Pub Date : 2025-05-26DOI: 10.1016/j.compind.2025.104313
Mingda Chen , Ruiyun Yu , Zhiyuan Liang , Kun Li , Haifei Qi
In the manufacturing industry, the demand for fault-prediction solutions is increasing to prevent unexpected downtimes and reduce maintenance costs. Although deep-learning methods have demonstrated excellent performance in this domain, the current methods typically overlook the analysis of variable and random processes within mixed-model production, which is a manufacturing strategy that offers flexibility and efficiency in satisfying diverse consumer demands. Hence, we propose the multiscale process-aware retention network (MPRNet), which segments a time series into multiscale patches, thus enabling the model to focus on local information within each production process and correlations across all production processes. Furthermore, the network incorporates a cross-channel interaction module designed to dynamically capture the interactions between various sensor data types using a graph attention network, as well as transmit fault information across processes using state equations. We validate our proposed model on the BBA stud welding gun dataset and four additional open case studies. Compared with other established fault-prediction and time-series models, the MPRNet demonstrates improved F1-score by 13.1% in the BBA case and consistently achieves the best or near-best results in the open case studies.
{"title":"A multiscale process-aware retention network for fault prediction in mixed-model production","authors":"Mingda Chen , Ruiyun Yu , Zhiyuan Liang , Kun Li , Haifei Qi","doi":"10.1016/j.compind.2025.104313","DOIUrl":"10.1016/j.compind.2025.104313","url":null,"abstract":"<div><div>In the manufacturing industry, the demand for fault-prediction solutions is increasing to prevent unexpected downtimes and reduce maintenance costs. Although deep-learning methods have demonstrated excellent performance in this domain, the current methods typically overlook the analysis of variable and random processes within mixed-model production, which is a manufacturing strategy that offers flexibility and efficiency in satisfying diverse consumer demands. Hence, we propose the multiscale process-aware retention network (MPRNet), which segments a time series into multiscale patches, thus enabling the model to focus on local information within each production process and correlations across all production processes. Furthermore, the network incorporates a cross-channel interaction module designed to dynamically capture the interactions between various sensor data types using a graph attention network, as well as transmit fault information across processes using state equations. We validate our proposed model on the BBA stud welding gun dataset and four additional open case studies. Compared with other established fault-prediction and time-series models, the MPRNet demonstrates improved F1-score by 13.1% in the BBA case and consistently achieves the best or near-best results in the open case studies.</div></div>","PeriodicalId":55219,"journal":{"name":"Computers in Industry","volume":"170 ","pages":"Article 104313"},"PeriodicalIF":8.2,"publicationDate":"2025-05-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144139287","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 : 2025-05-24DOI: 10.1016/j.compind.2025.104319
Fan Yang, Xiaofeng Liu, Chunbing Zhang, Lin Bo
As rotating machinery often operates under complex and variable harsh conditions, domain generalization-based fault diagnosis has been adopted to tackle the challenge of distribution shifts and unseen data in target domains. However, most existing methods depend on fully labeled data from multiple source domains to learn domain-invariant representations. In practice, collecting comprehensive labeled data across diverse working conditions is often impractical, resulting in data insufficiency and distribution inconsistencies. To address the challenging scenario in which only a single fully labeled source domain is available, this article proposes a multi-style adversarial variational self-distillation (MSAVSD) framework based on domain randomization for single-domain generalized fault diagnosis. First, a domain-randomized generation module is developed to adaptively generate samples following randomized distributions by integrating adaptive noise and multi-scale style learning, thereby enriching the synthetic data with diverse and informative fault representations. Next, a scale-enhanced feature extraction module is introduced to extract rich domain-invariant features, thereby maximizing the utilization of fault-related information under limited training conditions. The method suppresses task-irrelevant noise and redundancy via variational self-distillation and employs contrastive learning to enhance the discriminability and consistency of task-relevant features. Extensive diagnostic experiments on three datasets, two self-collected and one publicly available, demonstrate that the proposed method outperforms state-of-the-art methods.
