Precise remaining useful life (RUL) estimation can effectively enhance the reliability and safety of industrial machinery. While deep learning methods have demonstrated strong potential in this area, their effectiveness is often limited by the scarcity of labeled data, which is both costly and time-consuming to obtain in industrial applications. Therefore, a novel RUL estimation method based on meta contrastive learning is proposed to construct an efficient semi-supervised learning process, which specifically targets the challenges of limited labeled data by deeply mining the underlying degradation information within the vast amounts of unlabeled data. First, a meta contrastive learning algorithm is proposed to ensure that the contrastive information derived from unlabeled data positively influence the model’s performance on labeled data. Meanwhile, a meta-update strategy is designed to mitigate overfitting and catastrophic forgetting, effectively improving the feature extraction potential and generalization ability of deep learning structures. Additionally, a weight optimization strategy is proposed to adaptively adjust the weight factor of the semi-supervised learning process based on the performance of the supervised task, thus enhancing the flexibility and effectiveness of semi-supervised learning. The proposed method is validated on the turbofan engine and wind turbine gearbox datasets, demonstrating its effectiveness and superiority.

Remaining useful life (RUL) prediction under unseen domains has made domain generalization (DG) a promising technique. However, existing DG studies, especially multi-source DG methods, typically assume multiple labeled source domains, which is rarely realistic because labels are scarce and right-censored unlabeled data dominate. Moreover, existing approaches primarily rely on global feature alignment, which neither exploits the weak supervision in right-censored data nor prevents degradation-stage misalignment between unlabeled and labeled domains. Rigid feature alignment further ignores heterogeneity in degradation processes, limiting generalization. To address these limitations, a novel evidential semi-supervised domain generalization (Evi-SSDG) method is proposed for RUL prediction under unseen domains utilizing only one labeled and multiple unlabeled right-censored source domains. Evi-SSDG integrates evidential quantile modeling and survival likelihood to learn multi-quantile RUL distributions and uncertainties, while transforming not-yet-failed censoring information into effective supervision. Moreover, a stage-aware contrastive and uncertainty-structure preservation module employs degradation-stages as alignment anchors, while promoting compact and continuous intra-stage features manifold and reducing inter-stage overlap to improve pseudo-stage reliability and prevent stage misalignment. Finally, a confidence-guided evidential alignment module enforces cross-domain consistency of stage-wise evidence distributions, enabling generalization without rigid feature alignment. Experiments across industrial and public datasets highlight the effectiveness and superiority of Evi-SSDG.

Accurate remaining useful life (RUL) prediction in industrial applications is often hindered by the scarcity of degradation data. Data-driven synthetic degradation data generation offers a promising solution to this challenge. However, current data-driven generative models for RUL prediction don’t consider the inherent degradation properties of equipment and the synthetic data generation process has no relation with the RUL predictive model training process, which will lead to that the degradation information are not fully exploited to improve the accuracy of the predictive model. To address these issues, a novel RUL prediction method with degradation data generation is proposed by developing degradation-informed variational autoencoder (DIVAE) and meta fine-tuning (MFT) technique. DIVAE incorporates a novel dual degradation properties long short-term memory (DDP-LSTM) to embed the irreversibility and acceleration properties of degradation into the long short-term memory structure, enhancing the consistency of the generated synthetic data with some degradation principles. MFT establishes a co-optimizing scheme where the predictive model provides second-order gradient flow to dynamically fine-tune DIVAE, encouraging the DIVAE to generate synthetic data that positively contributes to prediction accuracy. Experiments on turbofan engine and wind turbine gearbox datasets consistently demonstrate the effectiveness and superiority of the proposed method under data-scarce scenarios.

Currently, deep learning methods are widely applied in predicting the remaining useful life (RUL) of equipment. These methods typically conduct training on a fixed dataset to obtain a static model. However, in industrial scenarios, equipment can continuously generate new degradation data. When utilizing the acquired new data to enhance the model’s predictive accuracy, retraining the model with all historical data is too time-consuming. Online learning can effectively address this issue. Nevertheless, existing studies face several challenges, including the model’s inability to retain key knowledge learned from historical data and the excessively complex structure resulting from model expansion. To address the abovementioned limitations, this paper proposes a novel Soft Actor-Critic (SAC) algorithm with hierarchical storage and experience mixing to achieve online learning for equipment RUL prediction. Within this algorithm, after each round of interaction, online and offline experiences are drawn from the buffer experience pool and the historical experience pool respectively, which are then mixed for updating the model parameters. The experience pools are hierarchically designed according to the stages of device degradation to ensure that the sampled experiences evenly cover the entire process of device degradation. The Online-Elastic Weight Consolidation (EWC) method is used to regulate the updates of the actor network parameters in the reinforcement learning model, thereby balancing the learning of new task data with the retention of previously learned key knowledge. Experiments demonstrate that the proposed method can steadily improve the RUL prediction accuracy during the online learning process, achieving higher accuracy than existing methods while significantly reducing computational resource consumption.

