In recent years, aging equipment in urban rail transit (URT) systems has led to frequent disruptions. Bus bridging services are an effective means of responding to such disruptions. Although prior studies have proposed bus bridging strategies to convert existing bus routes into bridging services, these strategies are inflexible with limited coverage. To address these limitations, this study proposes a novel bus bridging method that not only enables bus stops adjacent to the disrupted URT stations to directly provide bridging services, but also introduces two complementary bridging strategies. The first strategy involves fine-tuning partial segments of existing bus routes without skipping any regular stops, allowing detours to serve the disrupted URT stations, namely F-BB. The second strategy extends routes from the terminal of the existing bus routes to detour to the disrupted URT stations, namely E-BB. To accommodate time-varying travel demands, a time-segmented optimization (TSO) method is adopted to adjust the bridging strategies across different periods. A passenger flow control strategy is further developed to assign stranded passengers to designated waiting areas of bridging routes. A mixed-integer nonlinear programming (MINLP) model is formulated to jointly consider the travel demands of regular and stranded passengers, which is solved using a rolling-horizon-based adaptive large neighborhood search (RH-ALNS) algorithm. The proposed model and algorithm are validated through a case study in Chengdu, China. Numerical results indicate that, compared with four scenarios excluding E-BB, F-BB, TSO, and all bridging strategies, the proposed method reduces total travel time by 0.26%, 55.62%, 1.67%, and 89.92%, respectively.
The rapid growth of on-demand e-commerce has shifted logistics toward fragmented, time-sensitive orders, often forcing carriers to prioritize delivery speed over vehicle utilization, resulting in low loading rates and high empty mileage. While the Physical Internet (PI) provides a foundation for collaboration, existing mechanisms remain hub-centric and treat each transport unit as indivisible, leaving the upstream order-grouping phase unoptimized and the idle capacity within partially loaded units untradeable. To bridge this gap, this research proposes an order-to-collaboration framework within the Cyber-Physical Internet (CPI) network. The framework integrates three components. First, an Out-of-Order (OOO) clustering rule consolidates fragmented orders at the intake stage. Second, a three-state Protocol Shipment Unit (PSU) loading classification links each loading condition to a specific carrier action: direct routing, auctioning idle capacity, or bidding for external capacity. Third, a multi-attribute combinatorial auction prices intra-PSU idle capacity using the loading, routing, and goods tables of CPI cyber space. Slot discretization substantially reduces the admissible bidding space by restricting bids to contiguous slot bundles, thereby reducing the size of the allocation problem and supporting real-time tractability in the tested configurations. Large-scale numerical experiments on a Greater Bay Area logistics network show that the framework raises the average PSU loading rate to 0.81 and reduces transportation cost by 29% and CO2 emissions by 23%, while achieving a 71% auction success rate with a 32% revenue gain for auctioneers and 28% cost saving for bidders. Response surface analysis further identifies an effective multi-objective operating region for joint parameter tuning. The framework advances theoretical understanding of carrier collaboration in CPI and offers a structured mechanism with potential to support efficient, profitable, and low-carbon on-demand logistics.
Recent technological advancements have facilitated various initiatives in the last-mile location routing problem (LM-LRP), yet existing LM-LRP literature largely evaluates these initiatives in isolation and lacks a unified framework. This study aims to address this limitation by conducting a systematic literature review of emerging initiatives in LM-LRP to synthesize their technological and operational implications. First, transportation nodes and modes are categorized to show how emerging initiatives change the spatial configuration of logistics facilities and affect sustainability dimensions, while increasing coordination requirements across LM-LRP networks. The literature is then assessed based on key LM-LRP features, including distribution network structure, objective-function design, and last-mile applications. A digital twin perspective is further proposed as a unified decision-support system to connect LM-LRP components with data-driven modeling and simulation approaches. Sustainability trade-offs are subsequently examined across emerging LM-LRP initiatives, showing that their economic, environmental, and social benefits remain context-dependent. Key research gaps are identified, including limited operational realism, fragmented technology integration, and insufficient governance and policy considerations. Based on these findings, a forward-looking framework is proposed to summarize initiative adoption and future development pathways. Finally, this review provides actionable research insights that can facilitate the implementation of emerging initiatives for more efficient last-mile distribution networks.
With the rapid expansion of e-commerce logistics, delivery delays have become a critical factor affecting fulfillment efficiency, customer satisfaction, and supply chain coordination. Existing delay prediction methods usually rely on centralized training, but in real logistics systems, order and fulfillment data are often distributed across platforms, warehouses, carriers, and regional delivery nodes. Due to privacy constraints, data ownership, and cross-organizational boundaries, raw data are difficult to aggregate, resulting in a typical data-silo problem. To address this challenge, this study develops a federated learning framework for logistics delay prediction based on the public DataCo Supply Chain dataset and constructs a non-IID multi-client environment to simulate heterogeneous logistics nodes.To improve client scheduling under heterogeneous data distributions, this study proposes a multi-dimensional fusion client selection strategy that jointly considers clients’ instant improvement potential, update-direction consistency, and historical participation. A compensation memory mechanism is further introduced to mitigate long-term client selection bias. Experiments are conducted using an Embedding-MLP local learner and compared with centralized training, local-only training, FedAvg, FedProx, and Power-of-Choice baselines. The results show that the proposed method achieves the best average performance on most key federated metrics, including Accuracy, AUC, Macro-F1, delayed-class recall, and delayed-class F1. Statistical tests further indicate that the proposed method significantly improves AUC compared with FedAvg and significantly improves client participation balance compared with Power-of-Choice. Convergence analysis also shows that the proposed method reaches key performance thresholds earlier than the baselines.Overall, the experimental results suggest that federated learning with multi-dimensional client selection offers a promising approach for cross-organizational logistics delay prediction under privacy constraints, with the potential to simultaneously improve predictive performance, convergence efficiency, and client participation balance.
Motivated by the monitoring of automobile complaint data, quality monitoring methods require effective models for analyzing over-dispersed integer-valued time series (INTS) with network dependence and heterogeneous node-level dynamics. Existing count time series models often have limitations in simultaneously characterizing over-dispersion, network interactions, group heterogeneity, and covariate effects, which may reduce their effectiveness in monitoring structural changes in complex networked count processes. To address these challenges, we propose a Grouped Negative-Binomial Network Auto-Regressive (G-NB-NAR) model, in which network nodes are divided into groups and group-specific parameters are introduced to characterize heterogeneous dynamic patterns. The proposed model adopts the negative binomial distribution to accommodate over-dispersion, uses an adjacency matrix to describe network dependence, and incorporates covariates to account for additional explanatory information.The stationarity and ergodicity of the G-NB-NAR model are established, and the consistency and asymptotic normality of the maximum likelihood estimator are derived. For the online monitoring, the fitted model is combined with a top-q CUSUM scheme to detect sparse or localized structural changes in the over-dispersed network count data. Extensive simulations show that the proposed framework performs well in both parameter estimation and change detection. Further ablation experiments and robustness analyses demonstrate that the grouping structure, network dependence, and covariate information all contribute to the monitoring performance, and that the proposed framework remains stable under alternative grouping schemes, mis-specified adjacency structures, unequal group sizes, different network densities, different choices of q in the CUSUM monitoring scheme, and various shift patterns. Finally, an empirical study based on the automobile complaint data illustrates the practical usefulness and interpretability of the proposed method.
