Fault diagnosis under unseen working conditions remains a significant challenge in intelligent machinery health monitoring. Current domain generalization methods tend to overfit to domain-specific factors, limiting their robustness in unseen environments. This entanglement disrupts the class-level structure in the feature space, thereby diminishing the discriminative power of learned representations under unseen conditions. To address this limitation, we propose a novel domain generalization method for fault diagnosis, termed collaborative bi-stream hierarchical contrastive learning. First, a bi-stream factorization module separates features into task-invariant and domain-specific components. The task-invariant stream is regularized using an adversarial entropy objective to suppress domain-related cues and a reconstruction loss to preserve task-relevant information. Second, a hierarchical contrastive learning strategy enhances the discriminability of task-invariant features by jointly optimizing global prototype compactness, instance-to-prototype alignment, and domain-modulated instance-level contrast. Experimental results demonstrate that the proposed method achieves an average diagnostic accuracy of 94.4% across the tested datasets, representing an improvement of approximately 8.3% compared to the best-performing baseline, demonstrating strong generalization capability in industrial scenarios.

Early screening is essential for reducing the incidence and mortality of cervical cancer, and artificial intelligence-based analysis of whole slide images (WSIs) enables large-scale automated screening. However, existing methods often ignore image quality variations and inter-individual morphological differences, which limits their robustness in clinical settings. This study proposes a quality-aware cervical WSI classification framework that integrates image quality assessment with pathologist-inspired normal-abnormal cell comparison. A quality evaluation module filters unreliable patches, while a cell comparison and enhancement strategy enlarges the feature discrepancy between normal and abnormal cells to mitigate individual variability. Supervised contrastive learning further strengthens abnormal cell discrimination, and patch-level quality scores are incorporated into an attention-based multiple instance learning framework to guide WSI classification. Experiments on 2,434 WSIs from five medical institutions demonstrate that our method achieves superior performance in real-world scenarios, significantly outperforming state-of-the-art methods by 1.93% in average overall accuracy.

Score normalization is a key back-end technique in speaker verification for improving score comparability across trials. As a convenient and widely used normalization method, Adaptive Symmetric Normalization (AS-Norm) standardizes raw scores using mean-and-standard-deviation normalization parameters estimated from an adaptive cohort comprising the most similar impostor scores. However, since the adaptive cohort retains only the top-ranked impostor scores, these central-moment statistics may not optimally characterize the empirical distribution of these scores, resulting in suboptimal speaker verification performance. In this letter, we propose Adaptive Symmetric Extreme Value Normalization (AS-EVNorm), which treats the adaptive cohort as upper-tail samples from the impostor-score distribution and models them under extreme-value theory for more accurate normalization parameters. Experiments on VoxCeleb and CN-Celeb show that AS-EVNorm consistently reduces both EER and minDCF compared with AS-Norm across a broad range of adaptive cohort configurations, while maintaining competitive normalization time.

Reference-based super-resolution (RefSR) technology holds significant potential for enhancing the spatial resolution of remote sensing images. However, existing methods still face notable limitations in the effectiveness of large-scale texture transfer and struggle to guarantee the fidelity of re-constructed content when local variations exist between low-resolution (LR) images and reference (Ref) images. To address these problems, we propose a novel Mamba and U-Net-based method for RefSR, named Mamba-UNet. First, the LR image is spectrally decomposed into high-frequency (HF) and low-frequency (LF) components. The HF component is then enhanced by a Mamba module with long-sequence modeling capability, which can adaptively transfer textures from any location in the Ref image, thereby effectively resolving the blurring problem in large-scale texture transfer. Second, an adaptive multi-source data fusion module is designed to dynamically evaluate the land-cover consistency between the LR and Ref images through a channel attention gating mechanism. Once a local mismatch is detected, this mechanism automatically reduces the weight of Ref features and strengthens the dominance of the LR image’s own low-frequency content, thus ensuring the fidelity of content reconstruction. Finally, the gated fused features are fed into an attention-augmented U-Net for high-quality image reconstruction. Experimental results demonstrate that compared to state-of-the-art RefSR methods, the proposed method exhibits superior reconstruction performance and stronger robustness in quantitative metrics and visual quality assessments on the SECOND and CNAM-CD datasets. The code is available at https://github.com/wzh226/AM-UNET.

