Pub Date : 2025-08-26DOI: 10.1016/j.jocs.2025.102693
Daniel Heinlein, Anton Akusok, Kaj-Mikael Björk, Leonardo Espinosa-Leal
In federated learning, multiple devices compute each a part of a common machine learning model using their own private data. These partial models (or their parameters) are then exchanged in a central server that builds an aggregated model. This sharing process may leak information about the data used to train them. This problem intensifies as the machine learning model becomes simpler, indicating a higher risk for single-hidden-layer feedforward neural networks, such as extreme learning machines. In this paper, we establish a mechanism to disguise the input data to a system of linear equations while guaranteeing that the modifications do not alter the solutions, and propose two possible approaches to apply these techniques to federated learning. Our findings show that extreme learning machines can be used in federated learning with an extra security layer, making them attractive in learning schemes with limited computational resources.
{"title":"Private linear equation solving: An application to federated learning and extreme learning machines","authors":"Daniel Heinlein, Anton Akusok, Kaj-Mikael Björk, Leonardo Espinosa-Leal","doi":"10.1016/j.jocs.2025.102693","DOIUrl":"10.1016/j.jocs.2025.102693","url":null,"abstract":"<div><div>In federated learning, multiple devices compute each a part of a common machine learning model using their own private data. These partial models (or their parameters) are then exchanged in a central server that builds an aggregated model. This sharing process may leak information about the data used to train them. This problem intensifies as the machine learning model becomes simpler, indicating a higher risk for single-hidden-layer feedforward neural networks, such as extreme learning machines. In this paper, we establish a mechanism to disguise the input data to a system of linear equations while guaranteeing that the modifications do not alter the solutions, and propose two possible approaches to apply these techniques to federated learning. Our findings show that extreme learning machines can be used in federated learning with an extra security layer, making them attractive in learning schemes with limited computational resources.</div></div>","PeriodicalId":48907,"journal":{"name":"Journal of Computational Science","volume":"92 ","pages":"Article 102693"},"PeriodicalIF":3.7,"publicationDate":"2025-08-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144908410","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Small datasets often lead to poor performance of data-driven prediction models due to uneven data distribution and large data spacing. One popular approach to address this issue is to use virtual samples during machine learning (ML) model training. This study proposes a Hamiltonian Circuit Virtual Sample Generation (HCVSG) method to distribute virtual samples generated using interpolation techniques while integrating the K-Nearest Neighbors (KNN) algorithm in model development. The Hamiltonian circuit is chosen because it doesn’t depend on the distribution assumption and provides multiple circuits that allow adaptive sample distribution, allowing the selection of circuits that produce minimum errors. This method supports improving feature-target correlation, reducing the risk of overfitting, and stabilizing error values as model complexity increases. Applying this method to three datasets in material research (MLCC, PSH, and EFD) shows that HCVSG significantly improves prediction accuracy compared to conventional KNN and eight MTD-based methods. The distribution of virtual samples along the Hamiltonian circuit helps fill the information gap and makes the data distribution more even, ultimately improving the predictive model's performance.
{"title":"An adaptive Hamiltonian circuit of virtual sample generation for a small dataset","authors":"Totok Sutojo , Supriadi Rustad , Muhamad Akrom , Wahyu Aji Eko Prabowo , De Rosal Ignatius Moses Setiadi , Hermawan Kresno Dipojono , Yoshitada Morikawa","doi":"10.1016/j.jocs.2025.102711","DOIUrl":"10.1016/j.jocs.2025.102711","url":null,"abstract":"<div><div>Small datasets often lead to poor performance of data-driven prediction models due to uneven data distribution and large data spacing. One popular approach to address this issue is to use virtual samples during machine learning (ML) model training. This study proposes a Hamiltonian Circuit Virtual Sample Generation (HCVSG) method to distribute virtual samples generated using interpolation techniques while integrating the K-Nearest Neighbors (KNN) algorithm in model development. The Hamiltonian circuit is chosen because it doesn’t depend on the distribution assumption and provides multiple circuits that allow adaptive sample distribution, allowing the selection of circuits that produce minimum errors. This method supports improving feature-target correlation, reducing the risk of overfitting, and stabilizing error values as model complexity increases. Applying this method to three datasets in material research (MLCC, PSH, and EFD) shows that HCVSG significantly improves prediction accuracy compared to conventional KNN and eight MTD-based methods. The distribution of virtual samples along the Hamiltonian circuit helps fill the information gap and makes the data distribution more even, ultimately improving the predictive model's performance.</div></div>","PeriodicalId":48907,"journal":{"name":"Journal of Computational Science","volume":"92 ","pages":"Article 102711"},"PeriodicalIF":3.7,"publicationDate":"2025-08-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144891889","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
In this work, Density Functional Theory (DFT) was employed to investigate the impact of SnS, GeS, SnSe, and GeSe quantum dots doped black phosphorene on the sensitivity of black phosphorene toward various adsorbed gas molecules namely NO2 and H2S. The interaction of H2O molecule with doped black phosphorene surface is also investigated to evaluate the impact of humidity on the sensing response. The results revealed the large electronic changes in bands distribution upon exposure to the selected gas molecules, giving rise to a variation in the electronic band nature from hole to electron doping which can promote the electrical conductivity and the sensing properties of the doped phosphorene structures.
