Automatic surface defect rating is vital for industrial quality control but faces challenges in handling diverse defect scales and learning ordinal relationships with ambiguous boundaries. To address the inherent contradiction between scale and ordinality processing, we propose CFG-Net (Coarse-to-Fine Grouping Network), which adopts a "divide-and-conquer" strategy to decouple multi-scale rating tasks. CFG-Net uses a dual-stream grouping-gating module with lossless pixel unshuffle for initial-scale grouping and localization. Defects are dynamically dispatched to specialized branches: a low-level branch using anchor difference to amplify weak anomaly signals, and a high-level branch using deformable convolutions to fit irregular contours. Features are integrated via an adaptive ordinal fusion module guided by gating probabilities. Furthermore, we introduce a Multi-Margin N-Pair (MMNP) contrastive loss to construct an ordinal manifold space that strictly aligns with defect severity levels. Experimental results on real-world industrial datasets show that CFG-Net effectively resolves mixed-scale conflicts, achieving 97.2% group classification accuracy and 65.7% fine-grained rating accuracy. With a Mean Absolute Error (MAE) of 0.45 and Mean Squared Error (MSE) of 0.58, the proposed architecture demonstrates high precision and engineering robustness for complex, multi-scale defect rating tasks. We also develop an intelligent surface defect rating system, which can effectively balance defect detection efficiency and accuracy on the production line, avoiding extensive manual re-inspection.

Welding defects in steel structures are difficult to control, leading to fatigue cracking of the weld. This study develops an improved AlexNet model for welding defect detection and classification of steel structures. An experimental study on welding defect detection in steel plates was conducted. An optimal image preprocessing method for steel welds was proposed. And the effectiveness of this method is proved by comparative experiments. Steel structural components were produced including six types of weld defects. Based on the characteristics of weld defects in steel structures, the framework of the network model was optimized and an improved AlexNet network model was constructed. The numerical results show that the convergence accuracy of the improved AlexNet model reaches 98

Traditional domain adaptation methods often assume balanced data distributions. However, this assumption is frequently violated in real-world industrial scenarios, where normal samples predominate while fault samples are inherently scarce. Under severe class imbalance, conventional decision boundaries tend to shift toward minority fault regions. This shift leads to persistently high misclassification rates for rare fault samples. To overcome this limitation, we propose the Dynamic Maximum Triple-View Classifier Discrepancy (DMTVCD) network, which integrates a Triple-View Classifier (TVC) Architecture and a Primary-Auxiliary Fused Cooperative Loss (PAFL). Specifically, the TVC employs auxiliary binary classifiers to aggregate fine-grained fault sub-classes into a unified "Fault Super-class." This constructs a robust "normal-fault" binary boundary that effectively counteracts class imbalance. Driven by the PAFL, this boundary acts as a hierarchical geometric constraint to suppress the primary classifier's tendency to misclassify faults as normal samples, thereby enhancing feature discriminability. Furthermore, a dynamic weighting strategy is introduced to assign large initial weights. This forces the model to bypass simple decision logic dominated by the majority class, ensuring a smooth transition from global exploration to fine-grained alignment. Extensive evaluations on the CWRU and JNU datasets demonstrate that DMTVCD consistently outperforms state-of-the-art approaches under high imbalance ratios (e.g., 20:1).

The scarcity of a large amount of labeled data for adequately training of deep learning models, along with their restricted generalization capabilities, persistently hinders the real-world practical application of data-driven deep learning in few-shot fault diagnosis and transfer task fault diagnosis. This paper proposes a self-supervised Wide Kernel Time-Frequency Fusion (WTFF) contrastive learning method that leverages extensive unlabeled signals to extract discriminative time-frequency fusion features, thereby enhancing fault diagnosis performance even with a limited number of labeled samples. Moreover, the WTFF integrates a multi-layer time-frequency wide convolutional neural network (TFCNN) encoder with a novel local and global time-frequency contrastive loss (LGTFCL) to capture time frequency consistency by facilitating the alignment of time-domain and frequency-domain feature embeddings across the shallow and deep network layers. In the fine-tuning phase, time frequency features across various levels learned from transferred pretrained model are fused to extract signal characteristics that exhibit both time and frequency discrimination. The proposed method demonstrates superior diagnostic accuracy and robustness in experiments involving few-shot and transfer learning-based fault diagnosis.

