Pub Date : 2026-01-12DOI: 10.1016/j.asoc.2026.114618
Changsen Yuan , Rui Lin , Cunhan Guo
Large language models (LLMs) demonstrate strong reasoning abilities in complex tasks, but they are limited by their inability to access up-to-date information and are prone to generating plausible, known as hallucinations. Knowledge graphs (KGs), which are structured and dynamically updated collections of facts, offer a solution by providing LLMs with verified, current information. This enhances the models’ reasoning accuracy and reduces the risk of hallucinations. However, existing methods focus more on the inferability and structure of KGs, while overlooking the varying levels of difficulty that LLMs encounter when learning and understanding different reasoning paths between two nodes in the KGs. In this paper, we propose a novel method called Reliable Reasoning (R2) that selects appropriate inference paths from KGs for LLMs to learn and understand more easily. Specifically, we present a reliable reasoning path search framework in which R2 first extracts appropriate candidate reasoning paths based on KGs. The candidate reasoning paths are then filtered, selecting the ones preferred by the LLM for further learning. Comprehensive experiments conducted on two benchmark KGQA datasets indicate that R2 attains good performance in KGQA tasks, producing accurate and interpretable reasoning outcomes.
{"title":"Reliable reasoning: Learning and inference based on the ability of large language models","authors":"Changsen Yuan , Rui Lin , Cunhan Guo","doi":"10.1016/j.asoc.2026.114618","DOIUrl":"10.1016/j.asoc.2026.114618","url":null,"abstract":"<div><div>Large language models (LLMs) demonstrate strong reasoning abilities in complex tasks, but they are limited by their inability to access up-to-date information and are prone to generating plausible, known as hallucinations. Knowledge graphs (KGs), which are structured and dynamically updated collections of facts, offer a solution by providing LLMs with verified, current information. This enhances the models’ reasoning accuracy and reduces the risk of hallucinations. However, existing methods focus more on the inferability and structure of KGs, while overlooking the varying levels of difficulty that LLMs encounter when learning and understanding different reasoning paths between two nodes in the KGs. In this paper, we propose a novel method called <strong>R</strong>eliable <strong>R</strong>easoning (R<sup>2</sup>) that selects appropriate inference paths from KGs for LLMs to learn and understand more easily. Specifically, we present a reliable reasoning path search framework in which R<sup>2</sup> first extracts appropriate candidate reasoning paths based on KGs. The candidate reasoning paths are then filtered, selecting the ones preferred by the LLM for further learning. Comprehensive experiments conducted on two benchmark KGQA datasets indicate that R<sup>2</sup> attains good performance in KGQA tasks, producing accurate and interpretable reasoning outcomes.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"190 ","pages":"Article 114618"},"PeriodicalIF":6.6,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146024167","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 : 2026-01-12DOI: 10.1016/j.asoc.2026.114611
Bum Jun Kim , Sang Woo Kim
Regularization is essential for improving the generalization of deep neural networks while mitigating overfitting. Although the popular method of Dropout provides a regularization effect, it causes inconsistent properties in the output, which may degrade the performance of deep neural networks. In this study, we propose a new module called stochastic average pooling, which incorporates Dropout-like stochasticity into pooling. We describe the properties of stochastic subsampling and average pooling and leverage them to design a module without any inconsistency problems. The stochastic average pooling achieves a regularization effect without any potential performance degradation due to the inconsistency issue and can be easily plugged into existing deep neural network architectures. Experiments demonstrate that replacing existing average pooling with stochastic average pooling yields consistent improvements across a variety of tasks, datasets, and models.