{"title":"Multi-style adversarial variational self-distillation in randomized domains for single-domain generalized fault diagnosis","authors":"Fan Yang, Xiaofeng Liu, Chunbing Zhang, Lin Bo","doi":"10.1016/j.compind.2025.104319","DOIUrl":"10.1016/j.compind.2025.104319","url":null,"abstract":"<div><div>As rotating machinery often operates under complex and variable harsh conditions, domain generalization-based fault diagnosis has been adopted to tackle the challenge of distribution shifts and unseen data in target domains. However, most existing methods depend on fully labeled data from multiple source domains to learn domain-invariant representations. In practice, collecting comprehensive labeled data across diverse working conditions is often impractical, resulting in data insufficiency and distribution inconsistencies. To address the challenging scenario in which only a single fully labeled source domain is available, this article proposes a multi-style adversarial variational self-distillation (MSAVSD) framework based on domain randomization for single-domain generalized fault diagnosis. First, a domain-randomized generation module is developed to adaptively generate samples following randomized distributions by integrating adaptive noise and multi-scale style learning, thereby enriching the synthetic data with diverse and informative fault representations. Next, a scale-enhanced feature extraction module is introduced to extract rich domain-invariant features, thereby maximizing the utilization of fault-related information under limited training conditions. The method suppresses task-irrelevant noise and redundancy via variational self-distillation and employs contrastive learning to enhance the discriminability and consistency of task-relevant features. Extensive diagnostic experiments on three datasets, two self-collected and one publicly available, demonstrate that the proposed method outperforms state-of-the-art methods.</div></div>","PeriodicalId":55219,"journal":{"name":"Computers in Industry","volume":"170 ","pages":"Article 104319"},"PeriodicalIF":8.2,"publicationDate":"2025-05-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144131011","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 : 2025-05-24DOI: 10.1016/j.compind.2025.104316
Anoop Kumar Sinha , Youngmi Christina Choi , David W. Rosen
User centric smart products prioritize the needs and preferences of users, enhancing their experience and satisfaction. Involving users in the design and assessment of smart products ensures that they meet real-world requirements, leading to more intuitive product design, user interface, and functionalities that truly resonate with users. Further, the capability of generating and evaluating many alternative designs early in product development is beneficial. However, the need to construct physical prototypes for user testing limits the number of designs that can be evaluated during early design stages. As such, our interest is in automated methods that support user centered design and usability and user experience assessment. In this review article, we look at at two decades of automation methods that have been employed in the design and development of user centric smart products. The focus of these automation methods is to incorporate user voice in early design stages rather than replacing the users. We have identified five key activities of the design cycle in which automated methods have been employed: design thinking, design ideation, prototype creation, user data collection for usability study, and user data analysis. Overall, 154 articles were identified across engineering, human-computer interaction, human factors, inclusive design, industrial design, and other disciplines that have incorporated automation methods to include the user’s voice in the design of user centric smart products. This review examines the effectiveness and limitations of different automation methods compared to conventional methods, offering valuable insights and suggestions to enhance the design processes of smart products with a focus on widespread usability issues. Our specific interest lies in developing assistive mobility and rehabilitation devices, where constraints such as limited development time and resources persist, yet the usability and user experience profoundly influence significant outcomes like perceived functionality, stigma, and device acceptance.