{"title":"Tuning sensitivity of black phosphorene surface doped SnS, SnSe, GeS, and GeSe quantum dots toward water molecule and other small toxic molecules","authors":"Mamori Habiba , Moatassim Hajar , El Kenz Abdallah , Benyoussef Abdelilah , Taleb Abdelhafed , Abdel Ghafour El Hachimi , Zaari Halima","doi":"10.1016/j.jocs.2025.102707","DOIUrl":"10.1016/j.jocs.2025.102707","url":null,"abstract":"<div><div>In this work, Density Functional Theory (DFT) was employed to investigate the impact of SnS, GeS, SnSe, and GeSe quantum dots doped black phosphorene on the sensitivity of black phosphorene toward various adsorbed gas molecules namely NO<sub>2</sub> and H<sub>2</sub>S. The interaction of H<sub>2</sub>O molecule with doped black phosphorene surface is also investigated to evaluate the impact of humidity on the sensing response. The results revealed the large electronic changes in bands distribution upon exposure to the selected gas molecules, giving rise to a variation in the electronic band nature from hole to electron doping which can promote the electrical conductivity and the sensing properties of the doped phosphorene structures.</div></div>","PeriodicalId":48907,"journal":{"name":"Journal of Computational Science","volume":"92 ","pages":"Article 102707"},"PeriodicalIF":3.7,"publicationDate":"2025-08-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144889883","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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 : 2025-08-16DOI: 10.1016/j.jocs.2025.102692
Emrah Inan
An impression in a typical radiology report emphasises critical information by providing a conclusion and reasoning based on the findings. However, the findings and impression sections of these reports generally contain brief texts, as they highlight crucial observations derived from the clinical radiograph. In this scenario, abstractive summarisation models often experience a degradation in performance when generating short impressions. To address this challenge in the summarisation task, our work proposes a method that combines well-known fine-tuned text classification and abstractive summarisation language models. Since fine-tuning a language model requires an extensive, well-defined training dataset and is a time-consuming task dependent on high GPU resources, we employ prompt engineering, which uses prompt templates to programme language models and improve their performance. Our method first predicts whether the given findings text is normal or abnormal by leveraging a fine-tuned language model. Then, we apply a radiology-specific BART model to generate the summary for abnormal findings. In the zero-shot setting, our method achieves remarkable results compared to existing approaches on a real-world dataset. In particular, our method achieves scores of 37.43 for ROUGE-1, 21.72 for ROUGE-2, and 35.52 for ROUGE-L.
