Highway tunnels are critical transport assets in which fires can cause cascading failures, mass casualties, and long service interruption. Traditional point-sensor and video-based systems frequently fail to provide timely early warnings, as they lack the sensitivity required to detect incipient fires under the complex lighting and visibility constraints of urban tunnels. This study develops a real-time vision-based warning module using two lightweight convolutional neural networks trained on a mixed dataset consisting of 281 Internet tunnel-fire images, 1552 open-source vehicle-fire images, and 592 small-scale tunnel-fire experimental images collected in collaboration with Jiandun Fire Technology Co., Ltd. The dataset is divided into training, validation, and testing subsets at an approximately 8:1:1 ratio, and the positive test samples are restricted to tunnel-fire images to evaluate tunnel-specific generalization. The lightweight models are compared with deeper CNN baselines in terms of both predictive performance and computational cost. On the tunnel-focused test set, the best model achieves 98.75% accuracy, 98.0% recall, and 8.24 ms latency on a Tesla T4 GPU, indicating that real-time deployment is feasible without substantial loss of reliability. Robustness tests with Gaussian and salt-and-pepper noise quantify sensitivity to degraded imagery and highlight the practical value of denoising and image-quality control. Grad-CAM visualizations further show that the model attends to physically meaningful flame and smoke regions, supporting interpretability and engineering acceptance in tunnel safety applications.

Recent years have witnessed the rapid development of deep-learning-based, graph-neural-network-based forecasting methods for modern intelligent transportation systems. However, most existing work focuses exclusively on capturing spatio-temporal dependencies from historical traffic data, while overlooking the fact that suddenly occurring transportation incidents, such as traffic accidents and adverse weather, serve as external disturbances that can substantially alter temporal patterns. We argue that this issue has become a major obstacle to modeling the dynamics of traffic systems and improving prediction accuracy, but the unpredictability of incidents makes it difficult to observe patterns from historical sequences. To address these challenges, this paper proposes a novel framework named the Incident-Guided Spatiotemporal Graph Neural Network (IGSTGNN). IGSTGNN explicitly models the incident's impact through two core components: an Incident-Context Spatial Fusion (ICSF) module to capture the initial heterogeneous spatial influence, and a Temporal Incident Impact Decay (TIID) module to model the subsequent dynamic dissipation. To facilitate research on the spatio-temporal impact of incidents on traffic flow, a large-scale dataset is constructed and released, featuring incident records that are time-aligned with traffic time series. On this new benchmark, the proposed IGSTGNN framework is demonstrated to achieve state-of-the-art performance. Furthermore, the generalizability of the ICSF and TIID modules is validated by integrating them into various existing models.

In recent years, rapid and reliable interpretation for emergency response to landslides and other geological hazards has become increasingly important. This paper presents DFmamba, an improved deformable dual-branch visual state-space network, to address engineering challenges such as missed large landslide bodies, boundary shifts, and loss of small-scale details. DFmamba mitigates the limited effective receptive field and window-partition constraints that often prevent existing methods from balancing large-area semantic consistency, multi-scale detection, precise boundary delineation, and computational efficiency. It employs a parallel encoder with a convolutional branch and a Visual State-Space Model (VSSM) branch to jointly capture local textures and global context. In the decoder, deformable residual blocks (DRB) enhance geometric modeling of irregular boundaries, while multi-scale feature alignment and a shallow high-frequency injection (MFP) mechanism strengthen boundary responses and preserve fine details. Experiments on the public CAS dataset against representative CNN-, Transformer-, and SSM-based baselines show that DFmamba achieves improved Precision, Recall, F1-score, and IoU, with stable performance across multi-scale scenarios, demonstrating strong robustness for landslide segmentation.

As Electronic Health Records (EHRs) proliferate, multimodal methods become increasingly valuable for clinical prediction. However, missing modalities lead to modality imbalance and degraded fusion performance. Recovering missing modalities has emerged as an effective way to mitigate this issue. Despite this, existing recovery strategies neither efficiently use globally similar conditions as modal-unique priors nor fuse them with the patient’s own available modality, leading to the loss of patient-specific information. To fill these gaps, we propose RAMCare, a two-step missing modality recovery framework that combines retrieval-augmented priors and mixture-of-experts offset modeling for incomplete EHR representation learning, which is realized by a three-stage training process: (1) a Relevant Case Retrieval Augmented Module that builds a global case library of cross-modal aligned condition pairs via co-clustering, enabling efficient retrieval of condition-matched modal-unique priors for preliminary recovery; (2) a Mixture of Experts Offset Module that selects condition experts based on retrieved weights and patient-specific features, mapping the retrieved deviation to personalized offsets of the missing modality to capture individual variation and refine the recovery; (3) a finetuning stage that further enhances recovery and multimodal fusion performance. Evaluated on two real-world datasets with missing modalities, RAMCare achieves state-of-the-art performance.

