In practical bearing fault diagnosis scenarios, fault label annotation usually relies on costly expert knowledge and manual analysis, resulting in a very limited number of labeled samples available for model training, while abundant unlabeled vibration signals remain underutilized. To address this issue, this paper proposes a cross-view multi-stage contrastive learning method with label-guided decision boundary calibration. First, time-frequency dual-view sample pairs are constructed, and multi-stage contrastive constraints are imposed at both the encoder output layer and the projection space to mine cross-view stable structures from unlabeled samples. Meanwhile, a small number of labeled samples are introduced as class anchors to participate in the discriminative mapping from feature representations to the category space, thereby performing label-guided decision boundary calibration for the latent class regions formed by the structural aggregation of unlabeled samples. This enables the mapping to establish decision boundaries with explicit class semantics and discriminative margins in the feature space. Subsequently, the time-domain view representation of each unlabeled sample obtains a soft class assignment through the calibrated discriminative mapping, which is then used as a soft supervisory signal to constrain the frequency-domain view in a cross-view manner to produce consistent discriminative outputs, thereby driving unlabeled features from structural aggregation toward class-consistent aggregation. Experimental results show that, using only 5% labeled samples, the proposed method achieves accuracies of 96.36 and 96.88% on the self-built dataset and the PU dataset, respectively, verifying its effectiveness.

Objective Fetal cardiac ultrasound examination plays a pivotal role in the prenatal diagnosis of congenital heart disease,which is one of the most prevalent and serious birth defects worldwide.However,clinical acquisition of fetal heart standard planes is highly challenging because of several physiological and technical constraints.The fetal heart is small in size,has a high beating rate,and its position in the maternal abdomen is easily influenced by fetal posture and movement.As a result,capturing the five clinically required standard planes,such as the four-chamber and outflow tract views,demands high temporal precision and operator experience.Moreover,the acquisition process suffers from poor repeatabil-ity,further hindering consistent and efficient screening outcomes.These challenges limit the overall effectiveness of fetal cardiac ultrasound screening in routine clinical workflows.To address these limitations and promote the automation of fetal cardiac ultrasound analysis,this study proposes FHSP-Net,a novel multitask deep learning model designed for accurate and comprehensive recognition of fetal heart standard planes.The model adopts a detection-then-classification strategy,which explicitly separates the anatomical structure detection task from the subsequent plane classification task,enabling effective feature extraction and decision-making for each subtask.This design reflects the clinical reasoning process in which a sonographer locates key anatomical structures and interprets them in the context of a specific standard plane.Method In the detection task,FHSP-Net incorporates several carefully designed components to handle the inherent com-plexities of fetal ultrasound imaging.A path-weaved network(PWN)is developed to improve the multiscale representation of features,particularly enhancing the detection of small and morphologically variable cardiac structures.In addition,a target aware enhancer(TAE)module is introduced to strengthen object-level context modeling,making the detection branch highly sensitive to anatomical structures that are often obscured by noise,artifacts,or low contrast.Equalization loss function version 2(EQLv2)is adopted to further mitigate the effect of data imbalance,which is common in clinical datasets where some structures or views are underrepresented.This loss formulation ensures that the model maintains stable learning dynamics across different categories and structure sizes,yielding robust and generalizable detection results.While the detection task ensures a strong foundation by locating relevant anatomical regions,the plane recognition task focuses on accurately classifying ultrasound views on the basis of the detected structures.Traditional algorithms that rely on whether certain anatomical structures are present in an image(i.e.,anatomical inclusion conditions)tend to have high precision but suffer from low recall.In other words,although they make only a few mistakes in their predictions,they often fail to identify all valid instances of each standard plane,resulting in missed diagnoses.To overcome this issue,we propose a candidate view scoring algorithm that evaluates and ranks multiple candidate views generated during the detection stage.This algorithm integrates structural cues and contextual features to refine classification decisions,substantially improving the model's ability to capture valid plane instances across varied imaging conditions.Result Experiments are conducted on a self-built fetal cardiac ultrasound dataset to evaluate the effectiveness of FHSP-Net.The dataset comprises a wide range of clinical scenarios and includes comprehensive annotations for anatomical structures and standard plane labels.Quantita-tive results demonstrate the superior performance of FHSP-Net across both subtasks.In the detection task,the model achieves a mean average precision of 0.962,representing a 0.027 improvement over a strong baseline model.This finding indicates that the proposed enhancements in feature fusion,object sensitivity,and loss balancing substantially benefit structure localization performance.In the plane recognition task,the candidate view scoring algorithm boosts classification accuracy by 0.10 compared with conventional structure-inclusion-based approaches.Ultimately,FHSP-Net attains an overall standard plane recognition accuracy of 0.959,representing a 0.194 improvement over the baseline model and the anatomical-inclusion-condition-based classification method.These results strongly validate the effectiveness and advantage of the proposed multitask design and its components.Conclusion The FHSP-Net multi-task model demonstrates strong capabilities in anatomical structure detection and fetal heart standard plane recognition.The integration of PWN,TAE,and EQLv2 enhances robustness and precision at the detection stage even under challenging imaging conditions.The candi-date view scoring algorithm effectively addresses the recall bottleneck of previous classification methods,ensuring that the model captures a large portion of valid standard plane instances without sacrificing accuracy.Together,these innovations make FHSP-Net a reliable and efficient solution for intelligent ultrasound analysis.FHSP-Net has potential for advancing computer-aided diagnosis in prenatal cardiology.