The advent of deep learning technologies has opened new pathways for the fault diagnosis problem in the industrial production process. However, existing fault diagnosis methods primarily depend on large volumes of high-quality labeled data, which are often difficult and costly to obtain. Moreover, these methods may yield lead to inaccurate and unstable diagnostic results. Inspired by unsupervised learning and clustering ensemble techniques, this paper explores an effective and label-independent fault diagnosis method, termed reliability evaluation-based deep fuzzy clustering ensemble (RE-DFCE). Specifically, the multi-head attention mechanism and random mask technology are introduced to extract advanced features. Subsequently, fuzzy clustering based on the extracted advanced features is employed to describe the ambiguity and uncertainty in diagnostic data. Furthermore, a reliability evaluation method based on entropy measures is proposed to quantitatively evaluate the reliability of each cluster result, reducing the impact of uncertainty and noise on the diagnostic results. Finally, to further improve the accuracy and robustness of the diagnostic model, an unsupervised cluster ensemble based on the fuzzy consistency function of the graph is introduced to integrate these reliability-evaluated results. Extensive experiments on the CWRU benchmark dataset demonstrate that RE-DFCE achieves superior performance compared to existing methods.

Long-term time series forecasting is widely applied in fields such as transportation, energy consumption, and disease control, yet existing Transformer-based methods often struggle to capture complex temporal dependencies and accurately model long-range patterns through a single branch. To address these issues, we propose a Dual-branch Aggregation Method (DAformer) for long-term time series forecasting, designed to enhance prediction accuracy through a modular framework. DAformer consists of three main components: a temporal feature extraction module, a correlation-based feature decomposition module, and a dual-branch aggregation module. The temporal feature extraction module utilizes a self-attention mechanism to capture both short-term and long-term dependencies between time points, regardless of their temporal distance, enabling effective modeling of complex temporal patterns. The correlation-based feature decomposition module decomposes time series into distinct seasonal and trend components and further analyzes key temporal features by evaluating interactions across various time intervals. Finally, the dual-branch aggregation module integrates the features extracted by the two branches, capturing feature similarities and interdependencies across multiple time series. Our method outperforms state-of-the-art (SOTA) methods in terms of predictive ability. It achieves the best forecasting performance across eight datasets. For prediction lengths of 96, 192, 336, and 720, the average MSE values are 0.258, 0.310, 0.375, and 0.499, respectively, with corresponding MAEs of 0.295, 0.334, 0.378, and 0.449.

Deep neural networks have achieved strong performance in 3D point cloud understanding, yet their decision process remains difficult to interpret. Existing explainability methods are mainly designed for images and are not well suited to irregular, sparse point clouds. In point cloud networks, Farthest Point Sampling (FPS) breaks the correspondence between deep features and original inputs, leading to spatial inconsistency. Some existing methods alter the original point cloud by removing points during explanation, which alters the input itself and undermines input consistency. We propose Hierarchical Progressive Fusion (HPF), which progressively fuses multi-level gradients and activations and projects them back to the original input space to achieve spatial consistency. Furthermore, HPF only requires a single forward and backward propagation to generate heatmaps, achieving input consistency. Moreover, HPF does not modify the point cloud network architecture as a plug-and-play tool. Experiments on ShapeNet-Part and SUN RGB-D datasets show that HPF yields faithful and effective explanations for point cloud models. Code is available at https://github.com/perfect979/HPF.

Physiological studies have shown that differences between depressed and healthy individuals are manifested in the audio and video modalities. Hence, some researchers have combined local and global information from audio or video modality to obtain the unimodal representation. Attention mechanisms or Multi-Layer Perceptrons (MLPs) are then used to complete the fusion of different representations. However, attention mechanisms or MLPs is essentially a linear aggregation manner, and lacks the ability to explore the element-wise interaction between local and global representations within and across modalities, which affects the accuracy of estimating the depression severity. To this end, we propose a Representation Interaction (RI) module, which uses the mutual linear adjustment to achieve element-wise interaction between representations. Thus, the RI module can be seen as an mutual observation of two representations, which helps to achieve complementary advantages and improve the model’s ability to characterize depression cues. Furthermore, since the interaction process generates multiple representations, we propose a Multi-representation Prediction (MP) module. This module implements multi-representation vectorization in a hierarchical manner from summarizing a single representation to aggregating multiple representations, and adopts the attention mechanism to obtain the estimation of an individual depression severity. In this way, we use the RI and MP modules to construct the Multimodal Local Global Interaction (MLGI) network. The experimental performance on AVEC 2013 and AVEC 2014 depression datasets demonstrates the effectiveness of our method.