{"title":"Stochastic subsampling with average pooling","authors":"Bum Jun Kim , Sang Woo Kim","doi":"10.1016/j.asoc.2026.114611","DOIUrl":"10.1016/j.asoc.2026.114611","url":null,"abstract":"<div><div>Regularization is essential for improving the generalization of deep neural networks while mitigating overfitting. Although the popular method of Dropout provides a regularization effect, it causes inconsistent properties in the output, which may degrade the performance of deep neural networks. In this study, we propose a new module called stochastic average pooling, which incorporates Dropout-like stochasticity into pooling. We describe the properties of stochastic subsampling and average pooling and leverage them to design a module without any inconsistency problems. The stochastic average pooling achieves a regularization effect without any potential performance degradation due to the inconsistency issue and can be easily plugged into existing deep neural network architectures. Experiments demonstrate that replacing existing average pooling with stochastic average pooling yields consistent improvements across a variety of tasks, datasets, and models.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"190 ","pages":"Article 114611"},"PeriodicalIF":6.6,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145980370","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 : 2026-01-12DOI: 10.1016/j.asoc.2026.114615
Jujie Wang, Xiawei Wu, Minghong Chen
Wind speed forecasting is essential for integrating renewable energy systems and advancing carbon neutrality objectives. As a key technology for wind power grid integration, accurate wind speed prediction enhances the grid’s capacity to absorb renewable energy, thereby reducing reliance on fossil fuels and mitigating environmental impacts from greenhouse gas emissions. However, wind speed’s dynamic evolutionary characteristics and the inherent data drift problem pose significant challenges to precise forecasting. This research develops an integrated wind speed prediction model incorporating adaptive real-time feature decoupling and a multi-head attention ensemble algorithm to address data drift. The enhanced successive variational mode decomposition, combined with a sliding window, is first used for adaptive real-time feature decoupling of original wind speed data, which dynamically tracks drift-induced variations by decoupling data into adaptive subsequences. A multi-dimensional quantitative optimal model matching strategy is adopted to achieve precise matching between each subsequence and the model. Multi-head attention adjusts integration weights to mitigate the impact of drift. Quantile regression with multi-level information fusion is utilized to additionally assess the uncertainty in wind speed variations and derive the ultimate forecasting outcomes. Experimental findings indicate that in one-step forecasting, the prosed model attained PICP values of 0.9645 and 0.9602, along with PINAW values of 0.2155 and 0.2296, at the two wind farms. While ensuring high coverage, it effectively controlled interval width, fully validating the system’s superior performance.
{"title":"Dynamic ensemble point-interval wind speed prediction system for data drift: Adaptive real-time feature decoupling and multi-level information fusion quantile regression","authors":"Jujie Wang, Xiawei Wu, Minghong Chen","doi":"10.1016/j.asoc.2026.114615","DOIUrl":"10.1016/j.asoc.2026.114615","url":null,"abstract":"<div><div>Wind speed forecasting is essential for integrating renewable energy systems and advancing carbon neutrality objectives. As a key technology for wind power grid integration, accurate wind speed prediction enhances the grid’s capacity to absorb renewable energy, thereby reducing reliance on fossil fuels and mitigating environmental impacts from greenhouse gas emissions. However, wind speed’s dynamic evolutionary characteristics and the inherent data drift problem pose significant challenges to precise forecasting. This research develops an integrated wind speed prediction model incorporating adaptive real-time feature decoupling and a multi-head attention ensemble algorithm to address data drift. The enhanced successive variational mode decomposition, combined with a sliding window, is first used for adaptive real-time feature decoupling of original wind speed data, which dynamically tracks drift-induced variations by decoupling data into adaptive subsequences. A multi-dimensional quantitative optimal model matching strategy is adopted to achieve precise matching between each subsequence and the model. Multi-head attention adjusts integration weights to mitigate the impact of drift. Quantile regression with multi-level information fusion is utilized to additionally assess the uncertainty in wind speed variations and derive the ultimate forecasting outcomes. Experimental findings indicate that in one-step forecasting, the prosed model attained PICP values of 0.9645 and 0.9602, along with PINAW values of 0.2155 and 0.2296, at the two wind farms. While ensuring high coverage, it effectively controlled interval width, fully validating the system’s superior performance.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"190 ","pages":"Article 114615"},"PeriodicalIF":6.6,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145980372","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 : 2026-01-12DOI: 10.1016/j.asoc.2026.114633