Objective In optical three-dimensional (3D) shape reconstruction, phase unwrapping is a critical step that directly affects reconstruction accuracy and efficiency. Traditional temporal phase unwrapping methods rely on multi-frame input, limiting their adaptability to dynamic scenes due to poor real-time performance. Meanwhile, spatial unwrapping techniques, though capable of single-frame processing, are prone to spectral aliasing, leading to significant phase jump errors at edges. Current deep learning-based methods face three core limitations: accuracy bottlenecks under single-frame input, with noticeable prediction errors in complex texture and jump regions of industrial parts; contradictions between interpretability and robustness, as "black-box" models hinder error tracing while physics-integrated models suffer from reduced convergence efficiency due to complex constraints; and the challenge of balancing efficiency and precision, where lightweight models sacrifice detail extraction and high-precision models fail to meet real-time requirements in industrial online detection. To address these issues, this study proposes a multi-task deep learning framework for high-precision absolute phase unwrapping under single-frame input conditions. Methods A multi-task architecture model (MTU-SEDD) featuring a shared encoder and dual decoders is developed. The model incorporates three key innovative designs: 1) a shared encoder-dual branch decoder structure, where both the encoder and decoder employ multi-branch dilated convolution modules. These modules utilize convolution kernels with different dilation rates (e. g., dilation rate 1 for focusing on local fringe deformation and rate 4 for capturing global order distribution patterns) to extract multi-scale fringe features, enabling simultaneous prediction of wrapped phase and fringe order from a single input frame. 2) A channel-spatial hybrid attention (CSHA) module is specifically designed for fringe projection images. In the frequency domain, this module performs precise channel weight calibration through global average pooling and fully connected layers; in the spatial domain, it introduces anisotropic Gaussian kernels to enhance feature responses in phase jump regions, effectively suppressing noise interference. 3) A dynamic gradient balancing strategy is proposed, which adaptively adjusts task weights by calculating the gradient norms of task losses is proposed. This strategy mitigates gradient competition between the wrapped phase regression task [with mean-square error (MSE) loss] and the fringe order classification task (with cross-entropy loss), ensuring balanced convergence of both tasks. Experiments were conducted on both simulated and real datasets. The simulated dataset was constructed using a random virtual object surface generator, producing height maps with 1-3 convex structures (including rectangles, ellipses, polygons, and noise-disturbed surfaces) with heights ranging from 5 mm to 60 mm and resolutions of 128x128 and 480x640. The real dataset consisted of 1120 groups of deformed fringe images of objects with diverse materials (plaster, metal, and plastic), covering complex surface deformations (grooves, edge mutations) and specular reflection interference. Performance was evaluated using metrics including structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), normalized root mean square error (NRMSE), and root mean square error (RMSE), with comparisons against traditional convolutional neural network (CNN), DL-SPU, MultiResHNet, and U-Net (double) models. Ablation experiments were also conducted to verify the contributions of key modules, including the multi-task balancing strategy, dual-branch structure, and multi-branch dilated convolution. Results and Discussions Experimental results clearly demonstrate the superiority of the proposed method. In simulated dataset tests, MTU-SEDD achieved significantly lower RMSE than the traditional CNN model across all samples-for instance, image 1 showed an RMSE of 0.9260 for MTU-SEDD versus 8.9656 for CNN. The model maintained stable performance across different resolutions, with an RMSE fluctuation of <= 0.15 between 128x128 and 480x640 simulations. In real scenario experiments, MTU-SEDD outperformed other single-frame models, achieving the highest SSIM (0.9885), PSNR (54.64 dB), and the lowest NRMSE (0.0019), with an inference speed of 75.1 ms/frame. Ablation experiments further validated the necessity of each module: removing the multi-task balancing strategy (Exp-1) reduced SSIM by 0.72% and increased edge errors due to the marginalization of the fringe order task; Independent training of dual branches (Exp-2) led to a 131 % increase in runtime and accuracy degradation due to the lack of shared feature learning and potential misalignment between branch outputs; replacing multi-branch dilated convolution with ordinary convolution (Exp-3) caused severe detail loss, dropping SSIM to 0.8847 as the model failed to capture multi-scale features. Specifically, the CSHA module reduced edge region errors by 68%, while the dynamic gradient balancing strategy decreased inference time by 40 %and avoided multi-frame registration errors. Conclusions This study successfully develops a multi-task learning-based deep learning framework for structured light phase unwrapping. The shared encoder-dual branch decoder structure enables joint optimization of wrapped phase regression and fringe order classification tasks, achieving high precision and real-time performance. The channel-spatial hybrid attention module effectively suppresses spectral aliasing by enhancing phase jump region features. The dynamic gradient balancing strategy ensures stable model performance across different resolutions, demonstrating strong generalization. Future work will focus on integrating topology repair algorithms with 3D attention mechanisms and achieving model lightweighting via knowledge distillation to further accelerate training and inference.

