Vision Transformers (ViTs) have made significant strides recently, but vanilla ViT models struggle with complex scenes, particularly multi-label images and occluded objects. Humans can extract specific visual cues from complex images to guide effective classification. Inspired by the observation that class-conditional saliency-typically used as a post-hoc explanatory tool-can be converted into an ante-hoc structural prior through distillation, we propose ZoomViT, a novel approach that introduces a Categorical Attention Prior (CAP) to guide adaptive zoom for efficient vision transformers. ZoomViT is based on two key observations: (1) Humans and advanced models can intelligently ignore class-irrelevant areas and focus on semantically important regions through categorical attention prior. (2) Standard ViTs can achieve superior classification accuracy when guided by adaptive zooming into regions that align with categorical attention prior. Our approach introduces the Zoomer, a lightweight adapter with only 0.8M parameters that generates categorical attention prior-guided score maps for image regions and dynamically adjusts patch sizes accordingly. This component leverages the CAP to increase patch density in class-decisive regions before classification inference, converting a traditionally post-hoc saliency signal into an ante-hoc input prior, while using larger patches for less critical areas. The categorical attention prior-guided adaptation enhances both efficiency and accuracy, especially in complex images. Experiments show ZoomViT, based on the DeiT-S framework, achieves 83.8%(+4.0%) top-1 accuracy on ImageNet-1k, surpassing existing efficient state-of-the-art (SOTA) ViTs in accuracy and efficiency. The code is publicly available at https://github.com/chenziwenhaoshuai/ZoomViT.

Inspired by the cilia structure of the human ear, we introduce flexible piezoelectric fibers as high-performance acoustic transducers with exceptional sensitivity across audio and ultrasonic frequency ranges, operating effectively in both air and water environments for the first time. Advances in thermal-drawn multi-functional fiber technology enabled the fabrication of kilometer-long piezoelectric fibers with micro-scale diameters, seamlessly combining structural flexibility with scalable manufacturing. By incorporating silica nanoparticles into piezoelectric materials to create multi-interface porous structures through the thermal drawing process, the fibers achieve a markedly enhanced piezoelectric electret effect, resulting in improved acoustic sensitivity. Leveraging interface engineering without changing the base material, this versatile approach not only enhances fiber functionality but also offers adaptability to diverse multi-material multi-functional fiber systems. The resulting multimaterial fibers demonstrate consistent and robust acoustic sensitivity across audio and ultrasonic frequency ranges (from 50 Hz to 30 MHz), achieving signal-to-noise ratios of approximately 30 dB in air and 60 dB in water. Additionally, acoustic fiber arrays highlight their potential as planar and three-dimensional sound source locators, underscoring their versatility for cutting-edge acoustic sensing applications.

The particle size distribution (PSD) of combustion smoke aerosols is a crucial microphysical parameter for combustion diagnostics, energy efficiency evaluations, and environmental monitoring. Existing methods for online aerosol PSD sensing typically assumes spherical particles and employs optical scattering methods. However, in practical scenarios, both the size and shape of particles jointly influence the optical scattering characteristics, and neglecting either factor results in inaccurate measurements. To address this issue, we propose a multidimensional scattering angle spectrum (MDSAS) sensing technique, integrated with a classification-regression concatenated (CRC) framework, for the simultaneous evaluation of aerosol shape and PSD. The MDSAS technique constructs a two-dimensional scattering intensity matrix by synthesizing data from three distinct incident wavelengths and 17 scattering angles. Thus, it effectively captures the coupled effects of particle shape, defined by ovality, and size. The CRC framework comprises two sequential sensing tasks. Initially, a classification model utilizing a convolutional neural network categorizes particle shapes into 10 distinct types. Subsequently, a PSD regression model is individually trained for each shape using the classification results. Furthermore, a 51-channel optical sensor is validated on an aerosol integration experimental platform using four typical combustion smokes as specified in the Chinese National Standard GB 4715-2025. The experimental results indicate that the average mean absolute percentage error of PSD measurements is only 10.76%, which is noticeably lower than that of a standard scanning mobility particle sizer. The proposed method achieves noncontact, online sensing with robust noise immunity, providing a reliable technical solution for real-time combustion monitoring and aerosol characterization.

