Pub Date : 2026-03-16DOI: 10.1016/j.autcon.2026.106875
Yu Yao, Yunhua Li, Liman Yang, Zhaoxiong Wang, Molei Peng, Xu Yang
Safe and efficient excavation trajectories are essential for autonomous operation of intelligent electric shovels in open-pit mining. However, irregular ore pile distributions, multi-objective requirements, and operational constraints pose a significant challenge to the real-time generation of high-performance trajectories. This paper formulates the excavation trajectory optimization as a Markov decision process and proposes a real-time multi-objective optimization surrogate model based on reinforcement learning, with the objectives of maximizing bucket fill rate, minimizing mass-specific energy consumption, and reducing excavation time. By embedding the solution evolution into reinforcement learning training process, the model achieves a 2.87 s runtime, 84.13% non-dominated solutions, and a hypervolume value of 0.9403, outperforming other multi-objective optimization algorithms. After optimization, an entropy-based decision-making method is designed to objectively select the final excavation trajectory from obtained non-dominated solutions. Simulations and experiments indicate that the surrogate model and decision-making method effectively enable efficient and stable excavation for electric shovels.
{"title":"Multi-objective optimization of electric shovel excavation trajectories using ore distribution perception and reinforcement learning","authors":"Yu Yao, Yunhua Li, Liman Yang, Zhaoxiong Wang, Molei Peng, Xu Yang","doi":"10.1016/j.autcon.2026.106875","DOIUrl":"https://doi.org/10.1016/j.autcon.2026.106875","url":null,"abstract":"Safe and efficient excavation trajectories are essential for autonomous operation of intelligent electric shovels in open-pit mining. However, irregular ore pile distributions, multi-objective requirements, and operational constraints pose a significant challenge to the real-time generation of high-performance trajectories. This paper formulates the excavation trajectory optimization as a Markov decision process and proposes a real-time multi-objective optimization surrogate model based on reinforcement learning, with the objectives of maximizing bucket fill rate, minimizing mass-specific energy consumption, and reducing excavation time. By embedding the solution evolution into reinforcement learning training process, the model achieves a 2.87 s runtime, 84.13% non-dominated solutions, and a hypervolume value of 0.9403, outperforming other multi-objective optimization algorithms. After optimization, an entropy-based decision-making method is designed to objectively select the final excavation trajectory from obtained non-dominated solutions. Simulations and experiments indicate that the surrogate model and decision-making method effectively enable efficient and stable excavation for electric shovels.","PeriodicalId":8660,"journal":{"name":"Automation in Construction","volume":"20 1","pages":""},"PeriodicalIF":10.3,"publicationDate":"2026-03-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147465519","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 : 2026-03-16DOI: 10.1016/j.autcon.2026.106880
Chikezie Chimere Onyekwena, Yunli Li, Chengkai Fan, Samuel J. Abbey, Wen-Ping Wu, Zhen-Song Chen
Soft soil stabilization poses a major challenge in geotechnical engineering, requiring solutions that balance performance with sustainability. This paper presents automated methodologies combining Machine Learning (ML) and optimization algorithms for designing cement-Supplementary Cementitious Material (SCM) blend binders for effective soil stabilization. Key design variables are investigated, highlighting their pivotal role in achieving optimal strength. Among the ML models, the Extreme Gradient Boosting (XGB) with Grey Wolf Optimization (GWO) achieved the highest predictive accuracy (R2 = 0.9798). Feature evaluations highlight the importance of curing time, binder content, cement proportion, and Ground Granulated Blast-furnace Slag (GGBS) content, while revealing the negative correlation of parameters like plasticity index and liquid limit. GGBS incorporation proves effective in enhancing soil strength. The proposed approach, validated against European standards, demonstrates superior mechanical performance and environmental benefits, with multi-criteria decision analysis identifying sustainable mix designs that balance economic and environmental factors without compromising mechanical performance.