{"title":"Survey of automated methods for design and assessment of smart products","authors":"Anoop Kumar Sinha , Youngmi Christina Choi , David W. Rosen","doi":"10.1016/j.compind.2025.104316","DOIUrl":"10.1016/j.compind.2025.104316","url":null,"abstract":"<div><div>User centric smart products prioritize the needs and preferences of users, enhancing their experience and satisfaction. Involving users in the design and assessment of smart products ensures that they meet real-world requirements, leading to more intuitive product design, user interface, and functionalities that truly resonate with users. Further, the capability of generating and evaluating many alternative designs early in product development is beneficial. However, the need to construct physical prototypes for user testing limits the number of designs that can be evaluated during early design stages. As such, our interest is in <u>automated</u> methods that support user centered design and usability and user experience assessment. In this review article, we look at at two decades of automation methods that have been employed in the design and development of user centric smart products. The focus of these automation methods is to incorporate user voice in early design stages rather than replacing the users. We have identified five key activities of the design cycle in which automated methods have been employed: design thinking, design ideation, prototype creation, user data collection for usability study, and user data analysis. Overall, 154 articles were identified across engineering, human-computer interaction, human factors, inclusive design, industrial design, and other disciplines that have incorporated automation methods to include the user’s voice in the design of user centric smart products. This review examines the effectiveness and limitations of different automation methods compared to conventional methods, offering valuable insights and suggestions to enhance the design processes of smart products with a focus on widespread usability issues. Our specific interest lies in developing assistive mobility and rehabilitation devices, where constraints such as limited development time and resources persist, yet the usability and user experience profoundly influence significant outcomes like perceived functionality, stigma, and device acceptance.</div></div>","PeriodicalId":55219,"journal":{"name":"Computers in Industry","volume":"170 ","pages":"Article 104316"},"PeriodicalIF":8.2,"publicationDate":"2025-05-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144131013","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 : 2025-05-21DOI: 10.1016/j.compind.2025.104318
Wen Peng , Cheng-yan Ding , Yu Liu , Jia-nan Sun , Zhen Wei , Wen-bo Wang , Dian-hua Zhang , Jie Sun
In the hot rolling industry, uneven roll wear significantly influences schedule free rolling and product quality, necessitating more precise wear prediction to improve the capabilities of hot rolling production. However, existing methods, laden with limitations, struggle to predict uneven roll wear precisely and transparently. To address these challenges, we present a novel paradigm that combines a computer simulation technique, classical wear theory and a data-driven approach for predicting uneven work roll wear in the hot rolling industry. Initially, a finite element model is constructed to simulate hot rolling processing. Subsequently, an Archard-theory-based work roll wear model is derived to calculate the theoretical wear loss using the simulation results. Following this, based on the theoretical wear loss, a deep ensemble model containing three base predictors is established. Notably, Shapley additive explanations (SHAP) and ensemble mechanism analysis are implemented to explain the predictive process of the wear loss. The comparative experimental results demonstrate the deep ensemble method achieves a 2 % accuracy improvement over other machine learning models. Additionally, the wear prediction results for a real case of a roll change period prove that, at the peak position of wear profile, the proposed paradigm surpasses the existing model by 7.2 %. Significantly, the feature contributions and process interpretable analysis based on SHAP make the proposed paradigm both transparent and reliable.
{"title":"A novel paradigm for predicting and interpreting uneven roll wear in the hot rolling steel industry","authors":"Wen Peng , Cheng-yan Ding , Yu Liu , Jia-nan Sun , Zhen Wei , Wen-bo Wang , Dian-hua Zhang , Jie Sun","doi":"10.1016/j.compind.2025.104318","DOIUrl":"10.1016/j.compind.2025.104318","url":null,"abstract":"<div><div>In the hot rolling industry, uneven roll wear significantly influences schedule free rolling and product quality, necessitating more precise wear prediction to improve the capabilities of hot rolling production. However, existing methods, laden with limitations, struggle to predict uneven roll wear precisely and transparently. To address these challenges, we present a novel paradigm that combines a computer simulation technique, classical wear theory and a data-driven approach for predicting uneven work roll wear in the hot rolling industry. Initially, a finite element model is constructed to simulate hot rolling processing. Subsequently, an Archard-theory-based work roll wear model is derived to calculate the theoretical wear loss using the simulation results. Following this, based on the theoretical wear loss, a deep ensemble model containing three base predictors is established. Notably, Shapley additive explanations (SHAP) and ensemble mechanism analysis are implemented to explain the predictive process of the wear loss. The comparative experimental results demonstrate the deep ensemble method achieves a 2 % accuracy improvement over other machine learning models. Additionally, the wear prediction results for a real case of a roll change period prove that, at the peak position of wear profile, the proposed paradigm surpasses the existing model by 7.2 %. Significantly, the feature contributions and process interpretable analysis based on SHAP make the proposed paradigm both transparent and reliable.</div></div>","PeriodicalId":55219,"journal":{"name":"Computers in Industry","volume":"170 ","pages":"Article 104318"},"PeriodicalIF":8.2,"publicationDate":"2025-05-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144099280","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 : 2025-04-30DOI: 10.1016/j.compind.2025.104302
Caihua Hao , Zhaoyu Wang , Xinyong Mao , Songping He , Bin Li , Hongqi Liu , Fangyu Peng , Weiye Li
