Pub Date : 2025-11-18DOI: 10.1016/j.compind.2025.104417
Keke Zha, Jiabin Yuan, Lili Fan, Yiyu Shen, Xu Liu
Rock segmentation, a crucial task in deep space exploration, demands high algorithmic accuracy. However, existing high-precision deep learning models often suffer from high computational complexity and energy consumption, which limit their deployment in resource-constrained space environments. To address these challenges, we present the spike temporal residual transformer network (ST-RTNet), the first spike-driven rock segmentation model directly trained with spiking neural network (SNN) architectures. ST-RTNet integrates convolutional layers and Transformer modules, introducing a novel attention mechanism that incorporates the temporal dimension of SNNs. By leveraging neuron voltage dynamics over time, ST-RTNet captures both temporal and spatial information, thereby enhancing the precision of segmentation. We evaluate ST-RTNet on three datasets and compare its performance with recent rock segmentation models. Experiments demonstrate that ST-RTNet achieves up to 90.13% energy reduction on INT-8 chips and 83.76% on Float-32 chips compared to artificial neural network models, while maintaining competitive segmentation accuracy. These findings demonstrate that ST-RTNet provides an efficient and accurate solution for rock segmentation in space exploration.
{"title":"ST-RTNet: An energy-efficient spike temporal residual transformer network for rock segmentation in deep space exploration","authors":"Keke Zha, Jiabin Yuan, Lili Fan, Yiyu Shen, Xu Liu","doi":"10.1016/j.compind.2025.104417","DOIUrl":"10.1016/j.compind.2025.104417","url":null,"abstract":"<div><div>Rock segmentation, a crucial task in deep space exploration, demands high algorithmic accuracy. However, existing high-precision deep learning models often suffer from high computational complexity and energy consumption, which limit their deployment in resource-constrained space environments. To address these challenges, we present the spike temporal residual transformer network (ST-RTNet), the first spike-driven rock segmentation model directly trained with spiking neural network (SNN) architectures. ST-RTNet integrates convolutional layers and Transformer modules, introducing a novel attention mechanism that incorporates the temporal dimension of SNNs. By leveraging neuron voltage dynamics over time, ST-RTNet captures both temporal and spatial information, thereby enhancing the precision of segmentation. We evaluate ST-RTNet on three datasets and compare its performance with recent rock segmentation models. Experiments demonstrate that ST-RTNet achieves up to 90.13% energy reduction on INT-8 chips and 83.76% on Float-32 chips compared to artificial neural network models, while maintaining competitive segmentation accuracy. These findings demonstrate that ST-RTNet provides an efficient and accurate solution for rock segmentation in space exploration.</div></div>","PeriodicalId":55219,"journal":{"name":"Computers in Industry","volume":"174 ","pages":"Article 104417"},"PeriodicalIF":9.1,"publicationDate":"2025-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145560051","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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 : 2025-11-18DOI: 10.1016/j.compind.2025.104415
Siyue Yang, Qi Sima, Liang Shen, Yukun Bao
Electricity market liberalisation has heightened competitive pressure among retailers, necessitating the accurate forecasting of retail consumer demand to support informed strategic decision-making in markets. Within this landscape, electricity retailers face two critical challenges: expanding, dynamic customer bases with limited historical consumption data from newly enrolled customers; and heterogeneous consumption patterns across diverse consumers, which necessitates tailored analytical approaches. However, conventional local forecasting methods, which require building individual models for each consumer, become operationally inefficient, and it is practically impossible to use these to meet such challenges. Hence, this study proposes a decomposition-based multi-sight convolutional neural network as a unified global method to generate predictions for multiple consumers. Given the inherent periodicity in electricity consumption profiles, this model incorporates three modules to handle both the commonality and diversity of periodic features among different consumers simultaneously: (1) a built-in decomposition module to recognise universal periodic patterns, enabling generalisation to new customers through shared temporal variations; (2) temporal transformation from one-dimensional input sequences to two-dimensional space according to daily periodicity, representing temporal dependencies along both intra- and inter-day dimensions; (3) a novel multi-sight convolutional neural network block comprising parallel convolution branches specialised for diverse subregions of the two-dimensional tensors, effectively detecting and modelling heterogeneous consumer-specific periodic sub-patterns across multiple series. Experiments using real-world datasets demonstrate that the proposed model achieves superior performance for global forecasting tasks in terms of both prediction accuracy and computational efficiency, compared with advanced methods. Ablation studies validate the effectiveness of the designed architecture.