Lin-Chieh Huang, Hung-Hsu Tsai, Yu-Che Chuang
This paper proposes a new pixel-reconstruction-based method that combines high-frequency sub-bands of Wavelets, Gaussian Fourier features (GFF) and Variational Embedding (VE) for one-class anomaly detection on medical images, which is called WGF-VAE hereafter. Traditional reconstruction-based methods reconstruct low-frequency information, leading to miss image details during image reconstruction, especially for medical image reconstruction. As a result, those methods often cause false-positives, regarding normal parts as anomalies. The WGF-VAE scheme can overcome these drawbacks mentioned above due to the use of high-frequency sub-bands of wavelets, GFF and VE in the design of image reconstruction process. High-frequency sub-bands conserve high-frequency information, making the decoder of the WGF-VAE scheme easier to learn and handle these details of images. Moreover, the decoder leverages GFF to cover a broader frequency spectrum by transforming coordinates of an input image into a higher-dimension space so as to enhance the learning of high-frequency functions. Meanwhile, the scheme can accurately capture and reconstruct high-frequency details of medical images by utilizing the localized frequency information from high-frequency sub-bands and the expanded frequency spectrum from GFF. Furthermore, a variational autoencoder (VAE) produces VE which is employed in the decoding phase to play a role as the latent feature of high-frequency sub-bands. It makes the decoder stable to yield normal images so as to precisely compute the difference between input and output images, resulting in promoting the recognition ability of the scheme. Hence, the WGF-VAE scheme possesses remarkably ability on detection and localization for anomalies because of taking a combination of three features as inputs of the decoder. Finally, massively experimental results show that the WGF-VAE scheme outstandingly surpasses state-of-the-art methods on anomaly detection for brain and liver images in two public benchmarks.
{"title":"One-class anomaly detection based on image reconstruction by Wavelet, Gaussian Fourier and variational Embedding for medical images","authors":"Lin-Chieh Huang, Hung-Hsu Tsai, Yu-Che Chuang","doi":"10.1016/j.asoc.2026.114633","DOIUrl":"10.1016/j.asoc.2026.114633","url":null,"abstract":"<div><div>This paper proposes a new pixel-reconstruction-based method that combines high-frequency sub-bands of <strong>W</strong>avelets, <strong>G</strong>aussian <strong>F</strong>ourier features (GFF) and <strong>V</strong>ariational <strong>E</strong>mbedding (VE) for one-class anomaly detection on medical images, which is called WGF-VAE hereafter. Traditional reconstruction-based methods reconstruct low-frequency information, leading to miss image details during image reconstruction, especially for medical image reconstruction. As a result, those methods often cause false-positives, regarding normal parts as anomalies. The WGF-VAE scheme can overcome these drawbacks mentioned above due to the use of high-frequency sub-bands of wavelets, GFF and VE in the design of image reconstruction process. High-frequency sub-bands conserve high-frequency information, making the decoder of the WGF-VAE scheme easier to learn and handle these details of images. Moreover, the decoder leverages GFF to cover a broader frequency spectrum by transforming coordinates of an input image into a higher-dimension space so as to enhance the learning of high-frequency functions. Meanwhile, the scheme can accurately capture and reconstruct high-frequency details of medical images by utilizing the localized frequency information from high-frequency sub-bands and the expanded frequency spectrum from GFF. Furthermore, a variational autoencoder (VAE) produces VE which is employed in the decoding phase to play a role as the latent feature of high-frequency sub-bands. It makes the decoder stable to yield normal images so as to precisely compute the difference between input and output images, resulting in promoting the recognition ability of the scheme. Hence, the WGF-VAE scheme possesses remarkably ability on detection and localization for anomalies because of taking a combination of three features as inputs of the decoder. Finally, massively experimental results show that the WGF-VAE scheme outstandingly surpasses state-of-the-art methods on anomaly detection for brain and liver images in two public benchmarks.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114633"},"PeriodicalIF":6.6,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039914","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 : 2026-01-12DOI: 10.1016/j.asoc.2026.114634
Layth Rafea Hazim , Oguz Ata
