This study proposes a multi-stage maintenance optimization framework for offshore wind farms under stochastic and fuzzy uncertainty. Unlike purely statistical approaches, it embeds physical degradation mechanisms—Miner's fatigue rule, Arrhenius thermal aging, and a power-law corrosion model—directly into a Markov Decision Process, preserving physical traceability. Component failures follow physics-calibrated Weibull distributions, while epistemic maintenance-cost uncertainty is handled via a 1-Wasserstein Distributionally Robust Optimization (DRO) framework that guarantees performance without exact probability assumptions. A multi-objective model minimizes worst-case cost, maximizes reliability, and minimizes downtime, solved by a two-layer exact method combining AUGMECON2 for Pareto-front generation with an inner Mixed-Integer Linear Programming (MILP) reformulation. Validated on synthetic data calibrated to Taiwan Strait conditions, the framework consistently outperforms Fuzzy Goal Programming, standalone MDP, and Genetic Algorithm benchmarks, confirming the value of physics informed, distributionally robust O&M decision-making.
While multi-kernel correntropy, as a powerful local similarity measure, has demonstrated superior robustness against non-Gaussian noise and outliers, existing approaches are limited by their reliance on homogeneous kernels and suboptimal treatment of free coefficients. To address this gap, this paper introduces the concept of heterogeneous multi-kernel correntropy (HMKC), which employs a mixture of two Gaussian Versoria functions with flexible, non-zero mean values as its kernel. Building on this criterion, we propose a unified unscented Kalman filter (UKF) framework, the robust heterogeneous multi kernel maximum correntropy UKF (R-HMKMC-UKF), designed to enhance both estimation performance and numerical stability, thereby providing reliable state estimates for state feedback control. The framework incorporates an enhanced particle swarm optimization algorithm to calculate and select free coefficients automatically. Simultaneously, to enhance numerical stability, a critical requirement for reliable implementation, we incorporate Moore-Penrose pseudoinversion and hyperbolic QR decomposition, which effectively address ill-conditioned matrix inversions and prevent loss of positive definiteness in covariance matrices. Finally, we validate the excellent accuracy and numerical stability of the proposed R-HMKMC-UKF in two challenging applications: IEEE 30-bus power system state estimation and land vehicle navigation, under multimodal noise and severe measurement outliers.
This paper investigates the problem of adaptive event-triggered (ET) formation collision avoidance control for high-order nonlinear multi-agent systems (MASs) under denial-of-service (DoS) attacks. First, a distributed resilient observer is designed to estimate the information of the leader when the communication is interrupted by DoS attacks. Subsequently, a distributed command governor is proposed to dynamically generate safe reference commands in real time. Moreover, an ET command regulation scheme is designed to ensure system safety performance while effectively saving communication resources from the controller to the actuator. Consequently, the adaptive ET resilient formation collision avoidance control scheme is formulated. Theoretical analysis demonstrates that the designed control scheme enables agents to accomplish formation collision avoidance even under DoS attacks. Finally, the effectiveness of the designed control scheme is verified by simulation results.
Bearing diagnosis under simultaneous operating-condition and device shifts is difficult because fault signatures, load-dependent responses, and sensor/platform effects are entangled. We propose TF-RiNO, a task-factorized Riemannian neural operator framework for cross-speed, cross-load, and cross-device bearing diagnosis. Its backbone is a finite-dimensional one-dimensional Fourier operator network: low temporal modes are transformed by learned complex channel maps, while a local convolution branch preserves transient content. We therefore use “neural operator” in this architectural sense and do not claim discretization-convergent approximation of an underlying PDE solution operator. The condition-aware features are lifted to an $SPD\times$Grassmann product representation that combines second-order statistics with basis-invariant dominant-subspace information. A task factorization head then separates a fault-shared embedding from a task-associated nuisance embedding; the orthogonality term is treated as a decorrelation surrogate rather than proof of identifiable disentanglement. Across five runs on CWRU and HUST, TF-RiNO attains strong cross-speed and cross-load results and improves over the strongest reported baseline on both HUST cross-device tasks. Welch tests with Holm correction support the gains for the 0-W and 200-W device-transfer tasks ($p_{\text{adj}}=0.0063$ and 0.0196), whereas two of the three cross-load gains are not statistically significant. A block-disjoint, source-selected rerun obtains $0.872\pm 0.036$ and $0.632\pm 0.055$ accuracy on the easy and hard device transfers, respectively. Embedding visualizations further show that the hard 200-W device shift reduces class separation and increases class-conditional domain discrepancy, with outer-race faults forming the principal failure mode.
