9 December 2025 in Neural Networks
9 October 2025 in IEEE Transactions on Instrumentation and Measurement
Deep learning methods have demonstrated impressive effectiveness in diagnosing bearing faults. Nevertheless, their performance tends to decline significantly in the presence of long-tailed distributions, where certain fault types are underrepresented, which is common in industrial environments where fault occurrences are rare. To address the challenges posed by long-tailed data in fault diagnosis, this paper proposes the Probabilistic Contrastive Quadratic Network (PCQNet), a novel framework designed to enhance the diagnostic capabilities of neural networks under class imbalance conditions. PCQNet consists of three core components: a quadratic convolutional residual network that strengthens feature representations through higher-order interactions; a Probabilistic Contrastive Loss (PCL) that models class features using the von Mises-Fisher (vMF) distribution, combining class expectations and prior probabilities to adaptively balance the contributions of each class, effectively mitigating long-tail bias; and a logit-adjusted cross-entropy loss that improves the discrimination ability for minority classes while enhancing classification fairness. The parabolic decay formula dynamically adjusts the dominance of these two loss functions during training, ensuring the model achieves both effective feature learning and accurate classification.The effectiveness of PCQNet is validated on one public dataset and two self-constructed datasets, demonstrating its superior performance in handling long-tailed data. Analysis of the attention mechanism derived from quadratic neurons further explains the excellent feature extraction capability of the quadratic network. Building on this analysis, ablation studies further verify the independent contributions and synergistic effects of each module, confirming their critical roles in improving diagnostic performance.
1 November 2024 in Physics of Fluids
Particle image velocimetry technology calculates velocity fields by capturing two consecutive frames of particle images, and it is widely used in the research of fluid mechanics, meteorology, aerospace, and other fields. Challenges arise from uncontrolled refraction and reflection of lasers in water, as well as reflections from moving objects in water, introducing noise in particle images. Changes in noise shape and its inconsistent motion direction with particles can affect the accuracy of velocity field results. A particular challenge addressed in this study is the removal of noise from particle images. Existing learning-based methods employ supervised models, often relying on synthetic datasets due to the difficulty in obtaining pairs of particle images (noisy images and corresponding noise-free images) for training, leading to a gap between training setups and real-world scenarios. In this paper, a hybrid model named dynamic partition histogram matching with PatchCore (DPHMP) is proposed. This model comprises two primary steps. Initially, it detects noise in particle images utilizing PatchCore, which establishes a memory bank for identifying noisy regions. Subsequently, noise removal is achieved through dynamic partition histogram matching. To validate the effectiveness DPHMP, a semi-synthetic dataset and a real dataset are generated containing real particle images with real noise. On the semi-synthetic dataset, the DPHMP method achieves a peak signal-to-noise ratio of 34.393 and a structural similarity index measure of 0.9722 between the denoised and real noise-free particle images, outperforming all existing methods. Moreover, on real datasets, the approach also surpasses other techniques.
21 November 2025 in Neural Networks
17 November 2025 in Neural Networks