Issue 08, 2026
Study on the impact of air traffic volume on controllers' attention quality based on eye movement data
Wang Lili;Zhu Mincong;Attention quality is typically characterized by three key indicators: attention span, transfer efficiency, and stability. Previous studies have seldom examined the specific impact of traffic flow on controllers' attention quality, and there has been a lack of quantitative criteria for determining the upper limit of control sessions based on attention quality. To address these research gaps, this study designed a dynamic traffic flow simulation experiment focused on regional air traffic control, progressively increasing the number of control flights. Eye movement data were collected from 22 experienced air traffic controllers throughout the experiment. The standard deviation of reaction time(σRT), saccade speed(VS), and standard deviation of sustained gaze duration(σSGD) were utilized as quantitative measures for attention span, transfer efficiency, and stability, respectively. By integrating the entropy weight method with the Criteria Importance Through Intercriteria Correlation(CRITIC) method, the weights of σRT, VS, and σSGD were established to construct an Attention Quality Deterioration Index(IAQD). The results revealed a significant cubic relationship between attention quality and the number of control flights. The model demonstrated that the IAQD value increased as the number of control flights(FC) rose, indicating a decline in attention quality with an increasing number of flights. Further analysis of the inflection point in the cubic function indicated that the rate of attention quality deterioration peaked at FC = 17.58 flights. An independent samples t-test was conducted on the IAQD values for two groups: FC ≤ 18 and FC > 18, revealing a statistically significant difference. Due to the small sample size, a Mann-Whitney U test was also performed, which indicated a highly significant difference between the two groups(U = 0.000, exact p = 0.003). The mean rank of the FC ≤ 18 group(4.00) was significantly lower than that of the FC > 18 group(10.00), with a large effect size(r = 0.82), consistent with the results of the parametric test. Based on these findings, it is proposed that the safety threshold for the number of aircraft a regional controller should manage simultaneously is 17 flights.
Research on influencing factors of confidential reporting behavior based on self-determination theory
Liu Junjie;Tian Pengcheng;Chen Ze;Based on Self-Determination Theory, this study constructed an indicator system comprising 25 evaluation metrics across five dimensions(external regulation, introjected regulation, identified regulation, integrated regulation, and intrinsic motivation) to investigate the multidimensional motivational composition and influencing mechanisms of aviation safety confidential reporting behavior among civil aviation practitioners. The objective is to optimize the effectiveness of confidential reporting systems and improve both the quantity and quality of reports. Through a combined methodology of questionnaire surveys and statistical analysis, the research empirically examined the influence of different motivational dimensions on reporting behavior among airline employees. The results demonstrate that:(1) the confidential safety reporting behavior among civil aviation personnel is significantly influenced by five distinct regulatory factors, revealing a sophisticated multi-level motivational framework that governs their reporting decisions and actions;(2) current reporting practices are predominantly driven by external regulation and introjected regulation mechanisms, while autonomous motivation forms—particularly identified regulation and intrinsic motivation—remain substantially underdeveloped, resulting in suboptimal outcomes in both reporting frequency and information quality;(3) a quality-oriented incentive mechanism, a non-punitive safety culture, and value guidance coupled with behavioral internalization constitute the key pathways for transforming employees' voluntary reporting behavior from passive compliance to active engagement. The findings reveal the complex interplay between controlled and autonomous motivations in influencing reporting behavior and highlight the need for organizations to develop targeted strategies that enhance autonomous motivations through systematic interventions. These include redesigning incentive structures to emphasize quality over quantity, fostering an organizational culture that genuinely values and acts upon safety reports without fear of reprisal, and implementing communication and training programs that help employees internalize the importance of reporting for safety improvement. Ultimately, these efforts will contribute to more effective safety management systems in the aviation industry through increased frequency and quality of confidential safety reports.
Analysis of online public opinion regarding occupational safety accidents among outdoor workers against the backdrop of extreme heat
Yang Ting;This study examines the complete life cycle of online discourse triggered by the heatstroke death of a delivery rider in Xinyang, Henan, under extreme conditions of 39 ℃. We collected 9 367 original posts and comments from two central trending topics on Sina Weibo between July 10(the day of the initial report) and July 21, 2025(when public attention began to wane). After cleaning the data to remove irrelevant content and stop words, we obtained 7 163 valid texts for subsequent analysis. Methodologically, the research employs a two-pronged analytical approach. First, we utilized the BERTopic model to dynamically identify and track the evolution of thematic focuses. This process involved generating text embeddings via a pre-trained Bidirectional Encoder Representations from Transformers(BERT) model, reducing dimensionality using Uniform Manifold Approximation and Projection(UMAP), and applying Hierarchical Density-Based Spatial Clustering of Applications with Noise(HDBSCAN) clustering to derive initial topics. The number of topics was optimized through a combination of the Calinski-Harabasz Index(CH Index), the elbow method based on Within-Cluster Sum of Squares(WCSS), and topic coherence scores(Coherence). Second, we constructed a specialized emotional lexicon for occupational safety accidents by enhancing the Dalian University of Technology Sentiment Dictionary(DUTSD) with domain-specific high-frequency words. This lexicon, integrated with emotion mappings from Weibo emoticons, was used to classify public sentiment into six categories: anger, sadness, sympathy, satire, disappointment, and encouragement. A fine-grained, time-series sentiment analysis was subsequently conducted. The results reveal a clear and progressive evolution of public discussion across three distinct phases. On the first day of the incident, public attention centered on immediate health risks. Over the next two days, the focus shifted towards criticism of inadequate labor protections and platform accountability mechanisms. Three days post-incident, discussions escalated to broader reflections on the social welfare system. Correspondingly, public sentiment initially displayed strong condemnation and fear. Following an emotional peak, a segment of the discourse notably transitioned towards rationality, characterized by constructive suggestions, active solution-seeking, and explicit demands for enhanced institutional safeguards. These findings highlight the dual nature of online discourse surrounding occupational accidents during extreme heat events, indicating both significant public concern for livelihoods and pointed systemic criticism. This research provides empirical evidence to support governmental efforts in guiding public opinion and informs policymaking related to high-temperature labor protection.