{"title":"Global electricity demand forecasting for multi-consumer retailers using a decomposition-based multi-sight convolutional neural network","authors":"Siyue Yang, Qi Sima, Liang Shen, Yukun Bao","doi":"10.1016/j.compind.2025.104415","DOIUrl":"10.1016/j.compind.2025.104415","url":null,"abstract":"<div><div>Electricity market liberalisation has heightened competitive pressure among retailers, necessitating the accurate forecasting of retail consumer demand to support informed strategic decision-making in markets. Within this landscape, electricity retailers face two critical challenges: expanding, dynamic customer bases with limited historical consumption data from newly enrolled customers; and heterogeneous consumption patterns across diverse consumers, which necessitates tailored analytical approaches. However, conventional local forecasting methods, which require building individual models for each consumer, become operationally inefficient, and it is practically impossible to use these to meet such challenges. Hence, this study proposes a decomposition-based multi-sight convolutional neural network as a unified global method to generate predictions for multiple consumers. Given the inherent periodicity in electricity consumption profiles, this model incorporates three modules to handle both the commonality and diversity of periodic features among different consumers simultaneously: (1) a built-in decomposition module to recognise universal periodic patterns, enabling generalisation to new customers through shared temporal variations; (2) temporal transformation from one-dimensional input sequences to two-dimensional space according to daily periodicity, representing temporal dependencies along both intra- and inter-day dimensions; (3) a novel multi-sight convolutional neural network block comprising parallel convolution branches specialised for diverse subregions of the two-dimensional tensors, effectively detecting and modelling heterogeneous consumer-specific periodic sub-patterns across multiple series. Experiments using real-world datasets demonstrate that the proposed model achieves superior performance for global forecasting tasks in terms of both prediction accuracy and computational efficiency, compared with advanced methods. Ablation studies validate the effectiveness of the designed architecture.</div></div>","PeriodicalId":55219,"journal":{"name":"Computers in Industry","volume":"174 ","pages":"Article 104415"},"PeriodicalIF":9.1,"publicationDate":"2025-11-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145560134","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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 : 2025-11-17DOI: 10.1016/j.compind.2025.104416
Jianbin Xin , Peiyan Guo , Hongbo Li , Andrea D’Ariano , Yanhong Liu
The robotic mobile fulfillment system (RMFS) has revolutionized the manufacturing and logistics industries by enhancing the efficiency of automated storage and order fulfillment through automated guided vehicles (AGVs). However, existing multi-AGV path planning methods in RMFS typically decouple path planning from conflict resolution, thereby simplifying the problem but limiting system performance, especially in dynamic and complex operational environments. To address this challenge, we introduce a novel learning-based hierarchical framework for lifelong multi-AGV path planning. Our framework integrates a dual-mode heuristic global guidance planner with a local reinforcement learning planner, leveraging asynchronous proximal policy optimization and a recurrent neural network to achieve fully decentralized, online navigation. Critically, our dual-mode guidance mechanism adapts to multi-phase transport tasks by enabling unloaded AGVs to travel beneath stationary pods—a key distinction from conventional methods. This approach mitigates congestion in narrow corridors and boosts overall system throughput. Experimental results demonstrate that our method outperforms state-of-the-art centralized and decentralized approaches in large-scale deployments, achieving higher success rates and throughput while significantly reducing computational costs. This research thus offers a scalable and efficient solution to the complex path-planning challenges inherent in RMFS.