{"title":"Making hierarchically aware decisions on short findings for automatic summarisation","authors":"Emrah Inan","doi":"10.1016/j.jocs.2025.102692","DOIUrl":"10.1016/j.jocs.2025.102692","url":null,"abstract":"<div><div>An impression in a typical radiology report emphasises critical information by providing a conclusion and reasoning based on the findings. However, the findings and impression sections of these reports generally contain brief texts, as they highlight crucial observations derived from the clinical radiograph. In this scenario, abstractive summarisation models often experience a degradation in performance when generating short impressions. To address this challenge in the summarisation task, our work proposes a method that combines well-known fine-tuned text classification and abstractive summarisation language models. Since fine-tuning a language model requires an extensive, well-defined training dataset and is a time-consuming task dependent on high GPU resources, we employ prompt engineering, which uses prompt templates to programme language models and improve their performance. Our method first predicts whether the given findings text is normal or abnormal by leveraging a fine-tuned language model. Then, we apply a radiology-specific BART model to generate the summary for abnormal findings. In the zero-shot setting, our method achieves remarkable results compared to existing approaches on a real-world dataset. In particular, our method achieves scores of 37.43 for ROUGE-1, 21.72 for ROUGE-2, and 35.52 for ROUGE-L.</div></div>","PeriodicalId":48907,"journal":{"name":"Journal of Computational Science","volume":"91 ","pages":"Article 102692"},"PeriodicalIF":3.7,"publicationDate":"2025-08-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144852163","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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 : 2025-08-15DOI: 10.1016/j.jocs.2025.102695
Sungyeop Lee , Jisu Ryu , Young-Gu Kim , Dae Sin Kim , Hiroo Koshimoto , Jaeshin Park
Simulation plays a crucial role in the semiconductor chip manufacturing. In particular, process simulation is primarily used to solve the dopant diffusion dynamics, which describes the temporal evolution of doping profiles during the thermal annealing process. The diffusion dynamics constitutes a multiscale problem, formulated as a set of coupled partial differential equations (PDEs) with respect to the concentration of dopants and point defects. In this paper, we demonstrate that Physics-Informed Neural Networks (PINNs) can accurately predict not only the evolution of the doping profile, but also the unknown physical parameters, specifically the diffusivities appearing as PDE coefficients. Furthermore, we propose a physics-informed calibration method, which performs PDE-constrained optimization by leveraging a pre-trained PINN model. We experimentally verify that this post-processing significantly improves the accuracy of coefficients fine-tuning. To the best of our knowledge, this is the first demonstration of an annealing simulation for the semiconductor diffusion process using a physics-informed machine learning approach. This framework is expected to enable more efficient calibration of simulation parameters based on measurement data.
{"title":"Solving the dopant diffusion dynamics with physics-informed neural networks","authors":"Sungyeop Lee , Jisu Ryu , Young-Gu Kim , Dae Sin Kim , Hiroo Koshimoto , Jaeshin Park","doi":"10.1016/j.jocs.2025.102695","DOIUrl":"10.1016/j.jocs.2025.102695","url":null,"abstract":"<div><div>Simulation plays a crucial role in the semiconductor chip manufacturing. In particular, process simulation is primarily used to solve the dopant diffusion dynamics, which describes the temporal evolution of doping profiles during the thermal annealing process. The diffusion dynamics constitutes a multiscale problem, formulated as a set of coupled partial differential equations (PDEs) with respect to the concentration of dopants and point defects. In this paper, we demonstrate that Physics-Informed Neural Networks (PINNs) can accurately predict not only the evolution of the doping profile, but also the unknown physical parameters, specifically the diffusivities appearing as PDE coefficients. Furthermore, we propose a physics-informed calibration method, which performs PDE-constrained optimization by leveraging a pre-trained PINN model. We experimentally verify that this post-processing significantly improves the accuracy of coefficients fine-tuning. To the best of our knowledge, this is the first demonstration of an annealing simulation for the semiconductor diffusion process using a physics-informed machine learning approach. This framework is expected to enable more efficient calibration of simulation parameters based on measurement data.</div></div>","PeriodicalId":48907,"journal":{"name":"Journal of Computational Science","volume":"92 ","pages":"Article 102695"},"PeriodicalIF":3.7,"publicationDate":"2025-08-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144863514","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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 : 2025-08-13DOI: 10.1016/j.jocs.2025.102696
Shupeng Gao , Qi Li , M.A. Gosalvez , Xi Lin , Yan Xing , Zaifa Zhou , Qianhuang Chen
Currently, the time and cost required to obtain large datasets limit the application of data-driven machine learning in nanoscale manufacturing. Here, we focus on predicting the nanoscale damage induced by helium focused ion beams (He-FIBs) on silicon substrates. We briefly review the most relevant atomistic defects and the partial differential equations (PDEs), or rate equations, that describe the mutual creation and annihilation of the defects, eventually leading to the amorphization of the substrate and, the nucleation and early growth of helium bubbles. The novelty comes from the use of a physics-informed neural network (PINN) to simulate quantitatively the evolution of the bubbles, thus bypassing the dataset availability problem. As usual, the proposed PINN learns the underlying physics through the incorporation of the residuals of the PDEs and corresponding Initial Conditions (ICs) and Boundary Conditions (BCs) in the network’s loss function. Meanwhile, the system of PDEs poses some challenges to the PINN modeling strategy. We find that (i) hard constraints need to be imposed on the network output in order to satisfy both BCs and ICs, (ii) all the inputs and outputs of the PINN need to be cautiously normalized to ensure convergence during training, and (iii) customized weights need to be carefully applied to all the PDE loss terms in order to balance their contributions, thus improving the accuracy of the PINN predictions. Once trained, the network achieves good prediction accuracy over the entire space-time domain for various ion beam energies and doses. Comparisons are provided against previous experiments and traditional numerical simulations, which are also implemented in this study using the Finite Difference Method (FDM). While the L2 relative errors for all collocated points remain below 10%, the accuracy of the PINN decreases at lower beam energies and larger ion doses, due to the presence of higher numerical gradients.
