Pub Date : 2026-01-16DOI: 10.1016/j.engappai.2025.113664
Milad Taleby Ahvanooey , Wojciech Mazurczyk
The Personal Identification Number (PIN) authentication scheme is still a broadly deployed standard security protocol for smart devices due to its simplicity and usability in Internet of Everything (IoE) environments. However, the classical PIN schemes are technically susceptible to side-channel attacks, where adversaries can capture the victims’ PINs through camera-based recording, keystroke logging, or visual watching of User-to-Device (U2D) interactions. To overcome this critical security flaw, we introduce an innovative Dynamical PIN Hiding (HDynPIN) multifactor authentication scheme for protecting IoE machines, which functions by concealing a Hidden-PIN (HP) under the guise of a Dynamic-Passcode (DPC) based on a Recurrent Neural Network (RNN)-generated hint item and a randomized entry pathway. HDynPIN requires the user to choose a 4- or 6-digit HP, a set of hint items, and their corresponding operators during the registration phase. Then, it displays a random hint item generated using a broad learning-based RNN algorithm, considering the user’s settings, which guides her/him through a randomized entry pathway by utilizing a one-time valid DPC during the authentication phase. By concealing the HP and randomizing the DPC entry pathway, HDynPIN provides a user-friendly and more secure U2D protocol that is robust against side-channel attacks. Our extensive experimental evaluation confirms that HDynPIN provides better performance compared to state-of-the-art schemes.
{"title":"An innovative user-to-device authentication scheme using broad learning-based dynamic hint generation","authors":"Milad Taleby Ahvanooey , Wojciech Mazurczyk","doi":"10.1016/j.engappai.2025.113664","DOIUrl":"10.1016/j.engappai.2025.113664","url":null,"abstract":"<div><div>The Personal Identification Number (PIN) authentication scheme is still a broadly deployed standard security protocol for smart devices due to its simplicity and usability in Internet of Everything (IoE) environments. However, the classical PIN schemes are technically susceptible to side-channel attacks, where adversaries can capture the victims’ PINs through camera-based recording, keystroke logging, or visual watching of User-to-Device (U2D) interactions. To overcome this critical security flaw, we introduce an innovative Dynamical PIN Hiding (HDynPIN) multifactor authentication scheme for protecting IoE machines, which functions by concealing a Hidden-PIN (HP) under the guise of a Dynamic-Passcode (DPC) based on a Recurrent Neural Network (RNN)-generated hint item and a randomized entry pathway. HDynPIN requires the user to choose a 4- or 6-digit HP, a set of hint items, and their corresponding operators during the registration phase. Then, it displays a random hint item generated using a broad learning-based RNN algorithm, considering the user’s settings, which guides her/him through a randomized entry pathway by utilizing a one-time valid DPC during the authentication phase. By concealing the HP and randomizing the DPC entry pathway, HDynPIN provides a user-friendly and more secure U2D protocol that is robust against side-channel attacks. Our extensive experimental evaluation confirms that HDynPIN provides better performance compared to state-of-the-art schemes.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"167 ","pages":"Article 113664"},"PeriodicalIF":8.0,"publicationDate":"2026-01-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145980120","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 : 2026-01-16DOI: 10.1016/j.engappai.2026.113822
Xuanyuan Su , Kaixin Jin , Yongzhe Ma , Chen Lu , Laifa Tao
Imbalanced data and diverse operating conditions (OCs) are two common issues in fault diagnosis, which are generally addressed by data-driven artificial intelligence (AI) variants focusing on data generation and transfer learning. However, applying these approaches to complex electromechanical systems (EMS) remains challenging, as extended faults and diverse OCs create harsh data situations, such as scarcity of fault data and unseen OCs, thus limiting the efficacy of purely data-driven paradigms. This paper proposes a data and knowledge-combined intelligent fault diagnosis framework. Firstly, a collaborative hierarchical modeling mechanism is proposed to construct a full-system digital twin (DT) for EMS, which generates two modalities of information: DT fault data and DT knowledge, enriching both the scale and type of the available dataset. Furthermore, a heterogeneous domain generalization network (HDGN) is proposed to achieve generalized fault diagnosis from both data and knowledge perspectives. By embedding prior DT knowledge, domain-invariance is stably retained from the data. Driven by the triplet specific similarity loss, domain-specific discriminative representations are adaptively learned by multi-channels from the knowledge-embedded data. The resulting HDGN progressively improves model generalization to unseen OCs with well-balanced stability and adaptiveness. The experimental results demonstrate the proposed method's effectiveness and superiority, providing a reference for AI applications in industrial scenarios with imbalanced data and unseen OCs.