{"title":"Dual-mode guided reinforcement learning for decentralized lifelong path planning of multiple automated guided vehicles in robotic mobile fulfillment systems","authors":"Jianbin Xin , Peiyan Guo , Hongbo Li , Andrea D’Ariano , Yanhong Liu","doi":"10.1016/j.compind.2025.104416","DOIUrl":"10.1016/j.compind.2025.104416","url":null,"abstract":"<div><div>The robotic mobile fulfillment system (RMFS) has revolutionized the manufacturing and logistics industries by enhancing the efficiency of automated storage and order fulfillment through automated guided vehicles (AGVs). However, existing multi-AGV path planning methods in RMFS typically decouple path planning from conflict resolution, thereby simplifying the problem but limiting system performance, especially in dynamic and complex operational environments. To address this challenge, we introduce a novel learning-based hierarchical framework for lifelong multi-AGV path planning. Our framework integrates a dual-mode heuristic global guidance planner with a local reinforcement learning planner, leveraging asynchronous proximal policy optimization and a recurrent neural network to achieve fully decentralized, online navigation. Critically, our dual-mode guidance mechanism adapts to multi-phase transport tasks by enabling unloaded AGVs to travel beneath stationary pods—a key distinction from conventional methods. This approach mitigates congestion in narrow corridors and boosts overall system throughput. Experimental results demonstrate that our method outperforms state-of-the-art centralized and decentralized approaches in large-scale deployments, achieving higher success rates and throughput while significantly reducing computational costs. This research thus offers a scalable and efficient solution to the complex path-planning challenges inherent in RMFS.</div></div>","PeriodicalId":55219,"journal":{"name":"Computers in Industry","volume":"174 ","pages":"Article 104416"},"PeriodicalIF":9.1,"publicationDate":"2025-11-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145560050","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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 : 2025-11-12DOI: 10.1016/j.compind.2025.104412
Shanshan Yu , Jian Zhang , Xiaoyuan He
In this study, a robust binocular stereo vision method based on modified three-dimensional digital image correlation is proposed to address challenging measurement conditions, including camera motion and poor correlation quality. To mitigate camera motion-induced errors, which can distort the external imaging geometry of binocular systems and generate false displacement measurements, a flexible self-calibration approach is introduced. This method employs an inverse-depth parameterized coordinate framework to overcome the planar constraints inherent in conventional techniques. For resolving poor correlation issues caused by sparse texture, surface reflections, and perspective differences, a feature-guided subset construction strategy is developed. This approach emphasizes key gray-scale features while integrating an M-estimator-based correlation criterion to effectively exclude subsets with abnormal gray variations. Experimental validation through a wing model loading test demonstrates the proposed method's superior capability in capturing comprehensive deformation fields, showcasing significant improvements over existing approaches.
