Accurate and robust fault diagnosis of aero-engine systems is essential for ensuring safety and reliability in complex operational environments. However, conventional deep learning approaches often fail to effectively capture both the spatial correlations among multi-sensor signals and their temporal evolution characteristics. To address these limitations, this paper proposes a novel Integrative Graph-LSTM Network (IGL-Net) for intelligent aero-engine fault diagnosis. The proposed network comprises two complementary modules: a graph-based spatial representation module that dynamically constructs inter-sensor relationships to extract spatial dependencies, and an attentionenhanced LSTM module that adaptively models temporal dynamics under varying operating conditions. Experimental evaluations on the DIRG and HIT datasets demonstrate that IGLNet achieves superior classification accuracy and generalization capability compared with conventional deep learning models, validating its effectiveness in complex mechanical system diagnostics.

Surface mount technology (SMT) is a core process in consumer electronics manufacturing, and its soldering quality directly affects product performance, stability, and service life. In post-reflow automated optical inspection (AOI) scenarios, complex and varying backgrounds, imbalanced defect category distributions, and structural interference caused by unmounted regions pose significant challenges to the accuracy of industrial online inspection. To address these issues, this paper proposes a structurally decoupled multi-stage cascaded solder joint defect detection framework for consumer electronics SMT production lines. The framework decomposes the detection task into three stages, namely feature-region segmentation, anomaly localization, and few-shot defect classification, so as to separately handle three subtasks with different statistical characteristics: structural recognition, anomaly screening, and category discrimination. Specifically, a lightweight context-aware semantic segmentation network is first constructed to accurately extract key structural regions. Then, image registration and multi-feature template matching are combined to achieve precise localization of candidate anomalous regions. Finally, a few-shot classification network that integrates physical information and metric learning is designed to improve the discrimination of different defect categories under limited defect samples. Experimental results show that the proposed method achieves a classification accuracy of 96.20%, a defect error rate of 0.14%, and an average inference time of 0.25 s per board. These results demonstrate that the proposed method can effectively reduce false alarms while satisfying industrial real-time requirements, and it exhibits strong stability, interpretability, and engineering deployment potential.

In recent years, Vision Transformers (ViTs) have demonstrated remarkable advantages in computer vision. However, their quadratic computational complexity limits applications in dense prediction tasks. Although efficient Transformers based on local attention mechanisms reduce computational costs, they sacrifice the critical global information interaction capability. To address this, we propose a plug-and-play Clustering-based Sparse Global Attention (CSGA) module, which intelligently integrates k-means clustering with attention mechanisms to achieve global interaction while maintaining local computational efficiency. The CSGA module offers three key advantages: (1) plug-and-play compatibility, (2) linear computational complexity, and (3) significant performance gains. Extensive experiments on ImageNet-1K classification, ADE20K, and Cityscapes segmentation, COCO object detection demonstrate that CSGA consistently improves performance across five mainstream architectures (SegFormer, PVT, Swin, NAT, and Twins-SVT). For classification task, SegFormer+CSGA achieves a +1.17 % top-1 accuracy gain on ImageNet-1K. For segmentation tasks, NAT+CSGA achieves a +2.55 % mIoU improvement on ADE20K and delivers stable improvements on Cityscapes. For object detection task, NAT+CSGA improves AP by +1.7 % on COCO. These results validate CSGA’s effectiveness in balancing computational efficiency and global interaction, offering a new direction for optimizing Vision Transformers. Code available: https://github.com/Jeff1996/CSGA.