To solve high-dimensional expensive problems, this paper proposes a two-stage surrogate-assisted bi-cooperative optimization algorithm (TSSABC) in which two collaborative frameworks of global and local surrogates-based optimization are effectively cooperated in the current promising and historical potential regions respectively. For the first one, local RBF-based JADE with historical position exclusion is combined with dynamic coordinate search using response surface (DYCORS) models for local and global search, effectively exploring the current promising region and accelerating convergence. For the second one, the local RBF-based JADE algorithm repositions the best solution within the historical potential region. Two approximate center solutions, i.e., dimension-based and space-based, are reconstructed using the Parzen window method. DYCORS-based global search then investigates these solutions to efficiently explore the historical trajectory from previous iterations. Additionally, an adaptive intervention mechanism adjusts the search focus by incorporating greedy information if the two frameworks fail to yield positive optimization progress over several successive iterations. Experimental results on three different test suites demonstrate that TSSABC outperforms classical and state-of-the-art algorithms in terms of performance.

The existing surrogate-assisted algorithms for computationally expensive multi-objective optimization problems (EMOPs) face three key challenges, i.e., the gradual loss of diversity of the population, the excessive randomness of local search, and the low adaptability to problems with complex PF shapes. This paper proposes an optimization state-driven adaptive evolution algorithm called OSAE to address EMOPs, where both the association and update states are employed to adjust the search directions adaptively. Specifically, two different types of evolution starting points are determined based on the association state. Thus, high-potential sub-populations, obtained by a two-step sub-population generation strategy, are employed as the starting populations of the designed RBF-based local search to accelerate the local exploitation for each sub-problem. Subsequently, the exact offspring can be obtained by a two-metric-driven selection, and both the exact and optimized populations are updated where the update state represented by inverted generational distance comparison is employed to determine whether to maintain or transform the reference point configuration. Therefore, OSAE achieves the adaptive adjustment of reference point configuration and adaptive search for each sub-problem, thus enhancing the adaptability to complex problems such as irregular PF shapes. Experimental studies on both classical test suites and real-world application verify the performance of OSAE.

Advances in sensor technology and the Industrial Internet of Things (IIoT) have enabled the collection of large-scale monitoring data, facilitating intelligent remaining useful life (RUL) prediction for industrial equipment. However, accurate RUL prediction in distributed environments faces two major challenges. First, the scarcity of data and limited computational resources at edge clients hinder the development of robust RUL models, while privacy constraints prohibit centralized data sharing. Second, distribution shifts across client machines severely limit the model’s ability to generalize to unknown operating conditions (OCs) and equipment. To address these challenges, this article proposes a cloud-edge collaboration (CEC) federated invariance and specificity domain generalization (DG) (CEC-FedISDG) method. CEC-FedISDG integrates both domain-invariant and domain-specific predictive knowledge within a unified cloud-edge federated learning (FL) framework. This design enables the model to exploit the broad generalizability of invariant features while retaining domain-specific predictive capabilities. Specifically, a progressive invariance refinement (PIR) module is developed to gradually strengthen domain-invariant features while preserving privacy through a two-stage learning process. Subsequently, a dynamic specificity selection (DSS) module is designed. It dynamically integrates the outputs of private-domain regressors that retain domain specificity utilizing a domain classifier, adapting weights to test samples, thereby improving RUL prediction accuracy. Experimental evaluations on two bearing datasets and a real-world industrial wind turbine gearbox (WTG) dataset demonstrate that the CEC-FedISDG achieves superior generalization performance while adhering to strict privacy preservation requirements.

Ensuring security is a critical requirement for production in smart factories. Unauthorized rotary-wing unmanned aerial vehicle (UAV) poses significant threats, as they may be used for surveillance, data theft, or even industrial sabotage. Thus, detecting these drones in a timely manner is crucial for safeguarding factory operations, while this conflicts with the ways to improve detection accuracy for complex tasks through the existing deep learning models built on stacked multi-layer architectures and advanced network designs. The computational and storage demands increase significantly with tack complexity, and typical network pruning face significant limitations in improving inference efficiency and reducing computational costs when applied to models with complex branching, nonlinear connections, and cross-layer dependencies. In this paper, we propose a model pruning approach specifically designed for infrared rotary-wing UAV detection, called Iterative Model Pruning with Sparsity Learning (IMPSL), to address these three issues. Our approach integrates dynamic pruning rate adjustment with sparsity learning, enabling the model to adaptively optimize its structure throughout the pruning process based on varying task demands. To validate the effectiveness of IMPSL, we introduce a specialized infrared dataset for UAV detection. Extensive experiments confirm the effectiveness of our method, demonstrating significant improvements in inference speed.