{"title":"Enhanced three-dimensional displacement monitoring: Integrating the KAZE detector with digital image correlation for scenario-free anti-disturbance analysis","authors":"Shanshan Yu , Jian Zhang , Xiaoyuan He","doi":"10.1016/j.compind.2025.104412","DOIUrl":"10.1016/j.compind.2025.104412","url":null,"abstract":"<div><div>In this study, a robust binocular stereo vision method based on modified three-dimensional digital image correlation is proposed to address challenging measurement conditions, including camera motion and poor correlation quality. To mitigate camera motion-induced errors, which can distort the external imaging geometry of binocular systems and generate false displacement measurements, a flexible self-calibration approach is introduced. This method employs an inverse-depth parameterized coordinate framework to overcome the planar constraints inherent in conventional techniques. For resolving poor correlation issues caused by sparse texture, surface reflections, and perspective differences, a feature-guided subset construction strategy is developed. This approach emphasizes key gray-scale features while integrating an M-estimator-based correlation criterion to effectively exclude subsets with abnormal gray variations. Experimental validation through a wing model loading test demonstrates the proposed method's superior capability in capturing comprehensive deformation fields, showcasing significant improvements over existing approaches.</div></div>","PeriodicalId":55219,"journal":{"name":"Computers in Industry","volume":"174 ","pages":"Article 104412"},"PeriodicalIF":9.1,"publicationDate":"2025-11-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145515614","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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 : 2025-11-12DOI: 10.1016/j.compind.2025.104413
Li Zhang, Yunjie He, Tao Diao, Ziming Liu
In recent years, the growing demand for laser welding products has led to higher quality requirements for welded products, making the precise and timely inspection of welding defects a critical issue. This paper focuses on detecting laser welding defects in the safety vent of electric vehicle batteries. We collected welding defect images from automated laser spot welding machines in manufacturing workshops and built a dataset for analysis. We propose a real-time lightweight defect detection algorithm, named Welding Defect Classification Network, based on a deep compression convolutional neural network. The proposed method employs an improved MobileNetV2 model combined with the Efficient Channel Attention module and Teacher-Free Knowledge Distillation technology. In experiments conducted on the welding defect dataset, our model achieved a prediction accuracy of 96.50%, outperforming some well-known lightweight networks while maintaining low model complexity. Furthermore, we deployed the model on a Raspberry Pi 4B with the support of Intel’s Neural Computing Stick, achieving an inference speed of 29.89 ms per defect prediction, meeting the real-time requirements of actual industries.
近年来，对激光焊接产品的需求不断增长，对焊接产品的质量要求也越来越高，焊接缺陷的精确、及时检测成为一个关键问题。本文主要研究了电动汽车电池安全孔激光焊接缺陷的检测。我们收集了制造车间自动激光点焊机的焊接缺陷图像，并建立了数据集进行分析。提出了一种基于深度压缩卷积神经网络的实时轻量级缺陷检测算法——焊接缺陷分类网络。该方法采用了一种改进的MobileNetV2模型，结合了高效通道关注模块和无教师知识蒸馏技术。在焊接缺陷数据集上进行的实验中，该模型的预测准确率达到96.50%，在保持较低模型复杂度的同时，优于一些知名的轻量级网络。在Intel的Neural Computing Stick的支持下，我们将模型部署在Raspberry Pi 4B上，每个缺陷预测的推理速度达到29.89 ms，满足了实际行业的实时性要求。
{"title":"A real-time lightweight laser welding defect inspection algorithm based on deep learning","authors":"Li Zhang, Yunjie He, Tao Diao, Ziming Liu","doi":"10.1016/j.compind.2025.104413","DOIUrl":"10.1016/j.compind.2025.104413","url":null,"abstract":"<div><div>In recent years, the growing demand for laser welding products has led to higher quality requirements for welded products, making the precise and timely inspection of welding defects a critical issue. This paper focuses on detecting laser welding defects in the safety vent of electric vehicle batteries. We collected welding defect images from automated laser spot welding machines in manufacturing workshops and built a dataset for analysis. We propose a real-time lightweight defect detection algorithm, named Welding Defect Classification Network, based on a deep compression convolutional neural network. The proposed method employs an improved MobileNetV2 model combined with the Efficient Channel Attention module and Teacher-Free Knowledge Distillation technology. In experiments conducted on the welding defect dataset, our model achieved a prediction accuracy of 96.50%, outperforming some well-known lightweight networks while maintaining low model complexity. Furthermore, we deployed the model on a Raspberry Pi 4B with the support of Intel’s Neural Computing Stick, achieving an inference speed of 29.89 ms per defect prediction, meeting the real-time requirements of actual industries.</div></div>","PeriodicalId":55219,"journal":{"name":"Computers in Industry","volume":"174 ","pages":"Article 104413"},"PeriodicalIF":9.1,"publicationDate":"2025-11-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145509642","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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 : 2025-11-12DOI: 10.1016/j.compind.2025.104410
Aitor Izuzquiza , Miguel A. Patricio , Juan J. Cuadrado-Gallego , Antonio Berlanga , José M. Molina
