Vessel arrivals at maritime ports are often modeled by Poisson processes with independent increments, yet real traffic data often exhibits short-term autocorrelation. Using daily arrival data derived from the Automatic Identification System (AIS), we detect significant positive autocorrelation at short lags even after removing weekly seasonality. This temporal dependence may reshape port congestion dynamics, affecting how congestion emerges, persists, and dissipates. We treat port congestion as a threshold-exceedance recurrent failure event and assess port reliability under uncertainty through failure recurrence, congestion duration, and recovery dynamics, recognizing that sustained congestion itself creates additional risk by propagating delays, slowing recovery, and increasing vulnerability to further disruption. This connects the temporally autocorrelated arrival process directly to port resilience assessment and capacity planning under real traffic. To account for autocorrelation, we develop a delay-induced arrival model that combines independent per-vessel and fractional block-wise delay components to reproduce observed short-lag correlations. Discrete-event simulations of parallel berth queues show that the proposed fractional delay mechanism introduces two distinct statistical effects. Under moderate loads, reduced marginal variance smooths daily transit times and lowers congestion frequency. Under near-capacity loads, positive autocorrelation yields longer and more severe congestion with slower recovery. Thus, temporal dependence has a regime-dependent impact on port reliability: it stabilizes performance under moderate utilization, but amplifies tail risk near saturation by prolonging congested episodes and increasing their severity. Recognizing and monitoring such dependence is essential for reliability assessment under uncertainty and for designing capacity policies that reduce extreme congestion and recovery time.

Domain generalization (DG) in time series poses significant challenges due to domain shift, particularly under the strict DG setting, where no target domain data are available during training. To address this, we propose AFM-CIR, a unified framework that integrates semantic-similarity-guided Adaptive Fourier Mixing (AFM) with Causality-Inspired Regression (CIR). Specifically, we construct a domain-invariant order-preserving guidance embedding that drives a similarity-based adaptive modulation of amplitude mixing and a bounded shortest-angle phase perturbation, thereby generating label-consistent and causally coherent augmented samples. CIR then enforces invariance and inter-dimensional independence through correlation factorization, while causal sufficiency is encouraged via adversarial masking. We further provide theoretical guarantees of the controllability of phase interventions, supported by mutual information analysis and Lipschitz-spectral norm bounds. Extensive experiments on four widely used benchmark industrial data sets demonstrate that AFM-CIR consistently achieves state-of-the-art performance, outperforming strong ERM, general DG, and task-specific DG baselines.

This paper studies the data-driven design of a smart emergency response system for out-of-hospital cardiac arrest (OHCA) that involves drones for automatic external defibrillator delivery and community responders alerted via a mobile application, in addition to ambulances. Our study is motivated by the widespread exploration of drones for delivery service, and the emergence of mobile applications that crowdsource community for emergency response. Based on a historical OHCA dataset with community responders' response records from Singapore, we develop a robust joint deployment model of drone and ambulance to maximize the survivability of the response system while accounting for data uncertainty in OHCA occurrence and responder behavior. We discretize the planning area into finite demand regions, and allow different regions to have different OHCA demand rates, alert response probabilities and alert response time distributions from responders. Each of these attributes is only known to reside in an uncertainty/ambiguity set constructed from historical data. Our objective is to maximize the worst-case demand-weighted survival rate in the presence of uncertainty. We reformulate the resulting robust deployment model as a mixed-integer linear program, which can be efficiently solved by a proposed row-and-column generation algorithm with convergence guarantee. We illustrate our model and solution approach using real data from Singapore. We find that (i) hedging against uncertainty leads to a higher survival rate of the response system, compared to a sample average approximation deployment approach; (ii) while adding more drones/ambulances to the system exhibits diminishing return, a few drones are sufficient to increase the survival rate dramatically; and (iii) the impact of the behavior of responders on survival outcomes is more significant than that of simply adding drones/ambulances. We also discuss several managerial insights from the numerical experiments.

Rotating machinery often operates under time-varying conditions. This can result in nonstationary behavior that violates the stationarity assumption underlying many fault monitoring methods. Cointegration analysis can address nonstationarity but assumes that all monitored variables are linearly related and integrated of order one. To overcome these limitations, this study reformulates the cointegration model in a functional-coefficient form. The coefficients therein are parameterized using neural networks to capture nonlinear relationships and accommodate variables with arbitrary integration orders. An iterative Bayesian inference algorithm is developed to estimate model parameters, with a stationarity constraint imposed on the cointegration error during estimation. This constraint mitigates overfitting to noise and enforces a consistent cointegration structure across varying noise levels. A heuristic fault isolation strategy is further introduced to eliminate the need for manual regularization. The effectiveness of the proposed method is validated through numerical simulations and two real-world cases involving large-scale rotating machinery.