Automated screening of geomorphologically defined landslide potential-hazard candidates from high-resolution topographic data remains challenging in mountainous regions where comprehensive field inventories are unavailable. This study proposes a two-stage framework for extracting rule-defined boot-shaped terrain candidates from airborne LiDAR digital elevation model (DEM) data. First, an expert-informed screening rule formalizes a steep-upper–gentle-lower terrain morphology using representative longitudinal profiles of slope units, and the screened units are converted into rule-derived reference masks. Second, semantic segmentation models are trained to approximate these reference patterns directly from DEM-derived raster inputs. Four architectures—U-Net, U-Net++, DeepLabV3+, and SegFormer-B0—were evaluated using 406 patches of 256 × 256 pixels at 2 m resolution from four LiDAR-covered subregions in Zhenxiong County, China. Under spatially grouped three-fold cross-validation, DeepLabV3+ with DEM + slope-gradient input and Dice + Focal loss achieved a mean pixel-level F1-score of 0.351, mIoU of 0.557, and Patch-F1 of 0.814. Input-feature experiments showed that slope-gradient information was particularly informative, whereas the incremental contribution of aspect was configuration-dependent; DEM + slope was retained as a parsimonious two-channel input. Sensitivity analysis showed that the rule-derived candidate definition changed materially with the screening parameters. Leave-one-subregion-out evaluation yielded a macro-averaged F1 of 0.321, indicating measurable within-county cross-subregion transfer. However, whole-area evaluation under natural candidate prevalence reduced the macro-average F1 to 0.072 at a fixed threshold and 0.095 using validation-derived operating thresholds. These results indicate that the proposed model is best interpreted as a raster-based surrogate for rule-derived geomorphological screening rather than as an independently validated landslide detector.

Deep learning-based modeling of multimodal Electronic Health Records (EHRs) has become an important approach for clinical diagnosis and risk prediction. However, due to diverse clinical workflows and privacy constraints, raw EHRs are inherently multi-level incomplete, including irregular sampling, missing modalities, and sparse labels. These issues cause temporal misalignment, modality imbalance, and limited supervision. Most existing multimodal methods assume relatively complete data, and even methods designed for incompleteness usually address only one or two of these issues in isolation. As a result, they often rely on rigid temporal/modal alignment or discard incomplete data, which may distort raw clinical semantics. To address this problem, we propose HealthPoint (HP), a unified clinical point cloud paradigm for multi-level incomplete EHRs. HP represents heterogeneous clinical events as points in a continuous 4D space defined by content, time, modality, and case. To model interactions between arbitrary point pairs, we introduce a Low-Rank Relational Attention mechanism that efficiently captures high-order dependencies across these four dimensions. We further develop a hierarchical interaction and sampling strategy to balance fine-grained modeling and computational efficiency. Built on this framework, HP enables flexible event-level interaction and fine-grained self-supervision, supporting robust modality recovery and effective use of unlabeled data. Experiments on large-scale EHR datasets for risk prediction show that HP consistently achieves state-of-the-art performance and strong robustness under varying degrees of incompleteness.

Urban pavement defects, such as cracks and potholes, pose significant challenges to road safety and maintenance. Traditional pavement defect detection methods rely heavily on manual inspection, which is labor-intensive and time-consuming. Recent advances in deep learning have opened new opportunities for automating this process, particularly through the use of convolutional neural networks (CNNs). This paper presents an improved deep learning-based approach for detecting pavement defects using street view imagery. The proposed method leverages a customized dataset constructed from high-resolution street view images, incorporating both common and hazardous defects. The detection algorithm is based on an enhanced YOLOv8 model, optimized for handling low-resolution images and small defect targets. The model improvements include the introduction of a spatial-depth convolutional layer to preserve fine-grained information, a generalized feature pyramid network for better feature fusion, and a dynamic head with multi-task awareness for improved detection accuracy in complex urban environments. Experimental results demonstrate that the proposed model achieves superior performance in detecting pavement defects, with a mean Average Precision (mAP) improvement of 4.7% over the baseline model, while maintaining high inference speed. These findings suggest that the enhanced YOLOv8 model can be effectively applied to urban pavement maintenance, providing a reliable and efficient solution for large-scale defect detection.