{"title":"Nature-inspired ML for strength estimation and multi-objective optimization of cement-supplementary material-stabilized soft soils","authors":"Chikezie Chimere Onyekwena, Yunli Li, Chengkai Fan, Samuel J. Abbey, Wen-Ping Wu, Zhen-Song Chen","doi":"10.1016/j.autcon.2026.106880","DOIUrl":"https://doi.org/10.1016/j.autcon.2026.106880","url":null,"abstract":"Soft soil stabilization poses a major challenge in geotechnical engineering, requiring solutions that balance performance with sustainability. This paper presents automated methodologies combining Machine Learning (ML) and optimization algorithms for designing cement-Supplementary Cementitious Material (SCM) blend binders for effective soil stabilization. Key design variables are investigated, highlighting their pivotal role in achieving optimal strength. Among the ML models, the Extreme Gradient Boosting (XGB) with Grey Wolf Optimization (GWO) achieved the highest predictive accuracy (R<ce:sup loc=\"post\">2</ce:sup> = 0.9798). Feature evaluations highlight the importance of curing time, binder content, cement proportion, and Ground Granulated Blast-furnace Slag (GGBS) content, while revealing the negative correlation of parameters like plasticity index and liquid limit. GGBS incorporation proves effective in enhancing soil strength. The proposed approach, validated against European standards, demonstrates superior mechanical performance and environmental benefits, with multi-criteria decision analysis identifying sustainable mix designs that balance economic and environmental factors without compromising mechanical performance.","PeriodicalId":8660,"journal":{"name":"Automation in Construction","volume":"13 1","pages":""},"PeriodicalIF":10.3,"publicationDate":"2026-03-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147465583","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 : 2026-03-16DOI: 10.1016/j.autcon.2026.106887
Hua Chai, Tianyi Gao, Junwang Yu, Sylvain Rasneur, Yige Liu, Denis Zastavni, Philip F. Yuan
Traditional timber joinery has been largely replaced by metal connectors due to industrial standardization, compromising the material's inherent low-carbon benefits and recyclability. To address this, this paper proposes a joint-informed computational-robotic workflow for timber-only spatial frames. The approach integrates vector-based graphic statics (VGS), geometric computation, and robotic toolpath generation into a continuous process. As a proof-of-concept, the workflow is demonstrated through the construction of a full-scale, 9.4-m-tall timber tower. While mechanical joint properties were not quantified through laboratory testing, the prototype confirms the system's geometric adaptability and construction feasibility under self-weight. Results indicate that the workflow successfully enabled robotic fabrication of 20 unique spatial nodes, achieved a 74% reduction in embodied carbon compared to steel-jointed equivalents, and facilitated a rapid 10-h reassembly process. This paper establishes a reproducible framework for materially coherent construction, contributing to the advancement of circular building practices and automated timber fabrication.
{"title":"Integrated computational-robotic workflow for complex timber-only structures","authors":"Hua Chai, Tianyi Gao, Junwang Yu, Sylvain Rasneur, Yige Liu, Denis Zastavni, Philip F. Yuan","doi":"10.1016/j.autcon.2026.106887","DOIUrl":"https://doi.org/10.1016/j.autcon.2026.106887","url":null,"abstract":"Traditional timber joinery has been largely replaced by metal connectors due to industrial standardization, compromising the material's inherent low-carbon benefits and recyclability. To address this, this paper proposes a joint-informed computational-robotic workflow for timber-only spatial frames. The approach integrates vector-based graphic statics (VGS), geometric computation, and robotic toolpath generation into a continuous process. As a proof-of-concept, the workflow is demonstrated through the construction of a full-scale, 9.4-m-tall timber tower. While mechanical joint properties were not quantified through laboratory testing, the prototype confirms the system's geometric adaptability and construction feasibility under self-weight. Results indicate that the workflow successfully enabled robotic fabrication of 20 unique spatial nodes, achieved a 74% reduction in embodied carbon compared to steel-jointed equivalents, and facilitated a rapid 10-h reassembly process. This paper establishes a reproducible framework for materially coherent construction, contributing to the advancement of circular building practices and automated timber fabrication.","PeriodicalId":8660,"journal":{"name":"Automation in Construction","volume":"407 1","pages":""},"PeriodicalIF":10.3,"publicationDate":"2026-03-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147465517","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 : 2026-03-16DOI: 10.1016/j.autcon.2026.106884
Zhiyu Zheng, Sylvain Marié, Sylvain Kubler
Semantic tagging of Building Management Systems (BMS) metadata is critical for interoperability but remains labor-intensive. This paper presents a Retrieval-Augmented Generation (RAG) framework, BMS-RAG, that automates point-type classification using Large Language Models (LLMs) with minimal supervision. This framework dynamically retrieves relevant examples to guide the LLM, adapting to heterogeneous naming conventions without model retraining. A lightweight rectification layer enforces compliance with a predefined ontology (e.g., Brick), mitigating hallucinations. Evaluated on six real-world datasets, BMS-RAG achieves state-of-the-art results, consistently outperforming static few-shot LLM baselines by up to 15% in F1 score, with several datasets reaching near- or full 100% accuracy using our minimal, quality-driven context size. Grounded in a systematic ablation study of key architectural components, this paper’s main contribution is the application of RAG to BMS metadata tagging, offering a scalable, accurate, and low-effort pathway toward semantic interoperability.
