Convolutional neural networks (CNNs) are widely used in medical image classification, yet their performance is often hindered by data imbalance, limited adaptability, and overfitting. This study aims to improve CNN robustness by introducing a reinforcement learning (RL) framework based on Deep Deterministic Policy Gradients (DDPG) to automatically optimize data augmentation parameters. We propose an RL-driven optimization framework in which a DDPG agent autonomously adjusts data augmentation parameters to improve classification performance. The framework is evaluated across several benchmark datasets, including breast cancer, brain tumor, and COVID-19 X-ray images. Seven publicly available CNN architectures and a transformer model are used for comparison. Explainable artificial intelligence (XAI) using Grad-CAM is employed to validate model attention and prediction reliability. The proposed RL-based approach enables CNNs to rapidly identify optimal augmentation parameters, yielding superior performance compared to seven state-of-the-art CNN baselines and existing RL-integrated models. The DDPG agent demonstrates significant capability in automatically adapting augmentation strategies, consistently improving classification accuracy across medical. Grad-CAM visualizations further confirm improved feature focus and enhanced prediction consistency. The RL-driven DDPG framework effectively enhances image classification performance by autonomously optimizing augmentation parameters and improving CNN adaptability. Validation on multiple datasets and architectures—including transformers—shows that RL contributes to varying degrees of performance gains, establishing the proposed method as a promising direction for robust medical image analysis. The code is accessible at https://github.com/AIPMLab/DDPG.
3D morphable models (3DMMs) reconstruct the structure and appearance of 3D objects from 2D images, particularly human faces and bodies. Conventional 3DMM pipelines combine identity and expression blendshapes with a template mesh and are controlled by parameters such as shape, texture, illumination, and camera pose. However, many 3D human reconstruction methods focus on global facial geometry and make limited use of semantic facial cues such as age, gender, and landmark structure, characterising facial boundaries, curves, dips, and wrinkles. In order to adapt these high-level facial characteristics, this paper presents SARS, a shape- and appearance- aware modular reconstruction system that extracts face and body information from a single image and reconstructs a full body 3D human model. SARS improves the accuracy of semantic surface editing to capture expression-driven facial muscle deformation and appearance changes for different genders, while maintaining identity. SARS has three transformation modules. The first module estimates high-level facial attributes, including age, gender, and facial landmarks. The second module combines 3DMM expression coefficients, displacement maps, signed distance fields, and semantic facial features in a latent space to generate an identity-aware 3D face mesh. The third module uses the Skinned Multi-Person Linear model (SMPL) to extract high-level descriptors of the full body and based on these parameters, generates a 3D body model. A mesh fusion module then integrates the reconstructed face with the body mesh. Experiments on MICC, 3DPW, and EHF indicate that SARS improves facial detail and face-body continuity while maintaining competitive full-body reconstruction accuracy. Our code is available at https://github.com/gulraizk94/SARS_Integration.
This paper investigates the fuzzy distributed reinforcement learning (RL) optimal resilient control problem for multiple unmanned surface vehicle (multi-USV) systems under denial-of-service (DoS) attacks. Firstly, the considered multi-USV systems are modeled by Takagi-Sugeno (T-S) fuzzy systems. Then, a resilient distributed formation observer is developed to estimate the leader’s states. A fuzzy optimal resilient formation control scheme is proposed by using the proposed resilient observer and game theory. Considering that the analytical solution of the optimized control is transformed into the game algebraic Riccati equations (GAREs), which are difficult to obtain analytically, a value iteration (VI) learning algorithm is designed to learn the approximate solutions of GAREs under a mild initial excitation (IE) condition. It is proved that the designed controller guarantees that multi-USV systems achieve the desired formation shape with DoS attacks. In addition, the proposed learning algorithm is demonstrated to be convergent and the persistence of excitation (PE) condition in the existing RL algorithms is relaxed. Finally, the simulation and comparison results confirm the feasibility of the proposed optimized control approach.
This paper introduces a family of FMAVT-Sort models based on a fuzzy extension of the Multi-Criteria Decision Analysis (MCDA) method MAVT, alongside a framework for Uncertainty and Distinction Quantification (UQ and DQ) of the model outputs. The models differ in approaches to evaluating functions of fuzzy numbers (FNs), employing both precise and approximate techniques, and utilize four fuzzy ranking methods: defuzzification-based approaches (Centroid Index and Integral of Means), and two methods designed for ranking dependent FNs. The models are applied to a case study on university teacher evaluation. Assessment of local (within the case study) UQ and DQ and global (across a space of independent multicriteria problems) DQ is performed using algorithms that process output values from a large number of fuzzy multicriteria sorting problems simulated via Monte Carlo techniques in a fuzzy environment. A set of UQ indices, combined with a pre-uncertainty analysis stage, is implemented in a General Algorithm for Multicriteria Sorting, designed to support the decision-making process. The results demonstrate the effectiveness of UQ indices for decision support in multicriteria sorting and reveal significant distinctions among FMAVT-Sort models, underscoring the impact of model selection on final sorting outcomes. These findings hold scientific, methodological, and practical significance.