Federated Learning (FL) is vulnerable to backdoor attacks due to its distributed nature, necessitating further study to understand and counter these threats. Unfortunately, existing studies suffer from limited backdoor durability, as their backdoor features insufficiently maintain the discriminability after attacks stop and benign model fusion continues. To tackle this, we propose a new backdoor attack framework on FL via discriminative feature pushing-pulling, namely Parasite, to perform durably effective backdoor attacks. Specifically, we first propose a target-aligned trigger generation module, which pulls backdoor features closer to target-class benign features, generating triggers as priors for backdoor injection to help plant stronger backdoors. Then, we propose a boundary separated backdoor injection module, which pushes target-class features away from others while pulling non-target shift features to align with the pre-poisoned ones, thereby improving backdoor durability with minimal impacts on utility. Extensive experiments demonstrate Parasite achieves high attack success rates and injects more durable backdoors than SOTA baselines by large margins (e.g., up to 4.16x and 16.06x Lifespan on CIFAR-10 and GTSRB, respectively).

Complete formal verification of neural networks is crucial for their deployment in safety-critical domains. A key bottleneck stems from encoding complexity: traditional methods assign one binary variable per unstable ReLU neuron. We propose the Sign-Absolute Reformulation Theory (SART), which fundamentally breaks the conventional one-to-one mapping between unstable neurons and binary variables by establishing formal reducibility criteria. This allows for finer-grained modeling, where each unstable neuron corresponds on average to fewer than one binary variable, thereby reducing verification complexity at its source. Based on SART, we derive a theoretical lower bound on the number of binary variables required for complete verification and, under the assumption that P not equal NP, prove that variables in the final layer can be compressed by 50%, while the number of variables in intermediate layers cannot be further reduced. To overcome the apparent "last-layer-only" limitation, we recast verification as a sequential process and, crucially, show that the gain lifts to the entire network: LayerABS, a SART-based progressive tightening verifier, iteratively treats intermediate layers as temporary final layers and propagates tight bounds that shrink the global search space and binary-variable counts. Furthermore, we reveal a structural law influencing verification complexity: when the signs of weights of unstable neurons satisfy numerical symmetry, with positive and negative weights equal or differing by at most one, the worst-case verification complexity achieves the theoretical optimum, offering theoretical guidance for the design of verification-friendly architectures. As a general-purpose underlying encoding, the value of SART is independent of specific algorithms. To comprehensively evaluate its effectiveness, we first evaluate the abstraction-free SART encoding, and then integrate it with abstraction techniques to construct the complete verifier LayerABS and its incomplete variant Incomplete-LayerABS. Across benchmarks, our methods surpass state-of-the-art baselines, validating SART's practical impact.

In the field of intelligent transportation systems, data scarcity severely constrains the generalization capability and precision of prediction models. Existing transfer learning methods often suffer from negative transfer due to discrepancies between source and target domains, compromising prediction performance. Although current LLMs exhibit strong feature learning capabilities, they primarily depend on pretrained knowledge and make insufficient use of source city data for targeted training. Moreover, they fail to conduct in-depth exploration of the characteristics inherent in limited datasets. To address this issue, we propose a Prompt and Sequential Selection Transfer based Large Language Model (PST-LLM) for traffic prediction with limited data, which represents the first effort to leverage LLMs specifically for traffic prediction under limited data without relying on conventional transfer learning frameworks. First, we develop a Source City Sequential Selection (SCSS) strategy for knowledge fusion. Combined with simulated fine-tuning on data-rich source cities, this approach precisely identifies optimal transfer paths to mitigate negative transfer induced by data scarcity. Second, we incorporate text prompts to enable deep fusion of semantic cues with spatio-temporal data. This integration guides the model to efficiently capture traffic patterns under limited data. Finally, a joint pruning and freezing strategy is adopted to streamline model parameters, suppressing overfitting during limited data training and significantly enhancing generalization capability. Experimental results show that PST-LLM consistently outperforms state-of-the-art(SOTA) across real-world datasets under identical settings. Furthermore, PST-LLM not only can generates predictive outputs but also can produces interpretable reasoning analysis and actionable travel recommendations. This is achieved through a dedicated reasoning branch that decodes the model’s high-level embeddings into natural language based on prompt engineering. The source code for PST-LLM is available at https://github.com/pku-smart-city/source_code/tree/main/PST-LLM.