The ThinPrep Cytologic Test (TCT) is the most widely used method for cervical cancer screening, and the sample quality directly impacts the accuracy of the diagnosis. Traditional manual evaluation methods rely on pathologist's observation under microscopes. These methods exhibit high subjectivity, high cost, long duration, and low reliability. With the development of computer-aided diagnosis (CAD), an automated quality assessment system that performs at the level of a professional pathologist is necessary. To address this need, we propose a fully automated quality assessment method for Cervical Cytopathology Whole Slide Images (WSIs) based on the Bethesda System (TBS) diagnostic standards, artificial intelligence algorithms, and the characteristics of clinical data. The method analyzes the context of WSIs to quantify quality evaluation metrics which are focused by TBS, such as staining quality, cell count and cell mass proportion through multiple models including object detection, classification and segmentation. Subsequently, the XGBoost model is used to mine the attention paid by pathologists to different quality evaluation metrics when evaluating samples, thereby obtaining a comprehensive WSI sample score calculation model. Experimental results on 402 WSIs demonstrate that the proposed evaluation method has significant advantages in terms of speed and consistency.

The mixup strategy combined with Angular Margin Loss (AML) has shown promising results in Anomalous Sound Detection (ASD). However, a comprehensive theoretical understanding of how mixup facilitates normal distribution modeling within the AML framework remains absent, which impedes further optimization and application in ASD. Through gradient analysis, this paper reveals how mixup provides explicit, additional directional guidance when optimizing the normal distribution through AML, with its mixing coefficients acting as dynamic weights for this guidance. Based on these insights, we propose a self-supervised ASD method named AML-Driven and Distance-Aware Multi-Sample Mixing (ADMM). ADMM employs a novel distance-aware learning strategy that ensures focused optimization towards the target distribution while enriching the diversity of directional guidance, particularly for harder (e.g., boundary) samples. Comprehensive experiments on real-world datasets demonstrate the effectiveness and superior performance of the proposed ADMM method, achieving new state-of-the-art results within the mixup-AML framework.

Self-supervised Anomalous Sound Detection (ASD) relies on proxy anomaly generation to train discriminative models without real anomalous samples. However, existing generation strategies largely rely on manual design and operate without principled guidance, producing proxy anomalies of unguaranteed quality. Statistical exchange partially mitigates this by preserving spectrogram structure through distribution-level perturbation, but still depends on preset dimensions and indiscriminate perturbation. This paper proposes Attention-Guided Statistics Exchange (AG-StatEx) to establish traceable guidance for principled proxy anomaly generation. To establish this guidance, four lightweight attention-based modules are designed to jointly evaluate the importance of time-frequency feature dimensions and sub-components, enabling adaptive selection of the signal-aggregated dimension and the most discriminative sub-components for targeted anomalization. Furthermore, a shrinkage mechanism progressively narrows the anomalization range during training, implicitly regularizing the model to focus on the most stable and discriminative signal structures. We also develop an attention heatmap-based visualization method to improve interpretability and propose a new metric to assess proxy anomaly quality. Extensive experiments on DCASE2022–2025 and a constructed ASD dataset demonstrate the superior effectiveness, generalization capability, and interpretability of AG-StatEx.