In engineering practice, it is common to observe simultaneous degradation of multiple performance characteristics in a system, in which these characteristics are correlated and exhibit differing degradation behaviors. This poses significant challenges to reliability modeling and analysis of multivariate degradation data. In this study, we propose a novel bivariate degradation model to meet the challenge. We employ the Wiener and inverse Gaussian processes to model the marginal processes, allowing for differing degradation patterns in the two dimensions. A shared frailty is then incorporated into the two marginal processes to capture their dependence structure. We derive the closed form of the reliability function for the proposed bivariate degradation model, and we develop an efficient Bayesian procedure for parameter estimation by combining the Gibbs sampler with the Metropolis-Hastings algorithm for posterior sampling. The performance of the Bayesian estimation method, along with the derived reliability formulas, is validated through comprehensive numerical simulations and a practical example involving a permanent magnet brake.

Agentic kernel optimization automates manual GPU kernel tuning via iterative generation, validation, and profiling with reasoning LLMs, casting the optimization task as feedback-guided search. However, our workload characterization reveals three system-level inefficiencies that limit search efficiency: (1) long generation latency due to LLM reasoning, (2) insufficient profiling feedback, and (3) underutilized validation/profiling resources. Our key insight is that the ongoing reasoning generation exposes a window for producing additional candidate kernels before it completes, allowing the system to terminate reasoning early once a satisfactory kernel appears. We present SpecGen, an agentic kernel optimization system with \emph{speculative generation}. First, SpecGen forks non-reasoning generations at well-chosen trigger points in the reasoning trace to yield kernels, increasing the candidate kernel count per iteration. These kernels are validated and profiled in parallel with the ongoing reasoning, increasing profiling feedback, and keeping resources busy during generation. When a kernel meets the termination criterion, SpecGen terminates the reasoning generation early to reduce the generation latency. Second, SpecGen dynamically reallocates validation and profiling GPU pools based on the arrival rate and prioritizes requests to reduce profiling feedback latency under bursty speculative generation load. Furthermore, SpecGen utilizes spare memory of the validation/profiling GPUs as remote KV cache storage to eliminate prefix recomputation of speculative generations under limited memory budget. Experiments with two reasoning LLMs on H200 show that SpecGen reduces end-to-end time over three baseline systems, while producing more profiling feedback, increasing resource utilization, and improving kernel speedup under a fixed time and token budget.

High-dimensional data with left-censored responses are increasingly common in modern applications, yet existing methods for analyzing them are limited. Classical Tobit models fail to handle nonlinear relationships or perform high-dimensional variable selection, whereas deep learning approaches often prioritize prediction performance but lack selection and interpretation capabilities. To address this gap, we propose an integrated deep learning framework, the Deep Tobit model, which employs the negative Tobit log-likelihood as its loss function to properly account for data censoring. A two-stage feature selection algorithm is further developed, with theoretical guarantees on convergence rate and selection consistency. Extensive simulation studies and real-data applications on left-censored aero-engine casing vibration data and HIV viral load data demonstrate that the proposed framework outperforms several state-of-the-art baselines in both variable selection and prediction accuracy.

We study multi-period risk-averse inventory control in a data-driven setting. In this problem, a risk-averse retailer makes periodic decisions on inventory levels based only on historical demand observations without full knowledge of the demand distribution. We adopt the popular nested formulation for risk-averse programs to formulate this multi-period problem and its data-driven counterpart under a coherent risk measure. Our objective is to study the sample complexity bound such that with high probability, the data-driven policy is near-optimal, that is, the relative error of risk under the data-driven policy compared with the optimal risk is arbitrarily small. Analysis of this problem is inherently challenging, because the multi-period nature requires solving the risk-averse program and its data-driven version recursively backward in time, while the (empirical) risk-to-go functions in this process do not have closed-form derivatives for most risk measures, which renders existing first-order methods for the risk-neutral newsvendor model invalid. In this study, we develop a zeroth-order framework to establish the complexity bound on sample sizes to guarantee near-optimality of the data-driven policy with given accuracy levels. Instead of using first-order derivative information on the risk-to-go function, our analysis directly examines the class of functions that underpins each cumulative risk function and derives maximum inequalities for this functional class by computing the covering numbers. Finite-sample complexity bounds are then used to establish asymptotic properties of the estimated risk, including consistency and convergence rate. Computationally, the time complexity for solving the data-driven policy, which is essentially an empirical dynamic programming (EDP) estimator of the optimal policy, increases exponentially in the length of the planning horizon. To speed up computation, we propose an approximation scheme that recursively approximates the empirical cumulative risk function with a convex piecewise linear function and then minimize it to obtain a modified data-driven inventory policy. We show that with proper control for approximation error, the modified data-driven policy is also near-optimal, and it has the same order of sample complexity bound as that for the original EDP policy.