Coronary artery disease (CAD) diagnosis in real clinical practice requires multi-source evidence integration across heterogeneous modalities, yet existing AI systems remain limited in cross-modal reasoning and workflow-level deployability. To address this gap, we develop Medi-Agent, a clinically aligned multimodal CAD agent that combines LLM/VLM/RAG capabilities within a task-schedulable multi-agent workflow. The framework supports end-to-end processing from image-based report digitization and structured extraction to multi-source evidence fusion, diagnosis and risk stratification, treatment suggestion generation, and longitudinal record management. An internal feasibility evaluation is conducted on a real-world CAD-related case set from a Grade III Class A hospital, covering complete multimodal report chains. Experimental results show promising module-level and diagnostic performance: information extraction success rate 95.37%, file archiving success rate 99.15%, and medical tool-calling success rate 98.76%; overall diagnostic accuracy reaches 94.12%. Stability testing in continuous multi-turn interactions further reports 0% error rate, 0% crash rate, 0% logical confusion rate, and 100% context-maintenance rate. These findings suggest that Medi-Agent provides an interpretable and workflow-compatible pathway for CAD-oriented preliminary clinical decision support, while larger external validation remains necessary.

Fire sound detection remains underexplored despite its practical importance, primarily because widely used public audio datasets contain few fire-related events, and no dedicated public benchmark currently provides a unified data curation and evaluation protocol for this task. We address three questions: how to build a reproducible pipeline for fire sound detection, which pretrained audio backbones transfer best, and when synthetic audio helps under limited real-data conditions. Our framework combines curated real fire recordings, challenging negative samples from FSD50K for in-domain training and testing, synthetic training samples from a text-to-audio pipeline, and external evaluation on ESC-50 and AudioSet-Fire. Under a unified adaptation protocol, we compare diverse backbones and additionally evaluate LoRA for Transformer models. Results show that indomain performance saturates quickly, while AudioSet-Fire is the most discriminative benchmark for generalization. CED-base performs best, achieving an AudioSet-Fire F1 of 56.82 with 40% real data plus synthetic augmentation, and 52.78 in the full-data regime. These findings suggest that external generalization depends not only on the amount of real training data but also on the choice of backbone and the way synthetic data interacts with its representation.

Fire incidents represent a significant global disaster concern. Smoke is one of the earliest combustion products to appear, leading to the widespread application of smoke detectors based on optical scattering theory. However, owing to the frequent occurrence of false alarms, many buildings are not equipped with smoke detectors, and even when such detectors are installed, they are often deliberately disabled. To analyze the causes of false alarms in the detectors, we compared the intrinsic physical properties affecting the optical scattering characteristics of combustion smoke and interfering aerosols. Moreover, an anti-interference detection method for combustion smoke based on the combined sensing of aerosol particle size, shape, and refractive index was proposed. In this method, particle size sensitivity, shape sensitivity, and refractive index sensitivity are defined to optimize the optical path parameters of smoke detectors, thereby guiding the design of a 4-channel detector with a dual-transmitter dual-receiver. Furthermore, a deep learning network model named Size-Shape-Refractive index Driven Convolutional Neural Network (SSRD-CNN) was designed for aerosol classification tasks. To train and test the SSRD-CNN model, we measured the combustion smoke of beech, cotton, polyurethane, and n-heptane, which are specified in Chinese national standards, as well as oil fumes, dust, and water vapor using a 4-channel smoke detector. After loading the trained network model onto the main control chip, smoke detectors were installed at the BYD factory for on-site testing and long-term operation. The results show that the smoke detectors normally alarm for fire smoke, and never false alarm for interference aerosols.

