IEEE transactions on artificial intelligence最新文献_第4页

Toward a Unified Framework for Consistency Generative Modeling 面向一致性生成建模的统一框架

IEEE transactions on artificial intelligence Pub Date : 2026-03-01 Epub Date: 2025-10-23 DOI: 10.1109/TAI.2025.3624330

Hongkun Dou;Junzhe Lu;Jinyang Du;Chengwei Fu;Wen Yao;Hongjue Li;Yue Deng

{"title":"Toward a Unified Framework for Consistency Generative Modeling","authors":"Hongkun Dou;Junzhe Lu;Jinyang Du;Chengwei Fu;Wen Yao;Hongjue Li;Yue Deng","doi":"10.1109/TAI.2025.3624330","DOIUrl":"https://doi.org/10.1109/TAI.2025.3624330","url":null,"abstract":"Consistency modeling, a novel generative paradigm inspired by diffusion models, has gained traction for its capacity to facilitate real-time generation through single-step sampling. While its advantages are evident, the understanding of its underlying principles and effective algorithmic enhancements remains elusive. In response, we present a unified framework for consistency generative modeling, without resorting to the predefined diffusion process. Instead, it directly constructs a probability density path that bridges the two distributions. Building upon this novel perspective, we introduce a more general consistency training objective that encapsulates previous consistency models and paves the way for innovative, consistency generation techniques. In particular, we introduce two novel models: Poisson consistency models (PCMs) and coupling consistency models (CCMs), which extend the prior distribution of latent variables beyond the Gaussian form. This extension significantly augments the flexibility of generative modeling. Furthermore, we harness the principles of optimal transport (OT) to mitigate variance during consistency training, substantially improving convergence and generative quality. Extensive experiments on the generation of synthetic and real-world datasets, as well as image-to-image translation tasks (I2I), demonstrate the effectiveness of the proposed approaches.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 5","pages":"2761-2773"},"PeriodicalIF":0.0,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147757143","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

FUBA: Backdoor Federated Learning via Federated Unlearning FUBA：通过联邦学习的后门联邦学习

IEEE transactions on artificial intelligence Pub Date : 2026-03-01 Epub Date: 2025-11-06 DOI: 10.1109/TAI.2025.3630110

Xinyi Sheng;Wei Bao;Yichen Guo;Sen Fu

{"title":"FUBA: Backdoor Federated Learning via Federated Unlearning","authors":"Xinyi Sheng;Wei Bao;Yichen Guo;Sen Fu","doi":"10.1109/TAI.2025.3630110","DOIUrl":"https://doi.org/10.1109/TAI.2025.3630110","url":null,"abstract":"Federated unlearning (FU) enables participants in federated learning (FL) to exercise the “right to be forgotten (RTBF)” by removing their contributions from a collaboratively trained model. While this RTBF enhances user data privacy, it also introduces new vulnerabilities, as unlearning requests can be exploited to compromise the unlearned model. This work presents a novel federated unlearning backdoor attack (FUBA) framework, which leverages malicious unlearning requests to backdoor the FL model after FU. FUBA models this attack as an adversarial game between two types of adversarial clients: the Adv-attacker and the Adv-defender. During FL, the Adv-attacker keeps injecting backdoor, whereas the Adv-defender acts to mitigate these injections. Then, in FU, the Adv-defender activates the backdoor by submitting an unlearning request to remove its contribution. In the Adv-attacker, we propose an innovative gradient reweighting mechanism and an adaptive model update scaling algorithm for stable and durable trigger embedding. In the Adv-defender, we introduce an advanced two-stage defense strategy and a model consistency loss for reliable backdoor activation. Extensive experiments demonstrate that FUBA effectively and reliably backdoors the model after FU, regardless of the FU method applied. Furthermore, FUBA’s unique design exhibits high stealth against existing backdoor defenses. The code is available at <uri>https://github.com/stcebra/FUBA</uri>.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 5","pages":"2892-2907"},"PeriodicalIF":0.0,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147757152","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Graph Attention Networks With Dual-Edge Connectivity for Alzheimer’s Disease Detection From Speech 基于双边缘连接的阿尔茨海默病语音检测图注意网络