Accurately predicting the future wear of cutting tools with variable geometric parameters remains a significant challenge. Existing methods lack the capability to model long-term temporal dependencies and predict future wear values—a key characteristic of world models. To address this challenge, we introduce the Tool-Multimodal Generative Pre-trained Transformer (Tool-MMGPT), a novel and scalable multimodal large language model (MLLM) architecture specifically designed for tool wear prediction. Tool-MMGPT pioneers the first tool wear world model by uniquely unifying multimodal data, extending beyond conventional static dimensions to incorporate dynamic temporal dimensions. This approach extracts modality-specific information and achieves shared spatiotemporal feature fusion through a cross-modal Transformer. Subsequently, alignment and joint interpretation occur within a unified representation space via a multimodal-language projector, which effectively accommodates the comprehensive input characteristics required by world models. This article proposes an effective cross-modal fusion module for vibration signals and images, aiming to fully leverage the advantages of multimodal information. Crucially, Tool-MMGPT transcends the limitations of traditional Large Language Models (LLMs) through an innovative yet generalizable method. By fundamentally reconstructing the output layer and redefining training objectives, we repurpose LLMs for numerical regression tasks, thereby establishing a novel bridge that connects textual representations to continuous numerical predictions. This enables the direct and accurate long-term forecasting of future wear time series. Extensive experiments conducted on a newly developed multimodal dataset for variable geometry tools demonstrate that Tool-MMGPT significantly outperforms state-of-the-art (SOTA) baseline methods. These results highlight the model's superior long-context modeling capabilities and illustrate its potential for effective deployment in environments with limited computational resources.
{"title":"A novel and scalable multimodal large language model architecture Tool-MMGPT for future tool wear prediction in titanium alloy high-speed milling processes","authors":"Caihua Hao , Zhaoyu Wang , Xinyong Mao , Songping He , Bin Li , Hongqi Liu , Fangyu Peng , Weiye Li","doi":"10.1016/j.compind.2025.104302","DOIUrl":"10.1016/j.compind.2025.104302","url":null,"abstract":"<div><div>Accurately predicting the future wear of cutting tools with variable geometric parameters remains a significant challenge. Existing methods lack the capability to model long-term temporal dependencies and predict future wear values—a key characteristic of world models. To address this challenge, we introduce the Tool-Multimodal Generative Pre-trained Transformer (Tool-MMGPT), a novel and scalable multimodal large language model (MLLM) architecture specifically designed for tool wear prediction. Tool-MMGPT pioneers the first tool wear world model by uniquely unifying multimodal data, extending beyond conventional static dimensions to incorporate dynamic temporal dimensions. This approach extracts modality-specific information and achieves shared spatiotemporal feature fusion through a cross-modal Transformer. Subsequently, alignment and joint interpretation occur within a unified representation space via a multimodal-language projector, which effectively accommodates the comprehensive input characteristics required by world models. This article proposes an effective cross-modal fusion module for vibration signals and images, aiming to fully leverage the advantages of multimodal information. Crucially, Tool-MMGPT transcends the limitations of traditional Large Language Models (LLMs) through an innovative yet generalizable method. By fundamentally reconstructing the output layer and redefining training objectives, we repurpose LLMs for numerical regression tasks, thereby establishing a novel bridge that connects textual representations to continuous numerical predictions. This enables the direct and accurate long-term forecasting of future wear time series. Extensive experiments conducted on a newly developed multimodal dataset for variable geometry tools demonstrate that Tool-MMGPT significantly outperforms state-of-the-art (SOTA) baseline methods. These results highlight the model's superior long-context modeling capabilities and illustrate its potential for effective deployment in environments with limited computational resources.</div></div>","PeriodicalId":55219,"journal":{"name":"Computers in Industry","volume":"169 ","pages":"Article 104302"},"PeriodicalIF":8.2,"publicationDate":"2025-04-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143886351","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 : 2025-04-26DOI: 10.1016/j.compind.2025.104301
Xingjun Dong , Changsheng Zhang , Shuaitong Liu , Dawei Wang
Stamped parts play a crucial role in industrial manufacturing, and it is particularly important to automatically inspect their surface cracks. Since crack is rare and diverse, supervised defect detection methods lack sufficient data and cannot achieve ideal results. Unsupervised anomaly detection algorithms, which do not require crack data, can identify unknown cracks. However, they tend to have high rates of missed detections and false positives when dealing with complex backgrounds in stamped parts. To address these problems, this paper proposes a network called simple and reliable semi-supervised anomaly detection, considering the presence of a small number of anomalous data in actual production. This network uses a large number of normal samples and a small number of anomalous samples to detect surface cracks in stamped parts. Firstly, a pre-trained feature extractor is used for feature extraction, coupled with a designed feature adaptation network to reduce domain bias. Secondly, by extracting normal features from normal images, adding noise to these normal features to generate abnormal features, and extracting abnormal features from abnormal images at multiple scales, a feature space is constructed. Finally, by training a simplified discriminator based on the constructed feature space, computational efficiency is enhanced, and the deployment process is simplified. In the experiments, we collaborated with a multinational company, using an actual production dataset for verification. The proposed algorithm can achieve the score of area under the receiver operating characteristic curve of 98.2% for detection and 97.9% for localization at a processing speed of 19 frames per second.