{"title":"Retrieval-augmented LLM with structured sampling for Building Management Systems point tagging under minimal context","authors":"Zhiyu Zheng, Sylvain Marié, Sylvain Kubler","doi":"10.1016/j.autcon.2026.106884","DOIUrl":"https://doi.org/10.1016/j.autcon.2026.106884","url":null,"abstract":"Semantic tagging of Building Management Systems (BMS) metadata is critical for interoperability but remains labor-intensive. This paper presents a Retrieval-Augmented Generation (RAG) framework, BMS-RAG, that automates point-type classification using Large Language Models (LLMs) with minimal supervision. This framework dynamically retrieves relevant examples to guide the LLM, adapting to heterogeneous naming conventions without model retraining. A lightweight rectification layer enforces compliance with a predefined ontology (e.g., Brick), mitigating hallucinations. Evaluated on six real-world datasets, BMS-RAG achieves state-of-the-art results, consistently outperforming static few-shot LLM baselines by up to 15% in F1 score, with several datasets reaching near- or full 100% accuracy using our minimal, quality-driven context size. Grounded in a systematic ablation study of key architectural components, this paper’s main contribution is the application of RAG to BMS metadata tagging, offering a scalable, accurate, and low-effort pathway toward semantic interoperability.","PeriodicalId":8660,"journal":{"name":"Automation in Construction","volume":"10 1","pages":""},"PeriodicalIF":10.3,"publicationDate":"2026-03-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147465518","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 : 2026-03-15DOI: 10.1016/j.autcon.2026.106885
Liuhong Zhang, Xiaogang Wang, Min Wang, Yu Liu, Zhiwei Yin, Xinyu Wu, Chao Sun
Automated Guided Vehicles face challenges of glass wall perception distortion and autonomous localization drift when operating in intelligent construction. This paper proposes a marker-free navigation framework that achieves co-optimization through tight coupling of the perception and localization layers. The framework's perception layer employs a Fresnel optics-based physical model to enable real-time detection and accurate reconstruction of glass walls. Within the localization layer, a rotation-translation decoupled matching algorithm accomplishes global cold-start localization, and a multi-resolution manifold optimization algorithm achieves accurate, robust autonomous positioning during navigation, effectively suppressing localization drift in degenerative built environments. When tested in intelligent construction containing glass walls, the framework achieved 96.80% SR, absolute trajectory error of 8.189 cm, glass reconstruction accuracy of 1.83 cm, and per-frame processing time of 31.7 ms. This work will validate this framework in larger and more diverse glass-rich construction environments.
{"title":"Robust AGV navigation in degenerative built environments with glass walls: Perception and localization co-optimization","authors":"Liuhong Zhang, Xiaogang Wang, Min Wang, Yu Liu, Zhiwei Yin, Xinyu Wu, Chao Sun","doi":"10.1016/j.autcon.2026.106885","DOIUrl":"https://doi.org/10.1016/j.autcon.2026.106885","url":null,"abstract":"Automated Guided Vehicles face challenges of glass wall perception distortion and autonomous localization drift when operating in intelligent construction. This paper proposes a marker-free navigation framework that achieves co-optimization through tight coupling of the perception and localization layers. The framework's perception layer employs a Fresnel optics-based physical model to enable real-time detection and accurate reconstruction of glass walls. Within the localization layer, a rotation-translation decoupled matching algorithm accomplishes global cold-start localization, and a multi-resolution manifold optimization algorithm achieves accurate, robust autonomous positioning during navigation, effectively suppressing localization drift in degenerative built environments. When tested in intelligent construction containing glass walls, the framework achieved 96.80% SR, absolute trajectory error of 8.189 cm, glass reconstruction accuracy of 1.83 cm, and per-frame processing time of 31.7 ms. This work will validate this framework in larger and more diverse glass-rich construction environments.","PeriodicalId":8660,"journal":{"name":"Automation in Construction","volume":"5 1","pages":""},"PeriodicalIF":10.3,"publicationDate":"2026-03-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147465520","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 : 2026-03-01Epub Date: 2026-01-14DOI: 10.1016/j.autcon.2026.106776
Seungkeun Yeom , Juui Kim , Seungwon Seo , Seongkyun Ahn , Choongwan Koo , Taehoon Hong
This paper investigates how personality traits and psychological-cognitive states influence task performance, safety, and physiological responses of novice tower crane operators through a virtual reality (VR) simulation integrated with continuous biometric monitoring. Fifty participants completed object lifting, obstacle navigation, and precision placement tasks while personality profiles and biosignals (ECG, EDA) were collected and analyzed using principal component analysis,cluster-based classification, and additional statistical methods. High extraversion and situational awareness enhanced speed and accuracy, whereas high openness, stress sensitivity, and acrophobia led to longer durations and reduced accuracy. High conscientiousness shortened task times by 19.12% but increased collisions by approximately threefold, revealing a trade-off between efficiency and safety. By integrating behavioral, cognitive, and physiological data, this work advances technology-enabled, data-driven safety management. The proposed approach enables automated operator risk profiling, intelligent task allocation, and proactive safety interventions for high-rise construction projects involving crane operations.