{"title":"Imbalanced fault diagnosis of electromechanical systems under unseen operating conditions: a heterogeneous domain generalization framework combining digital twin knowledge and data","authors":"Xuanyuan Su , Kaixin Jin , Yongzhe Ma , Chen Lu , Laifa Tao","doi":"10.1016/j.engappai.2026.113822","DOIUrl":"10.1016/j.engappai.2026.113822","url":null,"abstract":"<div><div>Imbalanced data and diverse operating conditions (OCs) are two common issues in fault diagnosis, which are generally addressed by data-driven artificial intelligence (AI) variants focusing on data generation and transfer learning. However, applying these approaches to complex electromechanical systems (EMS) remains challenging, as extended faults and diverse OCs create harsh data situations, such as scarcity of fault data and unseen OCs, thus limiting the efficacy of purely data-driven paradigms. This paper proposes a data and knowledge-combined intelligent fault diagnosis framework. Firstly, a collaborative hierarchical modeling mechanism is proposed to construct a full-system digital twin (DT) for EMS, which generates two modalities of information: DT fault data and DT knowledge, enriching both the scale and type of the available dataset. Furthermore, a heterogeneous domain generalization network (HDGN) is proposed to achieve generalized fault diagnosis from both data and knowledge perspectives. By embedding prior DT knowledge, domain-invariance is stably retained from the data. Driven by the triplet specific similarity loss, domain-specific discriminative representations are adaptively learned by multi-channels from the knowledge-embedded data. The resulting HDGN progressively improves model generalization to unseen OCs with well-balanced stability and adaptiveness. The experimental results demonstrate the proposed method's effectiveness and superiority, providing a reference for AI applications in industrial scenarios with imbalanced data and unseen OCs.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"167 ","pages":"Article 113822"},"PeriodicalIF":8.0,"publicationDate":"2026-01-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145980122","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 : 2026-01-16DOI: 10.1016/j.engappai.2025.113669
Shanshan Liu , Liang Chang , Guanyu Hu , Guanghai Li
Predictive methods for complex engineering systems are critical for decision-making, but most rely on data fitting and fail to capture causal relationships amid ignorance and uncertainty. Traditional evidential reasoning (ER) addresses such system characteristics yet lacks causal mining and predictive capabilities. This paper proposes a novel temporal sequential ER prediction model that retains ER's strengths in handling uncertainty while mining and quantifying causality between evidence and conclusions via the improved Peter–Clark (PC) algorithm and transfer entropy. Incorporating a time parameter t enables predicting subsequent system states, enhancing accuracy and reliability. Validated across three industrial system domains, the model achieves remarkable performance: 93.78 % accuracy in short-term power load prediction (8.3 %–16.87 % improvement over baselines), 96.5 % in software defined networking (SDN) security prediction (10 %–14.96 % enhancement), and 93.65 % in flywheel system fault prediction (19.65 %–29.65 % improvement). These results confirm its practical value in boosting grid efficiency, strengthening network security, and improving equipment reliability for complex engineering systems.