IEEE transactions on artificial intelligence Pub Date : 2026-03-01 Epub Date: 2025-11-14 DOI: 10.1109/TAI.2025.3632821

Anass El Hallani;Adil Chakhtouna;Abdellah Adib

{"title":"Graph Attention Networks With Dual-Edge Connectivity for Alzheimer’s Disease Detection From Speech","authors":"Anass El Hallani;Adil Chakhtouna;Abdellah Adib","doi":"10.1109/TAI.2025.3632821","DOIUrl":"https://doi.org/10.1109/TAI.2025.3632821","url":null,"abstract":"Early and noninvasive screening tools for Alzheimer’s disease (AD) are critical to large-scale clinical management. However, current speech-based systems often require transcripts, language specific resources, or handcrafted features that overlook long range dependencies in spontaneous discourse. We propose a graph representation, where nodes correspond to short, overlapping speech segments encoded with acoustic–prosodic features, and edges comprise temporal connections and k-nearest neighbor (kNN) similarity links weighted by Euclidean distance. The adjacency matrix integrates these dual-edge structures, enabling the capture of both local dynamics and long-range dependencies relevant to AD, such as perseveration and prolonged pauses. We implement two graph neural networks (GNNs): a graph convolutional network (GCN) and a graph attention network (GAT). Both models are trained using stratified fivefold cross validation. Experimental results demonstrate superior performance with the full model (GCN: 0.910, GAT: 0.928 accuracy), with ablations revealing the critical role of similarity edges and attention mechanisms. Removing similarity or temporal edges reduces accuracy, while mean pooling and single-head attention slightly degrade GAT performance (0.917 and 0.922, respectively). This approach outperforms conventional sequential baselines such as LSTM and Transformer models (typically 0.80–0.90 accuracy), and provides a robust, scalable early AD screening tool. Future work will extend it to multimodal graphs to further enhance detection and generalization.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 5","pages":"2956-2966"},"PeriodicalIF":0.0,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147757123","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Multimodal Human Pose Estimation: A Wi-Fi-Driven Approach With Adaptive Kernel Selection 多模态人体姿态估计：一种wi - fi驱动的自适应核选择方法

IEEE transactions on artificial intelligence Pub Date : 2026-03-01 Epub Date: 2025-11-12 DOI: 10.1109/TAI.2025.3631005

Toan D. Gian;Dung T. Tran;Quoc-Viet Pham;Le-Nam Tran;Van-Dinh Nguyen

{"title":"Multimodal Human Pose Estimation: A Wi-Fi-Driven Approach With Adaptive Kernel Selection","authors":"Toan D. Gian;Dung T. Tran;Quoc-Viet Pham;Le-Nam Tran;Van-Dinh Nguyen","doi":"10.1109/TAI.2025.3631005","DOIUrl":"https://doi.org/10.1109/TAI.2025.3631005","url":null,"abstract":"Wireless fidelity (Wi-Fi)-based human pose estimation (HPE) has emerged as a promising alternative to vision-based techniques, enabling human pose detection and movement interpretation while ensuring privacy. However, high computational costs and performance limitations hinder widespread adoption, particularly on resource-constrained devices. This article introduces HPE-Li++, a novel approach leveraging multimodal sensors (e.g., camera and Wi-Fi) to achieve accurate 3-D skeletal HPE with lightweight computation. We develop an efficient deep neural network featuring a multibranch convolutional neural network (CNN) enhanced by selective kernel attention (SKA), which dynamically adjusts kernel sizes based on input characteristics, improving adaptability with negligible complexity increase. To enhance efficiency and robustness, we incorporate a Transformer module to capture inter-domain correlations for effective feature extraction and a stacked autoencoder (SAE)-based denoiser to improve accuracy through latent representations while reducing computational cost. Extensive experiments on MM-Fi and WiPose datasets demonstrate that HPE-Li++ outperforms state-of-the-art (SOTA) methods, achieving 85.58% and 94.27% at <inline-formula><tex-math>$text{PCK}_{50}$</tex-math></inline-formula>, respectively, with minimal computational overhead. Notably, it remains robust under noise, maintaining 80% <inline-formula><tex-math>$text{PCK}_{50}$</tex-math></inline-formula> under AWGN noise with an error variance of 0.5.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 5","pages":"2941-2955"},"PeriodicalIF":0.0,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147757146","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