The swift growth of the gargantuan language models has made it even harder to tell apart writings done by humans and AI especially in academia which led to the need for the setup of trustworthy detection frameworks that will satisfactorily balance accuracy, interpretability, and efficiency. In this paper, we walk you through the HQML-NLP, a detection system using a hybrid quantum-classical machine learning framework for the detection of AI-generated academic content. The system onboard the merging of Sentence-BERT semantic embeddings with quantum feature encoding that is supported by a 6-qubit, two-layer parameterized quantum circuit, thus resulting in a 390-dimensional hybrid representation which is classified through a lightweight multilayer perceptron. The framework has been tested on three benchmark datasets AI-GA, HWAI, and HAGT-1M to check its scalability and generalization to different academic writing situations. The results of the experiments show the framework has consistent and good discriminative capability achieving AUROC scores of more than 0.96 on all datasets and excellent performance (AUROC = 1.000, ACC = 99.98 %) on the large-scale data. Moreover, probability calibration by means of temperature scaling raises the trustworthiness of predicted confidence scores, leading to a 60 % reduction in the Expected Calibration Error (ECE) and without affecting the performance of the discrimination. When compared against the transformer-based and ensemble-learning detectors, HQML-NLP comes out with an equivalent and competitive detection accuracy and calibration quality yet demands more than 2000× less trainable parameters. These findings imply that hybrid quantum-classical representations act as an effective and compact alternative for the detection of AI-text in scholarly journals.
{"title":"HQML-NLP: A hybrid quantum machine learning framework for scholarly AI-text detection","authors":"Layth Rafea Hazim , Oguz Ata","doi":"10.1016/j.asoc.2026.114634","DOIUrl":"10.1016/j.asoc.2026.114634","url":null,"abstract":"<div><div>The swift growth of the gargantuan language models has made it even harder to tell apart writings done by humans and AI especially in academia which led to the need for the setup of trustworthy detection frameworks that will satisfactorily balance accuracy, interpretability, and efficiency. In this paper, we walk you through the HQML-NLP, a detection system using a hybrid quantum-classical machine learning framework for the detection of AI-generated academic content. The system onboard the merging of Sentence-BERT semantic embeddings with quantum feature encoding that is supported by a 6-qubit, two-layer parameterized quantum circuit, thus resulting in a 390-dimensional hybrid representation which is classified through a lightweight multilayer perceptron. The framework has been tested on three benchmark datasets AI-GA, HWAI, and HAGT-1M to check its scalability and generalization to different academic writing situations. The results of the experiments show the framework has consistent and good discriminative capability achieving AUROC scores of more than 0.96 on all datasets and excellent performance (AUROC = 1.000, ACC = 99.98 %) on the large-scale data. Moreover, probability calibration by means of temperature scaling raises the trustworthiness of predicted confidence scores, leading to a 60 % reduction in the Expected Calibration Error (ECE) and without affecting the performance of the discrimination. When compared against the transformer-based and ensemble-learning detectors, HQML-NLP comes out with an equivalent and competitive detection accuracy and calibration quality yet demands more than 2000× less trainable parameters. These findings imply that hybrid quantum-classical representations act as an effective and compact alternative for the detection of AI-text in scholarly journals.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114634"},"PeriodicalIF":6.6,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080200","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 : 2026-01-12DOI: 10.1016/j.asoc.2026.114607
Peng Zhao , Zhen-Yu Li
Open set recognition aims to classify known classes and reject unknown classes simultaneously. This study is often performed based on machine learning and deep learning. In machine learning based schemes, the classification boundary of known classes must be contracted and refined accurately, which is sometimes hard to solve in practice. In deep learning based schemes, a large dataset is required to train deep neural networks. These networks can only process images whose size should be sufficiently large. The spectral curve dataset cannot be processed by these networks usually. In this article, a novel open set recognition scheme is proposed based on a revised fuzzy rule classifier with application to spectral curve classification. A small spectral dataset is required to train this fuzzy rule classifier. After fuzzy rule training, the used fuzzy rules can be used to classify known classes, whereas the unused rules to reject unknown classes. Therefore, the contraction and refinement of classification boundary for known classes are not required. When a sample is sent into this revised fuzzy classifier, we get some fuzzy rules and their corresponding nonzero scores. The probabilistic distribution is then evaluated by the Entropy or Gini index. If this probabilistic distribution is certain, one class corresponding to the maximal score is final output class. Otherwise, this sample is rejected as unknown class. The comparison experimental results on mango, melamine and wood datasets demonstrate that our proposed scheme achieves approximately mean 42.30 % improvement in terms of overall recognition accuracy compared to baseline models.