{"title":"Helium focused ion beam damage in silicon: Physics-informed neural network modeling of helium bubble nucleation and early growth","authors":"Shupeng Gao , Qi Li , M.A. Gosalvez , Xi Lin , Yan Xing , Zaifa Zhou , Qianhuang Chen","doi":"10.1016/j.jocs.2025.102696","DOIUrl":"10.1016/j.jocs.2025.102696","url":null,"abstract":"<div><div>Currently, the time and cost required to obtain large datasets limit the application of data-driven machine learning in nanoscale manufacturing. Here, we focus on predicting the nanoscale damage induced by helium focused ion beams (He-FIBs) on silicon substrates. We briefly review the most relevant atomistic defects and the partial differential equations (PDEs), or rate equations, that describe the mutual creation and annihilation of the defects, eventually leading to the amorphization of the substrate and, the nucleation and early growth of helium bubbles. The novelty comes from the use of a physics-informed neural network (PINN) to simulate quantitatively the evolution of the bubbles, thus bypassing the dataset availability problem. As usual, the proposed PINN learns the underlying physics through the incorporation of the residuals of the PDEs and corresponding Initial Conditions (ICs) and Boundary Conditions (BCs) in the network’s loss function. Meanwhile, the system of PDEs poses some challenges to the PINN modeling strategy. We find that (i) hard constraints need to be imposed on the network output in order to satisfy both BCs and ICs, (ii) all the inputs and outputs of the PINN need to be cautiously normalized to ensure convergence during training, and (iii) customized weights need to be carefully applied to all the PDE loss terms in order to balance their contributions, thus improving the accuracy of the PINN predictions. Once trained, the network achieves good prediction accuracy over the entire space-time domain for various ion beam energies and doses. Comparisons are provided against previous experiments and traditional numerical simulations, which are also implemented in this study using the Finite Difference Method (FDM). While the L2 relative errors for all collocated points remain below 10%, the accuracy of the PINN decreases at lower beam energies and larger ion doses, due to the presence of higher numerical gradients.</div></div>","PeriodicalId":48907,"journal":{"name":"Journal of Computational Science","volume":"92 ","pages":"Article 102696"},"PeriodicalIF":3.7,"publicationDate":"2025-08-13","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144863513","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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 : 2025-08-08DOI: 10.1016/j.jocs.2025.102690
Himanshu Rai , Sanjeev K. Tomer
Bayesian competing risks analysis in presence of masked data often leads to an intractable posterior, for which Markov chain Monte Carlo (MCMC) methods are frequently utilized to evaluate various estimators of interest. However, while analyzing several risks simultaneously, MCMC methods may consume substantial amount of computation time. This paper introduces Variational Bayes, a machine learning technique, as an efficient alternative to MCMC for Bayesian analysis of competing risk data. Variational Bayes demonstrates faster convergence than MCMC, particularly in the context of extensive competing risk datasets. We compare the performance of variational Bayes over Gibbs sampling with respect to the number of considered risks through a simulation study. Additionally, we apply the two methods to analyze a real data set of computer hard drives.