To improve contour accuracy of two-degree-of-freedom parallel mechanisms for complex high-curvature trajectories under strong coupling, nonlinear friction, and nonrepetitive disturbances, a dual-loop contour error control method combining Active Disturbance Rejection Control (ADRC) and task-space-based iterative learning control (TS-ILC) is proposed. In the inner loop, a joint-space velocity-loop ADRC is designed to estimate and compensate lumped uncertainties, including inertial coupling, friction, and external disturbances, thereby enhancing disturbance rejection and weakening inter-joint coupling. In the outer loop, the shortest normal contour error is estimated in task space using the Newton iterative method and mapped into a joint-space learning error through the inverse Jacobian. A PD-type ILC with zero-phase filtering is then applied for repetitive error compensation. A MATLAB/Simulink and Simscape Multibody co-simulation platform is built to evaluate the method under white noise and random step load disturbances. Results on heart-shaped and five-leaf clover trajectories show that the proposed method achieves good robustness and convergence, reducing the maximum contour error by 74.15% and 54.48%, respectively, compared with the method that combines ADRC with joint-space tracking-error ILC. This study provides an effective solution for complex-trajectory contour control of high-speed, high-precision parallel mechanisms.

In the field of intelligent fault diagnosis, digital twins are increasingly utilized to generate controllable simulation data, thereby reducing the reliance on large-scale labeled measurements from practical engineering systems. However, simulation-to-reality transfer often encounters partial domain adaptation (PDA) scenarios. This label-space mismatch can induce negative transfer during the adaptation process, consequently degrading diagnostic performance. To address this limitation, this paper proposes a novel framework termed Independent Adversarial Reweighted Partial Domain Adaptation (IARPDA). This approach effectively mitigates negative knowledge transfer by filtering irrelevant source class information through a decoupled strategy. To extract domain-invariant features, a multi-scale deep frequency-filtering architecture is first designed to account for varying spectral scales, thereby bridging the simulation-to-measurement gap. Central to this approach is a Wasserstein-distance-guided instance reweighting module, which identifies and suppresses irrelevant source-domain classes without relying on biased classification heuristics. By isolating the weighting process from the primary classification task, the model suppresses the propagation of classification errors, enhancing both the stability and the ability to distinguish shared classes from anomalous ones. Additionally, target-domain conditional entropy regularization is incorporated to improve discriminative stability under unlabeled measurement conditions. Simulation-to-measurement transfer experiments conducted on the MCC5-THU gearbox demonstrate that the proposed IARPDA method consistently outperforms several comparison methods across multiple partial domain tasks, offering an innovative digital twin-guided paradigm for industrial fault diagnosis.