In highly competitive market of Internet service platforms, identifying and retaining potential churners through customer churn prediction techniques is crucial for maintaining platform vitality. The sequences of interaction behaviors between customers and platforms are closely related to churn prediction results. However, existing methods focus only on capturing the temporal dependencies in dynamic behavior sequences while ignoring the correlations between different behaviors. Moreover, classical methods apply only to static data, while deep learning-based methods focus on dynamic data, neither leveraging the complementary information between static and dynamic data. To address these issues, we propose a multi-modal customer churn prediction model based on Transformer with multi-scale Time-Behavior attention, TBformer, which adaptively fuses static and dynamic data. Time-Behavior module can capture multi-scale temporal dependencies and behavioral correlations in behavioral time series across time and behavior dimensions. We perform behavior-independent multi-scale dynamic feature fusion through bidirectional connection paths. Furthermore, the multi-modal fusion module based on the attention mechanism adaptively controls the fusion weights of static and dynamic features to improve performance. Extensive experiments on two publicly available datasets, KKBox and KDD, and a private dataset, HOF, demonstrate that our TBformer achieves an average AUC of 91.2% (+2.47%), outperforming the state-of-the-art customer churn prediction methods.

Predicting the Remaining Useful Life (RUL) of lithium-ion batteries is essential for ensuring the reliability and safety of devices in applications like electric vehicles, mobile devices, and renewable energy storage systems. Current RUL prediction methods perform well during certain phases of battery degradation. However, their overall performance remains limited due to complex data patterns and changing statistical properties over time. Models relying on local features or standard Transformer attention often fail to capture the non-stationary nature of degradation data. To address these challenges, we introduce AutoDS, a framework designed to forecast reconstructed degradation series using De-stationary attention. AutoDS incorporates a specialized attention map within the Transformer architecture to better capture temporal variations in non-stationary degradation data. An auto-correlation module refines degradation series at the sub-series level, removing redundancies while retaining key patterns. The framework then uses a De-stationary Transformer to model global non-stationary patterns. It restores the original attention map using statistical metrics from normalization, preserving both global relevance and non-stationarity in predicted and historical data. AutoDS employs an end-to-end design to ensure accurate and efficient forecasting. Experimental results show that AutoDS consistently outperforms leading RUL prediction and time series forecasting methods across two datasets.

The identification of anomalies in multivariate time series data from Internet of Things systems is essential for intelligent operation and maintenance. However, most research focuses solely on anomaly detection task after a fault has occurred, during which time equipment or economic losses may have already occurred. To provide early warnings before anomalies manifest, researchers investigate the more challenging task of time series anomaly prediction. Owing to the weak signs of anomaly precursors, existing anomaly detection models perform poorly in anomaly prediction tasks. To address this issue, a novel Time-Frequency contrastive framework with context modeling for time series Anomaly Prediction (TFAP) is proposed. TFAP introduces a time-frequency contrastive structure that uses a Transformer network to align cross-view semantic representations based on time-frequency consistency, thereby capturing discriminative features in time series data. Different views of the subtle precursor signals exhibit distinct representation discrepancies in the learned embedding space. Furthermore, a context modeling module is proposed to learn the contextual dependencies between current and future data, thereby enhancing the sensitivity of the model to precursors. Extensive experiments on five real-world datasets demonstrate that the proposed TFAP achieves an average F1 score improvement of 8.43%, significantly outperforming other state-of-the-art methods.

Optical scattering methods are more suitable for aerosol particle size distribution (PSD) measurements in portable, real-time sensors compared with existing aerodynamics- and electromobility-based methods. However, they still face challenges in estimating information-rich PSD in the absence of both the prescient distribution models and a large amount of incident light. To address this problem, we propose an optimal optical computational scattering scanning (CSS) method to measure the PSD in a model-free manner. In CSS, the scattering information expands from a few wavelengths of the incident light to a sufficient number of wavelengths. Thus, the full-rank scanning matrix and the expanded scattered light intensity vector are rebuilt to calculate the finegrained PSD with a simple transformation. Simulation results indicate that when five incident wavelengths are used for measurement, the inversion error based on the CSS method can be controlled to 20 % both in single- and double-peak PSDs. Furthermore, a sensor is developed and tested in experiments in conjunction with our method. Experimental results show that the proposed CSS method is suitable for arbitrary aerosols with uncertain distributions using only a small number of measurement channels and hardware components.

