International Journal of Intelligent Systems最新文献

To address the limitations of traditional computational fluid dynamics (CFD) simulations, such as high computational cost, long processing times, and limited scalability, this study identifies the inefficiencies of existing data-driven prediction methods, which often lack spatial–temporal coordination mechanisms and fail to capture fine-grained dynamic features of UAV airflow fields. We propose a novel deep learning model, VAN-ConvLSTM, for rapid and accurate prediction of UAV downwash airflow. Unlike conventional ConvLSTM-based frameworks, which struggle with modeling long-range dependencies and detailed spatial variations, our model introduces a visual attention unit (VAN) to enhance spatiotemporal sensitivity. The model architecture combines a convolutional encoder for spatial feature extraction, a VAN module for attention-guided temporal modeling, and a ConvLSTM decoder for sequence generation. This synergistic design improves both the accuracy and interpretability of airflow prediction. Experimental results show that the VAN-ConvLSTM model achieves an SSIM score of 0.96, demonstrating high consistency with CFD simulations. Compared to baseline methods, our model reduces error while improving stability and spatial fidelity. Ablation studies further validate the individual contributions of VAN and ConvLSTM modules. The results, verified through three representative case studies, confirm that VAN-ConvLSTM outperforms state-of-the-art approaches across multiple evaluation metrics, while offering significantly enhanced computational efficiency. This demonstrates its strong potential as a reliable and scalable alternative to traditional CFD methods in rotor airflow prediction scenarios.

In e-commerce, customer reviews wield significant influence over business strategies. This study proposes a robust sentiment analysis (SA) model tailored to e-commerce recommendations. It aims to address the key limitations of existing methods, including challenges in generalizability, feature extraction, class imbalance, and hyperparameter tuning. Our process uses an autoencoder (AE) to extract key features from the input sentence. We employ DistilBERT for word embedding, which performs faster than the standard BERT model (bidirectional encoder representations from transformers). The proposed architecture integrates an AE with a transductive long short-term memory (TLSTM) unit, which is trained with a modified generative adversarial network (GAN). TLSTM leverages transductive learning to emphasize training samples that closely resemble those in the test distribution, enhancing the flexibility and predictive accuracy of the model. Within the GAN, the generator is designed to exclude gradients from dominant batch instances, encouraging greater output diversity and generalization. Once the AE is trained, its compressed feature representations are fed into a multilayer perceptron (MLP) classifier. To tackle class imbalance issues during classification, we implement a reinforcement learning (RL) mechanism. This strategy prioritizes the minority class by applying a reward mechanism to balance the classification outcomes. Moreover, we use the Bayesian optimization hyperband (BOHB) algorithm to fine-tune the hyperparameters of the model. Experimental results on the AIV, AA, and Yelp datasets demonstrate superior performance, with F-measure scores of 91.603%, 89.504%, and 90.397%, respectively. These outcomes validate the robustness of the model and its potential to significantly enhance recommendation quality in dynamic e-commerce environments.

Parkinson’s disease (PD) is a chronic neurological condition causing an assortment of motor and cognitive prodromes. Each individual’s PD symptoms develop differently due to the variability of the ailment. This study aims to introduce the KNN Imputed Spearman’s Rank and Jaccard Convolutional Deep Neural Learning (KISRJCDNL) technique for automating early PD diagnosis depending on speech analysis. This work enhances disease diagnosis performance through preprocessing and early, precise PD detection. Several information collected from the given dataset are initially taken as input. Then, the preprocessing stage converts raw data into a structured format. Afterward, Spearman’s Rank Feature Selective and Jaccard Index–based Convolutional Deep Neural Learning Classifier with four layers, one input layer, one output layer, and two hidden layers, are deployed for diagnosing PD by efficiently performing the data classification. Experimental evaluation uses the Early Biomarkers of the PD dataset by different factors. Findings support the claim that the proposed KISRJCDNL technique enhances accuracy by 14%, reducing feature selection time, error rate, overall time, and space complexity by 16%, 43%, 36%, and 22% compared to the existing deep learning methods.