IGDNet: Zero-Shot Robust Underexposed Image Enhancement via Illumination-Guided and Denoising 基于光照引导和去噪的零镜头鲁棒曝光不足图像增强

IEEE transactions on artificial intelligence Pub Date : 2026-03-01 Epub Date: 2025-11-18 DOI: 10.1109/TAI.2025.3633653

Hailong Yan;Junjian Huang;Tingwen Huang

{"title":"IGDNet: Zero-Shot Robust Underexposed Image Enhancement via Illumination-Guided and Denoising","authors":"Hailong Yan;Junjian Huang;Tingwen Huang","doi":"10.1109/TAI.2025.3633653","DOIUrl":"https://doi.org/10.1109/TAI.2025.3633653","url":null,"abstract":"Current methods for restoring underexposed images typically rely on supervised learning with paired underexposed and well-illuminated images. However, collecting such datasets is often impractical in real-world scenarios. Moreover, these methods can lead to over-enhancement, distorting well-illuminated regions. To address these issues, we propose IGDNet, a zero-shot enhancement method that operates solely on a single test image, without requiring guiding priors or training data. IGDNet exhibits strong generalization ability and effectively suppresses noise while restoring illumination. The framework comprises a decomposition module and a denoising module. The former separates the image into illumination and reflection components via a dense connection network, while the latter enhances nonuniformly illuminated regions using an illumination-guided pixel-adaptive correction method. A noise pair is generated through downsampling and refined iteratively to produce the final result. Extensive experiments on four public datasets demonstrate that IGDNet significantly improves visual quality under complex lighting conditions. Quantitative results on metrics like PSNR (20.41 dB) and SSIM (0.860 dB) show that it outperforms 14 state-of-the-art unsupervised methods.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 5","pages":"2995-3005"},"PeriodicalIF":0.0,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147757159","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

MGLAN: A Mamba-Like Gated Linear Attention With Normalization for Solving CTSP 一类带归一化的类mamba门控线性注意解CTSP

IEEE transactions on artificial intelligence Pub Date : 2026-03-01 Epub Date: 2025-11-10 DOI: 10.1109/TAI.2025.3630620

Chengda Wen;Xiangping Xu;Xinli Shi;Jinde Cao;Liang Hua

{"title":"MGLAN: A Mamba-Like Gated Linear Attention With Normalization for Solving CTSP","authors":"Chengda Wen;Xiangping Xu;Xinli Shi;Jinde Cao;Liang Hua","doi":"10.1109/TAI.2025.3630620","DOIUrl":"https://doi.org/10.1109/TAI.2025.3630620","url":null,"abstract":"Recent neural network models, particularly those based on reinforcement learning (RL) and supervised learning, have shown great success in solving various combinatorial problems such as the traveling salesman problem (TSP). The colored TSP (CTSP) is a variant of the multiple TSP that uses colors to indicate the accessibility of cities to different salesmen. This makes CTSP an important combinatorial optimization problem (COP) with many practical applications. In this article, we propose a novel end-to-end method for solving CTSP. Specifically, we introduce a Mamba-like gated linear attention with normalization (MGLAN) model and leverage the REINFORCE algorithm to train diverse policies capable of efficiently exploring the optimal solutions. The proposed MGLAN replaces standard attention scaling with normalization, combined with a gating mechanism that significantly enhances training stability in the RL setting. Our model’s effectiveness is demonstrated through extensive experiments on random CTSP instances. MGLAN achieves an optimality gap of 0.27% for CTSP70 and 0.59% for CTSP100, while achieving faster inference and more stable convergence. Compared to the current state-of-the-art (SOTA) neural solver for CTSP (i.e., GEIAM), the proposed MGLAN achieves substantial performance gains, reducing the optimality gap by 86.2% and 77.1% on CTSP70 and CTSP100, respectively, while maintaining comparable inference times across all problem scales.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 5","pages":"2908-2919"},"PeriodicalIF":0.0,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147757125","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

DARKAN: Daily Activity Recognition Using Optimized Wavelet-Based Kolmogorov-Arnold Networks 基于优化小波的Kolmogorov-Arnold网络的日常活动识别

