
Abstract The increasing complexity of automotive clutch systems and the nonlinear interactions among thermal, mechanical, and operational parameters make accurate fault diagnosis a challenging task in modern vehicles. Most existing studies rely on conventional machine learning techniques or physics-based threshold methods, which often fail to capture cross-parameter dependencies and lack interpretability and adaptability under varying driving conditions. To address these limitations, this study proposes FusionCAT-X (fusion cognitive attention-infused transformer with boosted ensemble learning), a hybrid diagnostic framework that combines a cognitive attention-infused transformer (CAT) with a gradient-boosted ensemble classifier to model complex feature interactions and improve clutch fault classification performance. The framework employs embedding-based representations to unify heterogeneous inputs and utilizes multihead self-attention to dynamically learn interdependencies among clutch parameters, followed by ensemble-based decision refinement to improve classification boundary separation. The model is implemented using Python with tensorflow and pytorch for deep learning, along with numpy, pandas, and scikit-learn for preprocessing and evaluation. Experiments are conducted on a publicly available Kaggle clutch failure dataset containing over 5000 samples under diverse operating conditions. The proposed model achieves an accuracy of 97.90%, outperforming baseline methods such as SVM, KNN, and CatBoost by approximately 0.4%–3.8%. The model outputs probabilistic confidence estimates; however, their calibration should be quantitatively evaluated. The results demonstrate that FusionCAT-X improves clutch fault classification by effectively learning nonlinear interactions among thermal, mechanical, and operational parameters. The achieved performance of 97.90% accuracy, 97.53% F1-score, and 98.10% Area Under the Curve (AUC) demonstrates effective discrimination between normal and faulty clutch conditions.
Abstract Railway track geometry is monitored by comparing measurements against regulatory thresholds, but this approach does not capture how the distribution of geometry parameters evolves over time or across track segments. Optimal transport (OT) provides a natural framework for such distributional comparisons. This study investigates whether quantum optimal transport (QOT) can provide additional diagnostic information. Track geometry measurements are encoded as quantum states via angle encoding on three qubits, and transport distances are computed through semidefinite programming under four quantum cost functions alongside classical OT, maximum mean discrepancy, energy distance, principal component analysis, and Mahalanobis distance. The framework is applied to three rail datasets. In a temporal analysis of monthly recordings, the quantum swap cost nearly matches classical squared Euclidean (r=0.998), while the Hamming cost ranks transitions differently (Spearman ρ=0.45). Comparison against a monthly track quality index reveals that Hamming QOT is the only metric with a statistically significant correlation after Bonferroni correction (Pearson r=0.701, p<0.001). In a cross-sectional comparison of healthy and degraded track sections, the eigenvalue spectra of density matrices reveal that degraded sections produce higher-rank quantum states (rank 3–4 versus rank 2 for healthy). Cross-validation on a second class 3 track replicates this pattern. Applied as a per-section anomaly score, Hamming QOT achieves an area under the curve of 0.756 for classifying section health. Sensitivity analyses confirm that these findings are largely preserved across parameter choices. These results suggest QOT offers a distinct but complementary diagnostic lens for infrastructure condition assessment.
Abstract Probabilistic risk analysis is a fundamental approach for safety assessment in engineering systems. However, its application to emerging domains such as hydrogen-blended pipeline systems faces new challenges due to limited fault data and scarce expert knowledge. Bayesian networks (BNs) are widely adopted for risk analysis because they illustrate the relationships between risk factors and facilitate probabilistic reasoning. Under these conditions, two major challenges arise. First, the network structure, including the relevant risk factors and their dependencies, is typically unavailable. Second, when risk factors have multiple causes, the parameter space of conditional probabilities grows exponentially, which requires extensive expert input for parameter specification. To address these problems, this study presents a lightweight framework for risk analysis that minimizes the need for expert judgment and does not rely on data. The framework centers on a BN for probabilistic computation. A large language model identifies risk factors and their relationships from technical documentation. Additionally, a parameter reduction method named pre-conditional probability table (CPT) is developed to compress the dimensionality of conditional probability tables from an exponential to a polynomial scale through weighted combinations. This reduces the need for expert elicitation. The framework is validated through a case study on hydrogen-blended natural gas pipeline risk assessment and benchmark and numerical comparison. The results show less than a 5% deviation in posterior probabilities and over 90% consistency in BN structures compared with benchmark models. The workload for parameter specification is reduced to about 15% of that required in conventional BN modeling.
Abstract Accurately assessing the vulnerability of urban rail transit networks requires identifying intervals whose disruption can substantially degrade network operation. Existing studies often treat rail segments as independent failure units, which may underestimate disruption impacts when turn-back operations are constrained by the location of turn-back tracks. This study develops a turn-back-interval-based vulnerability assessment framework. Under the proposed partial-failure rule, failure of any segment within a turn-back interval makes the entire interval between adjacent turn-back-capable stations unavailable for train operation. Three indicators are proposed to measure vulnerability from structural and service perspectives: network connectivity loss, network efficiency loss, and passenger travel speed loss. A multi-objective optimization model solved by NSGA-II is then used to identify critical turn-back intervals. The framework is applied to the Wuhan Metro network. Results show that accounting for turn-back operations produces a more conservative and operationally realistic vulnerability assessment than conventional segment-based assumptions, especially from the passenger service perspective. Critical intervals are concentrated on high-demand corridors and often include transfer stations, with line 2 playing a particularly important role in maintaining network-wide service. The proposed framework supports emergency planning, maintenance prioritization, and infrastructure improvement for vulnerable turn-back intervals.
As a core component in steel production, the steel strip is prone to surface defects during the continuous production process due to its complex conditions. To achieve high-precision real-time monitoring of its surface quality and efficient quality control, the steel strip surface defect detection method is proposed based on uncertainty region fusion and digital twin (DT) model establishment. Specifically, the You Only Look Once version 8 (YOLOv8) and Single Shot MultiBox Detector (SSD) models pretrained on public defect datasets are used for independent surface detection, obtaining initial defect regions with confidence scores and uncertain detection results. To address single-model detection uncertainty and enhance detection accuracy, the decision-level uncertainty region fusion method is designed to fuse initial regions and uncertainty information based on region overlap and confidence, acquiring maximum reliable defect regions for surface defect recognition and localization. Different from existing state-of-the-art (SOTA) steel strip defect detection methods, the proposed method innovatively introduces uncertainty perception and diagonal equal-area decision-level fusion. It effectively reduces unreliable bounding box predictions and single-model detection bias and further realizes high-confidence defect data interaction and virtual-real synchronization with digital twin framework. To evaluate the performance of the proposed method, the public datasets and experimental steel strip defect datasets are utilized with several defects. The results demonstrate that the proposed method yields consistent and significant improvements in detection accuracy, recall, and overlap rate on both the public Northeastern University (NEU) dataset and the self-built dataset, which verifies the value for steel strip quality control and intelligent manufacturing.