A rapid prediction of track deterioration states under earthquakes for random complex structure scenarios is identified as a core task for post-earthquake emergency command and train speed limit scheme formulation. In this paper, the continuous prediction of seismic track irregularity is reduced to the discrete prediction of girder end misalignment, and the complex response prediction for random structures is reduced to the single response prediction combination of characteristic structures. Based on dual dimensionality reduction, the rapid construction of seismic track irregularity under random complex structure scenarios is realized using a bidirectional long short-term memory neural network. The conclusions indicate that the girder joint area of track alignment irregularity is the main factor causing wheel-rail impact, and both the girder joint and girder deck areas contribute to the lateral vibration of car body; the continuous seismic track alignment irregularity can be regarded as a spliced body of a series of mutually decoupled sub-segments, and the train response excited by continuous irregularity can be approximately decomposed into the superposition of the independent train response to each sub-segment; by aggregating seismic responses from multiple characteristic structures and applying an equivalent correction factor, the equivalent structural scenario can yield girder misalignments that envelope those of random structural scenarios at the 99 % confidence level.
Facial sketch synthesis is important for cross-modal face analysis and digital forensics, yet existing models often suffer from structural distortions and identity inconsistency under limited paired training data. Traditional methods, primarily based on generative adversarial networks, often suffer from training instability and insufficient detail reconstruction. To address these limitations, this study proposes a diffusion-based framework with a stage-wise multi-condition guidance mechanism that enhances both structural and textural fidelity. Specifically, (i) during the downsampling phase, semantic segmentation features are fused to guide the model with accurate structural information, such as facial part locations; and (ii) during upsampling, a hybrid cross-attention mechanism is employed to integrate coarse image textures with denoised noise, refining fine-grained details. Additionally, we incorporate the Vision Transformer within the U-Net backbone to better capture global contextual information in low-frequency regions, further enhancing image realism. Experiments on multiple benchmark datasets demonstrate strong overall performance in SSIM and FSIM, indicating improved structural similarity and feature-level fidelity under the evaluated settings. Moreover, our method obtains competitive LPIPS results, indicating improved perceptual similarity. We further compare with recent image-to-image translation and diffusion-based baselines and observe competitive performance in both visual coherence and identity preservation.
The prevailing rheological master curve models and time-temperature superposition techniques were not originally developed with consideration of physical hardening. Consequently, their applicability and sensitivity under such conditions remain unclear and warrant further investigation. This study presents the first systematic evaluation of the applicability and sensitivity of three master curve models (CAM, GLSM, and 2S2P1D) and five kinds of the time-temperature superposition techniques (Arrhenius, WLF, Kaelble, Modified Kaelble, and Non-Linear Least Squares) under physical hardening effects. Quantitative assessments were conducted by performing frequency sweeps on base asphalt, SBS-modified asphalt, and asphalt mastic under two low-temperature storage modes, followed by the construction of complex modulus master curves. The goodness-of-fit for the time-temperature superposition techniques followed the order: Arrhenius < WLF < Kaelble < Modified Kaelble < Nonlinear least squares, with the latter three demonstrating superior fitting performance. In contrast, the combination of the CAM model and Arrhenius yielded a R2 of only 0.911, revealing significant limitations. Based on the principles of optimal goodness-of-fit and conservative safety, the combination of the GLSM and the Modified Kaelble is recommended for constructing rheological master curves of asphalt materials under physical hardening. The findings of this study provide critical references and theoretical support for the accurate selection and effective characterization of asphalt material performance in cold regions.
To address the low local consumption and high operating costs caused by temporal mismatch between traction load and photovoltaic (PV) generation in electrified railways with distributed PV, this paper proposes a train operation optimization method considering source–load matching. The train operation process is simplified into a two-stage model (constant-speed traction and constant braking), and a multi-train power model incorporating regenerative braking energy balance is established based on average inter-station speeds. After analyzing the conflict between regenerative braking and source–load matching, a multi-objective optimization model is formulated with three objectives: source–load matching degree, total operation time, and system economy. Decision variables include departure time offsets, inter-station running time adjustment coefficients, and braking intensity adjustment coefficients. A differentiated braking intensity strategy is introduced to resolve the identified conflict. The NSGA-II algorithm is applied to obtain the Pareto-optimal timetable. Case studies using operational data show that the optimized timetable improves the local PV consumption rate by 20.5% over the baseline, and the variable braking intensity further reduces the traction system operating cost by 4.6% compared with the constant braking intensity under the same optimized timetable. These results confirm the effectiveness of the proposed method in enhancing PV utilization and the coordinated use of regenerative braking energy.
Vehicle trajectory prediction is fundamental to autonomous driving environment perception and decision control, directly impacting driving safety and traffic efficiency in complex interactive scenarios. However, existing deep learning-based trajectory prediction methods face three major challenges in real open road environments: (1) The lack of explicit modeling of multi-agent game interactions leads to a high rate of misjudgment in long-term intention prediction. (2) The coupling of multimodal uncertainty sources, including perceptual noise, behavioral ambiguity, and cognitive blind spots, prevents quantifiable risk assessment for downstream planning. (3) These models exhibit weak generalization ability in long-tail scenarios and cross-domain migration, while being constrained by the difficulty of real-time deployment on in-vehicle computing resources. To address these issues, this paper systematically reviews deep learning-driven vehicle trajectory prediction methods and highlights their application value in expert systems and intelligent transportation. The main contributions of this review are as follows: (1) Proposing a unified technical taxonomy that systematically analyzes the design principles, evolutionary logic, and inherent limitations of recurrent neural networks, convolutional neural networks, graph neural networks, Transformers, and deep generative models. (2) Revealing six major engineering gaps between simulation and real-world environments, namely the lack of game-theoretic interaction, uncertainty coupling, insufficient long-tail generalization, perception-prediction discrepancy, real-time constraints, and weak cross-scenario adaptability. (3) Summarizing mainstream public datasets and evaluation protocols to provide an empirical basis for algorithm validation. (4) Envisioning five future directions for next-generation expert systems: joint optimization of prediction and planning, decoupling of uncertainty quantification, game-theoretic multi-agent inference, lightweight adaptive deployment, and scenario-driven generalization evaluation. This review aims to provide a systematic reference for researchers and engineers, thereby promoting the development of safer, more interpretable, and deployable trajectory prediction for autonomous driving in open road environments.