In electronics manufacturing, surface mount technology (SMT) defect detection, encompassing component anomalies in pick-and-place machines and printed circuit board (PCB) defects identified by automated optical inspection (AOI), is critical for yield and quality control. However, deep segmentation models require massive annotated datasets, which are costly to acquire for rare production defects. This paper proposes a training-free multimodal information co-refinement framework for few-shot electronics manufacturing defect segmentation, built upon the frozen Segment Anything Model 3 (SAM3). Unlike unimodal approaches that rely solely on visual information or generic text queries, the proposed method introduces a multimodal complementary disambiguation mechanism. Multi-depth knowledge-guided prototype retrieval with morphological refinement localizes fine-grained defects at sub-patch resolution, while a knowledge-constrained information compressor distills manufacturing standards into SAM3-compatible structured information queries. The spatial query and distilled concept query jointly guide the decoder, where spatial information suppresses background interference and semantic information resolves visual ambiguity. Experiments on public benchmarks and the SMT-Inspection dataset demonstrate competitive segmentation accuracy, with mIoU margins of 1.5% and 1.8% over the strongest baseline under 4-shot settings on the SMT-Inspection and VisA datasets, respectively, along with effective cross-domain generalization under few-shot settings, offering a knowledge-driven approach for electronics manufacturing quality assurance.

Generalist industrial anomaly detection aims to train a unified model on auxiliary domains and directly apply it to anomalous defect detection in unseen target domains. Existing methods mainly rely on patch-level residual features between query samples and normal support samples to improve model generalization. However, such residual features are essentially derived from local correspondences and, therefore, lack sufficient contextual information. On the other hand, directly introducing contextual information may interfere with the original anomaly representation and consequently cause response diffusion. To address this issue, this work establishes a connection between the residual representation process and the PID controller, and explicitly decomposes anomaly residual representations into proportional, integral, and derivative components. Specifically, the proportional branch adopts the original residual feature to preserve the most direct anomalous deviation signal. The integral branch models the low-frequency context of the query feature to provide stable scene-level constraints for local residuals. The derivative branch captures the local variation trend of the residual through high-frequency residual details, thereby suppressing the overshoot effect caused by contextual injection. On this basis, a controlled PID fusion mechanism is designed, in which residuals serve as the dominant cue, contextual information acts as modulation, and high-frequency residuals provide controlled correction, enabling adaptive regulation of both anomaly response intensity and spatial extent. Extensive experiments on multiple public datasets and a real-world printed circuit board defect dataset demonstrate that the proposed method outperforms existing state-of-the-art approaches in cross-domain generalist anomaly detection tasks.

Automatic optical inspection (AOI) systems are crucial for quality control in surface mount technology (SMT). Their detection accuracy directly affects the yield and reliability of high-density printed circuit boards (PCBs). However, existing methods still face typical industrial challenges, such as high false negative rates and inconsistent responses to similar components when detecting miniature elements. This article addresses the specific requirements for PCB component detection in industrial AOI, particularly in scenarios involving small targets and densely parallel distributions. We propose an adaptive integration and optimization of existing frequency-domain and coordinate-aware methods, introducing a frequency-coordinate collaborative network (FC-YOLO). The model features a dual-branch detection framework, which processes image information in parallel across two dimensions. The coordinate branch uses the coordinate prior information embedding (CPIE) mechanism for high-frequency coordinate encoding, enhancing the model's spatial invariance and fine detail perception, thereby improving the consistency of responses for parallel components. The frequency-domain branch introduces frequency-domain features and designs the Fourier pixel difference convolution (FPDConv) module, employing learnable differential convolution kernels to adaptively suppress frequency-domain noise and amplify key frequency components, thus improving detection performance for small targets. The features from both branches are efficiently fused via the frequency-coordinate feature fusion (FCFF) module, and their synergistic effect enhances the detection accuracy and response consistency of miniature components. Experimental results on the public PCBWACV dataset and the self-constructed EAPCB dataset show that the proposed method achieves a 4.5% improvement in mAP50 and a 4.1% improvement in mAP50-95 compared to YOLO11. After deployment on an enhanced automated optical inspection (EAOI) system with an RTX 4060Ti GPU, the system achieves real-time detection at 206 FPS. It outperforms existing methods in both false positive and false negative rates, showcasing its potential for engineering applications.