IEEE transactions on artificial intelligence Pub Date : 2026-03-01 Epub Date: 2025-11-06 DOI: 10.1109/TAI.2025.3627523

Jiawei Li;Meng Xu;Wanqing Tu;Yifeng Zeng;Zhao Huang;Mikko Valkama;Chaoyun Song

{"title":"DARKAN: Daily Activity Recognition Using Optimized Wavelet-Based Kolmogorov-Arnold Networks","authors":"Jiawei Li;Meng Xu;Wanqing Tu;Yifeng Zeng;Zhao Huang;Mikko Valkama;Chaoyun Song","doi":"10.1109/TAI.2025.3627523","DOIUrl":"https://doi.org/10.1109/TAI.2025.3627523","url":null,"abstract":"Human activity recognition (HAR) plays a crucial role in intelligent healthcare, smart environments, and elderly monitoring. Traditional deep learning-based HAR methods often function as black-box models, limiting their interpretability. The recently proposed Kolmogorov–Arnold network (KAN) utilizes explicit, mathematically defined basis functions, which clarify its operation and enhance interpretability. However, these methods still face challenges, such as slow training speed, high computational costs, and suboptimal performance. Here, we propose the daily activity recognition with optimized Wavelet-based KAN (DARKAN), a lightweight architecture that leverages wavelet decomposition to boost performance, and simplifies KAN structure to lower model parameters and computational complexity. Specifically, low- and high-frequency inertial measurement unit (IMU) signals are extracted by a wavelet transform, while time-domain features are incorporated to enrich feature representation. Subsequently, the B-Spline is replaced by the wavelet function as the activation function in KAN (wav-KAN), and the network depth of wav-KAN is reduced to two layers. Finally, the optimized wav-KAN is utilized to classify daily activities by fusing the extracted time-frequency features. Extensive experiments on three open-source datasets demonstrate that DARKAN outperforms state-of-the-art methods, achieving 98.82%, 97.11%, and 98.57% in classification accuracy, respectively while reducing the number of model parameters by 1.45<inline-formula><tex-math>$boldsymbol{times}$</tex-math></inline-formula> and FLOPs by 4<inline-formula><tex-math>$boldsymbol{times}$</tex-math></inline-formula>.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 5","pages":"2841-2857"},"PeriodicalIF":0.0,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147757129","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

ArMA: Mitigating Catastrophic Forgetting using Attention-Regularized Model Averaging in Continual Fine-tuning Large Language Models ArMA：在持续微调的大型语言模型中使用注意正则化模型平均来减轻灾难性遗忘

IEEE transactions on artificial intelligence Pub Date : 2026-03-01 Epub Date: 2025-11-10 DOI: 10.1109/TAI.2025.3630623

Xihe Qiu;Leijun Cheng;Teqi Hao;Xiaoyu Tan

{"title":"ArMA: Mitigating Catastrophic Forgetting using Attention-Regularized Model Averaging in Continual Fine-tuning Large Language Models","authors":"Xihe Qiu;Leijun Cheng;Teqi Hao;Xiaoyu Tan","doi":"10.1109/TAI.2025.3630623","DOIUrl":"https://doi.org/10.1109/TAI.2025.3630623","url":null,"abstract":"Recent advancements in continual fine-tuning (CF) have aimed to enhance the instruction-following capabilities of large language models (LLMs) within domain-specific contexts. However, these models often suffer from catastrophic forgetting, manifesting as a significant decline in performance on general domain tasks. This presents a substantial challenge for developers who seek to improve performance on a specific domain without compromising efficacy across previously established tasks. To mitigate the negative impact of fine-tuning on generalization and enable the model to adapt to new tasks while preserving its ability on general-domain tasks, we propose a novel framework, attention-regularized model averaging (<sc>ArMA</small>). This framework addresses catastrophic forgetting in the CF of LLMs through <sc>ArMA</small>. Unlike the typical model average, which utilizes various metrics to measure and balance the general capabilities, our proposed <sc>ArMA</small> framework employs a heuristic based on attention regularization to universally consider all general instruction tasks. This approach is predicated on the hypothesis that models exhibiting similar attention distributions over the input instructions will likely yield comparable outcomes. Our evaluation across multiple benchmarks demonstrates that <sc>ArMA</small> balances the tradeoffs between different task capabilities in the CF of LLMs and effectively reserves the general domain capabilities, which will potentially benefit complex industrial domain-specific applications.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 5","pages":"2920-2930"},"PeriodicalIF":0.0,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147757134","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Integrating GAN and Dynamic Identity Convolution for Enhanced Radar Image Reconstruction From Geostationary Satellite Observations 基于GAN和动态同一性卷积的地球静止卫星雷达图像增强重建