This paper focuses on an approach to address large-scale data gathered from heterogeneous sources by integrating static and dynamic data in hierarchical clusterization, and its application to the analysis of retail branches. Traditionally, branch clustering analysis has relied on static information and the utilization of statistical measures to extract relevant features from the dynamic data and incorporate them into the static dataset; however, the application of this approach presents several challenges. This research proposes a solution that addresses these disadvantages while aiming to maintain the success achieved when applying unsupervised machine learning algorithms. The paper presents an approach based on the integration of static attributes and time series data in a hierarchical clustering manner that enables the identification of key performance indicators and offers insight into factors that influence branch performance over time. The results show the potential to optimize resource allocation, inventory management, and customer service strategies. The proposed approach is demonstrated using retail shop data from a Spanish telecommunications company (Grupo Masmovil), highlighting its effectiveness in enhancing cluster profiling and offering meaningful insights beyond the prevailing approaches. This method presents significant enrichment for clustering analysis that can be applied to different domains.
{"title":"Integrating static and dynamic hierarchical clustering and its application to retail segmentation","authors":"Aitor Izuzquiza , Miguel A. Patricio , Juan J. Cuadrado-Gallego , Antonio Berlanga , José M. Molina","doi":"10.1016/j.compind.2025.104410","DOIUrl":"10.1016/j.compind.2025.104410","url":null,"abstract":"<div><div>This paper focuses on an approach to address large-scale data gathered from heterogeneous sources by integrating static and dynamic data in hierarchical clusterization, and its application to the analysis of retail branches. Traditionally, branch clustering analysis has relied on static information and the utilization of statistical measures to extract relevant features from the dynamic data and incorporate them into the static dataset; however, the application of this approach presents several challenges. This research proposes a solution that addresses these disadvantages while aiming to maintain the success achieved when applying unsupervised machine learning algorithms. The paper presents an approach based on the integration of static attributes and time series data in a hierarchical clustering manner that enables the identification of key performance indicators and offers insight into factors that influence branch performance over time. The results show the potential to optimize resource allocation, inventory management, and customer service strategies. The proposed approach is demonstrated using retail shop data from a Spanish telecommunications company (Grupo Masmovil), highlighting its effectiveness in enhancing cluster profiling and offering meaningful insights beyond the prevailing approaches. This method presents significant enrichment for clustering analysis that can be applied to different domains.</div></div>","PeriodicalId":55219,"journal":{"name":"Computers in Industry","volume":"174 ","pages":"Article 104410"},"PeriodicalIF":9.1,"publicationDate":"2025-11-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145515667","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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 : 2025-11-10DOI: 10.1016/j.compind.2025.104411
Bo Zhu , Tao Geng , Baoyi Wang , Haoxuan Li , Xianhong Zhang
Accurate recognition of material properties (such as hardness, texture, and strength) is essential for enabling robots to successfully interact with diverse objects. This task becomes particularly challenging in complex, dynamic, and extreme environments, where robots must continuously adapt to fluctuating conditions to perceive their surroundings accurately and perform tasks efficiently. Ultrasonic signals are well-suited for such settings due to their robustness against environmental interference and reliable data transmission under harsh conditions. However, ultrasonic echoes collected in dynamic scenarios often exhibit complex multi-scale and multi-semantic characteristics, which present significant challenges for conventional signal processing methods. To address these issues, we propose a novel material recognition method based on ultrasonic echoes using a Global Frequency Filter-based Pyramidal Dynamic Convolutional Network (GFF-PDCN). The proposed model incorporates three specialized modules: a pyramidal dynamic convolution module, a global frequency filter module, and a non-local attention module, which work collaboratively to capture and process intricate features in ultrasonic signals. Extensive experiments are conducted using a robotic system, including real-world validation in extreme environments for identifying wall material properties. The results demonstrate that our GFF-PDCN model achieves an average recognition accuracy of 95%. Our approach significantly enhances a robot’s capability to acquire, interpret, and process critical environmental information under complex operational conditions. The implementation code is available at: https://github.com/drama-bo/GFF-PDCN.