High-precision land-cover classification (LCC) from remote sensing imagery provides valuable semantic information. This information is critical for intelligent navigation and environmental understanding in autonomous unmanned systems. Optical imagery provides rich spectral and texture information, while synthetic aperture radar (SAR) offers all-weather sensing and strong structural characterization. The effective fusion of optical and SAR data plays a vital role in LCC, but it is hindered by heterogeneous imaging mechanisms that introduce geometric misalignment and frequency-domain discrepancies in the fusion process. To address these challenges, this paper proposes a geometry–frequency synergistic fusion network (GFS-Net) for accurate optical–SAR land-cover classification. By jointly integrating a Swin Transformer-based hierarchical dual-stream encoder with a carefully designed decoder that incorporates semantic flow alignment and local-context embedding, GFS-Net enables accurate multi-scale feature alignment and fine-grained detail recovery. At the core of the network, a geometry–frequency synergistic fusion module explicitly combines deformable geometric alignment with frequency-domain amplitude-phase fusion. This module corrects cross-modal spatial offsets, suppresses SAR speckle noise, and preserves discriminative optical structures. Furthermore, a dual-domain gated recalibration module is introduced to adaptively suppress modality-specific redundancy and enhance complementary semantic responses through global gating and joint spatial–channel rectification. Extensive experiments on the WHU-OPT-SAR and DFC2025 datasets show that GFS-Net achieves 84.96% OA and 58.64% mIoU on WHU-OPT-SAR and 87.02% OA and 74.66% mIoU on DFC2025, outperforming the strongest compared method CFFormer by +1.14% and +0.44% mIoU, respectively.

To improve the convergence performance of the Polyline-based Core Sandwich Structures (PCSSs), this paper proposes a Surrogate-Assisted Fine Selected Sparse Search algorithm called SAFSSS within the Moving Morphable Components framework, in which both sparse centers-driven global exploration (SCGE) and optimization state-driven sparse local search (OSSLS) are well integrated for balancing convergence and diversity. In SCGE, sparse centers are dynamically selected from the database by comprehensively considering the convergence, feasibility, and diversity, and the targeted search strategies for different centers are designed. Specifically, the high-potential regions can be fully exploited by distinguished selecting base individuals for centers to generate candidate offspring individuals, and the diversity of infeasible individuals are employed to rank infeasible individuals efficiently for emphasizing the search around the constraint boundary. In OSSLS, when the optimization state is updated, the sparse local search is designed where the search centers are two excellent but no adjacent solutions for speeding up the convergence. Numerical experiments validate that the SAFSSS obtains the best structure with minimal compliance in the design of PCSSs. More importantly, Physical experiments demonstrate that the structure obtained by SAFSSS have the highest maximum bending load and the highest strength, and SAFSSS also delivers the highest normalized flexural rigidity.

Selective maintenance problems are intensively investigated by both scholars and practitioners. Nonetheless, existing research on such problems usually overlooks the comprehensive consideration of diverse decision criteria and uncertain maintenance environments in terms of stochastic mission and break durations. This work proposes a stochastic multi-objective selective maintenance problem that considers stochastic mission and break durations and puts forward the corresponding optimisation method. First, a stochastic multi-objective mixed-integer nonlinear chance-constrained programming model is constructed to maximise mission reliability and minimise maintenance cost under maintenance time constraints. Second, a learning-driven multi-objective artificial bee colony algorithm is tailored to tackle the proposed model. In the designed approach, a Q-learning method in the employed bee phase and an iterative local search method in the onlooker phase are deployed to enhance the exploration and exploitation abilities. Finally, through conducting comparison experiments between the customised method and five popular methods on several real-world cases, the experimental results confirm its feasibility and superiority in addressing the problem of interest.