This paper reviews large-area metal paste sinter joining, focusing on the progress and challenges related to delamination, warpage, and reliability—findings that are crucial for the development of highly reliable power electronics.

Single-stage depth detectors, represented by YOLO, operate quickly and have been widely applied in industrial inspection. However, their generalization is often limited when handling domain shift caused by significant differences in the appearance and background noise of surface defects under complex working conditions. Domain adaptation technology can enhance the adaptability of detection models. To address challenges such as intricate defects in precision components, the prevalence of small defects, and strong background noise interference, a cross-domain defect location method-based consistency conditional adversarial network is developed. Firstly, the structural characteristics of the defect features are explored to construct an image-instance multi-level adversarial network, reducing the data distribution differences between domains. Secondly, the discrimination results output from the adversarial network are analysed to establish a consistent adversarial strategy, enhancing the detection adaptability. Then, a weakly supervised localization method is studied to design an image-level predictor, which preliminarily locates the instance features of the region of interest to mitigate background noise interference. Finally, the prediction similarity between the image-level predictor and the bounding box predictor is measured, which can adaptively adjust the proportion of hard-to-align instances during training, thereby improving the robustness of the detection model in locating complex defects.

While deep learning has advanced industrial fault diagnosis, its reliance on massive labeled datasets remains a major obstacle to real-world application. In response to the scarcity of labeled fault samples and ineffective utilization of abundant unlabeled data, this paper proposes a novel semi-supervised fault diagnosis method based on supervision-guided contrastive learning (SGCL). First, a supervision-guided mechanism is introduced to leverage a small number of labeled samples to construct semantic class centers. This explicitly guides unlabeled data toward their respective categories, thereby alleviating the mismatch between different optimization objectives and enhancing intra-class compactness and inter-class separability. Second, a consistency regularization strategy is introduced to improve the model’s representation capability via pseudo-label constraints, enabling unlabeled representations to directly participate in updating the classifier parameters. By jointly optimizing these mechanisms, the proposed framework seamlessly translates raw unlabeled data into highly discriminative diagnostic features. Comprehensive experiments on a self-built dataset and the Paderborn dataset demonstrate that SGCL achieves diagnostic accuracies of 91.74% and 91.42%, respectively, under a 5% label ratio. This represents an average improvement of approximately 8%-10% over compared generative and contrastive semi-supervised methods, verifying the effectiveness of the proposed approach in addressing fault diagnosis tasks under severely limited label conditions.

Dynamic task assignment in multi-UAV swarms is a challenging problem due to partial observability, strongly coupled state information, and limited communication resources in complex environments. Existing multi-agent reinforcement learning methods often suffer from representation collapse when processing highly entangled observations, while communication-based coordination strategies may introduce excessive communication overhead and thus limit practical applicability. To address these issues, this paper proposes a context-separated and semantic-gated multi-agent proximal policy optimization (CSSG-MAPPO) framework for dynamic UAV task assignment. First, the problem is formulated as a partially observable Markov decision process to characterize the observation constraints encountered in realistic UAV swarm operations. Second, a context-separated representation (CSR) strategy is introduced to decompose observations into environment-oriented and agentoriented components, which improves task-relevant feature extraction and alleviates representation collapse during training. Third, a semantic-gated communication (SGC) mechanism is designed to adaptively regulate inter-agent message passing, enabling UAVs to communicate only when necessary and thereby reducing redundant communication under bandwidth constraints. Extensive experiments are conducted to compare the proposed method with several representative algorithms. The results show that CSSG-MAPPO achieves faster convergence, higher cumulative rewards, and lower communication ratios than the compared methods. In addition, ablation studies further verify that both the CSR strategy and the SGC mechanism make substantial contributions to training efficiency, communication efficiency, and overall task performance. These results demonstrate the effectiveness of the proposed framework for dynamic task assignment in resource-constrained UAV swarm systems.