Deep generative models are now capable of generating synthetic images with very high visual realism, often indistinguishable from real-world photographs. Such AI-generated images (AIGIs) can pose serious security concerns if used maliciously. Conventional AIGI detection methods are based on supervised learning and may have limited generalization ability. In this paper, we build a novel universal detector of AIGIs without the need to perform training on these images. Starting with a study on the effectiveness of various pretrained image models for the AIGI detection task, we then chose to build our detector based on the features of the popular CLIP model. Unlike existing methods, we use a small number of real images and their carefully processed counterparts as AIGI proxies during training, combined with a novel margin-based loss to promote generalization. Extensive experiments demonstrate the effectiveness of our method, outperforming existing supervised methods while not using any AIGI for training.

Time-series classification is a relevant step supporting decision-making processes in various domains, and deep neural models have shown promising performance in this respect. Despite significant advancements in deep learning, the theoretical understanding of how and why complex architectures function remains limited, prompting the need for more interpretable models. Recently, the Kolmogorov–Arnold Networks (KANs) have been proposed as a more interpretable alternative to deep learning. While KAN-related research is significantly rising, to date, the study of KAN architectures for time-series classification has been limited. In this paper, we aim to conduct a comprehensive and robust exploration of the KAN architecture for time-series classification utilizing 117 datasets from UCR benchmark archive, from multiple different domains. More specifically, we investigate (a) the transferability of reference architectures designed for regression to classification tasks, (b) the hyperparameter and implementation configurations for an architecture that best generalizes across 117 datasets, (c) the associated complexity trade-offs, and (d) KANs interpretability. Our results demonstrate that (1) the Efficient KAN outperforms MLPs in both performance and training times, showcasing its suitability for classification tasks. (2) Efficient KAN exhibits greater stability than the original KAN across grid sizes, depths, and layer configurations, especially when lower learning rates are employed. (3) KAN achieves competitive accuracy compared to state-of-the-art models such as HIVE-COTE2 and InceptionTime, while maintaining smaller architectures and faster training times, highlighting its favorable balance of performance and transparency. (4) The interpretability of the KAN model, as confirmed by SHAP analysis, reinforces its capacity for transparent decision-making.

Existing facial image inpainting methods demonstrate high reliance on the precision of prior knowledge. However, the acquisition of precise prior knowledge remains challenging, and the incorporation of predicted prior knowledge in the restoration process often leads to error propagation and accumulation, thereby compromising the reconstruction quality. To address this limitation, we propose a novel facial image inpainting framework that leverages knowledge distillation, which is specifically designed to mitigate error propagation caused by imprecise prior knowledge. More specifically, we develop a teacher network incorporating accurate facial prior information and establish a knowledge transfer mechanism between the teacher and student networks via knowledge distillation. During the training phase, the student network progressively acquires the prior information encoded in the teacher network, thus improving its restoration capability for missing or corrupted regions. Additionally, we introduce a Coordinate Attention Gated Convolution (CAG) module, which enables effective extraction of both structural and semantic features from intact regions. Experiments conducted on the public facial datasets (CelebA-HQ and FFHQ) show that our method achieves performance improvements over existing approaches in terms of multiple quantitative evaluation metrics, including PSNR, SSIM, MAE, and LPIPS. Thus, the knowledge transfer from teacher to student network via knowledge distillation significantly reduces the dependence on prior knowledge characteristic of existing methods, facilitating more precise and efficient facial image inpainting.