{"title":"Variational Bayes for analysis of masked data","authors":"Himanshu Rai , Sanjeev K. Tomer","doi":"10.1016/j.jocs.2025.102690","DOIUrl":"10.1016/j.jocs.2025.102690","url":null,"abstract":"<div><div>Bayesian competing risks analysis in presence of masked data often leads to an intractable posterior, for which Markov chain Monte Carlo (MCMC) methods are frequently utilized to evaluate various estimators of interest. However, while analyzing several risks simultaneously, MCMC methods may consume substantial amount of computation time. This paper introduces Variational Bayes, a machine learning technique, as an efficient alternative to MCMC for Bayesian analysis of competing risk data. Variational Bayes demonstrates faster convergence than MCMC, particularly in the context of extensive competing risk datasets. We compare the performance of variational Bayes over Gibbs sampling with respect to the number of considered risks through a simulation study. Additionally, we apply the two methods to analyze a real data set of computer hard drives.</div></div>","PeriodicalId":48907,"journal":{"name":"Journal of Computational Science","volume":"91 ","pages":"Article 102690"},"PeriodicalIF":3.7,"publicationDate":"2025-08-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144842213","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Datasets classification accuracy depends on their features. The presence of irrelevant and redundant features in the dataset leads to the reduction of classification accuracy. Identifying and removing such features is the main purpose in feature selection, which is an important step in the data science lifecycle. The objective of the Wrapper feature selection method is to reduce the number of selected feature (NSF) while improving the classification accuracy by working on a set of features. The feature selection is a challenging and computationally expensive problem that falls under the NP-complete category, so it requires computationally cheap and efficient algorithm to solve it. The artificial hummingbird algorithm (AHA) is a biological inspired optimization technique that mimics the unique flight capabilities and intelligent foraging tactics of hummingbirds in nature. Since feature selection is inherently a binary problem. In this paper, the binary form of the AHA meta-heuristic algorithm is proposed to show that binarizing the AHA meta-heuristic algorithm improves its performance for solving feature selection problems. The proposed method is tested on a standard benchmark dataset and compared with four state-of-the-art feature selection algorithms: Automata-based improved equilibrium optimizer with U-shaped transfer function (AIEOU), Whale optimization approaches for wrapper feature selection (WOA-CM), Ring theory-based harmony search (RTHS), and Adaptive switching gray-whale optimizer (ASGW). The results show the effectiveness of the proposed algorithm in searching for optimal features subset. The source code for the algorithm being proposed is accessible to the public on https://github.com/alihamdipour/baha.
{"title":"BAHA: Binary artificial hummingbird algorithm for feature selection","authors":"Ali Hamdipour , Abdolali Basiri , Mostafa Zaare , Seyedali Mirjalili","doi":"10.1016/j.jocs.2025.102686","DOIUrl":"10.1016/j.jocs.2025.102686","url":null,"abstract":"<div><div>Datasets classification accuracy depends on their features. The presence of irrelevant and redundant features in the dataset leads to the reduction of classification accuracy. Identifying and removing such features is the main purpose in feature selection, which is an important step in the data science lifecycle. The objective of the Wrapper feature selection method is to reduce the number of selected feature (NSF) while improving the classification accuracy by working on a set of features. The feature selection is a challenging and computationally expensive problem that falls under the NP-complete category, so it requires computationally cheap and efficient algorithm to solve it. The artificial hummingbird algorithm (AHA) is a biological inspired optimization technique that mimics the unique flight capabilities and intelligent foraging tactics of hummingbirds in nature. Since feature selection is inherently a binary problem. In this paper, the binary form of the AHA meta-heuristic algorithm is proposed to show that binarizing the AHA meta-heuristic algorithm improves its performance for solving feature selection problems. The proposed method is tested on a standard benchmark dataset and compared with four state-of-the-art feature selection algorithms: Automata-based improved equilibrium optimizer with U-shaped transfer function (AIEOU), Whale optimization approaches for wrapper feature selection (WOA-CM), Ring theory-based harmony search (RTHS), and Adaptive switching gray-whale optimizer (ASGW). The results show the effectiveness of the proposed algorithm in searching for optimal features subset. The source code for the algorithm being proposed is accessible to the public on <span><span>https://github.com/alihamdipour/baha</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":48907,"journal":{"name":"Journal of Computational Science","volume":"92 ","pages":"Article 102686"},"PeriodicalIF":3.7,"publicationDate":"2025-07-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144863515","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Anomaly detection is the process of identifying unusual patterns in data that may indicate a deviation from the expected norm. This paper proposes a semi-supervised deep learning solution to detect anomalies of a YANMAR energy device that produces heat and power utilizing an internal combustion engine supplied with natural gas. The main equipment of the analysis is a 20 micro-cogeneration unit installed in the energy plant of a facility school. More in detail, the dataset considered in this work consists of 12 features temporally acquired every 15 min. The authors exploit a deep learning architecture, an autoencoder with 1-D convolutional layers to retain temporal correlations, trained to learn the normal behavior of the cogenerator and report unseen operations. In consideration of the fact that autoencoders tend to yield false positives, a Fast-Fourier-Transform-based technique has been applied to filter spurious detections and improve the algorithm’s robustness. As the last contribution, a naive methodology to address the root cause of the anomalies has been explained and its effectiveness has been proved in a real malfunctioning of the CHP.