{"title":"A novel causal relationship-based evidential reasoning prediction method with a time parameter","authors":"Shanshan Liu , Liang Chang , Guanyu Hu , Guanghai Li","doi":"10.1016/j.engappai.2025.113669","DOIUrl":"10.1016/j.engappai.2025.113669","url":null,"abstract":"<div><div>Predictive methods for complex engineering systems are critical for decision-making, but most rely on data fitting and fail to capture causal relationships amid ignorance and uncertainty. Traditional evidential reasoning (ER) addresses such system characteristics yet lacks causal mining and predictive capabilities. This paper proposes a novel temporal sequential ER prediction model that retains ER's strengths in handling uncertainty while mining and quantifying causality between evidence and conclusions via the improved Peter–Clark (PC) algorithm and transfer entropy. Incorporating a time parameter <em>t</em> enables predicting subsequent system states, enhancing accuracy and reliability. Validated across three industrial system domains, the model achieves remarkable performance: 93.78 % accuracy in short-term power load prediction (8.3 %–16.87 % improvement over baselines), 96.5 % in software defined networking (SDN) security prediction (10 %–14.96 % enhancement), and 93.65 % in flywheel system fault prediction (19.65 %–29.65 % improvement). These results confirm its practical value in boosting grid efficiency, strengthening network security, and improving equipment reliability for complex engineering systems.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"167 ","pages":"Article 113669"},"PeriodicalIF":8.0,"publicationDate":"2026-01-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145980152","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 : 2026-01-15DOI: 10.1016/j.engappai.2026.113809
Michael Schäfer , Ulrike Faltings , Björn Glaser
Scrap is the most important secondary raw material in the transformation to low carbon dioxide () steel. However, the suitable use of different scrap types for producing high quality steels with the right chemical composition is non-trivial. It requires process control and detailed knowledge of all input materials used. SHapley Additive exPlanations (SHAP), a game-theoretic approach, is often used to interpret machine learning models through visualizations and feature attributions. In this paper, we present a novel application of SHAP values. This enables more precise control of material composition in steel production without the need for additional sensors. This makes it extremely practical for real steel production environments and enables better control of the materials used in the steel production process.
As a basis for this approach, various machine learning models were trained and the respective SHAP values computed. To validate the approach, the results were compared with the values from the steel plant. Comparing the calculated values with the historical estimates, the results agree for most input materials and target elements. The key innovation lies in using SHAP values not only for model interpretability, but also as a quantitative tool to estimate the chemical content of input materials (e.g., steel scrap) based on process data. The framework enables chemical composition estimation, relying solely on routinely collected process data. This is a novel application of SHAP and allows the back-calculation of predicted values and can be used in a wide range of applications in industry and academia.
{"title":"Artificial Intelligence-based back-calculation model for scrap compiling optimization","authors":"Michael Schäfer , Ulrike Faltings , Björn Glaser","doi":"10.1016/j.engappai.2026.113809","DOIUrl":"10.1016/j.engappai.2026.113809","url":null,"abstract":"<div><div>Scrap is the most important secondary raw material in the transformation to low carbon dioxide (<span><math><msub><mrow><mi>CO</mi></mrow><mrow><mn>2</mn></mrow></msub></math></span>) steel. However, the suitable use of different scrap types for producing high quality steels with the right chemical composition is non-trivial. It requires process control and detailed knowledge of all input materials used. SHapley Additive exPlanations (SHAP), a game-theoretic approach, is often used to interpret machine learning models through visualizations and feature attributions. In this paper, we present a novel application of SHAP values. This enables more precise control of material composition in steel production without the need for additional sensors. This makes it extremely practical for real steel production environments and enables better control of the materials used in the steel production process.</div><div>As a basis for this approach, various machine learning models were trained and the respective SHAP values computed. To validate the approach, the results were compared with the values from the steel plant. Comparing the calculated values with the historical estimates, the results agree for most input materials and target elements. The key innovation lies in using SHAP values not only for model interpretability, but also as a quantitative tool to estimate the chemical content of input materials (e.g., steel scrap) based on process data. The framework enables chemical composition estimation, relying solely on routinely collected process data. This is a novel application of SHAP and allows the back-calculation of predicted values and can be used in a wide range of applications in industry and academia.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"167 ","pages":"Article 113809"},"PeriodicalIF":8.0,"publicationDate":"2026-01-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145979801","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 : 2026-01-15DOI: 10.1016/j.engappai.2026.113863
Wei Wang , Mengyi Ma , Hongjun Zhang , Yun Zhou , Guangsheng Wu
Drug combination therapy demonstrates more significant efficacy than monotherapy in cancer treatment. Despite the proposal of several computational approaches aimed at effectively identifying synergistic drug combinations, challenges persist due to inadequate multi-level learning within multimodal data. Furthermore, existing models still struggle to adequately capture the complex biological network interactions between drug combinations and cell lines. To overcome these issues, we propose a novel hypergraph neural network method for synergistic drug combination prediction. This method integrates multi-view feature learning and enhanced hypergraph neural networks to improve drug combination prediction. First, multi-view learning is independently applied to the multimodal data of drugs and cell lines. This framework employs a fine-tuned ChemBERTa model enhanced by contrastive learning to effectively capture the contextual information of drug SMILES. Second, enhanced hypergraph neural networks equipped with a multi-head attention mechanism are designed to capture the complex topological information between drugs and cell lines and to address the limited ability of the hypergraph to capture global information. Third, the similarity-based multi-task supervision module further stabilizes the model. The experimental results show that our method outperforms state-of-the-art methods in various scenarios, including leave-drug-combination-out, leave-cell-out, and leave-drug-out scenarios. Specifically, in the leave-drug combination-out scenario, our method achieves a Mean Squared Error of 163.635, a Root Mean Squared Error of 12.792, and a Pearson Correlation Coefficient of 0.751. Finally, a case study demonstrates the efficacy of the model in predicting novel synergistic drug combinations.