To enhance the cyberattack resilience, performance optimization, and communication efficiency of multi-player Nash equilibrium (NE) seeking, this paper develops an adaptive fuzzy secure NE control framework for nonlinear multiagent systems (MASs) using reinforcement learning (RL). The multilink cy berattack scenario considered herein involves a composite threat that integrates denial-of-service and false data injection attacks. To mitigate the adverse effects of such attacks, the proposed strategy proceeds through three main components. First, a fuzzy state observer is constructed to reconstruct unmeasurable states corrupted by cyberattacks. Second, a distributed NE estimator is introduced to address the non-differentiability issue of intermit tent signals. Third, an optimized secure controller is designed, ensuring the NE seeking objective of the MASs. Leveraging fuzzy logic system approximations, RL is incorporated within an actor-critic framework to achieve optimal control. Furthermore, a logarithmic quantizer is embedded to realize a dynamic event triggered quantization mechanism, thereby reducing communication load. Finally, the effectiveness and resilience of the proposed method are demonstrated through simulations of three marine surface vehicles.
Due to manufacturing techniques, the initial performance characteristic (IPC) is often correlated with the degradation rate. Most existing degradation modeling methods address this issue by assuming a linear dependence and a normally distributed degradation rate. However, these assumptions overlook potential nonlinear dependence and inherent asymmetry in the degradation rate, leading to biased parameter estimation and reliability assessment results. To fill this gap, a novel degradation modeling method is developed, in which a candidate copula function is employed to characterize the dependence between the IPC and the degradation rate, while the degradation rate is allowed to follow a candidate one-sided skewed distribution. On this basis, a modified Wiener process-based model is constructed, and the corresponding reliability function is derived. Maximum likelihood–based statistical inference is then performed to identify the model structures and estimate the model parameters. Furthermore, the effectiveness of the proposed method is demonstrated through a comprehensive simulation study and two practical applications.
The complexity of automotive wiring harness systems increases constantly. Upcoming automated driving systems (ADS) and X-by-Wire technologies increase the Safety-Related Availability (SaRA) requirements of associated systems, as the driver cannot act as an immediate fallback. In particular, short-circuit events in lower-ASIL or QM load paths may cause an undervoltage in parallel safety-relevant paths, thereby potentially violating Freedom from Interference and SaRA requirements according to ISO 26262. This paper proposes a simulation-based method to systematically analyze interference caused by short-circuit events in automotive wiring harness topologies. The method combines iterative fault injection into potential short-circuit paths, Monte-Carlo-based parameter screening, and a subsequent classification of influencing parameters. The effectiveness of the method is demonstrated using a simplified wiring-harness topology modeled in MATLAB/Simulink. The topology comprises a safety-relevant path containing a steering system, four potential short-circuit paths, and a base load representing the remaining consumers. The investigated parameters include ambient temperature, 12 V battery state of charge, short-circuit resistance, aging condition, and fuse tripping characteristics. A Monte Carlo study with 160 simulations was used to identify statistically significant parameter influences on fuse tripping time and minimum voltage. The results show that the proposed method can identify critical short-circuit paths, derive worst-case parameter combinations, and distinguish between relevant and non-relevant interference scenarios. This method could complement conventional safety analysis and lead to improved design efficiency while supporting ISO 26262-compliant Freedom-from-Interference argumentation.
Analyzing the spread of malicious information in social networks is crucial for network security. This paper proposes a malicious information propagation model based on Generative Adversarial Network (GAN) and sparse representation to address the challenges posed by the temporal dynamics of malicious information propagation, data sparsity and incompleteness, and the complexity of the propagation space. Firstly, to tackle the dynamics and timeliness of malicious information, the lifecycle of malicious information is divided into time slices. Static and dynamic features within each time slice are extracted and combined into comprehensive user features to construct a propagation space for malicious information. Secondly, to address the issue of missing samples, we utilize GANs for homomorphic compensation of data and integrate an attention mechanism into the GAN to enhance feature learning and improve the handling of missing data. Finally, to manage the complexity of the propagation space, we leverage sparse representation to capture essential characteristics, combining it with Graph Convolutional Networks (GCNs) to develop a Dynamic Sparse Representation based GCN (DSR-GCN) model for predicting user propagation behavior. Experiments show the model effectively captures malicious information propagation trends.