Multi-modal Passenger Flow Forecasting at Hub Airports aims to predict the short-term origin-destination (OD) flow of outbound passengers from airports to various urban regions and its distribution across multiple transport modes. Accurate prediction is essential for efficient airport operations and for maintaining the stability of surrounding transportation systems. While recent deep learning approaches have shown potential in OD forecasting tasks, most focus on single-mode flow and neglect heterogeneous data sources, limiting their ability to model passenger flow and modal allocation in hub airports. To address these limitations, we propose M2F-Net, a Multi-Source and Multi-Modal Flow Forecasting Network that integrates a cross-modal encoder with a Universal Opportunity Model (UOM)-based decoder. Within the M2F-Net, a time-aware allocation matrix jointly models the spatiotemporal and modal flow distribution, while the encoder learns short-term temporal patterns and regional traffic states from historical OD data and road speeds. The decoder generates a dynamic travel-probability matrix to mask the encoder output, mitigating dynamic sparsity and guiding flow prediction. We construct a comprehensive benchmark based on real-world datasets collected from Beijing Capital International Airport and the Beijing Municipal Commission of Transport (June-August 2023), including total outbound flow, multi-modal OD flow, and roadnetwork speed. Experiments on this real-world dataset show that M2F-Net consistently outperforms strong baselines, and ablation results confirm the benefits of multi-source integration and our architectural design.

Accurate forecasting of passenger flows is critical for maintaining the efficiency and resilience of airport operations. Recent advances in patch-based Transformer models have shown strong potential in various time series forecasting tasks. However, most existing methods rely on fixed-size patch embedding, making it difficult to model the complex and heterogeneous patterns of airport passenger flows. To address this issue, this paper proposes a deformable temporal-spectral transformer named DTSFormer that integrates a multiscale deformable partitioning module and a joint temporal-spectral filtering module. Specifically, the input sequence is dynamically partitioned into multiscale temporal patches via a novel window function-based masking, enabling the extraction of heterogeneous trends across different temporal stages. Then, within each scale, a frequency-domain attention mechanism is designed to capture both high- and low-frequency components, thereby emphasizing the volatility and periodicity inherent in airport passenger flows. Finally, the resulting multi-frequency features are subsequently fused in the time domain to jointly model short-term fluctuations and long-term trends. Comprehensive experiments are conducted on real-world passenger flow data collected at Beijing Capital International Airport from January 2023 to March 2024. The results indicate that the proposed method consistently outperforms state-of-the-art forecasting models across different prediction horizons. Further analysis shows that the deformable partitioning module aligns patch lengths with dominant periods and heterogeneous trends, enabling superior capture of sudden high-frequency fluctuations.