{"title":"Virtual reality-based experimental analysis of personality and cognitive traits on task performance and safety in novice tower crane operators","authors":"Seungkeun Yeom , Juui Kim , Seungwon Seo , Seongkyun Ahn , Choongwan Koo , Taehoon Hong","doi":"10.1016/j.autcon.2026.106776","DOIUrl":"10.1016/j.autcon.2026.106776","url":null,"abstract":"<div><div>This paper investigates how personality traits and psychological-cognitive states influence task performance, safety, and physiological responses of novice tower crane operators through a virtual reality (VR) simulation integrated with continuous biometric monitoring. Fifty participants completed object lifting, obstacle navigation, and precision placement tasks while personality profiles and biosignals (ECG, EDA) were collected and analyzed using principal component analysis,cluster-based classification, and additional statistical methods. High extraversion and situational awareness enhanced speed and accuracy, whereas high openness, stress sensitivity, and acrophobia led to longer durations and reduced accuracy. High conscientiousness shortened task times by 19.12% but increased collisions by approximately threefold, revealing a trade-off between efficiency and safety. By integrating behavioral, cognitive, and physiological data, this work advances technology-enabled, data-driven safety management. The proposed approach enables automated operator risk profiling, intelligent task allocation, and proactive safety interventions for high-rise construction projects involving crane operations.</div></div>","PeriodicalId":8660,"journal":{"name":"Automation in Construction","volume":"183 ","pages":"Article 106776"},"PeriodicalIF":11.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145957799","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 : 2026-03-01Epub Date: 2026-01-21DOI: 10.1016/j.autcon.2026.106786
Insoo Jeong , Junghoon Kim , Seungmo Lim , Jeongbin Hwang , Seokho Chi
This paper proposes a lightweight domain adaptation framework for construction safety monitoring by fine-tuning a pretrained text-to-image diffusion model using Low-Rank Adaptation (LoRA). To simulate high-risk construction environments underrepresented in training data, the model was adapted to environmental features and specific hazards, focusing on visually dominant scenarios including falls, struck-by, and caught-in incidents. To address data scarcity, Multi-LoRA fine-tuning was conducted using 20 images per hazard type (totaling 60 across three hazards) and 30 background images, enabling both contextual and hazard-specific adaptation. The generated images achieved the highest semantic consistency, yielding the top mean Contrastive Language-Image Pre-training (CLIP) scores with minimal variance, and improved visual realism by reducing the Fréchet Inception Distance (FID) by 86.72 points. Furthermore, a YOLOv8 model trained exclusively on these synthetic images achieved a mean average precision ([email protected]:0.95) of 94.1% on real-world frames, comparable to a baseline model trained on real data.