Single image super-resolution (SISR) seeks to reconstruct high-resolution images from low-resolution inputs under the tight latency and memory budgets of edge devices. Recent visual state space models enable linear-time sequence modeling, but they often approximate two-dimensional dependencies through multiple directional scans, which increases computation and memory. We propose a Multi-View Single-Scan Visual State Space Network (MVSSN) for efficient image super-resolution. MVSSN contains three main components. A shuffled input stacking module (SISM) organizes replicated and shuffled input channels before a lightweight projection, where the learned projection filters help produce less redundant shallow responses. A multi-view single-scan block (MSSB) uses an alternating scan axis and invertible geometric transforms to change the serialization order observed by one selective scan in each block. Across stacked blocks, this provides a lightweight cross-layer approximation to multi-view context modeling. A multi-scale local feature block (MLFB) complements global aggregation with depthwise convolutions of complementary receptive fields and a compact MLP to restore local details. Experiments on standard SISR benchmarks show that MVSSN achieves competitive or better PSNR and SSIM with fewer than one million parameters and low FLOPs. Additional ablations, RealSR evaluations, and downstream detection and segmentation studies further examine its efficiency and practical behavior.We also discuss the limitation under unknown real degradation, where bicubic-trained models may still suffer from domain gaps.

Surface mount technology (SMT) has become a cornerstone in manufacturing printed circuit boards (PCBs) for diverse consumer electronics, where reliable and stable processes are critical for product quality. Within SMT production lines, pick-and-place (P&P) machines demand accurate measurement of chip geometric parameters to ensure the final quality of consumer electronics. However, conventional visual inspection algorithms exhibit degraded performance under varying illumination conditions and diverse package geometries. To address these challenges, we propose a vision-based measurement system tailored for P&P machines in consumer electronics manufacturing. The system employs a dual-stream encoder-decoder architecture that integrates convolutional neural network (CNN) and Swin Transformer to improve measurement accuracy. To address the inherent semantic misalignment between local texture details and global topological structures in hybrid architectures, we incorporate a multi-level feature fusion module (FFM). This module facilitates adaptive cross-stream interaction to align CNN and Swin Transformer representations. Furthermore, addressing the need for metrological precision, we design a boundary guidance module (BGM) that injects explicit, deterministic edge priors directly extracted from raw images to preserve fine-grained high-frequency industrial details. The system accurately extracts key chip geometric parameters essential for precise chip placement in SMT processes. We validate performance on public and custom datasets, demonstrating the effectiveness of the proposed approach.

Industrial printed circuit board (PCB) defect detection plays a vital role in ensuring the quality and reliability of electronic products. However, accurately identifying small defects in high-resolution PCB images remains a highly challenging task. To address this challenge, we propose MSAD-Net, a small-defect detection framework designed for PCB inspection. By integrating mask-driven learning with spatially adaptive downsampling, MSAD-Net significantly enhances the saliency and detection reliability of small targets. First, a mask-Gaussian pixel-level auxiliary supervision mechanism is introduced, aligning intermediate-layer target masks with anisotropic Gaussian distributions derived from bounding-box annotations, thereby improving the geometric representation accuracy of tiny objects. Second, a mask-guided attention module (MGAM) is developed to suppress background interference and substantially boost the foreground signal-to-noise ratio, further strengthening the spatial saliency of small targets. Finally, a spatially adaptive downsampling module (SADM) is constructed, which employs content-aware strategies to preserve semantic and edge details while reducing feature-map resolution. Extensive experiments on two public PCB defect datasets and a self-built PCBA component dataset demonstrate that MSAD-Net significantly outperforms mainstream detection models in small-defect detection. Moreover, it maintains a lightweight architecture while achieving efficient inference performance, validating its potential for high-precision PCB defect inspection.

Lightweight super-resolution models have made notable progress in reducing computational complexity, yet they often struggle to recover fine structures and high-frequency details under constrained capacity. To address this challenge, we propose FASMNet-a compact and efficient super-resolution network that incorporates implicit frequency-aware representation learning and structural feature modulation. Specifically, FASMNet comprises three synergistic modules: (1) a Dynamic Frequency-aware Channel Modulation (DFCM) module that adaptively splits features into high-, mid-, and low-frequency branches based on channel activation strength; (2) a Multiscale Frequency Fusion (MFF) module that aggregates diverse receptive fields to enhance structural reconstruction; and (3) a Channel Distillation Gate (CDG) that suppresses redundant activations while preserving informative responses. Extensive experiments on natural and biomedical microscopy image datasets demonstrate that FASMNet achieves competitive or superior reconstruction quality compared to recent state-of-the-art methods, with only 0.133M parameters and 7G FLOPs. These results highlight the model's strong generalization capability and practical value for real-time and resource-constrained image enhancement tasks.