{"title":"Robust non-contact material recognition for robots in extreme and dynamic environments","authors":"Bo Zhu , Tao Geng , Baoyi Wang , Haoxuan Li , Xianhong Zhang","doi":"10.1016/j.compind.2025.104411","DOIUrl":"10.1016/j.compind.2025.104411","url":null,"abstract":"<div><div>Accurate recognition of material properties (such as hardness, texture, and strength) is essential for enabling robots to successfully interact with diverse objects. This task becomes particularly challenging in complex, dynamic, and extreme environments, where robots must continuously adapt to fluctuating conditions to perceive their surroundings accurately and perform tasks efficiently. Ultrasonic signals are well-suited for such settings due to their robustness against environmental interference and reliable data transmission under harsh conditions. However, ultrasonic echoes collected in dynamic scenarios often exhibit complex multi-scale and multi-semantic characteristics, which present significant challenges for conventional signal processing methods. To address these issues, we propose a novel material recognition method based on ultrasonic echoes using a Global Frequency Filter-based Pyramidal Dynamic Convolutional Network (GFF-PDCN). The proposed model incorporates three specialized modules: a pyramidal dynamic convolution module, a global frequency filter module, and a non-local attention module, which work collaboratively to capture and process intricate features in ultrasonic signals. Extensive experiments are conducted using a robotic system, including real-world validation in extreme environments for identifying wall material properties. The results demonstrate that our GFF-PDCN model achieves an average recognition accuracy of 95%. Our approach significantly enhances a robot’s capability to acquire, interpret, and process critical environmental information under complex operational conditions. The implementation code is available at: <span><span>https://github.com/drama-bo/GFF-PDCN</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":55219,"journal":{"name":"Computers in Industry","volume":"174 ","pages":"Article 104411"},"PeriodicalIF":9.1,"publicationDate":"2025-11-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145484973","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Organisations increasingly use data-driven artificial intelligence (AI) systems in their decision-making processes. These AI systems may operate autonomously, support human decision-makers or increasingly act as collaborative team members. However, data-driven AI systems often function as black boxes, lacking interpretability. This poses a challenge in decision-making, as stakeholders involved in or impacted by the decision-making process frequently need to understand the rationale behind decisions. Moreover, data-driven AI systems operate without leveraging structured domain knowledge. As a result, data-driven AI systems may generate outputs that are misaligned with the decision context, objectives, or constraints, potentially leading to poor decisions or reduced trust in AI systems among users. Consequently, recent years have seen an increasing interest in integrating domain knowledge with data-driven AI. This is evident in neuro-symbolic AI, a subfield of AI that combines neural networks with symbolic AI. While this approach shows promise for enhancing the trustworthiness of AI systems in decision-making, the specific mechanisms by which domain knowledge integration contributes to dimensions of trustworthiness remain insufficiently explored. Therefore, this study reviews and integrates recent knowledge- and data-driven AI literature, along with relevant concepts for decision-making. Building on this foundation, it proposes a lifecycle framework for integrated knowledge- and data-driven AI systems for decision-making, and demonstrates its application through a healthcare application example. It further analyses the dimensions of trustworthiness for knowledge- and data-driven AI systems using the proposed lifecycle framework and application example. In doing so, this study advances the discourse on trustworthy AI for decision-making.