{"title":"Addressing data scarcity in construction safety monitoring using low-rank adaptation (LoRA)-tuned domain-specific image generation","authors":"Insoo Jeong , Junghoon Kim , Seungmo Lim , Jeongbin Hwang , Seokho Chi","doi":"10.1016/j.autcon.2026.106786","DOIUrl":"10.1016/j.autcon.2026.106786","url":null,"abstract":"<div><div>This paper proposes a lightweight domain adaptation framework for construction safety monitoring by fine-tuning a pretrained text-to-image diffusion model using Low-Rank Adaptation (LoRA). To simulate high-risk construction environments underrepresented in training data, the model was adapted to environmental features and specific hazards, focusing on visually dominant scenarios including falls, struck-by, and caught-in incidents. To address data scarcity, Multi-LoRA fine-tuning was conducted using 20 images per hazard type (totaling 60 across three hazards) and 30 background images, enabling both contextual and hazard-specific adaptation. The generated images achieved the highest semantic consistency, yielding the top mean Contrastive Language-Image Pre-training (CLIP) scores with minimal variance, and improved visual realism by reducing the Fréchet Inception Distance (FID) by 86.72 points. Furthermore, a YOLOv8 model trained exclusively on these synthetic images achieved a mean average precision ([email protected]:0.95) of 94.1% on real-world frames, comparable to a baseline model trained on real data.</div></div>","PeriodicalId":8660,"journal":{"name":"Automation in Construction","volume":"183 ","pages":"Article 106786"},"PeriodicalIF":11.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146014826","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 : 2026-03-01Epub Date: 2026-01-21DOI: 10.1016/j.autcon.2026.106779
Mohamed Elrifaee , Tarek Zayed , Ahmed Mansour , Eslam Ali
Construction sites remain among the most hazardous work environments, where the lack of non-intrusive, worker-independent monitoring systems limits proactive safety management. Compared to existing approaches that rely heavily on wearables, RFID tags, or bespoke infrastructure, this paper presents a passive and non-intrusive framework leveraging WiFi probe request tracking for safety monitoring in semi-open areas with static hazards. Using low-cost TP-Link routers, the proposed system localizes workers without requiring active participation or additional equipment. To improve robustness beyond conventional fingerprinting models, a joint Autoencoder–Transformer architecture is employed to capture latent dependencies among access points, significantly reducing localization uncertainty. The resulting position estimates are integrated into a modified Zonal Safety Analysis (mZSA) framework adapted for semi-open construction zones. Unlike deterministic approaches that overlook error variability, the proposed method incorporates distribution-specific error modeling, enabling confidence-aware risk buffers. The framework provides a scalable, uncertainty-aware pathway for real-time risk detection in semi-open construction environments.
{"title":"Uncertainty-aware risk mapping with passive WiFi and modified Zonal Safety Analysis (mZSA) in BIM for construction","authors":"Mohamed Elrifaee , Tarek Zayed , Ahmed Mansour , Eslam Ali","doi":"10.1016/j.autcon.2026.106779","DOIUrl":"10.1016/j.autcon.2026.106779","url":null,"abstract":"<div><div>Construction sites remain among the most hazardous work environments, where the lack of non-intrusive, worker-independent monitoring systems limits proactive safety management. Compared to existing approaches that rely heavily on wearables, RFID tags, or bespoke infrastructure, this paper presents a passive and non-intrusive framework leveraging WiFi probe request tracking for safety monitoring in semi-open areas with static hazards. Using low-cost TP-Link routers, the proposed system localizes workers without requiring active participation or additional equipment. To improve robustness beyond conventional fingerprinting models, a joint Autoencoder–Transformer architecture is employed to capture latent dependencies among access points, significantly reducing localization uncertainty. The resulting position estimates are integrated into a modified Zonal Safety Analysis (mZSA) framework adapted for semi-open construction zones. Unlike deterministic approaches that overlook error variability, the proposed method incorporates distribution-specific error modeling, enabling confidence-aware risk buffers. The framework provides a scalable, uncertainty-aware pathway for real-time risk detection in semi-open construction environments.</div></div>","PeriodicalId":8660,"journal":{"name":"Automation in Construction","volume":"183 ","pages":"Article 106779"},"PeriodicalIF":11.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146014921","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 : 2026-03-01Epub Date: 2026-01-20DOI: 10.1016/j.autcon.2026.106789
Mingchao Li , Zuguang Zhang , Qiubing Ren , Yantao Yu , Jingyue Yuan , Jiamei Ma
Substantial crack imagery is hard to acquire in dam structural inspection due to high costs and risks. Crack image generation, as a crucial yet challenging visual task, still struggles with the quality-diversity trade-off under data scarcity. This paper thus presents CrackFSGAN, a few-shot Generative Adversarial Network (GAN) adaptation method for generating realistic, diverse dam crack images from limited samples. It incorporates the Cross-Scale Channel Interaction (CSCI) module to ensure robust gradient flow across network weights for efficient training, and the Self-Supervised Discriminator (SSDr), a redesigned feature-encoder with an additional decoder, to learn more discriminative, region-extensive feature maps. Extensive experiments on multiple damage datasets against state-of-the-art GANs validate CrackFSGAN's superiority in few-shot image synthesis quality and diversity, and its effectiveness in data augmentation for downstream crack detection tasks. Notably, it supports high-resolution (1024 × 1024 pixel2) crack image generation, offering a promising solution to data scarcity and advancing intelligent structural damage detection.