The structured illumination microscopy (SIM) technique, when applied under low photon efficiency, provides an effective solution for rapid live-cell imaging, thereby enabling the investigation of dynamic cellular processes. However, noise interference during the acquisition process significantly hinders the reconstruction of SIM images, leading to substantial artifacts. To address this challenge, we propose a zero-shot learning-based SIM image denoising method (ZS-SIM). This approach relies solely on a single acquisition of noisy SIM data and achieves accurate denoising through neural network training. The original SIM image stack is downsampled and interpolated to complete the resampling process, while the traditional Wiener-SIM reconstruction method is integrated to ensure physical fidelity. We introduce a symmetric reconstruction loss and a mutual constraint SSIM loss that jointly enhance training stability and accelerate convergence, as demonstrated by our convergence analysis. ZS-SIM further achieves a favorable balance between denoising quality and computational efficiency, with low model complexity and fast inference speed, making it well-suited for practical deployment in microscopy workflows. Experimental results demonstrate that ZS-SIM efficiently and rapidly achieves artifact-free, high-fidelity denoising reconstruction, making it particularly well-suited for low-photon efficiency live-cell imaging and scenarios with limited computational resources. Furthermore, by extending the method to scanning electron microscopy (SEM) data, we validate the effectiveness of ZS-SIM for SEM data denoising, significantly enhancing the performance of downstream segmentation tasks. We anticipate that ZS-SIM will play a pivotal role in low-photon efficiency imaging, driving advancements in this field and providing crucial support for rapid validation in biomedical research, thereby overcoming the challenges posed by acquisition noise.

This article presents a dual-branch CNN-Mamba super-resolution (SR) framework tailored for precision chip localization in surface mount technology (SMT) pick-and-place (P&P) machines. Unlike conventional SR methods that operate as independent preprocessing modules, our approach organically integrates a lightweight convolutional neural network (CNN)-Mamba architecture with distance transformation-based template matching (DTTM) to establish a task-driven metrological pipeline. To ensure metrological fidelity, a geometric mapping mechanism is introduced to ensure that subpixel localization results in the SR domain maintain strict spatial consistency with the original camera coordinate system across various scaling factors. The proposed network employs a dual-branch design that synergistically combines CNNs to restore sharp geometric edges essential for matching, with Mamba to preserve the global periodic structural patterns inherent in SMT chips while maintaining linear computational complexity. This architecture incorporates a multiscale edge enhancement module (MEEM) designed to preserve critical geometric integrity. Furthermore, a cross-attention fusion mechanism facilitates bidirectional information exchange between branches, enabling a comprehensive feature representation. Experimental evaluations on a custom chip dataset incorporating industrial degradation characteristics demonstrate that the proposed framework significantly improves subpixel positioning accuracy and robustness compared to conventional methods. The system meets the strict real-time inference requirements, validating its suitability for high-precision SMT manufacturing.

Inindustrial vision systems, image degradation due to noise, illumination variation, and optical aberrations undermines the reliability of downstream tasks, such as defect detection and parameter measurement. This article proposes PhySISR, a physics-consistent self-supervised single-image super-resolution framework tailored for industrial applications. It integrates region-aware noise simulation, Airy-disk point spread function-based degradation modeling, and structure-aware loss design to restore fine details and edge structures without any labeled data. A lightweight network is further designed to reduce parameters and inference latency, ensuring deployment efficiency. To validate the method, we construct and release SMT-ImageSet, a real-world industrial dataset captured from surface-mount equipment under diverse imaging conditions. Experiments demonstrate that PhySISR outperforms representative supervised and self-supervised methods in structural recovery, edge clarity, and downstream tasks, such as binarization and parameter extraction, showing strong applicability for practical industrial image enhancement.