As the problem of surface garbage pollution becomes more serious, it is necessary to improve the efficiency of garbage inspection and picking rather than traditional manual methods. Due to lightness, unmanned aerial vehicles (UAVs) can traverse the entire water surface in a short time through their flight field of view. In addition, unmanned surface vessels (USVs) can provide battery replacement and pick up garbage. In this paper, we innovatively establish a system framework for the collaboration between UAV and USVs, and develop an automatic water cleaning strategy. First, on the basis of the partition principle, we propose a collaborative coverage path algorithm based on UAV off-site takeoff and landing to achieve global inspection. Second, we design a task scheduling and assignment algorithm for USVs to balance the garbage loads based on the particle swarm optimization algorithm. Finally, based on the swarm intelligence algorithm, we also design an autonomous obstacle avoidance path planning algorithm for USVs to realize autonomous navigation and collaborative cleaning. The system can simultaneously perform inspection and clearance tasks under certain constraints. The simulation results show that the proposed algorithms have higher generality and flexibility while effectively improving computational efficiency and reducing actual cleaning costs compared with other schemes.

In response to improve the measurement accuracy of Multi-Parameter Aerosol Scattering Sensor (MPASS) in detecting fire smoke concentrations, this study proposes a method using dual-wavelength detection technology. Simulation results prove that the proposed method can effectively improve the measurement accuracy. Based on this, the effect of polarization angle on the measurement accuracy is discussed and the optimal polarization angle is selected. This method helps to reduce the false alarm rate of fire detectors.

Deep-learning-based fault diagnosis models perform well when the training and test sets have the same label set. However, these models are invalid in practical applications because they misclassify any unknown faults into existing known classes. An effective diagnosis model for practical industrial applications requires the ability to detect unknown faults as well as maintain high classification accuracy on known faults. To address this challenge, this article proposes a generic open-set classification method for vibration signals. We propose a variational encoder-classifier structure to extract the robust latent features that have different specific distributions with respect to their classes. According to the distances between the latent feature distributions, the samples from unknown faults are rejected using extreme value theory (EVT) and empirical threshold. In addition, we devised an EVT-based instance-level regularization weight function to allow the model to enhance the regularization on the samples that around the known and unknown decision boundaries, which can reduce the risk of bias in the empirical threshold setting caused by the hard training samples. Experimental results on five public rotating machinery vibration datasets reveal that the proposed method achieves the best performance for each dataset. This demonstrates the effectiveness and superiority of the proposed method for practical application scenarios.

Recently, the application of vibration signals to the fault diagnosis of industrial equipment has attracted increased attention. Open set recognition allows deep networks to detect unknown faults while maintaining a high-classification accuracy for known faults, which is necessary for practical industrial applications. Existing distance- and probability-based open set recognition methods show significant potential in handling this challenging task. In this study, we combine a deep backbone network and probabilistic model to construct a class-relevant feature density estimator (CRFDE) that integrates the advantages of distance- and probability-based methods. Specifically, a loss function is developed to enable the backbone network to extract the class-relevant features and probabilistic model to estimate them, where the employed probabilistic model is improved from variational auto-encoder (VAE) to class-conditional VAE (CCVAE) that can estimate the class-conditional data densities. The output class-relevant feature densities of CCVAE are treated as special distances to determine the class belongingness of the input samples. We validated the proposed CRFDE on two public motor bearing datasets and one laboratory gas-insulated switchgear vibration signal dataset. The results revealed that the proposed CRFDE achieved higher F1 scores than the existing methods.

Emissions from thermal power plants have always been the central consideration for environmental protection. Existing optical sensors in thermal power plants usually measure the total mass concentration of the particulate matter (PM) by a single-wavelength laser, bearing intrinsic errors owing to the variation in particle size distribution (PSD). However, the total mass concentration alone cannot characterize all the harmful effects of the air pollution caused by the power plant. Therefore, it is necessary to measure the mass concentration and PSD simultaneously, based on which we can obtain multi-particle-size channel mass concentration. To achieve this, we designed an optical sensor based on the three-wavelength technique and tested its performance in a practical environment. Results showed that the prototype cannot only correctly measure the mass concentration of the emitted PM but also determine the mean diameter and standard deviation of the PSDs. Hence, the mass concentrations of PM10, PM2.5, and PM1 are calculated, and the air pollutants emission by a thermal power plant can be estimated comprehensively.

