted from using high-performance models hosted on external servers due to security concerns. To address these challenges, we propose BearGen, a novel framework that combines the strong generative capabilities of Large Language Models (LLMs) with the precise data distribution learning of diffusion models to synthesize high-quality signal data in on-premise environments. BearGen first employs an LLM to generate descriptions of existing signals and then conditions a description-guided diffusion model on these descriptions to generate high-quality synthetic signals. We evaluated BearGen on eight publicly available bearing fault diagnosis datasets, and the results showed superior performance compared to existing approaches. In addition, we experimentally validated the reliability and usefulness of the generated signal descriptions. Further experiments under conditions simulating real industrial environments — such as limited data availability and severe data imbalance — verified the practical applicability of the framework. By operating in on-premise environments, BearGen resolves data security concerns while alleviating data scarcity and imbalance. Furthermore, by providing natural language descriptions, it enhances interpretability and offers significant potential for decision support in real-world industrial applications.
{"title":"BearGen: LLM-guided signal generation framework for bearing fault diagnosis","authors":"Jaeyoung Lee , Hyuna Jeon , Uiin Kim , Misuk Kim","doi":"10.1016/j.aei.2026.104400","DOIUrl":"10.1016/j.aei.2026.104400","url":null,"abstract":"<div><div>Signal data are essential for condition monitoring, fault diagnosis, and decision-making across industrial domains, and research leveraging signal data has been actively pursued in areas such as healthcare and manufacturing. However, acquiring such data is costly and difficult due to factors such as the risk of equipment damage, the need for expert labeling, and the scarcity of fault data. Moreover, collected data often contain sensitive operational information, making sharing difficult, and enterprises are restricted from using high-performance models hosted on external servers due to security concerns. To address these challenges, we propose <span>BearGen</span>, a novel framework that combines the strong generative capabilities of Large Language Models (LLMs) with the precise data distribution learning of diffusion models to synthesize high-quality signal data in on-premise environments. <span>BearGen</span> first employs an LLM to generate descriptions of existing signals and then conditions a description-guided diffusion model on these descriptions to generate high-quality synthetic signals. We evaluated <span>BearGen</span> on eight publicly available bearing fault diagnosis datasets, and the results showed superior performance compared to existing approaches. In addition, we experimentally validated the reliability and usefulness of the generated signal descriptions. Further experiments under conditions simulating real industrial environments — such as limited data availability and severe data imbalance — verified the practical applicability of the framework. By operating in on-premise environments, <span>BearGen</span> resolves data security concerns while alleviating data scarcity and imbalance. Furthermore, by providing natural language descriptions, it enhances interpretability and offers significant potential for decision support in real-world industrial applications.</div></div>","PeriodicalId":50941,"journal":{"name":"Advanced Engineering Informatics","volume":"71 ","pages":"Article 104400"},"PeriodicalIF":9.9,"publicationDate":"2026-02-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146188611","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-04DOI: 10.1016/j.aei.2026.104422
Maximilian-Peter Radtke , Marco F. Huber , Jürgen Bock
Combining system knowledge with deep learning for fault diagnosis in industrial applications offers the potential to reduce the dependency of deep learning algorithms on extensive labeled datasets. However, existing methods often rely on highly specialized, problem-specific knowledge or demand detailed physical insights into the system, which limits their generalizability. Additionally, inconsistencies in knowledge representation hinder the ability to compare and build upon prior approaches. In this work, we address these challenges by leveraging commonly available knowledge about the phase structure of systems and the hierarchical organization of condition spaces. This information is systematically represented using knowledge graphs (KGs) based on the Resource Description Framework (RDF). To integrate this knowledge into deep learning, we transform the input data and the corresponding labels based on the KGs, and employ a graph neural network (GNN) trained with a semantic loss function informed by the knowledge about the condition space. The proposed approach is evaluated on three diverse datasets with varying characteristics under the two scenarios of domain generalization and novel fault detection.