{"title":"Anomaly detection and root cause analysis using convolutional autoencoders: A real case study","authors":"Piero Danti , Alessandro Innocenti , Sascha Sandomier","doi":"10.1016/j.jocs.2025.102685","DOIUrl":"10.1016/j.jocs.2025.102685","url":null,"abstract":"<div><div>Anomaly detection is the process of identifying unusual patterns in data that may indicate a deviation from the expected norm. This paper proposes a semi-supervised deep learning solution to detect anomalies of a YANMAR energy device that produces heat and power utilizing an internal combustion engine supplied with natural gas. The main equipment of the analysis is a 20 <span><math><mrow><mi>k</mi><msub><mrow><mi>W</mi></mrow><mrow><mi>e</mi></mrow></msub></mrow></math></span> micro-cogeneration unit installed in the energy plant of a facility school. More in detail, the dataset considered in this work consists of 12 features temporally acquired every 15 min. The authors exploit a deep learning architecture, an autoencoder with 1-D convolutional layers to retain temporal correlations, trained to learn the normal behavior of the cogenerator and report unseen operations. In consideration of the fact that autoencoders tend to yield false positives, a Fast-Fourier-Transform-based technique has been applied to filter spurious detections and improve the algorithm’s robustness. As the last contribution, a naive methodology to address the root cause of the anomalies has been explained and its effectiveness has been proved in a real malfunctioning of the CHP.</div></div>","PeriodicalId":48907,"journal":{"name":"Journal of Computational Science","volume":"91 ","pages":"Article 102685"},"PeriodicalIF":3.7,"publicationDate":"2025-07-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144739414","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
The traditional K-means algorithm has several limitations, including sensitivity to initial center, unstable clustering results, local optimal clustering results, and a large number of iterations. In this paper, we propose an improved clustering algorithm called PH-K-means that utilizes the persistent homology to identify k clusters in the data set. The algorithm calculates the length of the longest Betti number to obtain k Betti numbers, which represent the k clusters respectively. The data is then output in k Betty numbers, and the average value of the data in each Betti number is used as the initialization center of k clusters. The algorithm iterates until the difference of the square sum of the errors in the adjacent two clusters is less than the threshold value. The PH-K-means algorithm is tested on seven common data sets, and the results show that it has high accuracy, stable clustering results, and requires fewer iterations than traditional K-means, K-means++, UK-means, and K-means algorithms.
{"title":"An improved K-means algorithm based on persistent homology","authors":"NingNing Peng, Shanjunshu Gao, Xingzi Yin, Xueyan Zhan","doi":"10.1016/j.jocs.2025.102680","DOIUrl":"10.1016/j.jocs.2025.102680","url":null,"abstract":"<div><div>The traditional K-means algorithm has several limitations, including sensitivity to initial center, unstable clustering results, local optimal clustering results, and a large number of iterations. In this paper, we propose an improved clustering algorithm called PH-K-means that utilizes the persistent homology to identify k clusters in the data set. The algorithm calculates the length of the longest Betti number to obtain k Betti numbers, which represent the k clusters respectively. The data is then output in k Betty numbers, and the average value of the data in each Betti number is used as the initialization center of k clusters. The algorithm iterates until the difference of the square sum of the errors in the adjacent two clusters is less than the threshold value. The PH-K-means algorithm is tested on seven common data sets, and the results show that it has high accuracy, stable clustering results, and requires fewer iterations than traditional K-means, K-means++, UK-means, and K-means algorithms.</div></div>","PeriodicalId":48907,"journal":{"name":"Journal of Computational Science","volume":"91 ","pages":"Article 102680"},"PeriodicalIF":3.7,"publicationDate":"2025-07-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144771517","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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