This study bridges a crucial gap in athletic performance analysis by introducing a novel machine learning (ML) framework that leverages integrated physiological signals (from the DB 2.0 database) towards Sport Artificial Intelligence of Things (S-AIoT). Understanding athletic performance is key to developing effective training programs and enhancing overall physical education. However, traditional methods often fall short in capturing the nuances of human movement. Our primary goal is to develop an innovative method for accurately classifying sports activities using advanced analytical techniques that consider various physiological signals. This study aims to improve classification accuracy and provide real-time analytics for sports performance. To achieve this, we employ spatial and temporal attention mechanisms to dynamically weight critical signals, enabling precise tracking of movement transitions across different sports. The model is trained on comprehensive datasets comprising respiration rate, ECG, and heart rate (HR), providing a multifaceted analysis of athletic performance. Extensive experiments validate the model, which achieves a remarkable accuracy of 90.32%. It is the first model of its kind, outperforming established models like 1D convolutional neural network (CNN), LSTM, BiLSTM, and 1D CNN-BiLSTM. The model demonstrates strong generalization ability on unseen data, proving its effectiveness in diverse scenarios, and exhibits moderate noise resilience, enhancing its practical applicability.

Time series forecasting is significant in market research and decision-making in the financial sector, but the complexity and uncertainty of financial data pose challenges to accurate forecasting. Although deep learning methods, including transformers, have significantly improved the forecasting effect, these methods still have limitations in dealing with the multiscale features of financial time series and their complex serial correlation. They fail to fully utilize the frequency domain’s multiscale features and spatial relationships. For this situation, this study proposes a time series forecasting method based on the multiscale fusion transformer for financial data, which aims to extract significant periodic patterns using frequency domain analysis effectively. Besides, the multiscale attention mechanism and graph convolution module are introduced to realize the detailed modeling of the time series simultaneously, effectively capture the spatial relationship, and obtain the correlation between different series on multiple frequency scales. In this study, experimental validation is carried out on several financial time series datasets, and the findings demonstrate that the proposed approach positively impacts predicting accuracy.

In order to block the spread of illegal stego-image and reduce the erasing traces of steganography attacks on images, this paper proposes a steganography attack network based on simulation of steganography information distribution. First, a strategy of simulating steganography noise was adopted, and the distribution of steganography noise was learned by convolutional neural network, and a small amount of noise was added to the position of the secret message accurately to complete the attack on the steganography information, while protecting the image content to the maximum extent. In addition, different image recovery modules are designed in the deep network, such as the shallow feature extraction module, progressive attention recovery module, and detail feature reconstruction module, which collectively leverage hierarchical pixel features to mitigate discrepancies between reconstructed and original images while preserving visual fidelity before and after image attacks. Through two kinds of loss functions, the deep network model continuously optimizes the network performance to achieve the minimum degree of damage to the image content and the maximum degree of recovery of the reconstructed image. Experimental results show that the proposed method is superior to other methods in erasing secret message and restoring image quality.

Lithium batteries are one class of key components in new-energy vehicles, and surface defects are easily generated during production, causing serious threats to safety. Most deep learning methods of surface defect detection heavily rely on lots of high-quality labeled samples. Unfortunately, it is very difficult and expensive to prepare defect datasets of lithium batteries in practice. To deal with this issue, this paper presents cross-object transfer learning (COTL)–based few-shot surface defect detection of lithium batteries by resort to massive defect samples of other objects. The COTL model is composed of image preprocessing, feature extraction, feature fusion, and contrastive learning-based defect detection modules. The ResNeXt-101 network is used as the backbone to enhance feature extraction capability. The path aggregation feature pyramid network (PAFPN) is used to realize multiscale feature fusion. The contrastive learning branch is added to improve the discrimination ability among different categories of region proposals under few defect samples and increase the generalization ability. Then, experiments are done to testify the proposed method, where base-class defect dataset from other objects and new-class defect dataset from soft-pack lithium batteries are adopted for training and testing. Furthermore, model comparison and ablation studies are performed. The results show that the recall rate, the AP50, the mAP, and the F1 values of the COTL model are much better than those of other existing models when only using few defect samples. In particular, when there are only 30 new-class defect samples, the above four metrics of the COTL model are already larger than 0.90. The results testify that the proposed COTL model provides a more effective solution for few-shot surface defect detection of lithium batteries.