IEEE transactions on artificial intelligence Pub Date : 2026-03-01 Epub Date: 2025-11-14 DOI: 10.1109/TAI.2025.3632822

Jianwei Si;Haonan Chen;Lei Han

{"title":"Integrating GAN and Dynamic Identity Convolution for Enhanced Radar Image Reconstruction From Geostationary Satellite Observations","authors":"Jianwei Si;Haonan Chen;Lei Han","doi":"10.1109/TAI.2025.3632822","DOIUrl":"https://doi.org/10.1109/TAI.2025.3632822","url":null,"abstract":"The radar reflectivity mosaic image derived from ground-based weather radar observations is crucial for monitoring severe weather events. But radar deployment is nonexistent in desert, oceanic, and many mountainous areas, making it impossible to acquire radar data in these regions. In this study, we propose DIC-GAN, a novel generative adversarial network (GAN)-based framework integrating dynamic convolution and mixed loss function, aiming to reconstruct radar mosaic images from multichannel satellite observation images. In the generator, nine dynamic identity convolution modules (DIMs) incorporating dynamic convolution and identity connections are introduced to adaptively adjust convolutional kernel weights based on input data and effectively enhance the generator’s representation capability. Furthermore, we develop a task-driven mixed loss function that combines binary cross-entropy (BCE) loss, mean squared error (MSE) loss, and weighted threshold loss to enable the generator to prioritize strong reflectivity regions, thereby improving the generative performance of inner storm structures in these areas. Experimental results demonstrate that DIC-GAN surpasses existing comparison methods. Generalization performance tests further reflect its effectiveness in reconstructing radar images across regions without radar coverage.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 5","pages":"2967-2979"},"PeriodicalIF":0.0,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147757138","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

An Optimized Neural Network for Efficient Resource Utilization and Enhanced Accuracy in Magnetic Field Prediction 一种有效利用资源和提高磁场预测精度的优化神经网络

IEEE transactions on artificial intelligence Pub Date : 2026-03-01 Epub Date: 2024-09-19 DOI: 10.1109/TAI.2024.3462301

Xinsheng Yang;Zining Wang;Lingyue Wang;Rentian Zhang;Guizhi Xu;Qingxin Yang

{"title":"An Optimized Neural Network for Efficient Resource Utilization and Enhanced Accuracy in Magnetic Field Prediction","authors":"Xinsheng Yang;Zining Wang;Lingyue Wang;Rentian Zhang;Guizhi Xu;Qingxin Yang","doi":"10.1109/TAI.2024.3462301","DOIUrl":"https://doi.org/10.1109/TAI.2024.3462301","url":null,"abstract":"The article presents a deep learning approach which enables numerical calculation of magnetic fields in various electromagnetic devices. In comparison to the finite element analysis (FEA) method, the trained model demonstrates a significantly faster computation speed. The accuracy of representing information within the solution domain is enhanced through the use of a bitmap technique. A shifted window-based self-attention (SW-MSA) mechanism is employed to analyze device information within the solution domain. Considering the nonnegativity property of magnetic flux density, the Softplus activation function is incorporated into the neural network model, resulting in the proposed Softplus-Enhanced Swin-Unet (SESU). Moreover, magnetic field prediction is conducted for three types of electromagnetic devices: coils, transformers, and motors. Compared with the commonly used convolutional neural network (CNN) and vision transformer (ViT) models, this approach achieves a minimum of 10-fold improvement in prediction accuracy while reducing computational resource consumption by 35%. The proposed method is validated through FEA and comparative experiments.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":"7 5","pages":"3006-3017"},"PeriodicalIF":0.0,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147757145","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0