Most of 3D single object trackers (SOT) in point clouds follow the two-stream multi-stage 3D Siamese or motion tracking paradigms, which process the template and search area point clouds with two parallel branches, built on supervised point cloud backbones. In this work, beyond typical 3D Siamese or motion tracking, we propose a neat and compact one-stream transformer 3D SOT paradigm from the novel perspective, termed as EasyTrack, which consists of three special designs: 1) A 3D point clouds tracking feature pre-training module is developed to exploit the masked autoencoding for learning 3D point clouds tracking representations. 2) A unified 3D tracking feature learning and fusion network is proposed to simultaneously learns target-aware 3D features, and extensively captures mutual correlation through the flexible self-attention mechanism. 3) A target location network in the dense bird's eye view (BEV) feature space is constructed for target classification and regression. Moreover, we develop an enhanced version named EasyTrack++, which designs the center points interaction (CPI) strategy to reduce the ambiguous targets caused by the noise point cloud background information. The proposed EasyTrack and EasyTrack++ set a new state-of-the-art performance (18%, 40% and 3% success gains) in KITTI, NuScenes, and Waymo while runing at 52.6fps with few parameters (1.3M). The code will be available at https://github.com/KnightApple427/Easytrack.

Deep neural networks have achieved impressive success in various applications, but they face challenges when deployed on mobile devices due to limited computational resources. One of the main reasons for this challenge is that these networks employ too many weight maps of convolution layers to generate redundant feature maps to ensure a thorough understanding of the input data, which leads to plenty of parameters and high computational costs. This paper proposes a method to exploit this redundancy efficiently and generate more lightweight convolutional neural networks. First, we employ the peak signal-to-noise ratio (PSNR) to quantitatively verify the similarity in feature maps of the convolution layer, which is also called redundancy. We also find the similarity correspondence between the feature maps and the weight maps. Second, inspired by the analysis, we propose a cost-efficient feature transform module (FTM) to generate redundant feature maps for replacing standard convolution layers and getting lightweight convolutional neural networks (CNNs). The FTM adopts a point-wise and a depth-wise convolution to obtain the prime weight maps which could extract distinguished feature maps based on the similarity correspondence. Then, these meaningful feature maps are broadcast by an efficient neighbor-1D convolution, ensuring the feature redundancy of the network information flow. Extensive experiments on image classification benchmark datasets (CIFAR-10, CIFAR-100, ImageNet) show that the FTM outperforms existing methods in terms of efficiency. In addition, if used as a drop-in replacement for standard architectures such as ResNet-50, the FTM variant could achieve a comparable score with only 21% parameters of the original layer. The code of the proposed method is available at: https://github.com/li-ju-bazhong/Efficient-feature-transform-module .

Effectively identifying surface defects in magnetic tiles has proven to be highly challenging due to limited sample availability and irrelevant background interference, which also plays a crucial role in significantly influencing the lifespan and reliability of permanent magnet motors. To address these challenges, our study draws inspiration from a comprehensive analysis of the retinal attention mechanism and proposes three guiding criteria: multi -level resolution, what to look for, and where to look at. These criteria are utilized as foundational principles to enhance the representation learning capability of designed neural network structures through the incorporation of the retinal attention mechanism. Subsequently, based on these guiding criteria, we introduce a novel convolutional retinal attention block (CRAB) to learn discriminative and robust feature representations for magnetic tile surface defect classification and detection. The proposed CRAB comprises three modules: multi -resolution module (MRM), global attention aggregation module (GAAM), and local attention aggregation module (LAAM), designed to extract discriminative and robust features by refining meaningful information and suppressing redundant ones. Comprehensive experimental results across image classification and object detection tasks demonstrate that the proposed CRAB outperforms existing methods such as SE, ECA, and CBAM, and can effectively amplify the representation power across various backbone networks, including VGG-16, GoogLeNet, ResNet-18, and ResNet-50. An evaluation on surface defect classification and detection tasks for industrial magnetic tiles further shows that CRAB achieves accuracies of 99.50% and 96.98%, respectively. These results emphasize the promising application prospects of the proposed method in detecting industrial surface defects amid expansive and inconsequential backgrounds. The code of the proposed method is available at: https://github.com/KWflyer/CRAB.

Focusing on the low accuracy and timeliness of traditional fault diagnosis methods for rolling bearings which combine massive amounts of data, a fault diagnosis method for rolling bearings based on Gramian angular field (GAF) coding technology and an improved ResNet50 model is proposed. Using the Graham angle field technology to recode the one-dimensional vibration signal into a two-dimensional feature image, using the two-dimensional feature image as the input for the model, combined with the advantages of the ResNet algorithm in image feature extraction and classification recognition, we realized automatic feature extraction and fault diagnosis, and, finally, achieved the classification of different fault types. In order to verify the effectiveness of the method, the rolling bearing data of Casey Reserve University are selected for verification, and compared with other commonly used intelligent algorithms, the results show that the proposed method has a higher classification accuracy and better timeliness than other intelligent algorithms.