{"title":"A simple and reliable semi-supervised anomaly detection network for detecting crack in stamped parts","authors":"Xingjun Dong , Changsheng Zhang , Shuaitong Liu , Dawei Wang","doi":"10.1016/j.compind.2025.104301","DOIUrl":"10.1016/j.compind.2025.104301","url":null,"abstract":"<div><div>Stamped parts play a crucial role in industrial manufacturing, and it is particularly important to automatically inspect their surface cracks. Since crack is rare and diverse, supervised defect detection methods lack sufficient data and cannot achieve ideal results. Unsupervised anomaly detection algorithms, which do not require crack data, can identify unknown cracks. However, they tend to have high rates of missed detections and false positives when dealing with complex backgrounds in stamped parts. To address these problems, this paper proposes a network called simple and reliable semi-supervised anomaly detection, considering the presence of a small number of anomalous data in actual production. This network uses a large number of normal samples and a small number of anomalous samples to detect surface cracks in stamped parts. Firstly, a pre-trained feature extractor is used for feature extraction, coupled with a designed feature adaptation network to reduce domain bias. Secondly, by extracting normal features from normal images, adding noise to these normal features to generate abnormal features, and extracting abnormal features from abnormal images at multiple scales, a feature space is constructed. Finally, by training a simplified discriminator based on the constructed feature space, computational efficiency is enhanced, and the deployment process is simplified. In the experiments, we collaborated with a multinational company, using an actual production dataset for verification. The proposed algorithm can achieve the score of area under the receiver operating characteristic curve of 98.2% for detection and 97.9% for localization at a processing speed of 19 frames per second.</div></div>","PeriodicalId":55219,"journal":{"name":"Computers in Industry","volume":"169 ","pages":"Article 104301"},"PeriodicalIF":8.2,"publicationDate":"2025-04-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143874165","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 : 2025-04-25DOI: 10.1016/j.compind.2025.104305
Jixiang Ding , Zhengding Zheng , Shayu Song , Long Bai , Jianfeng Xu , Jianguo Zhang , Wenjie Chen
Particle-reinforced metal matrix composites (PRMMCs) are widely used because of their exceptional material properties. Online control of the laser field to soften and modify the reinforcing particles on the machined surface of the composites is an effective way to improve the machinability and machining quality of PRMMCs. A real-time segmentation method for reinforcing particles in PRMMCs is proposed. First, real-time acquisition of reinforcing particle images along the processing path is achieved using machine vision, and cutting region images are determined. Next, to improve the model’s ability to effectively segment the reinforcing particles in low-resolution images of the machining region, a reinforcing particle segmentation network (RPSNet) is proposed, incorporating a multimodal fusion and space-to-depth convolution module. Subsequently, position signals along the cutting direction are obtained by using a sliding window method. The effectiveness of each module and the performance of the model are analyzed and verified through comparative and ablation experiments. The results demonstrated that the proposed RPSNet achieved a mean average precision (mAP) of 95.4 % in segmenting reinforcing particles, with an inference time of 5.8 ms. In comparison to other methods, it demonstrated better real-time performance and accuracy. Additionally, the proposed method can convert image information into position signals, thus enabling real-time control of the laser for softening and modifying the reinforcing particles.