In order to solve the common occlusion and complex nonlinear motion interference problems in drone target tracking, this paper proposes an improved algorithm based on YOLOv12 and BoT-SORT. This paper systematically improves the target tracking algorithm, introduces altitude state information as a potential clue, and introduces the target’s motion information into the correlation matrix for motion modeling, significantly optimizing the prediction of the target’s motion trend. To address the challenge of trajectory interruption caused by target occlusion, virtual observation interpolation is used to effectively restore trajectory continuity during occlusion, and for short-term occlusion, historical observations are used to reduce the risk of target identity switching (ID Switch). The performance of this method is verified on the MOT20, Dance Track and Visdrone2019 datasets. Experimental results demonstrate that the proposed method outperforms the original BoT-SORT across all three datasets, with particularly significant gains on Dance Track. Without ReID integration, HOTA and IDF1 improve by 1.1 and 0.9 percentage points, respectively. Integrating the ReID module further increases these gains to 2.6 and 2.2 percentage points.

Structured illumination microscopy (SIM) has been widely adopted for live-cell super-resolution imaging. However, under low-photon conditions, severe image noise can lead to prominent artifacts and the loss of fine structural details. Existing self-supervised denoising methods often rely on masking or downsampling operations, which make it difficult to simultaneously suppress noise and preserve high-frequency features. To address this limitation, we propose NDR-SIM, a self-supervised denoising framework tailored for SIM images. By introducing a physically inspired strategy of noise decomposition and recorruption, NDR-SIM constructs training image pairs without requiring clean ground truth labels. The method leverages the Poisson-Gaussian noise characteristics inherent to fluorescence microscopy to generate a pair of statistically complementary noisy views from raw SIM images, and utilizes the high-fidelity reconstruction algorithm HiFi-SIM to provide effective supervisory signals. Experimental results on both synthetic and real biological datasets demonstrate that NDR-SIM achieves superior robustness and structural fidelity across various noise types and signal-to-noise ratio levels, significantly outperforming existing classical and deep learning-based denoising methods. Notably, NDR-SIM restores super-resolved details without relying on clean images, showing strong generalization ability and broad applicability to biological imaging tasks that demand both high spatial resolution and high reconstruction fidelity.

Accurate pose estimation of surface mount components (SMCs) is essential for surface mount equipment (SME) in printed circuit board (PCB) assembly. The continuing trend toward higher component density and integration in PCB products places stringent demands on both estimation throughput and accuracy. Although surface-mount components (SMCs) are rigid targets, the images captured during manufacturing are susceptible to illumination variations and background interference, which can lead to translational and rotational measurement errors when conventional template matching algorithms are employed. Moreover, component outline dimensions are typically stored in databases as parametric descriptions rather than sample images suitable for feature extraction. This further increases the complexity of pose estimation. To address these challenges, we propose a modular template generation scheme that converts component dimensional parameters into templates. A hybrid similarity measure is developed to fuse grayscale and gradient features in a parallel framework, enabling efficient evaluation of candidate poses and reducing translational measurement error. To improve real-time performance, we introduce a edge-aware candidate region reduction method that exploits the imaging characteristics of SME. For box-type components, the template sampling strategy was optimized through the use of an anti-tilt template, effectively reducing rotational measurement error. In addition, a subpixel precision compensation module is incorporated to enhance the reliability of the measurement. In the box-type component multi-illumination dataset experiment, the translational measurement uncertainty of our proposed algorithm reached 0.00699 pixels, and the rotational measurement uncertainty reached 0.00475°. The code and dataset are available at: https://github.com/Dazeham/CompMatch.