Fine-tuning generic ASR models with large-scale synthetic personal data can enhance the personalization of ASR models, but it introduces challenges in adapting to synthetic personal data without forgetting real knowledge, and in adapting to personal data without forgetting generic knowledge. Considering that the functionally invariant path (FIP) framework enables model adaptation while preserving prior knowledge, in this letter, we introduce FIP into synthetic-data-augmented personalized ASR models. However, the model still struggles to balance the learning of synthetic, personalized, and generic knowledge when applying FIP to train the model on all three types of data simultaneously. To decouple this learning process and further address the above two challenges, we integrate a gated parameter-isolation strategy into FIP and propose a knowledge-decoupled functionally invariant path (KDFIP) framework, which stores generic and personalized knowledge in separate modules and applies FIP to them sequentially. Specifically, KDFIP adapts the personalized module to synthetic and real personal data and the generic module to generic data. Both modules are updated along personalization-invariant paths, and their outputs are dynamically fused through a gating mechanism. With augmented synthetic data, KDFIP achieves a 29.38% relative character error rate reduction on target speakers and maintains comparable generalization performance to the unadapted ASR baseline.

Currently, many studies use Fourier amplitude spectra of speech signals to predict depression levels. However, those works often treat Fourier amplitude spectra as images or sequences to capture depression cues using convolutional neural networks or multilayer perceptrons. Therefore, they ignore the complex element composition and time-frequency attributes of Fourier spectra, which is not conducive to capturing the differences among individuals with different depression levels. For this reason, we construct a Time-Frequency Self-Embedding (TFSE) module, which not only stores the correlation relationship among real (imaginary) parts of Fourier spectra of different subjects from the time-frequency perspective, but also maintain the physical properties of data through the weight embedding process. Besides, Global Average Pooling (GAP) or linear layers are difficult to balance both temporal and frequency dimensions in the vectorization process. Therefore, we construct a Time-Frequency Tensor Vectorization (TFTV) module, which summarizes each channel along time and frequency dimensions, and then generates the vectorization result by integrating various channels. In this way, we combine TFSE and TFTV modules to form our SpectrumFormer model for predicting depression levels. Evaluation indicators on AVEC 2013 and AVEC 2014 depression databases imply the progressiveness of our model.

Anomalous Sound Detection (ASD) requires modeling a compact and discriminative normal sound distribution. Recently, angular margin loss with multiple sub-centers has been shown to be effective for ASD by extending sub-centers to capture intra-class diversity and maximizing their orthogonality to enhance model discriminability. However, existing methods do not consider that the orthogonality of intra-class sub-centers needs to be optimized based on the inherent data structure to avoid over-extension of the representation space due to overorthogonality. To address this issue, we propose a Dual Orthogonality Sub-Center Loss (DOSCL) that enforces strict orthogonality of inter-class sub-centers to improve anomaly discrimination while applying relaxed constraints on intra-class sub-centers to capture the data structure. Experiments on the DCASE2023 Challenge Task2 dataset show that DOSCL achieves 1.74% AUC and 0.62% pAUC improvements over a strong baseline, validating its effectiveness.

In recent years, the zero-shot image recognition with semantic knowledge has achieved good performance due to vision-language models. However, because of the complexity of 3D shapes, the model cannot fully use the semantic knowledge of 3D shapes, which results in low accuracy of zero-shot 3D shape recognition. To address this problem, we propose a Semanticenhanced ULIP for Zero-shot 3D Shape Recognition (SE-ULIP). This method utilizes the contrastive learning to fine-tune the text encoder in two stages, including the domain adaptation finetuning and the triplets-based text encoder fine-tuning. In the domain adaptation fine-tuning, we fine-tune the image encoder and the text encoder using the views and the Semantic Descriptive Text (SDT) of each view generated by the Visual Question Answering (VQA) model, which aims to align the view features with the semantic knowledge. In the triplets-based text encoder fine-tuning, we propose an Adaptive Conditional Adjustment Context Optimization (ACACoOp) to learn the optimal context vectors. The optimal context vectors are used as the input to fine-tune the text encoder again, which enhance SE-ULIP to understand the semantic knowledge of 3D shapes. Experiments show that our method achieves the state-of-the-art performance through the fine-tuned text encoder on three 3D backbone networks for both zero-shot and standard 3D shape recognition.