Reliable crack recognition is a prerequisite for condition assessment and maintenance decision-making in transportation infrastructure such as pavements, bridges, and tunnels, and this study evaluates implemented artificial intelligence approaches for this task, including convolutional neural networks, vision transformer models, and Transformer-based object detectors. Real-world deployment is still hindered by four recurring bottlenecks: heterogeneous acquisition scenarios, inconsistent label granularity, fragmented evaluation protocols, and inadequately characterized cross-scenario performance degradation. To address these issues, this paper develops a model-centric and scenario-aware benchmarking methodology for systematically characterizing model performance and engineering applicability across representative inspection conditions. Using a multi-scenario dataset spanning Close-up Static (CS), Contextual Static (CTS), Dynamic Vehicle (DV), and Dynamic Aerial (DA) acquisition conditions, we establish a controlled classification-detection benchmark under shared infrastructure-level settings, including common data splits, input resolution, evaluation metrics, hardware platform, and comparable detection backbones where applicable. Architecture-dependent training components are retained from official or recommended configurations to preserve each model's intended design. Eighteen representative models are evaluated, including nine fine-grained classifiers and nine defect detectors. Performance is reported across three engineering-relevant dimensions: accuracy, efficiency, and resource footprint. Accuracy is measured using per-class and macro metrics for classification, together with mean average precision (mAP) at multiple intersection over union (IoU) thresholds for detection. Efficiency is assessed by end-to-end frames per second (FPS), pure-inference FPS, and floating-point operations (FLOPs), while resource footprint is characterized by model weight size and memory consumption. The main findings are as follows. For fine-grained crack classification, Shifted Window Transformer (Swin Transformer) achieves the strongest observed performance in static scenarios, while MobileNet v3 delivers a favorable accuracy-efficiency trade-off under dynamic and resource-constrained conditions. For crack detection, Mask Region-based Convolutional Neural Network (Mask R-CNN) and Cascade Region-based Convolutional Neural Network (Cascade R-CNN) attain the highest mAP, especially under stricter IoU criteria. Among detectors with competitive source-domain mAP, Transformer-based detectors, including Deformable Detection Transformer (Deformable DETR) and Detection Transformer with Improved Denoising Anchor Boxes (DINO)-show smaller normalized mAP drops in cross-scenario transfer evaluations without retraining, suggesting relative cross-scenario stability under the present benchmark conditions. Overall, the proposed benchmark offers a reproducible benchmarking protocol, quantitative robustness indicators, and practical model-selection guidance spanning diverse artificial intelligence architectures, infrastructure inspection scenarios, and annotation granularities.

Time series forecasting, particularly for data collected via Internet of Things (IoT) sensing innovations in transportation and energy sectors, plays a crucial role in understanding complex social behaviors and enhancing social intelligence. Accurate prediction of these sensor-driven signals is of great significance for optimizing societal resource allocation, proactively managing social dynamics, and solving complex urban challenges. However, because of the high nonlinearity, dynamism, and long-term dependence of time series, accurate prediction has become a challenging task. The existing models based on graph neural networks are challenging to adapt to dynamic dependencies that change over time, particularly when using static graph structures to represent dependency relationships. In addition, traditional structures have limitations in capturing long-term dependencies. Although attention-based methods have alleviated the above problems to some extent, their high computational complexity and memory overhead limit their practical application in large-scale scenarios. To this end, we propose a hybrid Attention-Mamba architecture named Trans2Mamba that combines attention and Mamba modules for efficient time series forecasting. This model can effectively capture the dynamic and remote dependency characteristics in time series. The model we propose consists of two main modules: an input embedding layer that adaptively characterizes the features of different time steps and variables through a learnable spatiotemporal embedding layer; a stacked ST Block consisting of an attention-based temporal block and a Mamba-based spatial block, aimed at accurately modeling spatiotemporal dependencies through division of labor and collaboration. We validated the effectiveness of the model on seven datasets covering electricity, meteorology, and transportation. In addition, the ablation experiment further verified the positive contribution and rational design of each key module in the model to the overall performance.

Monitoring and early warning systems are essential for ensuring the long-term stability of tunnels. However, existing methods often overlook the impact of multiple external factors, such as water pressure and temperature, on tunnel structural behavior, limiting their intelligence and accuracy. To address this gap, we propose a Dynamic Pre-Warning model (DPWNet), constructed as a Deep Probabilistic Autoregressive Architecture, which integrates deep learning to predict structural responses by considering spatiotemporal correlations and dynamic external loads. DPWNet incorporates Markov Chain Monte Carlo sampling during the decoding phase to simulate a range of potential working conditions, accounting for the complex interactions between environmental and structural factors. This probabilistic framework allows real-time adaptive adjustment of pre-warning thresholds based on scenario likelihoods. DPWNet is applied to an underwater shield tunnel, demonstrating significant improvements in accuracy. Specifically, the model calculates the 90% confidence interval of structural responses under multiple factors, translating probabilistic forecasts into actionable thresholds. Experimental results show that DPWNet reduces mean absolute error by 42.0%, root mean square error by 29.2%, and improves Pearson correlation coefficient by 1.3% over 30 days compared to existing methods. These results highlight the model's reliability and its potential to advance the intelligent monitoring of underwater shield tunnels.