{"title":"Open set recognition based on fuzzy rule classifier and probabilistic distribution analysis with application to spectral material classification","authors":"Peng Zhao , Zhen-Yu Li","doi":"10.1016/j.asoc.2026.114607","DOIUrl":"10.1016/j.asoc.2026.114607","url":null,"abstract":"<div><div>Open set recognition aims to classify known classes and reject unknown classes simultaneously. This study is often performed based on machine learning and deep learning. In machine learning based schemes, the classification boundary of known classes must be contracted and refined accurately, which is sometimes hard to solve in practice. In deep learning based schemes, a large dataset is required to train deep neural networks. These networks can only process images whose size should be sufficiently large. The spectral curve dataset cannot be processed by these networks usually. In this article, a novel open set recognition scheme is proposed based on a revised fuzzy rule classifier with application to spectral curve classification. A small spectral dataset is required to train this fuzzy rule classifier. After fuzzy rule training, the used fuzzy rules can be used to classify known classes, whereas the unused rules to reject unknown classes. Therefore, the contraction and refinement of classification boundary for known classes are not required. When a sample is sent into this revised fuzzy classifier, we get some fuzzy rules and their corresponding nonzero scores. The probabilistic distribution is then evaluated by the Entropy or Gini index. If this probabilistic distribution is certain, one class corresponding to the maximal score is final output class. Otherwise, this sample is rejected as unknown class. The comparison experimental results on mango, melamine and wood datasets demonstrate that our proposed scheme achieves approximately mean 42.30 % improvement in terms of overall recognition accuracy compared to baseline models.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"190 ","pages":"Article 114607"},"PeriodicalIF":6.6,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145980368","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 : 2026-01-12DOI: 10.1016/j.asoc.2026.114626
Yuan Zhou, Xiaofeng Yue
Integrating rich information from diverse source domains significantly enhances cross-domain knowledge transfer capabilities in mechanical fault diagnosis, which is critical for addressing fault diagnosis demands under complex and varying operating conditions. However, existing methods typically perform domain alignment at the global feature level while neglecting the local domain shift accumulation effects in time-frequency features, resulting in inadequate suppression of multi-level domain discrepancies. To address the aforementioned issues, a local-to-global multi-modal domain adversarial learning framework for multi-source domain mechanical fault diagnosis is proposed in this paper. First, a multi-scale time-frequency feature learning network is designed to achieve effective learning from multi-modal local features to unified global representations through parallel heterogeneous feature encoding and adaptive feature aggregation. To fundamentally eliminate multi-level domain bias, a hierarchical local-to-global domain adversarial learning strategy is further proposed. Through constructing a multi-level progressive domain discrimination system to achieve cross-domain collaborative adversarial training. On this basis, a globally-guided local neighborhood consistency learning mechanism is constructed, which generates high-quality pseudo-labels through joint adaptive cross-domain semantic association modeling and multi-level entropy-weighted confidence evaluation, effectively achieving cross-domain knowledge transfer. Extensive experiments on three datasets demonstrate that the proposed method achieves an average diagnostic accuracy of 93.46 %, outperforming the best baseline by 5.34 % across all 10 cross-domain transfer scenarios.