Multinomial logistic regression (MLR) is a prevalent method for modeling categorical outcomes, but it often encounters issues with parameter nonidentifiability. To combat this, various models, such as reference-based MLR (RMLR), constraint-based MLR (CMLR), and simplex-based MLR (SMLR), have been proposed. When dealing with highdimensional groupwise covariates, the application of group Least Absolute Shrinkage and Selection Operator-penalized MLRs becomes essential. However, group RMLR presents challenges due to its subjective reference selection, whereas group CMLR is computationally intensive. We introduce the group SMLR (GSMLR) in this paper, emphasizing its efficient parameterization and computational advantages. GSMLR not only eliminates the need for reference selection seen in group RMLR, but also matches the results of group CMLR. We provide a definitive oracle error bound for GSMLR estimators and further optimize its performance with a rapid algorithm that combines groupwise coordinate descent strategy with enhanced safe screening rules. Our numerical results demonstrate that GSMLR excels in both prediction accuracy and computational efficiency, making it a valuable tool for high-dimensional data analysis and decision support in complex domains such as operations management and financial risk assessment.

Marine anomaly detection is critical for sea traffic safety, and the detection is mostly based on regularly transmitted data from the Automatic Identification System (AIS) installed in the vessel, which include location, velocity, course, and safety-related information. Most existing detection methods rely solely on the sailing vessel’s most recent AIS information without fully utilizing the historical AIS data from the same voyage, and they fall short of providing a comprehensive assessment of the vessel’s sailing state, particularly in terms of identifying deviations in its course or unusual accelerations. To fill the gap, we propose a novel two-stage profile monitoring framework for real-time anomaly detection to make full use of historical sailing information. The first stage aims to extract distinct routes from all the trajectories in the historical data, where a novel trajectory-based route extraction method is developed. The second stage then proposes an anomaly detection algorithm by treating each trajectory as a profile and carefully constructing the monitoring statistics and control charts. We compare the proposed trajectory-based method with existing approaches that rely solely on the most recent vessel information, using comprehensive simulations. The results reveal the superiority of our proposal in detecting anomalies, with fewer false alarms and the ability to detect a wider range of anomalous behaviors. An AIS dataset covering the Singapore Strait is used for illustration throughout the study.

Railway monitoring data frequently contain missing values due to hardware failures, communication breakdowns, and harsh operating conditions. This data incompleteness degrades diagnostic accuracy and undermines predictive maintenance reliability, posing critical challenges for prognostics and health management (PHM) systems. To address this challenge, an adaptive trend-seasonal conditional diffusion (ATSCD) model is proposed that integrates time series decomposition principles with diffusion-based generation. The framework comprises three core components. A global-local trend module (GLTM) combines parametric polynomial modeling for macroscopic evolution with local linear modeling for fine-grained fluctuations. Meanwhile, a learnable Fourier seasonal module (LFSM) adaptively identifies optimal frequency combinations to extract periodic patterns through end-to-end optimization. A multi-layer transformer denoising network (MLTDN) captures long-range dependencies via the encoder while progressively decomposing and synthesizing trend-seasonal features through the decoder. During the denoising process, local observations serve as constraints at each iteration via Langevin dynamics, guiding reconstruction in missing regions. The model is validated on heterogeneous train bogie monitoring data including current, voltage, vibration, and acoustic signals. ATSCD outperforms state-ofthe-art imputation methods across all metrics. In downstream PHM tasks, the imputed data yield higher diagnostic accuracy than incomplete data, demonstrating practical effectiveness for railway condition monitoring.

For highly reliable products that degrade over time, it is not uncommon to observe an initial period during which no failure occurs. In this paper, a degradation model that gives rise to a first passage time distribution with failure-free life is proposed as a more plausible model compared to many common models. We call it Beta process as its increments are approximated by three-parameter Beta distributions. We start with a discrete time process as a degradation process is commonly measured at regular time intervals. We derive a closed-form first passage time distribution under a constant failure threshold. Closed-form maximum likelihood (ML)-type estimators are developed for the parameters of the Beta process. Then, the discrete time process is extended to the case of a continuous time process. Comprehensive simulations and case studies show that the Beta processes outperform many common degradation processes in reliability estimation.

