Applied Intelligence最新文献

{"title":"Predefined-time adaptive neural control for electro-hydraulic actuator systems with parameter uncertainties","authors":"Chen Wang,&nbsp;Jinping Ye,&nbsp;Xin Chen,&nbsp;Zhenlei Chen,&nbsp;Dan Jiang,&nbsp;Qing Guo","doi":"10.1007/s10489-026-07535-0","DOIUrl":"10.1007/s10489-026-07535-0","url":null,"abstract":"<div><p>Electro-hydraulic actuator systems (EHASs) are widely used in industrial automation. Existing EHAS controllers mainly improve transient and steady-state performance, but guaranteeing tracking-error convergence within a designer-specified time remains challenging under parameter uncertainties and unknown external loads. To address this problem, a predefined-time adaptive neural backstepping controller is developed. The parameter uncertainties are reformulated as two lumped nonlinear terms and compensated using radial basis function neural networks (RBFNNs), while twice continuously differentiable (<span>(C^2)</span>) segmented functions regularize the low-order fractional-power terms to avoid singularities in the virtual controllers. Rigorous theoretical analysis establishes practical predefined-time stability: the tracking error and adaptive-parameter estimation errors converge to adjustable neighborhoods of zero within a designer-specified time. Simulation and comparative experimental results validate the method and demonstrate improved tracking accuracy over finite-time and conventional backstepping controllers.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149189121","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}

{"title":"Intelligent building information modeling applications in real estate development: a scientometric mapping with implications for decision support and implementation","authors":"Qian Ji,&nbsp;Zhenping Liu,&nbsp;Han Zhou,&nbsp;Liang Li,&nbsp;Zhigang Wang","doi":"10.1007/s10489-026-07516-3","DOIUrl":"10.1007/s10489-026-07516-3","url":null,"abstract":"<div><p>Research on intelligent Building Information Modeling (BIM) applications in real estate development spans construction engineering, energy and carbon assessment, geospatial analysis, property valuation, project delivery, and asset management, yet its decision-support functions and implementation conditions remain fragmented. This study synthesizes English-language publications indexed in Web of Science Core Collection and Scopus from 2020 to 2025 using a BIM-centred, rather than title-restricted, eligibility framework. Direct BIM studies formed the core evidence, while adjacent digital or spatial studies were retained only when screening identified a clear information, sensing, modelling, analytical, environmental, infrastructure, or decision-support function relevant to built-asset decisions; these records were not treated as direct evidence of BIM adoption or effectiveness. Of 2,641 records, 1,758 were retained after deduplication and staged screening. Analyses combined publication and citation trends, year-normalized citation weighting, geographic and disciplinary mapping, keyword co-occurrence, author-keyword biclustering, thematic timelines, highly cited studies, and citation-weighted content coding. Six external experts assessed thematic relevance and coding validity, yielding ordinal Krippendorff’s alpha values of 0.830 and 0.815. Three overlapping periods emerged: digital foundations in 2020–2021, broader technological and life-cycle integration in 2022–2023, and increasing emphasis on intelligent analytics and governance in 2024–2025. Machine and deep learning mainly supported prediction, classification, recognition, and generation, whereas rule-based methods and optimization supported checking and alternative selection. Major barriers involved interoperability, data quality, organizational capability, cost, accountability, and SME adoption. A BIM-centred “1 + 3 + N” framework synthesizes these relationships while requiring project-level validation.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149189017","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}

{"title":"MCC-ED: a multi-module collaborative framework for event detection via contextual interaction, closed-loop graph fusion, and event contrastive learning","authors":"Xinyu He,&nbsp;Yijin Zhu","doi":"10.1007/s10489-026-07527-0","DOIUrl":"10.1007/s10489-026-07527-0","url":null,"abstract":"<div><p>Event detection (ED) is a fundamental task in textual information extraction and underpins a wide range of downstream applications, including public opinion monitoring, knowledge graph construction, and intelligent decision-making. Despite recent progress, existing ED models remain limited by insufficient contextual modeling, severe event-semantic ambiguity, and limited adaptability to heterogeneous texts in complex scenarios. To address these challenges, this paper proposes MCC-ED, a multi-module collaborative event detection framework. Built upon pretrained contextual representations, MCC-ED employs dataset-specific decoding heads, including a conditional random field (CRF) decoder for sequence labeling and a span prediction decoder for extractive question answering. The framework integrates three mutually reinforcing modules: a multi-layer fine-grained contextual interaction module that strengthens deep feature associations; a closed-loop self-evaluation multi-source graph fusion module that combines contextual semantics with valid-token sequence structure through iterative feedback; and an event contrastive learning module that constructs category-level contrastive samples to enlarge decision margins and mitigate semantic confusion. Experiments on ACE-2005 yield 86.76% F1 for Trigger Identification (Tri-ID) and 85.56% for Trigger Identification Classification (Tri-CLS), while MAVEN results reach 76.12% and 75.75%, respectively. These results demonstrate strong detection accuracy and consistent effectiveness across the two benchmark settings without conflating their task-specific decoding protocols.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149189018","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}