{"title":"Local-to-global multi-modal domain adversarial learning framework for multi-source domain mechanical fault diagnosis","authors":"Yuan Zhou, Xiaofeng Yue","doi":"10.1016/j.asoc.2026.114626","DOIUrl":"10.1016/j.asoc.2026.114626","url":null,"abstract":"<div><div>Integrating rich information from diverse source domains significantly enhances cross-domain knowledge transfer capabilities in mechanical fault diagnosis, which is critical for addressing fault diagnosis demands under complex and varying operating conditions. However, existing methods typically perform domain alignment at the global feature level while neglecting the local domain shift accumulation effects in time-frequency features, resulting in inadequate suppression of multi-level domain discrepancies. To address the aforementioned issues, a local-to-global multi-modal domain adversarial learning framework for multi-source domain mechanical fault diagnosis is proposed in this paper. First, a multi-scale time-frequency feature learning network is designed to achieve effective learning from multi-modal local features to unified global representations through parallel heterogeneous feature encoding and adaptive feature aggregation. To fundamentally eliminate multi-level domain bias, a hierarchical local-to-global domain adversarial learning strategy is further proposed. Through constructing a multi-level progressive domain discrimination system to achieve cross-domain collaborative adversarial training. On this basis, a globally-guided local neighborhood consistency learning mechanism is constructed, which generates high-quality pseudo-labels through joint adaptive cross-domain semantic association modeling and multi-level entropy-weighted confidence evaluation, effectively achieving cross-domain knowledge transfer. Extensive experiments on three datasets demonstrate that the proposed method achieves an average diagnostic accuracy of 93.46 %, outperforming the best baseline by 5.34 % across all 10 cross-domain transfer scenarios.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"190 ","pages":"Article 114626"},"PeriodicalIF":6.6,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145980293","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 : 2026-01-12DOI: 10.1016/j.asoc.2026.114600
Yanxin Wang , Jing Yan , Zhengrun Zhang , Jianhua Wang , Zhiyuan Liu , Yingsan Geng , Dipti Srinivasan
Deep learning based multi-source fusion has shown significant potential in diagnosing insulation defects in gas-insulated switchgear (GIS). However, its applicability in real engineering scenarios remains limited. Existing fusion frameworks struggle to model the heterogeneous sensing characteristics of optical and electrical channels, often relying on rigid or shallow interaction schemes that fail to capture modality complementarity. In addition, field data are typically scarce and distribution-shifted, making it difficult for conventional models to learn discriminative and generalizable features under small-sample conditions. To address these challenges, we propose a novel multi-sensor fusion few-shot learning network (MSFFLN) for GIS insulation defect diagnosis. First, a deep fusion network is developed to construct comprehensive representations of insulation defects. Specifically, a feature weighting fusion module is employed to improve robustness, while an adaptive attention-based fusion block suppresses redundant and aliased information, emphasizing the most discriminative features. Second, a contrastive learning-based few-shot strategy is introduced. By computing global and local contrastive losses and using contrastive learning as an auxiliary task, the model learns more accurate and generalizable feature representations. In addition, salient region mixing across samples is applied to decouple class-level and instance-level feature correlations. Finally, field experiments validate the effectiveness of the MSFFLN. Results show that the MSFFLN achieves a diagnostic accuracy of 95.06% with only 10 support samples, significantly outperforming baseline and ablation models in small-sample GIS insulation defect diagnosis.