{"title":"Multi-view feature learning and enhanced hypergraph neural networks for synergistic prediction of drug combination","authors":"Wei Wang , Mengyi Ma , Hongjun Zhang , Yun Zhou , Guangsheng Wu","doi":"10.1016/j.engappai.2026.113863","DOIUrl":"10.1016/j.engappai.2026.113863","url":null,"abstract":"<div><div>Drug combination therapy demonstrates more significant efficacy than monotherapy in cancer treatment. Despite the proposal of several computational approaches aimed at effectively identifying synergistic drug combinations, challenges persist due to inadequate multi-level learning within multimodal data. Furthermore, existing models still struggle to adequately capture the complex biological network interactions between drug combinations and cell lines. To overcome these issues, we propose a novel hypergraph neural network method for synergistic drug combination prediction. This method integrates multi-view feature learning and enhanced hypergraph neural networks to improve drug combination prediction. First, multi-view learning is independently applied to the multimodal data of drugs and cell lines. This framework employs a fine-tuned ChemBERTa model enhanced by contrastive learning to effectively capture the contextual information of drug SMILES. Second, enhanced hypergraph neural networks equipped with a multi-head attention mechanism are designed to capture the complex topological information between drugs and cell lines and to address the limited ability of the hypergraph to capture global information. Third, the similarity-based multi-task supervision module further stabilizes the model. The experimental results show that our method outperforms state-of-the-art methods in various scenarios, including leave-drug-combination-out, leave-cell-out, and leave-drug-out scenarios. Specifically, in the leave-drug combination-out scenario, our method achieves a Mean Squared Error of 163.635, a Root Mean Squared Error of 12.792, and a Pearson Correlation Coefficient of 0.751. Finally, a case study demonstrates the efficacy of the model in predicting novel synergistic drug combinations.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"167 ","pages":"Article 113863"},"PeriodicalIF":8.0,"publicationDate":"2026-01-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145979874","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 : 2026-01-15DOI: 10.1016/j.engappai.2026.113798
Kai Sun, Dongzhe Yang, Dasong Wang, Fangfang Zhang
Accurate forecasting of wind power is essential to prevent grid overload and minimize power wastage, thereby optimizing dispatch and reducing the operational costs of power systems. However, the intermittent and unpredictable nature of wind energy poses significant challenges in achieving timely and precise predictions. To address these challenges, this study proposes a hybrid wind power forecasting model integrating physical knowledge with a bidirectional long short-term memory (BiLSTM) network. First, data collected from a practical wind farm are preprocessed and resampled to mitigate the impact of measurement outliers stemming from sensor faults and turbulence. Second, mechanistic model identification for the studied wind turbine is conducted to encode the relevant physical knowledge into the BiLSTM model. Third, a wind-speed-based probabilistic penalty term is designed to address physically implausible predictions under low-wind-speed conditions. Moreover, an improved leaky rectified linear unit activation function is proposed to refine the BiLSTM model, preventing both negative power predictions and those exceeding the rated capacity. Finally, the developed model is applied to real-world wind turbines. Experimental results demonstrate that the proposed model can effectively eliminate physically implausible predictions, and exhibit superior robustness and enhanced prediction accuracy compared with other advanced algorithms.