{"title":"An adaptive multi-window approach to evolving fuzzy prediction in data streams with concept drift","authors":"Kangjian Pu,&nbsp;Tao Zhao","doi":"10.1007/s10489-026-07526-1","DOIUrl":"10.1007/s10489-026-07526-1","url":null,"abstract":"<div><p>Concept drift in datal streams is a major challenge in the field of machine learning. Existing detection methods based on sliding window struggle to simultaneously achieve timely detection and accurate drift localization, while existing evolving fuzzy system (EFS) often adapts slowly to abrupt drift and lack an explicit mechanism for coordinating drift detection with model evolution. To address these issues, this paper proposes a collaborative prediction framework based on a multi-window adaptive model and an EFS, that can accurately identify concept drift and trigger targeted model adaptation. In the concept drift detection stage, a multi-window adaptive model is designed, employing a three-stage process of coarse detection, interval merging, and fine detection. An adaptive indicator based on the statistical characteristics of a historical energy distance sequence is used to precisely locate drift points, which are then passed for ensemble structure updates. In the model adaptation stage, a dynamic adaptive threshold method is proposed to improve the rule generation mechanism of the EFS. Experimental results on several benchmark datasets demonstrate the effectiveness of the proposed combined strategy in handling non-stationary data streams.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149189120","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}

{"title":"MATE-Hard and the count-invariance problem in multi-agent active perception","authors":"Huy Hoang Son Tran","doi":"10.1007/s10489-026-07524-3","DOIUrl":"10.1007/s10489-026-07524-3","url":null,"abstract":"&lt;div&gt;&lt;p&gt;Cooperative multi-camera active perception sits at the intersection of multi-agent reinforcement learning (MARL), set-based representation learning, and active sensing. We introduce MATE-Hard, an extension of the Multi-Agent Tracking Environment (MATE) that adds three orthogonal sources of difficulty — heterogeneous camera types, finite energy budgets, and dynamic optical occlusions — via composable Gym wrappers. To establish that these modifications genuinely raise task difficulty, we adapt the four rule-based agents shipping with MATE to MATE-Hard and observe a coverage degradation of 49–84% across four scenarios (best MATE-Hard entry vs. best plain-MATE entry per scenario; the &lt;span&gt;(84%)&lt;/span&gt; worst case is &lt;span&gt;2v4&lt;/span&gt;, the &lt;span&gt;(49%)&lt;/span&gt; best case is &lt;span&gt;8v8&lt;/span&gt;). We then benchmark six cooperative-MARL methods on MATE-Hard-4v8-9 with three training seeds and 270 evaluation episodes per cell: Double-DQN (independent Q-learning), MAPPO (centralised critic), VDN, vanilla QMIX, our proposed TC-QMIX (a QMIX variant whose mixing hypernetwork is conditioned on per-agent camera-type embeddings), and a types-only ablation. All six learned methods cluster between 0.177 and 0.189 coverage and all sit 12–18% below the best MATE-Hard-adapted rule-based agent. TC-QMIX achieves the highest mean coverage but is statistically indistinguishable from Double-DQN (&lt;span&gt;(p = 0.12)&lt;/span&gt;, Cohen’s &lt;span&gt;(d = 0.14)&lt;/span&gt; negligible); the mixer-hyper-input ablation is significant: removing the per-agent type signal from the mixer drops coverage (&lt;span&gt;(p = 0.034)&lt;/span&gt;), supporting type-conditioning as a useful inductive bias for heterogeneous-team value mixing. A zero-shot transfer matrix across MATE-Hard-{2v4,4v4,8v8}-9 reveals the headline finding. Measured against the rule-based agents of the &lt;i&gt;same&lt;/i&gt; environment, the learned policies retain about a third of the Random-to-Greedy range in distribution (normalised skill 0.34 on &lt;span&gt;4v8&lt;/span&gt;) and on the larger teams (0.32 on &lt;span&gt;4v4&lt;/span&gt;, 0.35 on &lt;span&gt;8v8&lt;/span&gt;), but only 0.18 when transferred to the two-camera scenario — their advantage over a random policy is roughly halved. Raw coverage falls four-fold on MATE-Hard-2v4-9 (&lt;span&gt;(0.189 rightarrow 0.046)&lt;/span&gt;), but a random policy falls almost exactly as far on the same scenario (&lt;span&gt;(0.176 rightarrow 0.042)&lt;/span&gt;), so the raw drop chiefly measures the scenario and the normalised figure is the count-invariance signal. The permutation invariance of the Set-Transformer encoder does &lt;i&gt;not&lt;/i&gt; imply count-invariance of the resulting policy — a failure shared by independent learning, value-decomposition, hyper-conditioned mixing, and centralised-critic methods alike. The loss is one-sided: enlarging the team costs nothing, but shrinking it halves the policy’s margin over a random controller. A team trained for a fixed number of cameras therefore cannot be relied on when cameras are lost — to maintenance, energy depletion, or","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149188956","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}