{"title":"Towards trustworthy artificial intelligence for decision-making: A lifecycle perspective on knowledge- and data-driven artificial intelligence systems","authors":"Emiel Miedema, Sabine Waschull, Christos Emmanouilidis","doi":"10.1016/j.compind.2025.104409","DOIUrl":"10.1016/j.compind.2025.104409","url":null,"abstract":"<div><div>Organisations increasingly use data-driven artificial intelligence (AI) systems in their decision-making processes. These AI systems may operate autonomously, support human decision-makers or increasingly act as collaborative team members. However, data-driven AI systems often function as black boxes, lacking interpretability. This poses a challenge in decision-making, as stakeholders involved in or impacted by the decision-making process frequently need to understand the rationale behind decisions. Moreover, data-driven AI systems operate without leveraging structured domain knowledge. As a result, data-driven AI systems may generate outputs that are misaligned with the decision context, objectives, or constraints, potentially leading to poor decisions or reduced trust in AI systems among users. Consequently, recent years have seen an increasing interest in integrating domain knowledge with data-driven AI. This is evident in neuro-symbolic AI, a subfield of AI that combines neural networks with symbolic AI. While this approach shows promise for enhancing the trustworthiness of AI systems in decision-making, the specific mechanisms by which domain knowledge integration contributes to dimensions of trustworthiness remain insufficiently explored. Therefore, this study reviews and integrates recent knowledge- and data-driven AI literature, along with relevant concepts for decision-making. Building on this foundation, it proposes a lifecycle framework for integrated knowledge- and data-driven AI systems for decision-making, and demonstrates its application through a healthcare application example. It further analyses the dimensions of trustworthiness for knowledge- and data-driven AI systems using the proposed lifecycle framework and application example. In doing so, this study advances the discourse on trustworthy AI for decision-making.</div></div>","PeriodicalId":55219,"journal":{"name":"Computers in Industry","volume":"174 ","pages":"Article 104409"},"PeriodicalIF":9.1,"publicationDate":"2025-10-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145404584","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Construction sites are inherently high-risk environments, making safety training for workers crucial to enhancing operational skills, reinforcing safety awareness, and reducing accident risks. Traditional centralized training often fails to engage workers due to monotonous nature and lack of relevance, leading to low efficiency. Moreover, critical resources such as operating instructions, safety guidelines, and accident reports are frequently mismanaged or underutilized. Therefore, this study proposes ConSTRAG, an innovative personalized construction safety training framework. By integrating large language model-empowered agents with knowledge graph reasoning, ConSTRAG generates tailored training materials, significantly improving the relevance and effectiveness of safety training. Validation tests conducted on a dataset of 11,020 questions achieved an average score of 81.25, exceeding the benchmark by 6.94. The generated personalized training materials were evaluated through an expert questionnaire survey, with an average score of 4.16 out of 5 across five dimensions. This research contributes to overcoming individual heterogeneity in construction safety training, enhances training efficiency and effectiveness, and holds potential for extension to other personnel training industries.