{"title":"Toward laser-assisted cutting: A real-time segmentation method for reinforcing particles in particle-reinforced metal matrix composites","authors":"Jixiang Ding , Zhengding Zheng , Shayu Song , Long Bai , Jianfeng Xu , Jianguo Zhang , Wenjie Chen","doi":"10.1016/j.compind.2025.104305","DOIUrl":"10.1016/j.compind.2025.104305","url":null,"abstract":"<div><div>Particle-reinforced metal matrix composites (PRMMCs) are widely used because of their exceptional material properties. Online control of the laser field to soften and modify the reinforcing particles on the machined surface of the composites is an effective way to improve the machinability and machining quality of PRMMCs. A real-time segmentation method for reinforcing particles in PRMMCs is proposed. First, real-time acquisition of reinforcing particle images along the processing path is achieved using machine vision, and cutting region images are determined. Next, to improve the model’s ability to effectively segment the reinforcing particles in low-resolution images of the machining region, a reinforcing particle segmentation network (RPSNet) is proposed, incorporating a multimodal fusion and space-to-depth convolution module. Subsequently, position signals along the cutting direction are obtained by using a sliding window method. The effectiveness of each module and the performance of the model are analyzed and verified through comparative and ablation experiments. The results demonstrated that the proposed RPSNet achieved a mean average precision (mAP) of 95.4 % in segmenting reinforcing particles, with an inference time of 5.8 ms. In comparison to other methods, it demonstrated better real-time performance and accuracy. Additionally, the proposed method can convert image information into position signals, thus enabling real-time control of the laser for softening and modifying the reinforcing particles.</div></div>","PeriodicalId":55219,"journal":{"name":"Computers in Industry","volume":"169 ","pages":"Article 104305"},"PeriodicalIF":8.2,"publicationDate":"2025-04-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143869739","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 : 2025-04-25DOI: 10.1016/j.compind.2025.104290
Xiangkai Shen , Jinhai Liu , Yifu Ren , Lin Jiang , Lei Wang , He Zhao , Rui Li
Defect diagnosis based on magnetic flux leakage (MFL) signals is an important process for assessing pipeline health, including defect detection and size quantification. However, existing studies suffer from poor consistency of results, because they regard defect detection and size quantification as separate tasks, lacking paradigm harmonization and interaction. In addition, the calibration of experts is required to achieve harmonization between the two, which increases the time cost of data analysis. To address the above challenges, our motivation is to synergistically learn two tasks within a unified framework and utilize their task properties for mutual benefit. Therefore, a novel defect diagnosis method based on a task-oriented physical collaborative network (TOPC-Net) is proposed, which is the first attempt at joint defect detection and size quantification in MFL inspection. First, a feature extraction subnetwork with a heterogeneous focus module is proposed to decompose initial task-specific features from shared spaces. Second, considering the strong correlation between the two tasks, a cross-task information awareness method is proposed to realize the information interaction between the two tasks, so that the task-specific features can be enhanced. Finally, a physical information-guided collaborative decision subnetwork is proposed to jointly optimize two tasks, where MFL domain knowledge is embedded into the subnetwork to provide expert guidance, ensuring the accuracy and stability of predictions. Experimental results show that the proposed method outperforms existing methods, with a detection accuracy of 96.0% and an average improvement of 7.5% in quantification accuracy, which makes it promising for industrial applications.