Acoustic tomography utilizes sensor arrays to collect sound wave signals, enabling non-contact measurement of physical parameters within an area of interest. Compared to optical technologies, acoustic tomography offers the advantages of low cost, low maintenance, and easy installation. Current research in acoustic tomography mainly focuses on reconstruction algorithms for temperature fields, while monitoring the composition and concentration of gases is significant for ensuring safety and improving efficiency, such as in scenarios like boiler furnaces and aviation engine nozzles. In excitable gases, the speed of sound exhibits an S-shaped curve that changes with frequency, a characteristic that could be potentially useful for acoustic tomography. Therefore, this study primarily discusses the quantitative calculation of gas concentration and temperature based on the dispersion of the speed of sound. By employing graphic processing and pattern matching methods, a coupled relationship of the dispersion of the speed of sound with gas concentration and temperature is established. The projection intersection method is used to calculate the concentration and temperature of binary and ternary gas mixtures. Combined with the inversion method, a joint reconstruction method for gas concentration fields and temperature fields based on the dispersion of the speed of sound is developed. The feasibility of the proposed simultaneous reconstruction method for temperature and concentration fields is validated using numerical simulations. Additionally, an acoustic tomography experimental system was set up to conduct reconstruction experiments for binary gas concentration fields and temperature fields, confirming the effectiveness of the proposed method.

Click-through rate (CTR) prediction plays a crucial role in online services and applications, such as online shopping and advertising. The performance of CTR prediction can have a direct impact on user experience and the revenue of the online platforms. For CTR prediction models, self-attention-based methods have been widely applied to this field. Recent works generally adopted the Transformer architecture, where the self-attention mechanism can capture the global dependencies of the user’s historical interactions and predict the next item. Despite the effectiveness of self-attention methods in modeling sequential user behaviors, most sequential recommenders hardly exploit feature interaction techniques to extract high-order feature combinations. In this paper, we propose a Feature-Interaction-Enhanced Sequence Model (FESeq), which integrates feature interaction and the sequential recommendation model in a cascading structure. Specifically, the interacting layer in FESeq is an automatic feature engineering step for the Transformer model. Then, we add a linear time interval embedding layer and a positional embedding layer to the Transformer in the sequence-refiner layer to learn both the time intervals and the position information in the user’s sequence behaviors. We also design an attention-based sequence pooling layer that can model the relevance of the user’s historical behaviors and the target item representation through scaled bilinear attention. Our experiments show that the proposed method beats all the baselines on both public and industrial datasets.

Social media platforms have seen an influx of multimodal data, leading to heightened attention on image-text multimodal classification. Existing methods for multimodal classification primarily focus on multimodal fusion from different modalities. However, owing to the heterogeneity and high-dimensionality of multimodal data, the fusion process frequently introduces redundant information and noise limiting the accuracy and generalization. To resolve the limitation, we propose a Compact Bilinear pooling and Multi-Loss network (CBMLNet). Compact bilinear pooling is used for feature fusion to learn low-dimensional and expressive multimodal representations efficiently. Furthermore, a multi-loss function is proposed to import the specific information carried by each single modality. Therefore, CBMLNet simultaneously considers the correlation between multimodality and the specificity of single modality for image-text classification. We evaluate the proposed CBMLNet on two publicly available datasets, Twitter-15 and Twitter-17, and on a private dataset, AIFUN. CBMLNet is compared with the advanced methods such as multimodal BERT with Max Pooling, Multi-Interactive Memory Network, Multi-level Multi-modal Cross-attention Network, Image-Text Correlation model (ITC), Target-oriented multimodal BERT and multimodal hierarchical attention model (MHA). Experimental results demonstrate that CBMLNet averagely improves F1_score by 0.28