Induction motors are widely used in the industrial field such as electric drive systems for new energy vehicles and synchronous condenser for improving the power factor of the power grid. The motor health condition often influences the operation of the entire mechanical system, so it is necessary to conduct a health assessment on it. Current induction motor fault diagnosis largely relies on expert knowledge, while many deep learning methods suffer from limited generalization and poor interpretability, leading to unreliable results. To address these issues, a fault mechanism-guided interpretable causal disentanglement domain generalization detection method (ICGN) is proposed for typical fault diagnosis of induction motor. Firstly, a primary feature extractor is constructed based on transformer, which adaptively screens causal and non-causal factors through the self-attention mechanism, and an attention score evaluation mechanism is constructed to visually demonstrate interpretability. Secondly, to further disentangle and refine causal features and non-causal features, the developed causal aggregation loss and causal decoupling loss are combined, ensuring the cross-domain consistency of causal factors and promote the domain generalization ability of the network. Finally, the proposed method is validated using vibration signals collected from two Spectra Quest test benches from University of Ottawa and the private laboratory. The cases of cross device motor fault diagnosis are included, and the ICGN is compared with several advanced domain generalization algorithms. The results demonstrate that the proposed method achieves superior performance both in interpretability and domain generalization capability.

Surface defect detection is important for ensuring industrial manufacturing quality. Convolutional neural networks significantly improve the detection performance. However, surface defects exhibit significant differences in appearance and background due to complex working conditions. Traditional deep models show limited generalization for domain shift caused by data differences. To address the domain shift from the defect appearance change, a task-oriented domain confusion network (TDCN) is developed, which attempts to explicitly link its parameter structure with the structural feature of the target data, guiding the network to form a transferable causal mapping system, so that the solved feature can uniquely represent the intrinsic information of the target data. Specifically, defect data generation and inverse solution methods are studied to establish a feature-recoding strategy, which is used as prior knowledge to drive an adversarial function, transforming causal estimation into a Nash equilibrium of the network to eliminate noise bias caused by hard-to-transfer features. To respond to the model deviation under Nash equilibrium, a feature semantic storage dictionary is established to align the local and global semantic features. Ultimately, the parameter gradient evolution logic guided by classification is studied, which adaptively fits an unbiased causal estimation model, improving the transfer stability and cross-domain detection accuracy of defect features between multiple environments. Extensive experiments have proved that our TDCN outperforms the most state-of-the-art methods, especially for detection tasks with scarce data.

In industrial operational environments, rolling bearing vibration signals not only contain fault related periodic impulse components, but also interference noises. The coupling of these signals attenuates the fault characteristics, particularly in cases of compound faults, significantly affecting the diagnostic accuracy of bearings. This paper proposes a rolling bearing fault separation and compound diagnosis method using a sparse Bayesian framework with adaptive prior knowledge. Firstly, the proposed method segments signals into a finite number of modes, employing filter banks to adaptively select decomposition modes with a new compound fault indicator, which reduces noise interference and eliminates irrelevant components while ensuring the preservation of fault-related information. Secondly, by employing an interpretable strategy enables fault frequency estimation, eliminating errors from empirical parameter settings and avoiding the computational complexity of parameter optimization. The feature frequencies are estimated based on the envelope harmonic product spectrum and used as a prior knowledge of sparse Bayesian learning. Finally, to avoid the restriction of current feature extraction methods that primarily focus on the selection and extraction of a single demodulation frequency band, a sparse Bayesian probabilistic model is particularly designed for compound fault diagnosis. Simulation and experimental results show that the proposed method can effectively extract and separate the fault features of each individual modes and realize the diagnosis of rolling bearings.