{"title":"A task-oriented physical collaborative network for pipeline defect diagnosis in a magnetic flux leakage detection system","authors":"Xiangkai Shen , Jinhai Liu , Yifu Ren , Lin Jiang , Lei Wang , He Zhao , Rui Li","doi":"10.1016/j.compind.2025.104290","DOIUrl":"10.1016/j.compind.2025.104290","url":null,"abstract":"<div><div>Defect diagnosis based on magnetic flux leakage (MFL) signals is an important process for assessing pipeline health, including defect detection and size quantification. However, existing studies suffer from poor consistency of results, because they regard defect detection and size quantification as separate tasks, lacking paradigm harmonization and interaction. In addition, the calibration of experts is required to achieve harmonization between the two, which increases the time cost of data analysis. To address the above challenges, our motivation is to synergistically learn two tasks within a unified framework and utilize their task properties for mutual benefit. Therefore, a novel defect diagnosis method based on a task-oriented physical collaborative network (TOPC-Net) is proposed, which is the first attempt at joint defect detection and size quantification in MFL inspection. First, a feature extraction subnetwork with a heterogeneous focus module is proposed to decompose initial task-specific features from shared spaces. Second, considering the strong correlation between the two tasks, a cross-task information awareness method is proposed to realize the information interaction between the two tasks, so that the task-specific features can be enhanced. Finally, a physical information-guided collaborative decision subnetwork is proposed to jointly optimize two tasks, where MFL domain knowledge is embedded into the subnetwork to provide expert guidance, ensuring the accuracy and stability of predictions. Experimental results show that the proposed method outperforms existing methods, with a detection accuracy of 96.0% and an average improvement of 7.5% in quantification accuracy, which makes it promising for industrial applications.</div></div>","PeriodicalId":55219,"journal":{"name":"Computers in Industry","volume":"169 ","pages":"Article 104290"},"PeriodicalIF":8.2,"publicationDate":"2025-04-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143869732","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 : 2025-04-24DOI: 10.1016/j.compind.2025.104304
Chunhao Jiang , Nian-Zhong Chen
Physics-informed neural networks (PINNs) face significant challenges to predict the vortex shedding in the flow past a two-dimensional cylinder, mainly due to complex loss landscapes, spectral bias, and a lack of inductive bias towards periodic functions. To overcome these challenges, a novel gradient-free PINN (GF-PINN) coupled with a U-Net+ + architecture is proposed. For optimizing the complex loss landscape, the skip pathways in U-Net+ + are redesigned to reduce the semantic gap between encoder and decoder feature maps. Then, the stream function instead of velocity, is used as the input and output for the neural network, ensuring flow incompressibility and reducing output dimensionality. This approach aims to overcome the inherent problems of spectral bias and the lack of inductive bias towards periodic functions in PINNs. Furthermore, gradient-free convolutional filters are employed to approximate the derivative terms in the loss function to further optimize the complex loss landscape. A series of numerical experiments and dynamic mode analyses are conducted and the results show that the vortex shedding in the wake of a square cylinder is successfully captured by the proposed model and the estimated drag coefficients and Strouhal numbers are in a good agreement with those predicted by traditional methods. In addition, numerical experiments also show that the model exhibits great capabilities of generalization and extrapolation. This work demonstrates the potential of PINN-based models to effectively solve complex fluid dynamics problems.
基于物理信息的神经网络（pinn）在预测流过二维圆柱体的流体中的涡落方面面临着重大挑战，这主要是由于复杂的损失、光谱偏倚和缺乏对周期函数的归纳偏倚。为了克服这些挑战，提出了一种新型的无梯度pin - n （GF-PINN）结合U-Net+ +架构。为了优化复杂的损失情况，重新设计了U-Net+ +中的跳过路径，以减少编码器和解码器特征映射之间的语义差距。然后，用流函数代替速度作为神经网络的输入和输出，保证了流不可压缩性，降低了输出维数。该方法旨在克服pinn中固有的频谱偏置和对周期函数缺乏归纳偏置的问题。此外，采用无梯度卷积滤波器对损失函数中的导数项进行近似，进一步优化复杂损失格局。通过一系列的数值实验和动力模态分析，结果表明，该模型成功地捕获了方形圆柱体尾迹的涡脱落，所估计的阻力系数和Strouhal数与传统方法预测的结果吻合较好。此外，数值实验还表明，该模型具有良好的泛化和外推能力。这项工作证明了基于pup模型有效解决复杂流体动力学问题的潜力。
