24 August 2020 in IEEE Access
Skin cancer is a common and deadly cancer. Dermoscopy is an effective tool for the observation of abnormal skin pigmentation. However, dermoscopy images are extremely complex and present great challenges for diagnosis. Therefore, we proposed a classification method based on the ensemble of individual advantage and group decision in dermoscopy images, including the ensemble strategy of group decision, the ensemble strategy of maximizing individual advantage, and the ensemble strategy of block-integrated voting. We used generative adversarial networks (GANs) to create a balanced sample space to better train convolutional neural networks (CNNs). Through transfer learning, the pretraining CNNs were used for fine-tuning, then the effects of different CNNs on the classification of different categories of dermoscopy images were compared, and the CNNs with better classification effect were selected for the ensemble of different strategies. This study is based on the ISIC 2018 dataset and ISIC 2019 dataset. Compared with the different individual CNNs and the frameworks, the proposed ensemble strategies achieve a better improvement in the evaluation criteria.
1 April 2025 in Physics of Fluids
The complex pore structure and weak logging responses in low-porosity, low-permeability reservoirs pose challenges for traditional methods to identify reservoir fluids. Additionally, the imbalance in fluid categories leads to classification bias toward majority classes, reducing recognition performance for minority categories. To address these challenges, this paper proposes a Reservoir Fluid Identification method that combines an improved synthetic minority over-sampling technique (SMOTE) with a revisiting mobile convolutional neural network (CNN) from vision transformer (ViT) perspective (RepViT)-transformer hybrid model. Logging data is processed using a sliding window to generate sequence samples that preserve stratigraphic information while optimizing sample structure for CNN. The improved SMOTE algorithm employs a symmetric generation strategy to expand Minority Class Samples while maintaining data distribution characteristics, mitigating the class imbalance problem. The hybrid model integrates the local feature extraction of RepViT and the global dependency capturing capabilities of transformer, enhancing Classification Accuracy and Stability. Applied to logging data from the Dagang Oilfield, the method outperformed bidirectional long short-term memory (BiLSTM) and transformer models in both overall recognition performance and minority class identification, achieving 94.41% accuracy on the test set—4.92% and 10.24% higher than BiLSTM and transformer, respectively. These results demonstrate the method's effectiveness in identifying fluids in low-porosity, low-permeability reservoirs, offering a novel approach for fluid evaluation using conventional logging data.
18 March 2025 in IEEE Access
Automatic Modulation Recognition (AMR) is an electronic signal processing technology designed to automatically identify and classify the modulation type of radio signals. Existing AMR methods suffer from a significant decrease in classification accuracy when the Signal-to-Noise Ratio (SNR) degrades. To enhance the accuracy of AMR methods under low SNR conditions, this paper proposes a novel AMR method based on Channel-Enhanced Convolution and a Linear-Angular Attention Mechanism, named CCLANN (Channel-enhanced Convolutional Linear-Angular Attention Neural Network). This method first utilizes a channel-enhanced deep convolutional module to extract spatial information from the feature maps of the input signal across different channel dimensions. Subsequently, a linear-angular attention mechanism is introduced to effectively extract time-series features from the modulated signals. Experimental results on the public datasets RML2016.10a and RML2016.10b demonstrate that the proposed AMR method improves the average classification accuracy by 2.4% and 2.1%, respectively, compared to existing schemes in the SNR range of -4 dB to 4 dB, while maintaining robustness.Slight improvements in classification accuracy are also observed at SNRs above 0 dB.