Domain generalization (DG)-based remaining useful life (RUL) prediction aims to train predictive models solely on multiple observed source domains while ensuring robust generalization to unseen target domains. Existing DG methods predominantly focus on extracting domain-invariant features to enhance model transferability. However, they often overlook domain-specific labelrelevant cues that can be critical for cross-domain generalization. Additionally, relying on statistical correlations for invariant representation learning can lead to the inclusion of spurious, label-irrelevant features, ultimately impairing performance in previously unseen operating conditions. To address these challenges, a causality and information-theory inspired domain generalization (CITDG) method is proposed. Specifically, CITDG employs a learnable binary mask to explicitly disentangle latent representations into label-relevant causal features and labelirrelevant non-causal features. To guarantee the sufficiency of information captured in each part, an information-theoretic causal sufficiency loss is introduced. Furthermore, a causal intervention purification module is proposed to remove domain-invariant label-irrelevant redundancy from the causal subspace while redirecting domain-specific label-relevant information into the causal features, mitigating potential competition or redundancy between causal and non-causal representations. To enhance the degradation characterization ability and generalizability of causal features, a causal feature factorization and alignment module is designed to enforce joint independence across the causal feature dimensions and align causal representations of the same subdomain across different source domains. Extensive experiments conducted on one real-world industrial dataset and two publicly available datasets demonstrate the effectiveness and superiority of the CITDG.

Recently, domain adaptation (DA) has been widely used in the remaining useful life (RUL) prediction of rotating machinery to effectively mitigate domain shift. Traditional DA methods for RUL prediction mainly focus on single-source domain adaptation (SDA) algorithms. However, labeled data can often be collected from multiple sources in practical scenarios. Directly applying SDA algorithms may degrade the model performance. Therefore, this paper proposes a novel multi-source adversarial distillation domain adaptation (MADDA) network for RUL regression problems. Specifically, a source feature extractor and regressor are pre-trained for each labeled source domain to capture source-specific representation. Then, a target encoder is learned to align target and source features via adversarial training to alleviate domain shift. Furthermore, a source distillation weighting mechanism is devised to utilize source samples that are more similar to target domains for fine-tuning the source regressor, thereby enhancing its performance on target tasks. Meanwhile, a source aggregation strategy is proposed to assign domain weights to the prediction results of various source regressors depending on the disparities between the source and the target domain, aiming to achieve the optimal combination of the final prediction. Case studies on two bearing datasets demonstrate the effectiveness and superiority of the proposed method.

With the rapid advancement of diagnostic technology, the ability to detect pathological areas such as tumors and polyps has significantly improved. This progress provides medical imaging specialists with more precise visual information to support anomaly identification, diagnosis, treatment planning, and patient monitoring. However, existing unsupervised and semi-supervised anomaly detection methods struggle with data privacy constraints, limited annotated medical datasets, and challenges in generalization. Zero-Shot Anomaly Detection (ZSAD), which enables the detection of unseen categories without requiring class-specific training, has emerged as a promising solution by leveraging the vision-language alignment capabilities of Vision-Language Models (VLMs), such as Contrastive Language-Image Pretraining (CLIP). Despite recent progress, ZSAD remains hindered by high noise levels, sparse targets, and poor adaptability in complex medical imaging scenarios. To address these issues, we propose a novel framework: DiffusionCLIP, a diffusion-based VLM for zero-shot anomaly detection in two-dimensional medical images. Specifically, DiffusionCLIP integrates diffusion models into the VLM to progressively denoise multi-level features extracted from the CLIP visual encoder, enhancing feature robustness and discriminability. A multi-level feature fusion strategy is designed to aggregate multi-scale representations from different depths of the visual encoder, ensuring complementary semantic alignment across layers. In addition, a dynamically modulated weight loss function is introduced to adaptively balance the learning of hard and easy samples, further improving model generalization. Extensive experiments on multiple benchmark medical imaging datasets, demonstrate that the proposed method significantly outperforms existing zero-shot anomaly detection approaches in terms of accuracy, robustness, and generalization.

This paper proposes a new infill criterion for the optimization of expensive black-box design problems. The method complements the classical Efficient Global Optimization algorithm by considering the distribution of improvement instead of merely the expectation. During the optimization process, we maximize a penalized expected improvement acquisition function from a specially collected infill candidate set. Specifically, the acquisition function is formulated by penalizing the expected improvement with the variation of improvement, and the infill candidate set is composed of some global and local maxima of the expected improvement function which are identified to be “mutually non-dominated”. Some conditions necessary for setting the penalty coefficient of the acquisition function are investigated, and the definition of “mutually non-dominated infill candidates” is presented. The proposed method is demonstrated with a 1-D analytical function and benchmarked using six 10-D analytical functions and an underwater vehicle structural optimization problem. The results show that the proposed method is efficient for the optimization of expensive black-box design problems.