With the development of industrial intelligence,impeller surface defect detection has played a key role in guaranteeing equipment safety and improving production efficiency.However,the problems of incomplete defect sample information,time-consuming sample labeling,and the poor adaptability of detection models in actual industrial environments constrain improvements in detection accuracy.To solve these problems,a method for impeller surface defect detection is proposed,based on a Fourier-based adaptive consistency regularization(FACR)network.Initially,a source model is pretrained with its parameters held constant.Then,a target model is initialized,following the principles of transfer learning,in which the feature extractor utilizes the parameters of the source model and the parameters of the classifier borrow from the concept of imprinting,thereby leveraging the knowledge of the feature extractor of the source model for informed initialization.An adaptive knowledge consistency strategy is designed to ensure that the features extracted by the target model closely resemble those extracted by the source model.Both labeled and unlabeled data serve as bridges for knowledge transfer.The Kullback-Leibler divergence is employed to measure the dissimilarity in feature extraction between the pretrained and target feature extractors,thereby preventing discrepancies that may lead to negative transfer.In addition,an adaptive distribution consistency strategy is devised by using the structural information contained in the unsupervised data to guide the training of the supervised data,thereby minimizing the gaps between the representations of features from the unsupervised and supervised data.Furthermore,a strategy based on Fourier transformation is explored to address the issue of the inaccurate extraction of defect feature information by existing deep transfer networks.Leveraging the inherent characteristics of the Fourier transformation,where the phase component of the Fourier spectrum retains high-order semantics and the amplitude component contains low-order statistical information,defect images from different domains undergo a Fourier transformation to obtain phase and amplitude information.Linear interpolation is then applied to enhance the amplitude of the sample information,thereby ensuring a phase-invariant exchange of the corresponding regions.This approach directs the focus of the model toward the phase information(i.e.,structural details)to enhance feature information and reduce the impacts of background environments on the detection model.Subsequently,the concept of curriculum learning is introduced to dynamically adjust the thresholds.Without introducing new hyperparameters or incurring additional computational costs,fixed thresholds are transformed into real-time adjusted thresholds based on the category difficulty.This efficient utilization of unlabeled data aims to enhance the model performance,improve the classification accuracy for various defect types,and address the challenge of low accuracy in complex defect detection caused by fixed thresholds that do not consider different learning states or difficulties for different categories.Finally,the classification results of the original and enhanced samples are used as metrics with which to measure their respective contributions to the network.These contribution performances are applied as weights to the classification loss,thereby allowing the model to thoroughly learn impeller surface defects from limited samples.This approach addresses the issue of potential discrepancies in the predictions between the original and enhanced images after amplitude interpolation through Fourier transformation.The goal is to enhance the generalizability of the model for impeller surface defect detection in a real industrial environment.ResNet18 is selected as the backbone network to validate the superiority of the proposed model.Under the same experimental conditions and with the number of supervised training samples in the target domain set to 5,10,and 20,comparisons are made with other mainstream domain adaptation methods.The experimental results demonstrate that,compared with other semi-supervised domain adaptation methods,the proposed method effectively enhances the accuracy of impeller surface defect detection.The method presented in this study can improve the extraction capability of common features in the two domains and thus reduces the requirements for actual impeller surface defect samples and labeling.Simultaneously,using a feature enhancement approach increased the classification accuracy of the model.A favorable detection outcome can be achieved when the defect features exhibit a certain degree of similarity,even for components with different lighting conditions and surface textures.Therefore,this method not only provides strong technical support for practical industrial applications but also enriches the technical means in the field of defect detection,which has important theoretical value and practical significance for guaranteeing equipment safety and promoting industrial intelligence.

To enhance the understanding of blade non-synchronous vibration from the blade tip timing perspective, a single-stage centrifugal compressor representative of industrial architecture has been investigated experimentally. The characteristic of the wall pressure and blade tip timing (BTT) signal is first measured and analyzed experimentally under rotating stall condition. The relationship between the rotor vibration in the stationery and rotor reference frame is then established. A combination of BTT and strain gauge gives a full interpretation of non-synchronous vibration phenomenon and identifies the fluid-structure interaction parameters. To further verify the rotating stall state and corresponding signal characteristics, a numerical simulation of unsteady flow and aerodynamic forcing function is also conducted. The numerical results match the experimental findings well and can interpret the underlying physics. These results contribute to an in-depth understanding of the aeroelastic phenomena in industrial centrifugal compressors.

Flip chips are widely used in electronic systems for defense, aerospace, and other applications where packaging reliability is critical. However, flip chip defect samples present a variety of defect types and few samples with labels in actual industrial applications. The paucity of labeled defect samples indicates that the existing data volume cannot be matched with deep learning detection models. Therefore, flip chip intelligent defect detection faces the problems of poor model adaptability and weak generalization performance. As a solution to these problems, a semi-supervised dual-constraint centroid contrastive prototypical network (SSDCPN) for flip chip defect detection under limited labeled data is proposed in this paper. First, a prototype-based supervised contrastive learning strategy is developed to construct the contrastive prototypical network, which increases the inter-class sparsity and intra-class compactness of features to acquire more discriminative features. Then, to address the susceptibility of the support set prototypes to outliers, dual constraints are imposed on the support set prototypes to calibrate and refine the prototypes. Finally, a pseudo-labeled sample selection mechanism based on epistemic uncertainty and entropy is proposed to obtain rich semi-supervised information to guide the model training. The mechanism can select high-confidence pseudo-labeled samples that can complement the training samples to further strengthen the generalization performance of the model. Defect detection experiments on flip chip vibration signals indicate that the present method is superior to other methods in the case of limited labeled samples.