Account lockout policies protect password-based services from brute-force attacks but can be weaponized for denial-of-service (DoS) against legitimate users. Recovery from such lockouts on Linux services that authenticate through Plug gable Authentication Modules (PAM) — Secure Shell (SSH), File Transfer Protocol (FTP), databases, custom daemons — typically requires administrator intervention or passive expiry of the lock out window, because punitive PAM modules such as pam faillock provide no self-service remediation path. We present the Account Lock Protection System (ACCLPS), a drop-in PAM module that adds a self-service recovery channel to PAM-mediated services on Linux without modifying the protected application. ACCLPS does not propose a new authentication primitive; its contribution is the system-level integration of three established components — email-delivered recovery links, JSON Web Token (JWT) bound temporary credentials, and PAM enforcement — into a minimally-stateful remediation layer for the specific gap left by lockout-prevention system. Because recovery is delivered over email, ACCLPS elevates the user's mailbox to the trust anchor of the recovery flow. We evaluate a prototype implementation on a t2.micro-class Linux host (Amazon Linux 2023, Ubuntu 22.04, CentOS Stream 9) under controlled workloads. The prototype adds approximately 4% CPU during lockout detection (leq0.3% in steady state) and a 30 MB resident footprint, sustains 50 concurrent recovery flows on a single vCPU, and reduces median recovery time from a 15-minute unlock time floor to 0.78 minutes in our test harness. ACCLPS does not claim broad operational security across heterogeneous deployments; the evaluation is sized for the legacy and resource-constrained PAM environments ACCLPS targets and is bounded by the assumptions enumerated in the security model.
Cross-domain fault diagnosis plays an important role in industrial production. Current methods are able to achieve promising performance when both source and target domains are available at training time. However, when oriented towards fault diagnosis where the target domain is unseen, they often fail to effectively extract domain-invariant features and suffer severe performance degradation. To mitigate this problem, we propose a Suppressing Domain-Sensitive Feature (SDSF) framework for fault diagnosis under unseen conditions. A novel masking domain-sensitive feature module is constructed to emphasize the domain-invariant components in features and close the domain gaps. A tuning domain-sensitive feature module is further introduced to reduce the effect of domain-sensitive features extracted from different domains on the classifier. In addition, a consistency constraint is proposed to constrain the classifiability of features. Extensive experiments on multiple benchmark databases demonstrate that our method is more robust and efficient in dealing with both cross-working condition and cross-machine fault diagnosis under unseen conditions compared to other methods. The codes will be released on https://github.com/vilab-code/SDSF.
Fully leveraging spatio-temporal correlations in multi-sensor data is crucial for achieving accurate fault diagnosis of mechanical equipment. Although spatio-temporal graphs (STGs) have demonstrated potential in modeling such correlations, existing methods still suffer from limitations such as a single-perspective feature extraction paradigm and a lack of prior knowledge guidance, which consequently restrict their performance. To address these challenges, this paper proposes a novel prior-aware spatio-temporal graph neural network (PA-STGNN). First, an adaptive multi-domain feature extractor is designed to fuse complementary information from multiple domains, thereby generating more informative node representations. Subsequently, a prior-aware spatio-temporal graph construction strategy is proposed. Specifically, an initial spatio-temporal graph is constructed to capture both spatial correlations via a Gaussian kernel k-nearest neighbors (k-NN) approach and temporal correlations using a multi-head attention mechanism. Furthermore, a prior matrix is designed to inject physical prior knowledge about temporal decay and sensor autocorrelation into the graph to optimize its topology. Finally, Chebyshev graph convolution (ChebConv) is employed to aggregate information from multi-hop neighborhoods, enabling the effective use of complex spatio-temporal correlations. Comprehensive experiments on a public dataset and a real-world coal mill dataset demonstrate that PA-STGNN achieves superior diagnostic accuracy, consistently outperforming state-of-the-art baselines, particularly in few-shot scenarios.