{"title":"Personalized safety training for construction workers: A large language model-driven multi-agent framework integrated with knowledge graph reasoning","authors":"Qihua Chen , Xianfei Yin , Beifei Yuan , Qirong Chen","doi":"10.1016/j.compind.2025.104399","DOIUrl":"10.1016/j.compind.2025.104399","url":null,"abstract":"<div><div>Construction sites are inherently high-risk environments, making safety training for workers crucial to enhancing operational skills, reinforcing safety awareness, and reducing accident risks. Traditional centralized training often fails to engage workers due to monotonous nature and lack of relevance, leading to low efficiency. Moreover, critical resources such as operating instructions, safety guidelines, and accident reports are frequently mismanaged or underutilized. Therefore, this study proposes ConSTRAG, an innovative personalized construction safety training framework. By integrating large language model-empowered agents with knowledge graph reasoning, ConSTRAG generates tailored training materials, significantly improving the relevance and effectiveness of safety training. Validation tests conducted on a dataset of 11,020 questions achieved an average score of 81.25, exceeding the benchmark by 6.94. The generated personalized training materials were evaluated through an expert questionnaire survey, with an average score of 4.16 out of 5 across five dimensions. This research contributes to overcoming individual heterogeneity in construction safety training, enhances training efficiency and effectiveness, and holds potential for extension to other personnel training industries.</div></div>","PeriodicalId":55219,"journal":{"name":"Computers in Industry","volume":"174 ","pages":"Article 104399"},"PeriodicalIF":9.1,"publicationDate":"2025-10-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145315230","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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 : 2025-10-10DOI: 10.1016/j.compind.2025.104397
Zhengyang Ling , Duncan McFarlane , Sam Brooks , Lavindra de Silva , Gregory Hawkridge , Alan Thorne
Low-cost digital solutions have been proposed as a means of helping Small and Medium-sized Enterprises (SMEs) in manufacturing. To reduce development costs and enable SMEs to create digital solutions for their specific requirements, workers should be able to configure their own solutions. However, such an approach can be problematic – and at times infeasible – as the SME may not have access to staff with the necessary software skills. Hence, this paper proposes an automated configuration approach for the preparation, customisation, and automatic generation of low-cost digital solutions. This approach was implemented in the development of an Automated Solution Configurator (ASC) platform. The ASC specifically makes use of a particular reference architecture (the so-called Shoestring approach) as a foundation for the design of low-cost digital solutions. These solutions are composed of modules of key functions (referred to as a “Service Module”), which themselves integrate “Building Blocks” (BBs) of low cost technology elements. The paper presents an overview of the ASC platform; its usefulness and usability are evaluated via (a) three industrial application studies, (b) a user study with seven participants and (c) a direct comparison between ASC-based and expert-prepared solutions. The evaluations demonstrate that users with a range of expertise can rapidly create low-cost solutions using the ASC platform. Comparing the ASC-generated code side by side with that written by an expert, the ASC code tends to be longer than a solution developed by an expert but still operates effectively. It is also demonstrated that the ASC approach can support simple solution reuse by reconfiguring technology BBs for different digital solutions.
{"title":"Automated configuration for cost-effective digital solutions","authors":"Zhengyang Ling , Duncan McFarlane , Sam Brooks , Lavindra de Silva , Gregory Hawkridge , Alan Thorne","doi":"10.1016/j.compind.2025.104397","DOIUrl":"10.1016/j.compind.2025.104397","url":null,"abstract":"<div><div>Low-cost digital solutions have been proposed as a means of helping Small and Medium-sized Enterprises (SMEs) in manufacturing. To reduce development costs and enable SMEs to create digital solutions for their specific requirements, workers should be able to configure their own solutions. However, such an approach can be problematic – and at times infeasible – as the SME may not have access to staff with the necessary software skills. Hence, this paper proposes an automated configuration approach for the preparation, customisation, and automatic generation of low-cost digital solutions. This approach was implemented in the development of an Automated Solution Configurator (ASC) platform. The ASC specifically makes use of a particular reference architecture (the so-called <em>Shoestring</em> approach) as a foundation for the design of low-cost digital solutions. These solutions are composed of modules of key functions (referred to as a “Service Module”), which themselves integrate “Building Blocks” (BBs) of low cost technology elements. The paper presents an overview of the ASC platform; its usefulness and usability are evaluated via (a) three industrial application studies, (b) a user study with seven participants and (c) a direct comparison between ASC-based and expert-prepared solutions. The evaluations demonstrate that users with a range of expertise can rapidly create low-cost solutions using the ASC platform. Comparing the ASC-generated code side by side with that written by an expert, the ASC code tends to be longer than a solution developed by an expert but still operates effectively. It is also demonstrated that the ASC approach can support simple solution reuse by reconfiguring technology BBs for different digital solutions.</div></div>","PeriodicalId":55219,"journal":{"name":"Computers in Industry","volume":"173 ","pages":"Article 104397"},"PeriodicalIF":9.1,"publicationDate":"2025-10-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145261909","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"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