{"title":"Gradient-free physics-informed neural networks (GF-PINNs) for vortex shedding prediction in flow past square cylinders","authors":"Chunhao Jiang , Nian-Zhong Chen","doi":"10.1016/j.compind.2025.104304","DOIUrl":"10.1016/j.compind.2025.104304","url":null,"abstract":"<div><div>Physics-informed neural networks (PINNs) face significant challenges to predict the vortex shedding in the flow past a two-dimensional cylinder, mainly due to complex loss landscapes, spectral bias, and a lack of inductive bias towards periodic functions. To overcome these challenges, a novel gradient-free PINN (GF-PINN) coupled with a U-Net+ + architecture is proposed. For optimizing the complex loss landscape, the skip pathways in U-Net+ + are redesigned to reduce the semantic gap between encoder and decoder feature maps. Then, the stream function instead of velocity, is used as the input and output for the neural network, ensuring flow incompressibility and reducing output dimensionality. This approach aims to overcome the inherent problems of spectral bias and the lack of inductive bias towards periodic functions in PINNs. Furthermore, gradient-free convolutional filters are employed to approximate the derivative terms in the loss function to further optimize the complex loss landscape. A series of numerical experiments and dynamic mode analyses are conducted and the results show that the vortex shedding in the wake of a square cylinder is successfully captured by the proposed model and the estimated drag coefficients and Strouhal numbers are in a good agreement with those predicted by traditional methods. In addition, numerical experiments also show that the model exhibits great capabilities of generalization and extrapolation. This work demonstrates the potential of PINN-based models to effectively solve complex fluid dynamics problems.</div></div>","PeriodicalId":55219,"journal":{"name":"Computers in Industry","volume":"169 ","pages":"Article 104304"},"PeriodicalIF":8.2,"publicationDate":"2025-04-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143869738","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 : 2025-04-24DOI: 10.1016/j.compind.2025.104303
Jieyang Peng , Andreas Kimmig , Simon Kreuzwieser , Zhibin Niu , Xiaoming Tao , Jivka Ovtcharova
In the fast-paced manufacturing industry, rapid and efficient product design is essential for meeting customer demands and maintaining a competitive edge. Despite advancements, transforming 2D design concepts into accurate 3D models remains a complex challenge, primarily due to the non-differentiability of traditional rendering processes that hinder gradient-based optimizations. To address this limitation, this paper introduces an innovative dual-decoder architecture that effectively separates the shape and color components of 3D models. By assigning separate decoders for vertex positions and color assignment, our proposed model enables targeted optimization of each, leading to more refined and authentic 3D reconstructions. Moreover, we have overcome the non-differentiability issue, enabling gradient-based learning through the incorporation of differentiable rendering techniques. These techniques facilitate gradient-based optimization, paving the way for data-driven enhancements in the design process. Our empirical research has demonstrated the effectiveness of our approach in generating high-fidelity 3D models from 2D inputs. Additionally, we have shed light on the sensitivity of hyperparameters within our framework, offering valuable insights for future model refinement and optimization. In summary, our research provides valuable insights into enhancing 3D modeling frameworks, thereby contributing to incremental progress in the field of computer-aided design and manufacturing.
{"title":"3D modeling from a single image via a novel dual-decoder framework for Agile design","authors":"Jieyang Peng , Andreas Kimmig , Simon Kreuzwieser , Zhibin Niu , Xiaoming Tao , Jivka Ovtcharova","doi":"10.1016/j.compind.2025.104303","DOIUrl":"10.1016/j.compind.2025.104303","url":null,"abstract":"<div><div>In the fast-paced manufacturing industry, rapid and efficient product design is essential for meeting customer demands and maintaining a competitive edge. Despite advancements, transforming 2D design concepts into accurate 3D models remains a complex challenge, primarily due to the non-differentiability of traditional rendering processes that hinder gradient-based optimizations. To address this limitation, this paper introduces an innovative dual-decoder architecture that effectively separates the shape and color components of 3D models. By assigning separate decoders for vertex positions and color assignment, our proposed model enables targeted optimization of each, leading to more refined and authentic 3D reconstructions. Moreover, we have overcome the non-differentiability issue, enabling gradient-based learning through the incorporation of differentiable rendering techniques. These techniques facilitate gradient-based optimization, paving the way for data-driven enhancements in the design process. Our empirical research has demonstrated the effectiveness of our approach in generating high-fidelity 3D models from 2D inputs. Additionally, we have shed light on the sensitivity of hyperparameters within our framework, offering valuable insights for future model refinement and optimization. In summary, our research provides valuable insights into enhancing 3D modeling frameworks, thereby contributing to incremental progress in the field of computer-aided design and manufacturing.</div></div>","PeriodicalId":55219,"journal":{"name":"Computers in Industry","volume":"169 ","pages":"Article 104303"},"PeriodicalIF":8.2,"publicationDate":"2025-04-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143869740","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