In semiconductor automation lines, vision-based measurement for integrated circuit text recognition (ICTR) is important for product traceability and online inspection. Most existing methods emphasize accuracy over efficiency, making real-time use difficult under complex industrial conditions, while the lack of public datasets makes fair comparisons across methods challenging. To address these issues, we present an efficient vision-based measurement method. We first establish a standardized pipeline for data acquisition and processing and construct a dedicated dataset, ChipText. We then develop a lightweight end-to-end network, termed Fast-ICTR. Fast-ICTR uses a hierarchical visual encoding architecture and a spatial linear projection module (SLP) to efficiently model global contextual information under the structural constraints of integrated circuit text while avoiding the high computational cost of self-attention. On ChipText, Fast-ICTR achieves a measurement accuracy of 91.7% and an inference latency of 11.7 ms. Tests on a real industrial platform further demonstrate strong real-time capability and deployment feasibility. These results indicate that the proposed method provides a practical solution for accurate and efficient ICTR measurement. The ChipText dataset is available at https://github.com/kunkun-qi/ChipText.

Willems' fundamental lemma provides a new paradigm for data-driven system modeling and control by enabling trajectory prediction through learning of system behavior. To mitigate the influence of measurement noises, existing data-driven control approaches mainly incorporate regularization terms into the optimization problem. Although effective to some extent, their efficacy is fundamentally limited, as they do not explicitly address the probabilistic nature of noise nor do they systematically quantify inherent modeling uncertainties. This study proposes a data-driven control method based on the expectation-maximization (EM) framework. This method first establishes a probabilistic model for the offline data and performs data denoising based on a batch EM algorithm. Subsequently, a recursive EM algorithm, integrated within a receding horizon optimization framework, is used to solve the data-driven implicit model. By introducing hidden variables into the probabilistic model, the proposed method effectively addresses optimization problems involving noisy measurements in a unified probabilistic framework. Validation results from numerical simulations, the public NASA JT9D aero-engine system, and a hardware-in-the-loop platform for a two-spool turbofan engine show that, compared with the benchmark subspace predictive control method, the proposed method reduces the root-mean-square error (RMSE) by 30$\%$, 28$\%$, and 30$\%$ and increases the coefficient of determination ($\mathrm{R^{2}}$) by 45$\%$, 7$\%$, and 6$\%$, respectively, thus achieving a significant improvement in control performance.

In industrial vision pipelines, automated inspection and precise analysis demand high-fidelity image restoration, while single-image super-resolution (SR) must recover fine structural details under strict computational constraints. Existing convolutional and Transformer-based approaches often struggle to balance structural fidelity and efficiency, while recent directional-scanning state-space model (SSM) architectures still rely on fixed scan paths, offering limited local structural adaptivity. To address this, we propose SAMambaSR, an efficient SSM-based SR framework that introduces tokenwise, correlation-guided readout modulation under fixed scan paths. Specifically, we design a spatially adaptive scanning blockSASB that augments Mamba-style selective scanning with a spatial-correlation-guided bias on the state readout, and integrate a lightweight local enhancement module for 2-D local detail refinement. These components are organized into a hierarchical feature refinement group for progressive reconstruction within a linear-complexity pipeline. Experiments on standard SR benchmarks show that SAMambaSR achieves competitive or better peak signal-to-noise ratio/structural similarity index with fewer parameters and lower FLOPs than representative CNN-, Transformer-, and SSM-based SR models. In addition, downstream evaluations on industrial inspection tasks under controlled settings show that SAMambaSR can improve segmentation accuracy as a plug-in SR preprocessing module, supporting its practical potential in industrial inspection scenarios.

In surface mount technology, solder paste printing quality critically affects product reliability. Accurate segmentation of solder paste regions is essential for defect detection, quantitative analysis, and process optimization. Traditional threshold-based methods lack robustness under varying surface textures and lighting, while deep learning approaches require large annotated datasets and expensive hardware, limiting their use in cost-sensitive manufacturing. We propose a fast, annotation-free segmentation framework based on parameteric multimodal learning, integrating RGB color with 3-D height data. Height priors generate an initial mask, followed by a lookup table-based parameteric color model that adapts to different printed circuit board types and batches. A convolutional feature fusion operator then constructs a joint height-color probability space, suppressing interference from substrate variations and uneven illumination, yielding a refined probability map for final segmentation. Tests on a 3D-solder paste inspection industrial dataset achieve 96.0% mean intersection over union and 98.8% pixel accuracy, matching state-of-the-art deep learning performance while greatly improving efficiency and suitability for real-world deployment without annotated data.