The existing cervical whole slide image classification methods ignore the influence of image quality, resulting in low classification accuracy. To address this, we propose a dual-path multiple instance learning classification method guided by image quality assessment. Specifically, a pre-trained quality assessment model assigns quality scores to patches, splitting them into high- and low-quality paths. In the high-quality path, patch features are weighted by their quality scores to emphasize reliable diagnostic regions. In the low-quality path, a key instance is selected using clustering and feature distance matching. Finally, a cross-attention module fuses features across quality levels. Our method achieves 94.64% accuracy and 91.74% AUC on a dataset of 2,434 WSIs collected from five medical centers, outperforming state-of-the-art methods.

Anomalous Sound Detection requires constructing a distribution using only normal sounds. However, collecting sufficient normal samples across diverse conditions is challenging, leading to sample imbalance within subclasses. Existing subcenter angular margin loss methods use multiple subcenters to capture intra-class diversity but still suffer from under-representation or overfitting. To address this issue, we propose Adaptive Across-Subcenter Representation Learning (AASRL). Unlike existing methods that use either a single or all subcenters, AASRL adaptively selects subcenters based on the representation quality of samples and optimizes their representation across the most relevant subcenters. This ensures efficient representation of each sample and prevents the majority subclass from dominating the representation space. Experiments on the DCASE2023 Challenge Task2 dataset and a constructed imbalanced dataset demonstrate the effectiveness of AASRL.

Overlapping sound events are ubiquitous in real-world environments, but existing end-to-end sound event detection (SED) methods still struggle to detect them effectively. A critical reason is that these methods represent overlapping events using shared and entangled frame-wise features, which degrades the feature discrimination. To solve the problem, we propose a disentangled feature learning framework to learn a category-specific representation. Specifically, we employ different projectors to learn the frame-wise features for each category. To ensure that these feature does not contain information of other categories, we maximize the common information between frame-wise features within the same category and propose a frame-wise contrastive loss. In addition, considering that the labeled data used by the proposed method is limited, we propose a semi-supervised frame-wise contrastive loss that can leverage large amounts of unlabeled data to achieve feature disentanglement. The experimental results demonstrate the effectiveness of our method.

Despite the success of self-supervised respiratory sound classification methods, they do not consider that respiratory sounds are quasi-periodic signals with repetitive patterns in successive breaths, which is vital for distinguishing respiratory sounds from non-quasi-periodic sounds like noises. Therefore, the existing methods may achieve limited improvement due to ignoring the quasi-periodic dependency. To this end, considering that the segments containing the same respiratory sound pattern should be similar in a sample, we extract the segment-wise representations and evaluate the similarity between the periodic-dependent representations via a sparse self-relation matrix. By defining a periodic consistency loss, we push the sparse self-relation matrixes of two clips of the same sample closer, encouraging a larger similarity between the representations. In this manner, the method can focus more on the respiratory sound-related quasi-periodic patterns that repeatedly recur in the periodic-dependent segments. Taking HF_Lung_V1 and COVID-19 Sounds as pre-training sets, the method exceeds the baseline by 7.67% on the ICBHI 2017 classification task.

Automatic abnormal cervical cell detection provides cytopathologists with reliable evidence, rapiddiagnosis and reduced workload. However, this task is often prone to interference factors such as impurities,microorganisms, and neutrophils. The morphological diversity of abnormal cells also increasesthe complexity of detection. Unfortunately, existing abnormal cervical cell detection methods havenot fully utilized diagnostic knowledge from cytopathologists to handle these challenges.We proposean abnormal cervical cell detection method that leverages image and textual diagnostic knowledgein The Bethesda System (TBS) to interpret cervical cytology results, boosting the performance ofthe foundation model for target detection. First, we build a TBS diagnostic knowledge base througha TDK-Bert module. In TDK-Bert, TBS diagnostic knowledge can be efficiently extracted throughBert to capture textual features, which can then be aligned and fused with image features through anattention mechanism. In addition, we propose an ODC-SPPF module to improve the Spatial PyramidPooling-Fast (SPPF) module in a YOLO-based foundation model. The ODC-SPPF employs omnidimensionaldynamic convolution to improve the ability of feature extraction, particularly for subtlecells, and alleviates the interference of complex backgrounds in Thinprep cytologic test images. Experimentalresults on a cervical cell detection dataset demonstrate that our method effectively combinesknowledge bases and foundation models for abnormal cervical cell detection and outperforms thestate-of-the-art methods.