1 December 2025 in Physics of Fluids
Formation resistivity logging curves serve as a critical parameter for characterizing subsurface reservoir fluid properties and their spatial distribution. Their response features effectively differentiate fluid types and provide vital support for dynamic reservoir monitoring. However, resistivity curves are affected by various interrelated factors such as formation lithology, pore structure, and fluid properties, leading to highly nonlinear behavior. Traditional deep neural networks (DNNs), which use fixed activation functions, often have difficulty accurately capturing abrupt responses that occur under complex reservoir conditions, such as sharp peaks at high-resistance thin-bed boundaries. To enhance the reconstruction accuracy of resistivity curves, this study introduces a hybrid Transformer–Kolmogorov–Arnold network (Transformer–KAN) architecture for formation resistivity reconstruction. Kolmogorov–Arnold network (KAN) is based on the Kolmogorov–Arnold representation theorem. By replacing fixed activation functions with learnable basis function compositions, KAN enables precise approximation of nonlinear, discontinuous features inherent in resistivity data. Compared to conventional deep learning methods, the proposed Transformer–KAN demonstrates significantly improved accuracy in reconstructing both deep resistivity (RD) and shallow resistivity (RS) curves. Experimental results show that the model achieves the highest R2 values of 0.9180 for RD and 0.9597 for RS, substantially outperforming standard models such as Transformer (0.8469 and 0.8806) and bidirectional long short-term memory (0.8100 and 0.8904). These findings highlight the superior fitting capability and reduced prediction error of Transformer–KAN, offering a robust data-driven foundation for analyzing subsurface fluid behavior under complex geological conditions.
1 July 2025 in Physics of Fluids
Well logging data provide critical parameters to support the characterization of subsurface fluid properties. However, the complex downhole environment, harsh operating conditions, and various uncontrollable factors often lead to missing segments in logging curves, posing challenges to accurate analysis of subsurface fluid behavior. To address these challenges, this paper proposes a multi-module fusion approach for reconstructing logging curves. First, logging curve sequences are organized into a multi-channel feature map suitable for two-dimensional convolutional neural network input, enabling the extraction of local spatial features. Next, a squeeze-and-excitation module is employed to adaptively recalibrate channel-wise weights, thereby enhancing feature representation. Then, a temporal convolutional network is introduced to capture sequential information in the logging curves, further improving the model's ability to learn long-range dependencies. Finally, a Transformer is used to perform fine-grained refinement of the deep features for each sampling point and each logging curve. Comparative experiments based on actual logging data from the Dagang oil field demonstrate that the proposed approach achieves a Root Mean Square Error (RMSE) of 1.7302 and the R2 of 0.6160; compared with a vanilla Transformer, the RMSE is reduced by 0.3818, and the R2 is increased by 0.1882; compared with a Bidirectional Long Short-Term Memory model, the RMSE is reduced by 0.3261, and the R2 is increased by 0.1584. These results fully validate the superior feature-extraction capability and strong generalization performance of the proposed model. This study provides powerful support for reconstructing missing logging curves and investigating subsurface fluid behavior under complex geological conditions.
3 February 2026 in IEEE Access
Deep learning models for clinical tabular data often struggle with high dimensionality, extreme sparsity, and the quadratic computational complexity (O(N2)) of standard attention mechanisms. To address these bottlenecks for efficient clinical risk assessment, this study proposes the Adaptive-Sparse-TabTransformer (AST2ransformer), a framework that reframes feature interaction as an adaptive topology learning task at the schema level. The AST2ransformer incorporates a Correlation Assessment Module (CAM) and a Dynamic Sparse Pattern Generator (DSPG) to selectively capture feature dependencies by constructing a probabilistic graph. Furthermore, an Optimized Attention Computation Unit (OACU) leverages block-sparse operations to significantly reduce computational complexity by limiting interactions to a bounded budget (O(k · N · d)). Extensive evaluations were conducted on a private Venous Thromboembolism (VTE) dataset and two public benchmarks: the cardiovascular disease (CVD) dataset and the high-sparsity ECHO-NOTE2NUM dataset derived from MIMIC-III. The proposed model achieves state-of-the-art AUROC scores of 0.9487 and 0.8466 on the high-dimensional VTE and ECHO-NOTE2NUM datasets, respectively. On the lower-dimensional CVD benchmark, while achieving a competitive AUROC (0.7958) comparable to Gradient Boosting Decision Trees, our model demonstrates superior sensitivity (0.7644), significantly outperforming baseline methods in identifying positive risk cases. Crucially, it reduces computational costs by approximately 32% compared to dense baselines, offering a robust engineering solution for deploying complex risk models in resource-constrained clinical environments.