{"title":"KVCB: Key-Value cache merging towards consistent behavior for efficient inference of large language models","authors":"Zhaohong Jia,&nbsp;Jie Yu,&nbsp;Xiangqian Wang,&nbsp;Kun Jing,&nbsp;Chunchun Li","doi":"10.1007/s10489-026-07389-6","DOIUrl":"10.1007/s10489-026-07389-6","url":null,"abstract":"<div><p>Although key-value (KV) cache eliminates redundant computations in autoregressive generation of large language models (LLMs), its memory cost scales linearly with sequence length, rendering long-context deployment prohibitively expensive. Existing KV cache compression methods primarily prune or merge low-importance KV pairs with long-tail information, yet aggressive compression inevitably excises critical knowledge embedded in the long tail. This degrades generation fidelity and task generalizability. We observe that neighboring KV pairs frequently exhibit highly similar attention patterns during inference, suggesting inherent redundancy. Building upon this insight, we propose KV cache merging towards consistent behavior (KVCB), a novel KV cache compression method for efficient LLM inference. KVCB identifies neighboring KV pairs with consistent behavior through historical attention scores and merges them with weights derived from cumulative historical attention scores. When high-consistency pairs are scarce, KVCB falls back to importance-weighted merging based on temporal position and historical attention scores. To mitigate semantic drift from excessive compression, KVCB incorporates a periodic recovery mechanism that periodically flushes accumulated artifacts for sustained long-context stability. Extensive experiments with Llama-2-7B, Llama-2-7B-Chat, Llama-2-13B-Chat, and CodeLlama-7B on language modeling, long-document question answering, multiple-choice reasoning, and code generation—complemented by ablation studies across model architectures and scales—demonstrate that KVCB consistently matches or outperforms state-of-the-art eviction- and merging-based baselines under fixed memory budgets, effectively balancing compression efficiency with generative fidelity across diverse domains and context lengths.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149189374","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}

{"title":"FedGPA: a multi-level aggregation framework for federated learning with collaborative optimization of generalization and personalization","authors":"Zigang Chen,&nbsp;Haichong Jiang,&nbsp;Deyang Xie,&nbsp;Ruitao Jin,&nbsp;Haihua Zhu","doi":"10.1007/s10489-026-07354-3","DOIUrl":"10.1007/s10489-026-07354-3","url":null,"abstract":"<div><p>Federated Learning (FL) is an innovative approach that balances data privacy protection and distributed model training, effectively addressing the challenges posed by data silos and the growing demand for privacy preservation. However, the heterogeneity of client data, including Non-Independent and Identically Distributed (Non-IID) data and imbalanced datasets, poses significant challenges in balancing global generalization with local personalization. Existing approaches to optimizing aggregation strategies for personalized federated learning remain insufficient, as they fail to fully account for the impact of data heterogeneity and hierarchical model features, thereby limiting personalization and model accuracy improvements. This paper introduces a novel framework, FedGPA, which integrates the global model, clustered models, individual optimal models, and individual local models through an aggregation mechanism leveraging the hierarchical structure of deep neural networks. The framework incorporates an adaptive weight matrix that dynamically adjusts feature contribution ratios across different layers. Furthermore, a hyperparameter is introduced to control the aggregation scope, enabling lower-layer features to emphasize generalization while higher-layer features focus on task-specific personalization. Component-wise ablation studies are further conducted to validate the contribution of each fusion branch. Experimental evaluations on benchmark datasets including MNIST, SVHN, and CIFAR-10, demonstrate that FedGPA outperforms existing methods.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149188957","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}