In real industrial scenarios, the limited available labeled fault data and existed significant data distribution differences between the source domain and the target domain emplace challenges and obstacles for cross-domain fault diagnosis. In this letter, we proposed a novel cluster contrastive learning method (CLCO) for few-shot learning (even one-shot learning) and cross-domain fault diagnosis under complex fault modes, fault severities, and variable working conditions. The proposed CLCO combines clustering strategy and contrastive learning to learn discriminate domain invariant feature representation. Moreover, the consistency in the characteristic distribution of data is viewed as the pseudo label information captured by k-means clustering, which are subsequently embedded into strengthen contrastive loss (SCL) function of proposed CLCO model. Then, the pseudo label information is exploited to guide self-supervised contrastive pre-training for learning domain invariant features based on feature similarity comparison of positive and negative pairs. Extensive experimental results demonstrate that our proposed CLCO outperforms existing state-of-the-art methods by a large margin.

Objective To establish machine learning (ML) prediction models for prostate cancer (PCa) using transrectal ultrasound videos and multi-parametric magnetic resonance imaging (mpMRI) and compare their diagnostic performance. Materials and methods We systematically collated the data of 383 patients, including 187 with PCa and 196 with benign lesions. Of them, 307 patients (150 with PCa and 157 with benign lesions) were randomly selected to train and validate the ML models, 76 patients were used as test set. B-Ultrasound videos (BUS), mpMRI T2 sequence (T2), and ADC sequence (ADC) were obtained from all patients. We extracted 851 features of each patient in the BUS, T2, and ADC groups and used a t-test, the Mann–Whitney U test, and LASSO regression to screen the features. Support vector machine (SVM), random forest (RF), adaptive boosting (ADB), and gradient boosting machine (GBM) models were used to establish radiomics models. In addition, we fused the features screened via LASSO regression from three groups as new features and rebuilt ML models. The performance of the ML models in diagnosing PCa in the BUS, T2, ADC, and fusion groups was compared using the area under the ROC curve (AUC), sensitivity, specificity, and accuracy. Results In the test cohort, the AUC of each model in the ADC group was higher than that of in the.BUS and T2 groups. Among the models, the RF model had the best diagnostic performance, with an AUC of 0.85, sensitivity of 0.78 (0.61–0.89), specificity of 0.84 (0.69–0.94), and accuracy of 0.83 (0.66–0.93). The SVM model in both the BUS and T2 groups performed best. Based on the features screened in the BUS, T2, and ADC groups fused to construct the models, the SVM model was found to perform best, with an AUC of 0.87, sensitivity of 0.73 (0.56–0.86), specificity of 0.79 (0.63–0.90), and accuracy of 0.77 (0.59–0.89). The difference in the results was statistically significant ( p <0.05). Conclusion The ML prediction models had a good diagnostic ability for PCa. Among them, the SVM model in the fusion group showed the best performance in diagnosing PCa. These prediction models can help radiologists make better diagnoses.

Siamese network-based visual tracking algorithms have achieved excellent performance in recent years, but challenges such as fast target motion, shape and scale variations have made the tracking extremely difficult. The regression of anchor-free tracking has low computational complexity, strong real-time performance, and is suitable for visual tracking. Based on the anchor-free siamese tracking framework, this paper firstly introduces balance factors and modulation coefficients into the cross-entropy loss function to solve the classification inaccuracy caused by the imbalance between positive and negative samples as well as the imbalance between hard and easy samples during the training process, so that the model focuses more on the positive samples and the hard samples that make the major contribution to the training. Secondly, the intersection over union (IoU) loss function of the regression branch is improved, not only focusing on the IoU between the predicted box and the ground truth box, but also considering the aspect ratios of the two boxes and the minimum bounding box area that accommodate the two, which guides the generation of more accurate regression offsets. The overall loss of classification and regression is iteratively minimized and improves the accuracy and robustness of visual tracking. Experiments on four public datasets, OTB2015, VOT2016, UAV123 and GOT-10k, show that the proposed algorithm achieves the state-of-the-art performance.