{"title":"A novel multi-sensor fusion method for diagnosing insulation defects in gas-insulated substations guided by adaptive-attention and contrastive-based few-shot learning","authors":"Yanxin Wang , Jing Yan , Zhengrun Zhang , Jianhua Wang , Zhiyuan Liu , Yingsan Geng , Dipti Srinivasan","doi":"10.1016/j.asoc.2026.114600","DOIUrl":"10.1016/j.asoc.2026.114600","url":null,"abstract":"<div><div>Deep learning based multi-source fusion has shown significant potential in diagnosing insulation defects in gas-insulated switchgear (GIS). However, its applicability in real engineering scenarios remains limited. Existing fusion frameworks struggle to model the heterogeneous sensing characteristics of optical and electrical channels, often relying on rigid or shallow interaction schemes that fail to capture modality complementarity. In addition, field data are typically scarce and distribution-shifted, making it difficult for conventional models to learn discriminative and generalizable features under small-sample conditions. To address these challenges, we propose a novel multi-sensor fusion few-shot learning network (MSFFLN) for GIS insulation defect diagnosis. First, a deep fusion network is developed to construct comprehensive representations of insulation defects. Specifically, a feature weighting fusion module is employed to improve robustness, while an adaptive attention-based fusion block suppresses redundant and aliased information, emphasizing the most discriminative features. Second, a contrastive learning-based few-shot strategy is introduced. By computing global and local contrastive losses and using contrastive learning as an auxiliary task, the model learns more accurate and generalizable feature representations. In addition, salient region mixing across samples is applied to decouple class-level and instance-level feature correlations. Finally, field experiments validate the effectiveness of the MSFFLN. Results show that the MSFFLN achieves a diagnostic accuracy of 95.06% with only 10 support samples, significantly outperforming baseline and ablation models in small-sample GIS insulation defect diagnosis.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"190 ","pages":"Article 114600"},"PeriodicalIF":6.6,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145980371","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 : 2026-01-12DOI: 10.1016/j.asoc.2026.114640
Yuteng Zhang , Leijun Shi , Qinkai Han , Xueping Xu , Hui Liu , Fulei Chu , Yun Kong
Reliable fault diagnosis is essential for maintaining the safety and operational efficiency of advanced industrial equipment. Diagnostic methods based on transfer learning techniques such as unsupervised domain adaptation have demonstrated considerable potential for engineering applications. However, existing methods rely on the predefined specific assumptions regarding inter-domain label relationships and domain configurations, which severely restrict their practical applications. To address these issues, this study proposes a unified cross-domain fault diagnosis framework for transfer diagnostic tasks under different label and domain configurations, including closed-set, partial-set, open-set, multi-source domain, and multi-target domain transfer diagnostics. The presented unified framework leverages a predictive class confusion bias shared across multiple scenarios to guide cross-domain knowledge transfer, thus enabling effective domain adaptation to various transfer diagnostic scenarios. To measure the tendency of class confusion accurately, a prototype similarity-based fault discrimination method is developed, which enhances classification robustness and provides reliable prediction distributions for predictive class confusion estimation. Then, a label smoothing-based probability calibration mechanism is designed for probability regularization, mitigating erroneous class confusion estimation caused by prediction bias. Additionally, an open-set cross-domain diagnosis method with an adaptive threshold is provided to handle potential unseen faults, which has a straightforward design and can be implemented easily within the unified cross-domain diagnosis framework. Extensive experiments on two transmission system datasets verify the general applicability of the proposed unified framework across five cross-domain diagnosis settings, and its performance is competitive with advanced scenario-specific transfer diagnosis methods, providing an effective tool for intelligent diagnosis in industrial scenarios.
{"title":"Minimizing predictive class confusion: A unified framework for cross-domain fault diagnosis under different label and domain configurations","authors":"Yuteng Zhang , Leijun Shi , Qinkai Han , Xueping Xu , Hui Liu , Fulei Chu , Yun Kong","doi":"10.1016/j.asoc.2026.114640","DOIUrl":"10.1016/j.asoc.2026.114640","url":null,"abstract":"<div><div>Reliable fault diagnosis is essential for maintaining the safety and operational efficiency of advanced industrial equipment. Diagnostic methods based on transfer learning techniques such as unsupervised domain adaptation have demonstrated considerable potential for engineering applications. However, existing methods rely on the predefined specific assumptions regarding inter-domain label relationships and domain configurations, which severely restrict their practical applications. To address these issues, this study proposes a unified cross-domain fault diagnosis framework for transfer diagnostic tasks under different label and domain configurations, including closed-set, partial-set, open-set, multi-source domain, and multi-target domain transfer diagnostics. The presented unified framework leverages a predictive class confusion bias shared across multiple scenarios to guide cross-domain knowledge transfer, thus enabling effective domain adaptation to various transfer