{"title":"Development of a physics-guided bidirectional long short-term memory for wind power forecasting","authors":"Kai Sun, Dongzhe Yang, Dasong Wang, Fangfang Zhang","doi":"10.1016/j.engappai.2026.113798","DOIUrl":"10.1016/j.engappai.2026.113798","url":null,"abstract":"<div><div>Accurate forecasting of wind power is essential to prevent grid overload and minimize power wastage, thereby optimizing dispatch and reducing the operational costs of power systems. However, the intermittent and unpredictable nature of wind energy poses significant challenges in achieving timely and precise predictions. To address these challenges, this study proposes a hybrid wind power forecasting model integrating physical knowledge with a bidirectional long short-term memory (BiLSTM) network. First, data collected from a practical wind farm are preprocessed and resampled to mitigate the impact of measurement outliers stemming from sensor faults and turbulence. Second, mechanistic model identification for the studied wind turbine is conducted to encode the relevant physical knowledge into the BiLSTM model. Third, a wind-speed-based probabilistic penalty term is designed to address physically implausible predictions under low-wind-speed conditions. Moreover, an improved leaky rectified linear unit activation function is proposed to refine the BiLSTM model, preventing both negative power predictions and those exceeding the rated capacity. Finally, the developed model is applied to real-world wind turbines. Experimental results demonstrate that the proposed model can effectively eliminate physically implausible predictions, and exhibit superior robustness and enhanced prediction accuracy compared with other advanced algorithms.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"167 ","pages":"Article 113798"},"PeriodicalIF":8.0,"publicationDate":"2026-01-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145979875","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}
Advanced Persistent Threat (APT) has greatly threatened global cybersecurity. In recent years, a promising approach for APT detection has been proposed based on Graph Neural Networks (GNN) and provenance graphs constructed from host logs, in which graph nodes and edges represent processes and interactions, respectively. However, current methods primarily focus on the individual behaviors of single nodes (attacking subjects), neglecting the in-depth analysis of interactions between collaborative attacking processes. This results in limited capability in detecting complex and composite APT attacks. In this paper, a new GNN-based APT detection method, Dinspector, is proposed. It utilizes a dual factor attention mechanism to aggregate the features between neighboring nodes and edges simultaneously. Furthermore, Dinspector combines GraphSAGE (Sample and Aggregate) and the graph attention layer into a two-layer Graph Neural Network structure. By integrating node features, structural features, and neighbor features, Dinspector is capable of extracting features of complex attack patterns, improving the detection performance of novel APT attacks. Experimental results on three public datasets demonstrated that Dinspector achieves an average precision of 98% and a recall rate of 99%, attaining state-of-the-art detection performance and outperforming them in certain aspects. The source code of Dinspector is publicly available at: https://github.com/Qc-TX/Dinspector.
高级持续威胁（APT）严重威胁着全球网络安全。近年来，基于图神经网络（GNN）和由主机日志构造的来源图提出了一种很有前途的APT检测方法，其中图节点和图边分别表示过程和交互。然而，目前的方法主要关注单个节点（攻击主体）的个体行为，忽略了对协同攻击过程之间相互作用的深入分析。这导致检测复杂和复合APT攻击的能力有限。本文提出了一种新的基于gnn的APT检测方法——Dinspector。它利用双因素关注机制同时聚合相邻节点和边缘之间的特征。此外，Dinspector将GraphSAGE （Sample and Aggregate）和图注意层结合成一个两层图神经网络结构。通过融合节点特征、结构特征和邻居特征，能够提取复杂攻击模式的特征，提高对新型APT攻击的检测性能。在三个公共数据集上的实验结果表明，Dinspector达到了98%的平均准确率和99%的召回率，达到了最先进的检测性能，并在某些方面优于他们。disspector的源代码可在https://github.com/Qc-TX/Dinspector公开获取。
{"title":"Dinspector: Dual factor graph attention mechanism for Advanced Persistent Threat detection","authors":"Hongchao Wang, Wen Chen, Linrui Li, Haoyang Pu, Yilin Zhang","doi":"10.1016/j.engappai.2026.113861","DOIUrl":"10.1016/j.engappai.2026.113861","url":null,"abstract":"<div><div>Advanced Persistent Threat (APT) has greatly threatened global cybersecurity. In recent years, a promising approach for APT detection has been proposed based on Graph Neural Networks (GNN) and provenance graphs constructed from host logs, in which graph nodes and edges represent processes and interactions, respectively. However, current methods primarily focus on the individual behaviors of single nodes (attacking subjects), neglecting the in-depth analysis of interactions between collaborative attacking processes. This results in limited capability in detecting complex and composite APT attacks. In this paper, a new GNN-based APT detection method, Dinspector, is proposed. It utilizes a dual factor attention mechanism to aggregate the features between neighboring nodes and edges simultaneously. Furthermore, Dinspector combines GraphSAGE (Sample and Aggregate) and the graph attention layer into a two-layer Graph Neural Network structure. By integrating node features, structural features, and neighbor features, Dinspector is capable of extracting features of complex attack patterns, improving the detection performance of novel APT attacks. Experimental results on three public datasets demonstrated that Dinspector achieves an average precision of 98% and a recall rate of 99%, attaining state-of-the-art detection performance and outperforming them in certain aspects. The source code of Dinspector is publicly available at: <span><span>https://github.com/Qc-TX/Dinspector</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"167 ","pages":"Article 113861"},"PeriodicalIF":8.0,"publicationDate":"2026-01-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145980150","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 : 2026-01-15DOI: 10.1016/j.engappai.2026.113788