Calcification is a common phenomenon in both benign and malignant thyroid nodules. However, the clinical significance of calcification remains unclear. Therefore, we explored a more objective method for distinguishing between benign and malignant thyroid calcified nodules. This retrospective study, conducted at two centers, involved a total of 631 thyroid nodules, all of which were pathologically confirmed. Ultrasound image sets were employed for analysis. The primary evaluation index was the area under the receiver-operator characteristic curve (AUROC). We compared the diagnostic performance of deep learning (DL) methods with that of radiologists and determined whether DL could enhance the diagnostic capabilities of radiologists. The Xception classification model exhibited the highest performance, achieving an AUROC of up to 0.970, followed by the DenseNet169 model, which attained an AUROC of up to 0.959. Notably, both DL models outperformed radiologists (P < 0.05). The success of the Xception model can be attributed to its incorporation of deep separable convolution, which effectively reduces the model’s parameter count. This feature enables the model to capture features more effectively during the feature extraction process, resulting in superior performance, particularly when dealing with limited data. This study conclusively demonstrated that DL outperformed radiologists in differentiating between benign and malignant calcified thyroid nodules. Additionally, the diagnostic capabilities of radiologists could be enhanced with the aid of DL.

Objective The aim of this study is to develop AI-assisted software incorporating a deep learning (DL) model based on static ultrasound images. The software aims to aid physicians in distinguishing between malignant and benign thyroid nodules with echogenic foci and to investigate how the AI-assisted DL model can enhance radiologists' diagnostic performance. Methods For this retrospective study, a total of 2724 ultrasound (US) scans were collected from two independent institutions, encompassing 1038 echogenic foci nodules. All echogenic foci were confirmed by pathology. Three DL segmentation models (DeepLabV3+, U-Net, and PSPNet) were developed, with each model using two different backbones to extract features from the nodular regions with echogenic foci. Evaluation indexes such as Mean Intersection over Union (MIoU), Mean Pixel Accuracy (MPA), and Dice coefficients were employed to assess the performance of the segmentation model. The model demonstrating the best performance was selected to develop the AI-assisted diagnostic software, enabling radiologists to benefit from AI-assisted diagnosis. The diagnostic performance of radiologists with varying levels of seniority and beginner radiologists in assessing high-echo nodules was then compared, both with and without the use of auxiliary strategies. The area under the receiver operating characteristic curve (AUROC) was used as the primary evaluation index, both with and without the use of auxiliary strategies. Results In the analysis of Institution 2, the DeepLabV3+ (backbone is MobileNetV2 exhibited optimal segmentation performance, with MIoU = 0.891, MPA = 0.945, and Dice = 0.919. The combined AUROC (0.693 [95% CI 0.595-0.791]) of radiology beginners using AI-assisted strategies was significantly higher than those without such strategies (0.551 [0.445-0.657]). Additionally, the combined AUROC of junior physicians employing adjuvant strategies improved from 0.674 [0.574-0.774] to 0.757 [0.666-0.848]. Similarly, the combined AUROC of senior physicians increased slightly, rising from 0.745 [0.652-0.838] to 0.813 [0.730-0.896]. With the implementation of AI-assisted strategies, the accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of both senior physicians and beginners in the radiology department underwent varying degrees of improvement. Conclusions This study demonstrates that the DL-based auxiliary diagnosis model using US static images can improve the performance of radiologists and radiology students in identifying thyroid echogenic foci.

The utilization of LiDAR point cloud in various applications, such as autonomous driving and intelligent transportation, has gained significant attention and become a research hotspot. However, existing 3D object detection methods face two main challenges. (1) random or farthest point sampling can lead to a loss of foreground points and missed detections, particularly for small objects. (2) the sparsity of LiDAR point clouds causes the object point clouds to be non-uniformly distributed. In this paper, we present a novel method for 3D object detection. Our approach is based on the PV-RCNN network architecture and incorporates Centroid-aware (CA) Sampling and Local Attention Feature Encoding techniques. The voxel branch in feature extraction uses the Focals Conv module to predict the importance of voxel features at different locations and enhances the representation of local spatial structure by attention convolution. The key point branch uses CA Sampling to obtain key points. We calculate mask weights based on the position relationship between each point and the 3D bounding boxes, and the K highest scoring points are retained as key points. We further improve the local feature expression capability by aggregating each key point with the original point cloud and voxel features in the neighborhood using the PointNet++ network. Finally, we combine the collocated features and feed them into the classification and regression network of PV-RCNN to complete the detection process. Our simulation experiments show that the detection performance of our proposed method outperforms the PV-RCNN on the KITTI dataset.