Deep learning has recently emerged as a promising method for fault diagnosis in industrial equipment. However, due to changes in production demands, industrial equipment typically exhibits multimode characteristics, which naturally results in the deep learning-based diagnosis methods designed for a single mode hardly working. Moreover, deep learning models are essentially black boxes, which are opaque in terms of interpretability. To tackle these problems, an interpretable fault diagnosis scheme is proposed. First, an improved autoencoder (AE) is constructed for feature extraction using the algorithm unrolling technique, which can be regarded as having a clear theoretical basis and an interpretable network architecture. Then, based on the extracted features, a mode identification model is designed, and its identification capability is enhanced by incorporating a discriminative mechanism. For the identified mode, an improved adversarial autoencoder with a memory module (AAEMM) is further designed to detect faults by jointly considering reconstruction error and feature distribution. Finally, an aero-engine model on the public T-MATS platform is provided to verify the effectiveness of the proposed method.
To address the uncertainty of fault signals in rotating machinery under strong noise and complex operating conditions, as well as the limited robustness of existing diagnostic models, this paper proposes an adaptive dual-branch feature enhancement network (ADB-FENet) with noise robustness and uncertainty awareness. The framework includes both a topological branch and a temporal branch. In the topological branch, a dynamic weighted progressive feature extraction (DWPFE) mechanism enhances feature representation, and a multi-criterion graph (MCG) is constructed to capture structural relationships among samples, followed by a graph information refinement network (GIRN) to strengthen graph representations. In the temporal branch, an uncertainty-aware enhancement module (UAEM) based on evidence theory models intrinsic uncertainty in vibration signals, improving robustness under noisy conditions. A shared cross-feature fusion (SCFF) module further promotes complementary interaction between the two branches. Experimental results demonstrate that the proposed method achieves superior diagnostic accuracy and adaptability compared with existing approaches, maintaining strong robustness under severe noise. Additionally, the proposed measure of uncertainty can be used to identify new classes of faults unseen before.
Partially observed systems are common in practice, where the exact operational state of each component is inaccessible or obscured, yet aggregate functioning component numbers can be obtained by inspection or monitoring. Existing methodologies that incorporate observation data for reliability evaluation often lack the versatility for diverse system architectures and component interdependencies. To bridge these gaps, this study establishes a reliability updating framework applicable to general system configurations and dependency structures. This approach fully leverages partially observed order statistics data – specifically aggregate survival counts. By introducing the concepts of conditional ordered lifetime distributions and a general survival signature, this framework facilitates the derivation of closed-form reliability expressions for partially observed systems. An illustrative example is presented to elucidate the methodology and to demonstrate the generality, interpretability, and practical applicability of the proposed approach.
Credible diagnosis at the system-level is essential for maintaining the reliability and the resilience of large-scale multiprocessor systems. To characterize the fault diagnosis capability for different requirements, various conditional diagnosabilities, including $g$-extra diagnosability and $r$-component diagnosability, have been proposed. So far, most existing research investigated the diagnosability of a system under a single isolated condition. In this paper, we determine the system's diagnosability with both the $g$-extra and $r$-component conditions required simultaneously. We first establish a novel relationship between $g$-extra $r$-component compound connectivity and $g$-extra $r$-component compound diagnosability (referred to as the $g$E–$r$C relationship) under PMC model. We then apply the obtained relationship to find out the $g$-extra $r$-component diagnosability for $S_{n}$ and $S_{n,k}$. We also review and disprove some earlier results, reported in [19] and [28], relating component connectivity and component diagnosability, by providing counterexamples and identifying the flaws in the proofs. Finally, we present a diagnostic algorithm, named DAPMC, which is tailored for the $g$-extra $r$-component compound condition. To evaluate the effectiveness and robustness of DAPMC, we have conducted experiments on star graph $S_{n}$, as well as a real-world network, the web-polblogs dataset. The experimental results show that DAPMC can diagnose faulty nodes with an accuracy greater than 95%, even if there are up to 50% faulty nodes present in the system.