This paper proposes a new latent variable Gaussian process modeling method for problems involving both qualitative and quantitative input variables. By exploring a full-dimensional latent space where the values of qualitative variables are appropriately represented by certain normalized points therein, we generalize the concept of distance to be applicable to qualitative variables and thereby adapt the mixed qualitative-quantitative inputs to any conventional quantitative-only correlation structures. Specifically, we present distinct treatments for ordinal and nominal qualitative variables. For each ordinal qualitative variable, the corresponding values are represented by intrinsically ordered points of a one-dimensional unit latent space. For each nominal qualitative variable, the corresponding values are represented by points residing in an axis-aligned latent space, with the dimensionality equal to the number of values, and each point anchored to a separate unit axis. The coordinates of the latent points, along with other hyper-parameters of the Gaussian process model, are jointly estimated via maximum likelihood estimation. An experimental study was conducted to compare four representative Gaussian process modeling methods for mixed qualitative and quantitative factor problems and the proposed method, using twelve analytical test functions from relevant literature and a dataset of cooling system noise. The effectiveness of the proposed method was validated by the results. We further demonstrated how to gain insights into the qualitative factors via the proposed LVGP-Full method.

With the extensive sensor data provided by the Industrial Internet of Things, data-driven remaining useful life (RUL) prediction methods are crucial for enhancing equipment reliability in industrial environments. To enhance prediction accuracy under unknown operating conditions (OCs), domain adaptation and domain generalization-based RUL prediction methods have emerged. However, when only a single-source domain is available, the lack of sample diversity, coupled with significant and unpredictable domain shifts (DSs) between the source and unknown target domains, hinders the model’s ability to generalize effectively to the unknown target domain. To address these challenges, a novel RUL prediction method based on adversarial contrastive learning for single-source domain generalization (ACL-SDG) under unknown OCs is proposed. First, a semantic embedding-based multi-pseudo-domain generation (SE-MPDG) module is designed, which generates diverse and valid pseudo-domain samples, guided by the developed subdomain-level supervised contrastive learning loss, subdomain continuity manifold regularization, and semantic consistency constraints to improve the model’s out-of-domain generalization capability. Subsequently, a domain-invariant feature-guided RUL prediction (DIF-RP) module is proposed to alleviate DS. This module compels the feature extractor to mine domain-invariant degradation features across different domains, constrained by label-level supervised contrastive learning loss. Finally, adversarial training is conducted between the SE-MPDG and DIF-RP modules to further enhance the diversity of pseudo-domains while ensuring the cross-domain invariance of degradation features. Extensive experimental validation of single-source cross-domain RUL prediction for one practical dataset and two public datasets, under unknown OCs, demonstrates the efficacy and superiority of the proposed method.

Complicated systems routinely serve to carry out a series of planned missions separated by predetermined breaks, with maintenance activities constrained by limited resources. To develop a more pragmatic maintenance model, a multi-mission selective maintenance optimization model is formulated, incorporating both learning and forgetting effects. It aims to reach the minimal phased-mission reliability maximization under maintenance time constraints. To address the proposed model, an enhanced artificial bee colony algorithm is specifically tailored. Applied to a real-life coal transportation system, the developed approach demonstrates superior effectiveness compared to two alternative methods. The experimental results provide well-informed and high-quality maintenance strategies for decision-makers.

The Industrial Internet of Things (IIoT) has greatly facilitated prognostics and health management of complex mechanical equipment by enabling seamless data collection from interconnected devices, providing rich data sets crucial for advancing data-driven prognostics methodologies. In data-driven prognostics field, multitask learning (MTL) is a prominent way to acquire robust remaining useful life (RUL) estimation models by extracting complementary degradation information from highly related auxiliary tasks. However, as the key component of MTL, the auxiliary task is often designed manually with substantial effort, while not considering how to build a reasonable auxiliary task that can provide more worthy information for the primary RUL estimation task. In response to the challenge, this article proposed a MTL approach with meta auxiliary generation network for RUL estimation. First, a novel meta auxiliary generation network is developed to perform auxiliary task design. The network can generate auxiliary task labels automatically. It is updated by meta learning strategy with the RUL estimation loss as the main objective function, ensuring that the performance of the RUL estimation task can be significantly enhanced. Additionally, soft labels are introduced to the automated design to reduce the label scale difference between regression and classification tasks, and a new label regularization strategy is designed for improving the degradation consistency between regression and classification labels. Experiments on the turbofan engine and wind turbine gearbox data sets verify the effectiveness and superiority of the proposed approach.