{"title":"Few-shot GAN adaptation for high-fidelity and diverse crack image generation in dam damage detection","authors":"Mingchao Li , Zuguang Zhang , Qiubing Ren , Yantao Yu , Jingyue Yuan , Jiamei Ma","doi":"10.1016/j.autcon.2026.106789","DOIUrl":"10.1016/j.autcon.2026.106789","url":null,"abstract":"<div><div>Substantial crack imagery is hard to acquire in dam structural inspection due to high costs and risks. Crack image generation, as a crucial yet challenging visual task, still struggles with the quality-diversity trade-off under data scarcity. This paper thus presents CrackFSGAN, a few-shot Generative Adversarial Network (GAN) adaptation method for generating realistic, diverse dam crack images from limited samples. It incorporates the Cross-Scale Channel Interaction (CSCI) module to ensure robust gradient flow across network weights for efficient training, and the Self-Supervised Discriminator (SSDr), a redesigned feature-encoder with an additional decoder, to learn more discriminative, region-extensive feature maps. Extensive experiments on multiple damage datasets against state-of-the-art GANs validate CrackFSGAN's superiority in few-shot image synthesis quality and diversity, and its effectiveness in data augmentation for downstream crack detection tasks. Notably, it supports high-resolution (1024 × 1024 pixel<sup>2</sup>) crack image generation, offering a promising solution to data scarcity and advancing intelligent structural damage detection.</div></div>","PeriodicalId":8660,"journal":{"name":"Automation in Construction","volume":"183 ","pages":"Article 106789"},"PeriodicalIF":11.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146014902","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 : 2026-03-01Epub Date: 2026-01-27DOI: 10.1016/j.autcon.2026.106802
Guoqiang Huang , Chengjin Qin , Jie Lu , Pengcheng Xia , Haodi Wang , Chengliang Liu
Accurately predicting muck particle size distribution (PSD) of Tunnel Boring Machine (TBM) is constrained by the cumbersome process of manual annotation and environmental noise. This paper investigates robust prediction of muck PSD curve under noisy TBM operation conditions, while reducing reliance on manual annotations. A noise-robust self-supervised learning method with frequency-bias decomposition is proposed, which integrates contrastive pre-training based on noise augmentation, frequency-domain bias decomposition, and hybrid edge-aware loss function. The experiments show that with only 10% annotation, it achieves performance comparable to existing models trained on 90% annotation, with a maximum particle size MAPE of 6.7% and Rosin-Rammler parameter errors between 10 and 20%. These results demonstrate a low-cost, accurate, and noise-robust approach for muck monitoring, substantially reducing the need for manual annotation and improving prediction reliability. Future work will combine muck PSD with multi-modal TBM excavation data to support intelligent tunneling decision-making.
{"title":"Noise-robust self-supervised learning with frequency-bias decomposition for TBM muck particle size distribution prediction","authors":"Guoqiang Huang , Chengjin Qin , Jie Lu , Pengcheng Xia , Haodi Wang , Chengliang Liu","doi":"10.1016/j.autcon.2026.106802","DOIUrl":"10.1016/j.autcon.2026.106802","url":null,"abstract":"<div><div>Accurately predicting muck particle size distribution (PSD) of Tunnel Boring Machine (TBM) is constrained by the cumbersome process of manual annotation and environmental noise. This paper investigates robust prediction of muck PSD curve under noisy TBM operation conditions, while reducing reliance on manual annotations. A noise-robust self-supervised learning method with frequency-bias decomposition is proposed, which integrates contrastive pre-training based on noise augmentation, frequency-domain bias decomposition, and hybrid edge-aware loss function. The experiments show that with only 10% annotation, it achieves performance comparable to existing models trained on 90% annotation, with a maximum particle size MAPE of 6.7% and Rosin-Rammler parameter errors between 10 and 20%. These results demonstrate a low-cost, accurate, and noise-robust approach for muck monitoring, substantially reducing the need for manual annotation and improving prediction reliability. Future work will combine muck PSD with multi-modal TBM excavation data to support intelligent tunneling decision-making.</div></div>","PeriodicalId":8660,"journal":{"name":"Automation in Construction","volume":"183 ","pages":"Article 106802"},"PeriodicalIF":11.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146071736","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}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：