{"title":"Residual gated AutoEncoder with patchwise nearest neighbor reconstruction for visual anomaly detection","authors":"Nighil Natarajan,&nbsp;Nithilan M,&nbsp;Raghav Sridharan,&nbsp;Chandrakala S","doi":"10.1007/s10489-026-07530-5","DOIUrl":"10.1007/s10489-026-07530-5","url":null,"abstract":"<div><p>Unsupervised Visual Anomaly Detection is vital in tasks such as industrial inspection, medical diagnosis, and autonomous perception, where anomalous patterns must be identified without explicit supervision. Existing reconstruction-based approaches often face a trade-off between global semantic consistency and local texture fidelity, resulting in blurred reconstructions and reduced accuracy of anomaly localization. We propose a dual-path architecture consisting of Residual Gated AutoEncoder (RGAE) module and a Patchwise Nearest Neighbor Reconstruction (PNNR) module. This proposed model forms a unified framework for unsupervised anomaly detection which is built on pretrained DINOv2 visual features. The RGAE enhances global semantic reconstruction while preserving the fine details and suppressing redundant feature propagation. In parallel, the PNNR module reconstructs fine-grained local structures through nearest-neighbor retrieval from a memory bank constructed using normal samples. Anomaly map is produced by computing pixel-wise reconstruction discrepancies between the outputs of the two modules, yielding interpretable localization of abnormal regions. Extensive experiments conducted on both the VisA and MVTec AD benchmark datasets demonstrate that the proposed RGAE+PNNR architecture achieves improved image-level and pixel-level AUROCs compared to state-of-the-art methods. The results highlight the framework’s ability to generalize effectively across diverse domains—from complex real-world industrial scenes (VisA) to controlled object surfaces (MVTec AD). Code is available at https://github.com/NithilanM23/RGAE-PNNR.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149132794","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}

{"title":"DACD: a dynamic adaptive concept drift learning method for streaming data","authors":"Hui Qi,&nbsp;Chang Liu,&nbsp;Xiaobo Qi,&nbsp;Ying Shi,&nbsp;Gaoxia Jiang","doi":"10.1007/s10489-026-07520-7","DOIUrl":"10.1007/s10489-026-07520-7","url":null,"abstract":"<div><p>Concept drift in streaming data presents a persistent challenge for constructing robust analytical models. Traditional algorithms often exhibit limitations in complex scenarios, performing suboptimally with diverse classification tasks and binary-featured data. This paper introduces the Dynamic Adaptive Concept Drift learning method (DACD) to address these issues. DACD employs a window-based Cumulative Sum (CUSUM) algorithm for drift detection, coupled with a neural network that is updated based on the detected drift type. An adaptive sample processing strategy further enhances model resilience by filtering and augmenting the training data under temporal constraints. The framework selects suitable loss functions optimized with class weights and integrates batch processing and learning rate schedulers to improve training stability. On various benchmark datasets, DACD demonstrates strong performance in accuracy, convergence, and robustness, providing a practical solution for real-world streaming data analysis.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149133065","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}

{"title":"Bridging free-form generation and structured pair prediction for emotion-cause pair extraction","authors":"Yan Xia,&nbsp;Amirrudin Kamsin,&nbsp;Zhuangzhuang Pan","doi":"10.1007/s10489-026-07505-6","DOIUrl":"10.1007/s10489-026-07505-6","url":null,"abstract":"<div><p>Emotion-cause pair extraction (ECPE) in conversations identifies emotional turns and their supporting cause turns as explicit turn-indexed pairs. Generative models often return free-form responses, making these decisions difficult to recover and score without post-hoc interpretation. We introduce Bridging Free-Form Generation and Structured Pair Prediction (BridgeProt), a score-aligned generative protocol for ECPE. The protocol first generates candidate emotion–cause records from the full dialogue, then re-evaluates the cause set for each proposed emotion turn within its local context. Valid local decisions selectively refine the initial proposal before deterministic reconstruction converts the resulting pairs into normalized JSON records for direct scoring. Structural validity is evaluated separately from Pair F1 to distinguish schema compliance from pair-prediction accuracy. Experiments span three public benchmarks, with four modality settings across four training regimes. In a three-seed matched SFT comparison against Plain-text post-hoc and Schema-single, BridgeProt  replace/refine attains the highest three-dataset macro Pair F1 of 51.80. Deterministic reconstruction yields 100% final structural validity in the reported analyses. Serialization studies reveal dataset-dependent evidence gains, while efficiency analysis shows latency overhead from repeated verification calls. These findings support score-aligned structured generation for conversational ECPE while exposing its accuracy–latency trade-off. The code is available at https://github.com/xiay9/BridgeProt.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149132621","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}