Extended warranties (EWs) constitute a significant source of revenue for capital-intensive products. Such products comprise multiple subsystems, enabling flexible EW design. For example, providers can bundle tailored sets of subsystems within different EW contracts, facilitating the creation of a service menu with differentiated warranty options. From the perspective of a third-party EW provider servicing multi-tier products, we develop a novel model to jointly optimize bundle design and pricing for EW options in order to maximize the expected total profit. Specifically, the problem involves determining which contracts-each containing a differentiated bundle of subsystems-to recommend for the multi-tier products and identifying the appropriate price for each contract. As the complexity of the joint optimization problem increases exponentially with the number of subsystems, we devise two solution approaches. The first approach leverages a mixed-integer second-order cone programming reformulation, which guarantees optimality but is applicable only for a small number of subsystems. The second approach utilizes an iterative two-step process, offering enhanced computational efficiency for scenarios involving a large number of subsystems. Numerical experiments validate the effectiveness of our model, particularly in scenarios characterized by high failure probabilities and a large number of subsystems.

Mitigating climate change requires the food industry to substantially reduce its greenhouse gas emissions. Food warehouses are key contributors to the industry’s carbon footprint, as they consume large amounts of energy for refrigeration. Their carbon emissions can be reduced by upgrading refrigeration units to energy-efficient inverter-based ones or by purchasing green electricity. However, long-term upgrade decisions face uncertainties such as evolving climate policies, energy efficiency targets, and varying electricity prices. To capture these uncertainties, this paper develops a two-stage stochastic optimization model that jointly optimizes refrigeration unit upgrades and green electricity procurement for food warehouses. The first stage determines near-term refrigeration unit upgrade schedules and green electricity purchase plans, while the second stage addresses long-term decisions. The resulting model is a large-scale mixed-integer program. To enhance computational efficiency, we tailor an L-shaped algorithm with linear warm-start and relaxation filtering techniques, achieving a fivefold speed improvement over solving the monolithic model and avoiding memory overflow.

Hybrid parallelism underpins large-scale LLM training across tens of thousands of GPUs. At such scale, hardware failures on individual devices lead to performance skew across devices, diminishing overall training efficiency. Existing resilient systems overlook sequence length variability in datasets and device performance skew under hybrid parallelism. As a result, (1) iteration time fluctuations induced by sequence length variability can trigger spurious fail-slow detections, and (2) failures are mitigated through individual adaptations in hybrid parallelism, leading to unnecessary detection overhead and inefficient resilient training. To respond, this paper presents ResiHP, a resilient system that enables robust failure detection and fine-grained adaptation for hybrid parallel training. First, we develop a Detector to accurately identify failures. In particular, it employs a workload-aware execution time predictor that disentangles failures from iteration time fluctuations while remaining lightweight for online detection. Second, we design a Scheduler that dynamically adapts parallelism group sizes, model partitioning, and workload scheduling policies to improve training efficiency under failures. Experiments show that ResiHP improves training throughput by 1.04-4.39× compared with state-of-the-art resilient training systems under diverse failure scenarios in a 256-GPU cluster.

Many remaining useful life (RUL) prediction models rely on neural networks. These models focus on predictive accuracy at the component level, but their impact on system-level maintenance remains underexplored. Quantifying this effect is difficult because maintenance decisions are often optimized from long-term simulations. These simulations do not provide real-time measurements for RUL prediction, making direct use of specific prognostic models challenging. To address this gap, we propose a maintenance optimization framework that uses generative modeling with a generative adversarial network (GAN) to create realistic prognostic trajectories for system-level cost assessment. Feature adaptation is added to the GAN to ensure that generated trajectories retain key statistical properties, such as accuracy, timeliness, and credibility, seen in actual model evaluations. This framework allows the benefits of prognostic models to be assessed probabilistically and offers guidance on whether to include prognostics in maintenance planning. A case study on wind farm maintenance demonstrates the effectiveness of the method. Although the wind farm example considers only limited subsystem dependencies, the framework is not restricted to this configuration. The approach is flexible and can be extended to systems with more complex dependencies by adjusting maintenance policies.

GPU checkpointing and restore promise to enable emerging tasks, such as deep learning, to benefit from functionalities like task scheduling and fault tolerance. However, existing GPU checkpointing/restore solutions suffer from runtime overhead, bloated checkpoint images, and correctness issues. This paper presents FlowGPU, a system-level GPU checkpointing/restore mechanism that overcomes all aforementioned limitations. Our key insight is that the limitations of prior mechanisms implicitly stem from their architectural design, which tightly couples checkpointing/restore with a legacy virtualization technique: API forwarding. In response, the design of FlowGPU decouples checkpointing/restore from virtualization with two key techniques: per-task interception and ghost process, thereby overcoming these limitations. Furthermore, FlowGPU comes with a set of novel techniques to further improve performance and ensure correctness under complex scenarios, such as a task operating on multiple GPUs. Our evaluation shows that FlowGPU outperforms prior mechanisms by up to 4.5 × .