Qiaosong Hu , Dujian Zou , Zhilin Bai , Tiejun Liu , Ao Zhou
Concrete shrinkage under non-stationary climatic forcing poses an increasing threat to the serviceability and longevity of infrastructure in low-pressure, arid and high-altitude regions. Current models neglect multi-environment interactions and climate-driven risk evolution. This study presents a hybrid modeling and assessment framework that coupled physics-informed empirical priors with optimized machine learning to predict shrinkage evolution, quantify structural risk, and map spatiotemporal vulnerability under future climate scenarios. A curated shrinkage database was fused with high-resolution meteorological projections and downscaled via filtering and cubic interpolation. The empirical CEB-FIP 2010 shrinkage formulation and air pressure parameters were embedded into feature engineering to create temperature-humidity-pressure coupled predictors. An XGBoost (Extreme Gradient Boosting) model was optimized through systematic hyperparameter tuning and physics-guided transfer learning. The optimized coupling model attained R2 = 0.92 to predict shrinkage evolution, and reduced long-term prediction divergence to within 15% against independent data from three-factor experiments. To translate material-level shrinkage into structural risk, multiphysics finite-element simulations of a representative reinforced-concrete pier incorporated eigenstrain shrinkage fields and reinforcement constraint to resolve strain–stress–damage progression. Four critical normalized strain thresholds were identified that demarcated initiation, stable propagation, accelerated expansion and through-crack stages. A five-tier risk zoning map across China was constructed, covering both historical data and mid-future climate scenario. Plateau and northwestern basins showed marked vulnerability. Using C60 concrete as a representative case study due to its prevalence, results showed the medium-to-high risk area increasing by 65%, with 31.1% of China's territory classified as medium–high risk by 2050.
{"title":"Hybrid machine learning and physical modeling framework for climate-driven risk zonation of concrete shrinkage damage","authors":"Qiaosong Hu , Dujian Zou , Zhilin Bai , Tiejun Liu , Ao Zhou","doi":"10.1016/j.engappai.2026.113788","DOIUrl":"10.1016/j.engappai.2026.113788","url":null,"abstract":"<div><div>Concrete shrinkage under non-stationary climatic forcing poses an increasing threat to the serviceability and longevity of infrastructure in low-pressure, arid and high-altitude regions. Current models neglect multi-environment interactions and climate-driven risk evolution. This study presents a hybrid modeling and assessment framework that coupled physics-informed empirical priors with optimized machine learning to predict shrinkage evolution, quantify structural risk, and map spatiotemporal vulnerability under future climate scenarios. A curated shrinkage database was fused with high-resolution meteorological projections and downscaled via filtering and cubic interpolation. The empirical CEB-FIP 2010 shrinkage formulation and air pressure parameters were embedded into feature engineering to create temperature-humidity-pressure coupled predictors. An XGBoost (Extreme Gradient Boosting) model was optimized through systematic hyperparameter tuning and physics-guided transfer learning. The optimized coupling model attained R<sup>2</sup> = 0.92 to predict shrinkage evolution, and reduced long-term prediction divergence to within 15% against independent data from three-factor experiments. To translate material-level shrinkage into structural risk, multiphysics finite-element simulations of a representative reinforced-concrete pier incorporated eigenstrain shrinkage fields and reinforcement constraint to resolve strain–stress–damage progression. Four critical normalized strain thresholds were identified that demarcated initiation, stable propagation, accelerated expansion and through-crack stages. A five-tier risk zoning map across China was constructed, covering both historical data and mid-future climate scenario. Plateau and northwestern basins showed marked vulnerability. Using C60 concrete as a representative case study due to its prevalence, results showed the medium-to-high risk area increasing by 65%, with 31.1% of China's territory classified as medium–high risk by 2050.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"167 ","pages":"Article 113788"},"PeriodicalIF":8.0,"publicationDate":"2026-01-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145979876","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 : 2026-01-15DOI: 10.1016/j.engappai.2026.113870
Yanshen Zhao, Yifan Zhao, Huayu Fu, Xiang Ji, Zhongzhi Han