Recent studies have revealed that multi-agent systems are vulnerable to adversarial attacks, where attackers can distort cooperation perception by sharing adversarial features. Though achieving promising results, current defenses are computation-expensive and limited in real-time scenarios due to their redundant consensus verification. To solve this problem, we propose Sieve to computationally efficiently detect adversarial features in multi-agent perception. Specifically, we first propose a coherence-based fused feature verification module, which preliminarily verifies fused features and bypasses unnecessary detection when they are benign. Then, we propose a perception discrepancy attribution detection module, which detects adversarial features by evaluating each feature’s contribution to discrepancies between cooperation and ego perception results, identifying adversarial features with abnormally high contribution. Extensive experiments show that our method achieves superior defense ability to baselines within the limited detection time (e.g., +9.4% AP@0.5 and +9.9% AP@0.7 on average for 100-ms detection time).

Early warning of landslides is vital for disaster prevention and risk reduction. However, most existing warning methods overlook the combined effects of geological and environmental factors and fail to capture the dynamic and nonlinear interactions between displacement trends and seasonal variations. To overcome these limitations, this study proposes a deep learning-based Dynamic Pre-warning and Frequency Enhancement Block Network (DPFEB-Net). The model integrates rainfall and reservoir water level variations to simulate environmental dynamics and incorporates geological features such as slope volume, thickness, gradient, and material type to represent structural stability. Fourier-based frequency selection and enhancement module enables extraction the nonlinear of periodic-trend relationships, while a Markov Chain Monte Carlo (MCMC) sampling strategy provides dynamic displacement-based early warnings with 90% confidence. Experimental validation on five typical landslides in Wushan County within the Three Gorges Reservoir area demonstrates that DPFEB-Net achieves high prediction accuracy, with Pearson correlation coefficients (PCC) exceeding 0.99 in both short-term (7-14 days) and long-term (30 days) forecasts, significantly outperforming LSTM and other baseline models. In addition, the influence of different feature combinations and frequency-domain components on landslide displacement characteristics is analyzed. Overall, DPFEB-Net offers a novel displacement-driven framework for landslide early warning, supporting risk assessment and disaster management applications.

The growing volume of electronic health records (EHRs) has made deep learning a vital tool for real-time treatment analysis. While early models focused on single-modality learning, their limited informational perspective constrained performance. Recently, multimodal approaches that integrate complementary data sources have shown performance gains. However, two major challenges persist: 1) A large amount of unlabeled data remains underutilized. Unlabeled medical data is prevalent due to privacy concerns and other factors. Most end-to-end models rely on labeled data and thus fail to exploit the full potential of available data sources. 2) Insufficient case-level semantic constraints in existing pretraining methods. Although a few pretraining models enable the utilization of unlabeled data, current methods primarily establish constraints at the modality level, neglecting the constraints on holistic representations. This limitation undermines their effectiveness in downstream clinical tasks. To fill this gap, we propose a dual-level contrastive semi-supervised representation learning framework that leverages both hourly-level and case-level constraints to utilize unlabeled data. First, a time-aware alignment module aligns different modalities and introduces fine-grained cross-modal contrastive learning at the hourly level, enhancing local consistency across modalities. Then, to generate case-level positive samples, we introduce a memory-based case-level augmentation strategy. We combine partially dropped current patient representations with those of other patients retrieved from the memory bank, forming case-level positive pairs with hidden state augmentation. By jointly optimizing representations at the local temporal scale and the global case level, our model generates more comprehensive representations. Experiments on real-world tasks demonstrate our method achieves state-of-the-art performance and maintains representation quality under sparse labeling conditions.

In mountainous regions with multiple fault zones, ballastless tracks were susceptible to damage and failure under the effect of fault displacement, which compromised their normal service performance and operational safety. In this study, a track-subgrade coupling model crossing fault zones was established using the finite element method and a concrete damaged plasticity model. The mechanical behavior of a novel short board track was compared with that of continuous and unit board tracks. The results showed that the vertical deformation and interlayer gap of the three tracks increased with increasing fault displacement. The short board track exhibited slightly greater rail deformation, with a vertical deformation amplitude approximately 1.12 times that of the other two tracks. In contrast, the maximum interlayer gap between the base board and the subgrade was larger for the continuous and unit board tracks than for the short board track. The concrete track structure of the short board track was not damaged under normal fault dislocation. In contrast, the damage variables for the concrete track structure of the continuous and unit board tracks increased with fault displacement. This study provided crucial guidance for selecting new ballastless track types for railways traversing multi-fault regions.