{"title":"RDF-based knowledge graph integration with deep learning for fault diagnosis","authors":"Maximilian-Peter Radtke , Marco F. Huber , Jürgen Bock","doi":"10.1016/j.aei.2026.104422","DOIUrl":"10.1016/j.aei.2026.104422","url":null,"abstract":"<div><div>Combining system knowledge with deep learning for fault diagnosis in industrial applications offers the potential to reduce the dependency of deep learning algorithms on extensive labeled datasets. However, existing methods often rely on highly specialized, problem-specific knowledge or demand detailed physical insights into the system, which limits their generalizability. Additionally, inconsistencies in knowledge representation hinder the ability to compare and build upon prior approaches. In this work, we address these challenges by leveraging commonly available knowledge about the phase structure of systems and the hierarchical organization of condition spaces. This information is systematically represented using knowledge graphs (KGs) based on the Resource Description Framework (RDF). To integrate this knowledge into deep learning, we transform the input data and the corresponding labels based on the KGs, and employ a graph neural network (GNN) trained with a semantic loss function informed by the knowledge about the condition space. The proposed approach is evaluated on three diverse datasets with varying characteristics under the two scenarios of domain generalization and novel fault detection.</div></div>","PeriodicalId":50941,"journal":{"name":"Advanced Engineering Informatics","volume":"71 ","pages":"Article 104422"},"PeriodicalIF":9.9,"publicationDate":"2026-02-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146188616","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2026-02-03DOI: 10.1016/j.aei.2026.104391
Jian Wei, Yue Pan, Jin-Jian Chen
Accurate prediction of excavation-induced multi-risk evolution is essential for intelligent construction and on-site safety management. This study proposes a Physics-Informed Edge-Enhanced Temporal Graph Convolutional Network (PI-ETGCN) that seamlessly integrates mechanistic priors with data-driven learning for spatiotemporal multi-risk forecasting in deep excavations. During graph modeling, multi-source construction information is encoded as node attributes to construct a Multi-Source Risk Graph (MRG) that links heterogeneous variables across space and time. Mechanical laws governing wall–soil–pipeline interactions are parameterized with learnable edge coefficients, enabling adaptive modeling of spatial–physical couplings under varying construction conditions. During graph learning, PI-ETGCN combines edge-enhanced graph convolution with temporal modules to capture coupled data–mechanics relationships and spatiotemporal dependencies within the MRG. A physics-informed regularization term is incorporated into the training objective to promote mechanically consistent risk-evolution patterns and improve robustness and generalization. Validated on a real-world Shanghai Rail Transit project, PI-ETGCN delivers superior predictive performance, reducing errors by over 20% relative to the second-best baseline. Ablation studies further confirm that the proposed physics–data fusion strategy maintains high accuracy under limited samples, noise, and outliers. Overall, PI-ETGCN provides interpretable, decision-ready risk information for real-time monitoring and reliability-aware on-site safety management.
{"title":"Physics-informed edge-enhanced temporal graph convolutional network for multi-risk evolution prediction in deep excavation","authors":"Jian Wei, Yue Pan, Jin-Jian Chen","doi":"10.1016/j.aei.2026.104391","DOIUrl":"10.1016/j.aei.2026.104391","url":null,"abstract":"<div><div>Accurate prediction of excavation-induced multi-risk evolution is essential for intelligent construction and on-site safety management. This study proposes a Physics-Informed Edge-Enhanced Temporal Graph Convolutional Network (PI-ETGCN) that seamlessly integrates mechanistic priors with data-driven learning for spatiotemporal multi-risk forecasting in deep excavations. During graph modeling, multi-source construction information is encoded as node attributes to construct a Multi-Source Risk Graph (MRG) that links heterogeneous variables across space and time. Mechanical laws governing wall–soil–pipeline interactions are parameterized with learnable edge coefficients, enabling adaptive modeling of spatial–physical couplings under varying construction conditions. During graph learning, PI-ETGCN combines edge-enhanced graph convolution with temporal modules to capture coupled data–mechanics relationships and spatiotemporal dependencies within the MRG. A physics-informed regularization term is incorporated into the training objective to promote mechanically consistent risk-evolution patterns and improve robustness and generalization. Validated on a real-world Shanghai Rail Transit project, PI-ETGCN delivers superior predictive performance, reducing errors by over 20% relative to the second-best baseline. Ablation studies further confirm that the proposed physics–data fusion strategy maintains high accuracy under limited samples, noise, and outliers. Overall, PI-ETGCN provides interpretable, decision-ready risk information for real-time monitoring and reliability-aware on-site safety management.</div></div>","PeriodicalId":50941,"journal":{"name":"Advanced Engineering Informatics","volume":"71 ","pages":"Article 104391"},"PeriodicalIF":9.9,"publicationDate":"2026-02-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146188617","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","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