High-precision tracking of robotic manipulators is challenged by nonlinear friction, model mismatch, external disturbances, and abrupt actuator faults. Conventional robust controllers often lump these effects into a single uncertainty, resulting in conservative gains and high control effort. This paper develops a novel fixed-time active fault-tolerant control (AFTC) framework that couples sparse physics-informed dynamics learning with fixed-time robust compensation. A physics-informed neural network with sparse regression (PINN–SR) learns the physics gap between the nominal model and the plant and identifies parsimonious, interpretable terms for friction and structured uncertainties. A fixed-time fault observer then reconstructs the remaining lumped perturbations, including actuator faults, disturbances, and residual approximation errors. Using these compensations, a novel nonsingular fixed-time sliding-mode controller with a state-dependent exponent and hyperbolic-secant gain guarantees convergence within a settling-time bound independent of initial conditions; fixed-time stability is proven via Lyapunov analysis. MATLAB/Simulink–Simscape simulations of a 3-DOF SAMSUNG FARA AT2 manipulator against five recent controllers show improved performance. In the fault-free case, convergence time is 0.276–0.322 s, and the integral absolute error is 0.0186–0.0506rad (38–42% and approximately 40% improvements over the best baseline results, respectively). Under actuator faults, the proposed method further reduces the tracking error by 41–51% while preserving smooth control inputs.
In complex industrial scenarios, due to equipment aging and changing operating conditions, rotating machinery often exhibits new fault modes. Existing deep learning-based fault diagnosis models suffer from catastrophic forgetting during incremental learning under distribution drift, leading to historical knowledge loss and overfitting to new tasks. To address this issue, we propose a Synergistic Weight Refocusing Mechanism (SWRM) for few-shot class-incremental fault diagnosis under varying operating conditions. SWRM coordinates feature adaptation, prototype construction, and metric calibration through sequential information transfer, thereby improving the stability-plasticity balance in incremental learning. Specifically, an efficient channel attention-enhanced refocusing convolutional module keeps the pretrained backbone fixed to preserve historical knowledge while introducing learnable adaptation layers to refocus features toward new faults. A gated multi-scale prototype fusion mechanism processes task-adaptive features, suppresses noise through dynamic aggregation, generates robust fault prototypes, and transfers the learned gating weights as scale priors for calibration. Finally, a scale-prior adaptive meta-calibration module utilizes the fused prototypes and inherited scale priors as adaptive constraints to align old-class classifiers and new-class prototypes into a unified metric space for classification. Experiments on the MBHM dataset demonstrate that SWRM outperforms baseline methods in diagnostic accuracy and anti-forgetting performance, and ablation studies further verify its effectiveness.
Self-supervised Heterogeneous Graph Representation Learning (HGRL) aims to extract robust node semantics from unlabeled multi-relational data. However, prevailing methods rely on representation-level alignment or single-view reconstruction, which fail to model structural semantic complementarity across multi-relation views and cannot bridge structural discrepancies via simple similarity constraints, leaving cross-relation synergies under-exploited. To address this, we propose DSSG, a framework that achieves Decoupled Structural Synergy for Cross-View Generative HGRL. Specifically, we introduce a relation-aware edge decoupling strategy to construct two structurally complementary augmented views. At its core, DSSG employs a bidirectional cross-view structural reconstruction mechanism that treats cross-view structural consistency as core self-supervised signal. A shared relation-aware encoder captures view-specific representations, while a consistency-driven decoder performs bidirectional generative tasks. This paradigm shifts the learning focus from implicit embedding alignment to explicit structural generative supervision, forcing the model to infer missing structural semantics from complementary views to learn invariant representations. Furthermore, meta-path guided reconstruction and masked attribute reconstruction are introduced as auxiliary tasks to enhance intra-view discriminability. We provide an information-theoretic justification proving that our cross-reconstruction objective maximizes the mutual information between complementary views. Extensive experiments show DSSG achieves superior or competitive performance against state-of-the-art methods, validating the effectiveness of cross-view structural generative supervision.
Unsupervised Domain Adaptation (UDA) aims to transfer knowledge from a labeled source domain to an unlabeled target domain. However, existing pseudo-label-based self-training UDA methods often struggle with data noise and prediction uncertainty during training, leading to confirmation bias caused by mislabeled samples and the under-utilization of informative target data. To address these challenges, we propose a UDA method named Memory-Augmented Robust Self-training via Dynamic Consistency (MARSD). MARSD introduces a label refinement strategy that combines a momentum-updated memory bank with a K-Nearest Neighbors (KNN) voting mechanism to generate reliable pseudo-labels. Second, a dynamic thresholding strategy is employed to guide the adaptation process, enabling the model to prioritize reliable high-potential samples and expand the selected target set progressively. Furthermore, we integrate dual-domain consistency regularization to enhance the model’s robustness and cross-domain generalization. Extensive experiments on public benchmark datasets demonstrate that MARSD achieves state-of-the-art performance, validating its effectiveness for robust unsupervised domain adaptation. Our code is available at https://github.com/Zoeyadan/MARSD.