Pesticide residues pose significant risks to food safety and human health, especially in crops like tomatoes, which rely heavily on pesticide use in greenhouse cultivation. Spectral detection techniques, although promising for their lossless nature, face challenges due to the limited feature information available from one-dimensional spectral data. To overcome this, we applied Gramian Angular Field (GAF) methods, including the Gramian Angular Difference Field (GADF) and Gramian Angular Summation Field (GASF), to transform the spectral data into two-dimensional representations, enhancing feature extraction. A Dual-Channel Convolutional Neural Network (DCCNN) was able to achieve an accuracy of 93.50 % on the tomato dataset. Additionally, using Principal Component Analysis (PCA) for key wavelength selection further improved performance: accuracy reached 89.77 % with 20 % of wavelengths and 93.68 % with 80 %. Generalizability tests conducted on apple and cucumber datasets resulted in accuracies of 91.28 % and 81.45 %, respectively. For the bacterial dataset and the aflatoxin B1 (AFB1) dataset, the model achieved performances of 90.83 % and 84.69 %, respectively. These findings highlight the effectiveness of GAF methods and DCCNN for pesticide residue detection in tomatoes, while also suggesting that further advancements in feature extraction and selection could broaden the application of these techniques to other agricultural crops.
{"title":"Dual-channel convolutional neural network for tomato pesticide residue detection using Gramian angular field transformations","authors":"Yanshen Zhao, Yifan Zhao, Huayu Fu, Xiang Ji, Zhongzhi Han","doi":"10.1016/j.engappai.2026.113870","DOIUrl":"10.1016/j.engappai.2026.113870","url":null,"abstract":"<div><div>Pesticide residues pose significant risks to food safety and human health, especially in crops like tomatoes, which rely heavily on pesticide use in greenhouse cultivation. Spectral detection techniques, although promising for their lossless nature, face challenges due to the limited feature information available from one-dimensional spectral data. To overcome this, we applied Gramian Angular Field (GAF) methods, including the Gramian Angular Difference Field (GADF) and Gramian Angular Summation Field (GASF), to transform the spectral data into two-dimensional representations, enhancing feature extraction. A Dual-Channel Convolutional Neural Network (DCCNN) was able to achieve an accuracy of 93.50 % on the tomato dataset. Additionally, using Principal Component Analysis (PCA) for key wavelength selection further improved performance: accuracy reached 89.77 % with 20 % of wavelengths and 93.68 % with 80 %. Generalizability tests conducted on apple and cucumber datasets resulted in accuracies of 91.28 % and 81.45 %, respectively. For the bacterial dataset and the aflatoxin B1 (AFB1) dataset, the model achieved performances of 90.83 % and 84.69 %, respectively. These findings highlight the effectiveness of GAF methods and DCCNN for pesticide residue detection in tomatoes, while also suggesting that further advancements in feature extraction and selection could broaden the application of these techniques to other agricultural crops.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"167 ","pages":"Article 113870"},"PeriodicalIF":8.0,"publicationDate":"2026-01-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145979877","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 : 2026-01-14DOI: 10.1016/j.engappai.2025.113681
Nguyen Quoc Anh , Tran Truong Tuan Phat , Ha Xuan Son , Thai Thi Thanh Nhan , Nguyen Ngoc Phien , Trung Phan Hoang Tuan , Ngan Nguyen Thi Kim
Financial time-series forecasting is challenged by non-linear, non-stationary dynamics driven by macroeconomic factors, market sentiment, and stochastic events. Traditional statistical models assume stationarity and linear dependencies, failing to capture complex temporal patterns, while deep learning approaches struggle with vanishing gradients and long-term dependencies. Standard Transformers incur high computational costs (quadratic complexity, , per layer) due to attention mechanisms and large parameter counts, where is the sequence length and is the model dimension. This study proposes LiteFormer, a lightweight, encoder-only Transformer for univariate stock price forecasting, leveraging encoder layers with multi-head self-attention and feed-forward networks (). Operating on sequences of closing prices (, ), LiteFormer employs sinusoidal positional encodings, a causal mask, dropout (), and layer normalization to model temporal dependencies and enhance generalization. With only 750,000+ parameters, LiteFormer reduces per layer complexity via compact design, thereby enabling low-latency inference (38 millisecond) and energy efficiency (96.894 Watt), which promises to offers scalable real-time inference for industrial fintech systems. Experiments across 30 stocks from the S&P 500, FTSE 100, and Nikkei 225 indices demonstrate Mean Absolute Error and Root Mean Square Error reductions of 3.45%–9.09% over vanilla Transformers and up to 48% over recurrence neural models for high-volatility stocks. LiteFormer’s efficient, interpretable architecture, driven by attention weights, offers a scalable solution with potential for multivariate extensions and real-world multi-modal applications in predictive domain.