The effective extraction of weak defect features is crucial to the reliability analysis of flip-chip microbumps. A novel dual-drive adaptive reweighted sparse (DDARS) framework is proposed, which can skillfully realize the time-varying noise suppression and unsupervised feature stabilization of flip-chip vibration signals under large-scale environmental disturbances. Specifically, the compact representation (CR) of patch operators. and its inverse Lambda(-1) are defined to segment the high-dimensional data into low-scale time-varying noisy patches. By integrating the prior knowledge of noise variance distribution and singular value sparse distribution, an adaptive knowledge prior reweighted matrix(ARM) is constructed to precisely reweight and match sparse regularization terms, thereby effectively mitigating time-varying noise interference. Quadratic sparse projection constraints are introduced under the multi-patch sample condition, utilizing quadratic sparsity to evaluate feature saliency and stabilize the potential representation of signal features. Based on the potential connection between patches and sweep-frequency signals, analyze the sparsity intensity of signals in different time periods to guide the optimal frequency range for ultrasonic excitation of chips. The validity of DDARS is verified through qualitative and quantitative analysis with simulated signals and data from engineering application examples, which can provide technical support for flip-chip microbumps reliability analysis.

This article proposes a sparse denoising method for ultrasonic signals based on the dual memory-enhanced deep unfolding network (DME-DUN), which aims to accurately reconstruct high-frequency ultrasonic signals in flip chip detection. To address information loss during the iterative process in deep unfolding networks, two memory enhancement mechanisms are designed, i.e., adjacent-stage and cross-stage. For adjacent-stage, the high-throughput information from the previous stage is introduced into the current stage, thus facilitating the transmission of information between adjacent stages. Additionally, the cross-stage memory enhancement mechanism is applied to promote information fusion between separated stages. This mechanism is based on Richardson extrapolation and calculates the output error between every two adjacent stages to update the sparse coefficient. Extensive experiments verify that the proposed method effectively captures weak echo features while removing noise.

Flip chip technology has been extensively adopted in integrated circuit (IC) packaging, and the integration of flip chip technology with solder joint interconnection methods has been utilized in the manufacturing of electronic devices universally. As the development of flip chip towards high density and ultra-fine pitch, the inspection of flip chips is confronted with great challenges. Aiming at the characteristics of the chip signal with strong interference noise and weak feature information, the sparse representation model under the reweighting constraint is proposed. We analyzed the low-rank prior of the signal feature matrix and explore the construction of a sparse representation dictionary to establish an adaptive singular value decomposition (SVD) dictionary for the variant features. Subsequently, we investigate the adaptive weight matrix that aligns with the time-varying structure of the noise and the distribution pattern of the feature singular values, and construct a sparse characterization model under the reweighting constraint. Simulation analysis shows that the proposed algorithm can effectively realize the filtering of time-varying noise in the signal, so as to reliably extract the weak defect features.

The widespread use of flip-chip technology in the field of microelectronics packaging makes defect detection technology face great challenges, which requires the development of detection technology with less manual intervention, a lightweight network architecture, and high precision. Aiming at the dual-stability of the neural architecture search (NAS), and explicitly associating the network architecture searched by NAS with the architecture features of vibration signals. This study proposes a dual-convergence sparse feature extractor (DSFE) for visualization vibration signals architecture feature searching, which attempts to learn a representation whose architecture feature can uniquely represent the intrinsic information of the target signal. DSFE analyzed and summarized three deficiencies of gradient-based NAS, namely, “huge GPU memory consumption,” “high collapse probability,” and “rigid number of node precursor operations.” The corresponding solution modules are proposed in turn, which are “primary and secondary search spaces,” “skip_connect coefficient modification,” and “dynamic sparse selection of precursor operations.” The effectiveness of DSFE is verified with flip-chip vibration signals excited by the air-coupled ultrasonic wave. By comprehensive comparative experiments with fixed-structure networks and gradient-based NAS methods, it is proved that DSFE cannot only achieve dual-stability of detection precision and architecture searched, but also ensure less resource consumption and multifarious architecture, which can broaden the application scope of DSFE in practical engineering.