The rapid evolution of information technology has led to a surge in radio frequency devices, exposing communication systems to multi-source complex electromagnetic environments (EMEs) and making reliability assessment crucial. However, existing methods face two key limitations: system link modeling struggles to represent real dynamic interference, while analytical assessments based on “sampling–playback” of real EMEs waveforms require excessive testing resources and contain significant feature redundancy, failing to isolate the core factors causing device performance degradation. From the perspective of interference effect equivalence, this paper proposes a host–guest generative cooperative reliability assessment framework based on a Generative Cooperative Network (G-CoopNet). First, a baseline signal construction method via feature distillation is developed, where multi-objective genetic algorithms refine massive heterogeneous emitter waveforms into four high-coverage baseline signals, enabling effective dimensionality reduction and EMEs characterization. Second, by incorporating real test data, the actual bit-error-rate (BER) degradation inversely constrains environmental modeling to accurately characterize the host device's response differences. Finally, G-CoopNet performs host–guest cooperative training to transform real EMEs into Electromagnetic Environment Configuration Graphs (EECGs), which preserve the performance degradation effects of real interference on the host device while removing redundant environmental features. Experimental results show that the proposed method improves reliability assessment accuracy by 5.4%, providing a new intelligent paradigm for low-cost, high-efficiency laboratory equivalent retesting and system reliability assessment under multi-source interference.
Vehicle battery reliability progresses through two stages: dealer storage and customer usage. Attrition may occur during storage under a replace-as-new policy, and ultimate failures are observed during use. Thus, effective assessment requires modeling both stage progression and stage-specific failure. In the reliability literature, however, two-stage analyses often rely on a shared frailty model or focus solely on the terminal stage, thereby obscuring early failures and cross-stage dependence. In this research, we propose a two-stage correlated frailty framework with transition-specific random effects, allowing estimation of cross-transition covariance between progression and failure risks. Model parameters are estimated semiparametrically using partial likelihood. To address computational complexity, a piecewise-exponential generalized linear mixed model (PEM/GLMM) is developed as a computationally simpler approximation. A case study on a real-world production cohort of vehicle batteries with time-stamped transitions, failures, and censoring shows that the proposed framework improves calibration and discrimination over shared-frailty and single-stage baselines. It further reveals a significant cross-stage correlation, underscoring its practical value and generalizability to other two-stage reliability settings.
Rolling bearings operate under varying working conditions, posing a significant challenge to achieving high-accuracy bearing fault diagnosis. To enhance the fault diagnosis performance of rolling bearings under cross-domain working conditions and noisy environments, this study proposes a similarity-based multiscale cluster region transfer network (SMCRTN), which integrates three core modules: similarity-based processing (SBP), multiscale concatenation U-Net (MCU-Net), and dynamic cluster region transfer (DCRT). Specifically, the SBP module selects source domain samples using entropy score and anomaly score, and optimizes target domain samples through statistical projection. Meanwhile, the MCU-Net incorporates Hilbert-transformed inputs, gated convolution (gated-conv) blocks, and standard convolution blocks to extract multiscale domain-invariant features via dynamic weight adjustment. Furthermore, the DCRT module achieves cross-domain alignment by leveraging cluster-based region transfer and minimizing dynamic entropy-weighted loss. To verify the feasibility and effectiveness of the proposed SMCRTN, comprehensive experiments are performed on the public CWRU dataset and the proprietary PT dataset. Experimental results under various transfer tasks and noise levels demonstrate that the SMCRTN outperforms other intelligent models in terms of diagnostic accuracy and transferability.
Universal Serial Bus (USB) Human Interface Devices (HIDs) remain vulnerable to adversarial keystroke injection attacks because host systems implicitly trust connected peripherals. Recent poisoning and evasion attacks further reduce detection reliability by mimicking benign human typing patterns, creating overlapping benign and malicious USB traffic distributions. This paper proposes LITE-RATE, a reliability-oriented lightweight transformer framework for real-time adversarial USB keystroke injection detection. LITE-RATE integrates kernel-level USB traffic acquisition, sliding-window feature representation, lightweight transformer models, layer-wise pruning, few-shot adaptation, and dual-mode explainability to jointly address detection accuracy, latency, robustness, and transparency. Experimental results show that LITE-RATE achieves up to 92% detection accuracy with an average inference latency of 10.64 ms, demonstrating its suitability for real-time deployment on resource-constrained host systems.