{"title":"LiteFormer: A lightweight encoder-only Transformer for efficient financial time series forecasting across global stock indices","authors":"Nguyen Quoc Anh , Tran Truong Tuan Phat , Ha Xuan Son , Thai Thi Thanh Nhan , Nguyen Ngoc Phien , Trung Phan Hoang Tuan , Ngan Nguyen Thi Kim","doi":"10.1016/j.engappai.2025.113681","DOIUrl":"10.1016/j.engappai.2025.113681","url":null,"abstract":"<div><div>Financial time-series forecasting is challenged by non-linear, non-stationary dynamics driven by macroeconomic factors, market sentiment, and stochastic events. Traditional statistical models assume stationarity and linear dependencies, failing to capture complex temporal patterns, while deep learning approaches struggle with vanishing gradients and long-term dependencies. Standard Transformers incur high computational costs (quadratic complexity, <span><math><mrow><mi>O</mi><mrow><mo>(</mo><msup><mrow><mi>n</mi></mrow><mrow><mn>2</mn></mrow></msup><mi>⋅</mi><mi>d</mi><mo>)</mo></mrow></mrow></math></span>, per layer) due to attention mechanisms and large parameter counts, where <span><math><mi>n</mi></math></span> is the sequence length and <span><math><mi>d</mi></math></span> is the model dimension. This study proposes LiteFormer, a lightweight, encoder-only Transformer for univariate stock price forecasting, leveraging <span><math><mrow><mi>N</mi><mo>=</mo><mn>4</mn></mrow></math></span> encoder layers with <span><math><mrow><mi>h</mi><mo>=</mo><mn>8</mn></mrow></math></span> multi-head self-attention and feed-forward networks (<span><math><mrow><msub><mrow><mi>d</mi></mrow><mrow><mtext>ff</mtext></mrow></msub><mo>=</mo><mn>512</mn></mrow></math></span>). Operating on sequences of closing prices (<span><math><mrow><mi>T</mi><mo>=</mo><mn>14</mn></mrow></math></span>, <span><math><mrow><msub><mrow><mi>d</mi></mrow><mrow><mtext>model</mtext></mrow></msub><mo>=</mo><mn>128</mn></mrow></math></span>), LiteFormer employs sinusoidal positional encodings, a causal mask, dropout (<span><math><mrow><mi>p</mi><mo>=</mo><mn>0</mn><mo>.</mo><mn>1</mn></mrow></math></span>), and layer normalization to model temporal dependencies and enhance generalization. With only 750,000+ parameters, LiteFormer reduces per layer complexity via compact design, thereby enabling low-latency inference (38 millisecond) and energy efficiency (96.894 Watt), which promises to offers scalable real-time inference for industrial fintech systems. Experiments across 30 stocks from the S&P 500, FTSE 100, and Nikkei 225 indices demonstrate Mean Absolute Error and Root Mean Square Error reductions of 3.45%–9.09% over vanilla Transformers and up to 48% over recurrence neural models for high-volatility stocks. LiteFormer’s efficient, interpretable architecture, driven by attention weights, offers a scalable solution with potential for multivariate extensions and real-world multi-modal applications in predictive domain.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"167 ","pages":"Article 113681"},"PeriodicalIF":8.0,"publicationDate":"2026-01-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145957667","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