Expressway traffic crash risk prediction based on an integrated approach of XGBoost-CNN
Liu Wenke;Wang Xuekai;Zhou Mo;Wan Qingsong;Hao Wenbang;Yan Ying;Using real-world operational data from an expressway in Shandong Province, China, this study developed a framework for short-term traffic crash risk prediction aimed at routine safety management. Traffic flow, meteorological conditions, and crash records from multiple monitoring sites along the expressway were aggregated into 5-minute intervals and spatiotemporally fused with crash occurrences. Candidate explanatory variables were constructed to characterize upstream and downstream traffic volume, speed, occupancy, visibility, rainfall, temperature, wind, time of day, and other temporal indicators. Recursive Feature Elimination(RFE) was employed to rank variable importance, resulting in the retention of the 15 most informative features within the pre-crash window, thereby highlighting the joint influence of traffic conditions and adverse weather. Within an ensemble learning framework, an expressway crash risk predictor was developed by integrating eXtreme Gradient Boosting(XGBoost) with a Convolutional Neural Network(CNN). XGBoost captures nonlinear relationships among heterogeneous tabular features, while the CNN extracts local temporal patterns from traffic flow sequences. Their probabilistic outputs were combined through an optimized weighting scheme, and model hyperparameters were tuned via cross-validation. For fair comparison, baseline models using only XGBoost and only CNN were trained on the same dataset following identical preprocessing and evaluation protocols. Experimental results indicate that the proposed integrated model achieves an accuracy of 0.96 and an AUC of 0.94 on the test set, outperforming both single models in terms of discrimination and robustness. Feature ablation analyses reveal that excluding meteorological variables results in the greatest performance degradation, with visibility and rainfall identified as the most influential factors. Prediction errors increase significantly under rainy or low-visibility conditions, confirming the critical role of meteorological factors in crash risk prediction. Furthermore, testing on a later time period of the same roadway yielded an accuracy of 87.4%, demonstrating effective out-of-time generalization. Overall, the XGBoost-CNN ensemble provides an accurate, robust, and interpretable tool for estimating short-term crash risk on expressways.
Research on intelligent prevention and control of safety risks in distribution rooms based on knowledge graph
Zhu Botao;Tian Kaichun;Li Yang;Wang Ronghua;Li Tai;To tackle the challenges of high labor costs and inefficient inspection processes in traditional operation and maintenance models for distribution rooms, this study proposes a solution involving the construction of a fault-disaster knowledge graph and its intelligent applications tailored specifically for these environments. Initially, a multi-source fusion strategy was utilized to create the knowledge graph corpus. Employing the “Human-Machine-Environment-Management”(HMEM) multi-factor coupling model, attribute definitions and association rules for various entities were established, resulting in a unified semantic framework for the knowledge graph construction. Following this, entities and relationships were manually annotated using the Colabeler annotation assistant in conjunction with the BIO annotation system, ultimately producing 200 valid data entries. During the knowledge extraction phase, the Bidirectional Long Short-Term Memory-Conditional Random Field(Bi-LSTM-CRF) algorithm was utilized for named entity recognition. Simultaneously, Text Convolutional Neural Networks(TextCNN) were employed to facilitate entity-relationship extraction. Additionally, using the Neo4j graph database for knowledge storage and visualization, a total of 174 entities and 1 140 entity relationships were imported into the database. This process enabled the creation of a dual-chain knowledge graph model consisting of the “risk evolution chain” and the “prevention and disposal chain”. Finally, utilizing the Dify framework, an intelligent question-answering system was developed for knowledge graph-enhanced retrieval. This system established a complete closed-loop process: “user query → intent parsing → graph retrieval → LLM generation”, facilitating real-time interaction between the knowledge graph and large language models. The results indicate that the constructed knowledge graph model effectively represents the risk evolution mechanism under the coupling of the four key elements of HMEM within distribution rooms. Furthermore, the knowledge extraction model demonstrates stable and reliable performance, achieving an overall F1 score exceeding 80%. Moreover, the intelligent question-answering system exhibits improved professional response capabilities. This research offers scientific and intuitive decision support for operations and maintenance personnel, while also serving as a reference for transforming the power operations and maintenance sector from “experience-driven” to “intelligence-driven” approaches.