Aiming at the problems of traditional salient object detection methods such as fuzzy boundary and insufficient information integrity, a salient object detection network composed of feature diversity enhancement module, global information guidance module and feature fusion module is proposed. Firstly, asymmetric convolution, cavity convolution and common convolution are spliced to form a feature diversity enhancement module to extract different types of spatial features corresponding to each feature layer. Secondly, the global information guidance module transmits the information captured by the coordinate attention mechanism to each feature layer through the global guidance stream, so as to learn the semantic relationship between different feature layers and alleviate the dilution effect; Finally, the feature fusion module receives the high-level features output from the previous layer, the low-level features of the corresponding layer and the global context information generated by the global information guidance module, and the cascade feature diversity enhancement module gradually integrates the multi-level features to refine the saliency feature map. Comparative experiments on five public data sets show that this method can achieve the highest values of 0.959 and 0.030 in F-measure and MAE. Compared with other seven advanced algorithms, it has better detection performance.

在强风化混合花岗岩地层中掘进时,盾构会遇到刀盘结泥饼、喷涌、掌子面失稳等问题.为解决该问题,需要对地层进行渣土改良.依托深圳地铁13号线留仙洞站—白芒站盾构区间项目,针对强风化混合花岗岩地层,通过自主研发的土壤-界面黏附力测试仪,研究含水率优化、注入分散剂、注入泡沫剂3种改良方法对强风化混合花岗岩地层的改良效果.研究结果表明:1)含水率优化能有效降低渣土的法向、切向黏附力,而分散剂和泡沫剂只对渣土的切向黏附力有改良作用,对渣土的法向黏附力反而有增大作用,且分散剂的增大幅度较泡沫剂大.2)泡沫剂渣土改良方案对盾构的掘进速度、推力、转矩以及掘进比能都有较好的改善作用.3)泡沫剂渣土改良方案并不能完全防止刀盘结泥饼,建议在采取泡沫剂改良方案的同时适当进行含水率优化,以最大限度降低刀盘结泥饼的风险.

The rapid growth of data and the requirement of designers to track massive data to obtain design stimuli have posed challenges to conceptual design, thereby promoting the development of data-driven design. Concept networks precisely capture design information from a large volume of unstructured and heterogeneous textual data and saliently decrease time and labor cost for designers to read texts, which creates new opportunities for developing a smart product design system. To advance data-driven design, this study proposes the novel function-structure concept network (FSCN) construction method, which combines sentence parsing and word/phrase extraction to integrate functional and structural information. Furthermore, a network analysis method is proposed to explore design information associations that contain both explicit and implicit associations together and thereby recommend them simultaneously to designers as inspirational stimuli to support design ideation. This approach can enhance designers' capabilities to build associations between design information, conceive new design ideas during conceptual design, and increase creativity for solving design problems. The proposed FSCN construction and analysis method can be used as an auxiliary tool to visualize associations among design information so as to inspire idea generation in the early stage of conceptual design. An illustrative example was used to validate the practicability of the proposed methodology. The code of the proposed method is available at https://github.com/KWflyer/FSCN .

In the real-world industrial application, all measured vibration signals are usually suffered from substantial spurious factors and characterized as large intra-class and intra-subclass variations due to changes in the operating conditions (i.e., working loads, shaft speeds), which seriously deteriorates most of the machine learning methods’ ability to learn the discriminative feature representations. In this work, we propose a novel subclass reconstruction network (SCRN) to learn discriminative feature representations from raw vibration signals under different working conditions by suppressing the intra-class and intra-subclass variations in the feature space. Specifically, a novel and simple average strategy is developed to represent the cluster centroid of each subclass information as an effective supervised representation which is effectively embedded into the SCRN model. Furthermore, a new cost function of SCRN is formulated by jointly minimizing the basic and subclass-level reconstruction errors. To better exploit the discriminative information and improve classification performance, we further develop a deep subclass reconstruction network (DSCRN) model by stacking multiple SCRN models together through non-linear transformations to learn better deep feature representations. Extensive comparative evaluations on three benchmarks CWRU, 2009 PHM, and MFPT demonstrate that the proposed methods achieve consistently better diagnosis performance than existing state-of-the-art methods. Especially, when only 1% training samples from CWRU are used, the proposed DSCRN has significant improvement with classification performance of 94.55% over compared method with a gain of at least 9%.

Object Tracking is a foundation in the field of vision, with powerful application capabilities. Discriminant-related tracking uses the technology of diagonalization of the circulant matrix to further improve the speed and accuracy of tracking. in the deep convolutional neural network based tracking, there is no clear explanation for the uncertainty of the target labeling. Relying heavily on manual labeling and the selection of loss function. We proposed alpha divergence-based Siamese network Tracking (alphaTK) to solve this problem. We minimized the alpha divergence between the conditional probability density output by the network and the conditional probability of annotations of the samples, void to choose the loss function. We modeled the noise generated by the label from the probabilistic perspective, and interpreting the uncertainty of the annotations. The Tracker is trained on a large number of data sets, and achieved excellent results on the OTB and UAV benchmark data sets.