diagnostic scenarios. To measure the tendency of class confusion accurately, a prototype similarity-based fault discrimination method is developed, which enhances classification robustness and provides reliable prediction distributions for predictive class confusion estimation. Then, a label smoothing-based probability calibration mechanism is designed for probability regularization, mitigating erroneous class confusion estimation caused by prediction bias. Additionally, an open-set cross-domain diagnosis method with an adaptive threshold is provided to handle potential unseen faults, which has a straightforward design and can be implemented easily within the unified cross-domain diagnosis framework. Extensive experiments on two transmission system datasets verify the general applicability of the proposed unified framework across five cross-domain diagnosis settings, and its performance is competitive with advanced scenario-specific transfer diagnosis methods, providing an effective tool for intelligent diagnosis in industrial scenarios.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"190 ","pages":"Article 114640"},"PeriodicalIF":6.6,"publicationDate":"2026-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145979792","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 : 2026-01-11DOI: 10.1016/j.asoc.2026.114629
Zhiyong Tan , Ruifen Cao , Pijing Wei , Chao Zhou , Yansen Su , Chunhou Zheng
Medical image segmentation is crucial for disease diagnosis and treatment, especially in oncology. However, current image-based unimodal methods are limited by data acquisition challenges, making it difficult to improve segmentation performance. The medical text annotation generated together with images provides rich semantic information at low cost, and the utilization of textual data presents a complementary approach to enhance analytical capabilities of image-based unimodal methods. However, medical image segmentation still encounters challenges, including complex background distributions, variable lesion shapes and sizes, and ambiguous boundaries. Accurate capture of edge information between the foreground (lesion or region of interest) and background (surrounding tissue) significantly influences segmentation outcomes. To address these challenges, we propose the Text-guided and Edge-guided Fusion Network (TEFNet), which integrates medical text knowledge and edge information to enhance segmentation performance. The text-guided strategy enriches the model's understanding of image content by leveraging semantic information from textual reports associated with medical images, enabling more accurate segmentation judgements. Edge-guided attention enhances the model's ability to identify anatomical structures and tissue boundaries by leveraging high-frequency edge information, enabling more reliable boundary delineation. Additionally, we introduce the Segmentation Anything Model (SAM), specifically tailored for the biomedical domain, to further enhance medical feature representation. Comprehensive evaluations across three established medical image segmentation benchmarks demonstrate that TEFNet, through a synergistic fusion of visual and textual features, achieves superior segmentation accuracy compared with current leading methods, validating the effectiveness of joint visual-text feature learning in medical image segmentation.
{"title":"Text-guided and Edge-guided fusion network for enhancing medical image segmentation","authors":"Zhiyong Tan , Ruifen Cao , Pijing Wei , Chao Zhou , Yansen Su , Chunhou Zheng","doi":"10.1016/j.asoc.2026.114629","DOIUrl":"10.1016/j.asoc.2026.114629","url":null,"abstract":"<div><div>Medical image segmentation is crucial for disease diagnosis and treatment, especially in oncology. However, current image-based unimodal methods are limited by data acquisition challenges, making it difficult to improve segmentation performance. The medical text annotation generated together with images provides rich semantic information at low cost, and the utilization of textual data presents a complementary approach to enhance analytical capabilities of image-based unimodal methods. However, medical image segmentation still encounters challenges, including complex background distributions, variable lesion shapes and sizes, and ambiguous boundaries. Accurate capture of edge information between the foreground (lesion or region of interest) and background (surrounding tissue) significantly influences segmentation outcomes. To address these challenges, we propose the Text-guided and Edge-guided Fusion Network (TEFNet), which integrates medical text knowledge and edge information to enhance segmentation performance. The text-guided strategy enriches the model's understanding of image content by leveraging semantic information from textual reports associated with medical images, enabling more accurate segmentation judgements. Edge-guided attention enhances the model's ability to identify anatomical structures and tissue boundaries by leveraging high-frequency edge information, enabling more reliable boundary delineation. Additionally, we introduce the Segmentation Anything Model (SAM), specifically tailored for the biomedical domain, to further enhance medical feature representation. Comprehensive evaluations across three established medical image segmentation benchmarks demonstrate that TEFNet, through a synergistic fusion of visual and textual features, achieves superior segmentation accuracy compared with current leading methods, validating the effectiveness of joint visual-text feature learning in medical image segmentation.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"190 ","pages":"Article 114629"},"PeriodicalIF":6.6,"publicationDate":"2026-01-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145980379","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