Research on the modular design of hazardous chemicals road accident rescue vehicles using FAHP-FQFD-TRIZ methodologies
Li Jingxiang;Gao Xuenong;Jing Guoxun;To improve the functional integration, scenario adaptability, and modular coordination of rescue vehicles for hazardous chemical road accidents, this study develops a modular design method integrating the Fuzzy Analytic Hierarchy Process(FAHP), Fuzzy Quality Function Deployment(FQFD), and the Theory of Inventive Problem Solving(TRIZ). First, the emergency rescue process for hazardous chemical road transportation accidents was analyzed, and a hierarchical demand system with 9 primary indicators and 21 secondary indicators was established. Expert judgments were represented by triangular fuzzy numbers, and FAHP was used to calculate the weights of demand indicators. The results show that rapid response capability(0.182), accident-disposal effectiveness(0.160), and personnel safety guarantee(0.158) are the three highest-weighted user demands. Second, FQFD was used to translate weighted demands into technical characteristics. The 21 technical-characteristic modules obtained from preliminary mapping were further refined into 32 evaluation-level characteristics in the House of Quality, allowing comprehensive modules to be decomposed into measurable and configurable design elements. A relationship matrix between the 21 secondary demands and the 32 evaluation-level characteristics was then constructed by 10 engineering experts. The FQFD results indicate that the integrated communication module(0.122), integrated detection module(0.087), and energy power support module(0.077) should be prioritized in design optimization. Third, correlations and key conflicts among technical characteristics were identified through the House of Quality, mapped into TRIZ engineering parameters, and reorganized into seven core design contradictions to derive corresponding inventive principles and implementation strategies. Based on these results, the 32 evaluation-level characteristics were integrated into five functional units: rapid response, core disposal, intelligent support, survival support, and human-machine interaction. The final design achieves a 28% reduction in overall vehicle weight, module switching within 5 min, leak response and disposal within 45 s, a comprehensive driving range of at least 800 km, and a gradeability of 45%. It also defines a Class Ⅲ blast-resistant protection target and reserves interfaces for multi-robot collaborative operation, providing methodological support for the standardization and intelligent upgrading of hazardous chemical rescue equipment.
Research on a unified field model for traffic risk in horizontal and vertical alignment combination sections of mountainous expressways
Liu Ze;Ma Siyue;Hu Liwei;To investigate vehicle driving safety and quantitatively assess traffic risk levels on mountainous expressways with combined horizontal and longitudinal alignments, this study develops a unified traffic risk field model based on human-vehicle-environment interaction theory. The model integrates the vehicle kinetic field, road environment field, and driver behavior field to comprehensively describe the coupling effects of vehicle dynamics, roadway geometry, and driver responses under complex terrain conditions. A complex environmental influence factor is introduced into the kinetic field to better represent the effects of varying slope, curvature, and sight distance typical of mountainous highways. Model parameters are calibrated using national standards and field survey data, employing elliptical or teardrop-shaped fields to characterize the spatial extent and directionality of risk. Real-vehicle experiments are conducted to extract and analyze speed and acceleration data from typical combined sections, enabling risk field simulation and visualization. The results reveal that traffic risk on mountainous expressways exhibits distinct spatial variability, with higher risk potentials concentrated on long downhill segments, crests, and sharp curves, while lower risks are found on gentle slopes and straight sections. Furthermore, variations in driver behavior amplify the evolution of risk, indicating that delayed perception and reaction are key intrinsic factors influencing safety in these complex roadway environments. To further understand driving safety patterns in horizontal-vertical combination zones of mountainous expressways, this paper constructs a unified traffic risk field model for such zones, considering driving risk characteristics from human-vehicle-environment perspectives. First, a unified traffic risk field is established by integrating the vehicle kinetic field, road environment field, and driver behavior field across the three dimensions of “person-vehicle-road.” Complex environmental influence factor parameters are included in the vehicle kinetic field to enhance the model's applicability to the unique driving scenarios of combined horizontal and vertical gradient zones. Second, parameters of the unified traffic risk field are calibrated based on national standards and field survey results, utilizing ellipse and teardrop shapes to describe the extent of the vehicle risk field. Finally, the calibrated model is applied to visualize the unified traffic risk field along the experimental route, with speed and acceleration data extracted from real-vehicle experiments on the combined horizontal-vertical gradient section analyzed to identify driving characteristics. Results indicate that risk distribution in horizontal-vertical combination zones of mountainous highways generally shows the highest risk in specific areas such as long downhill slopes, summit regions, and sharp curves, while gentler slopes and straight sections demonstrate relatively lower risk.
Consequence prediction of offshore platform fire accidents based on PSO-SVM
Liu Yue;Chen Guoming;Fire incidents rank among the most frequent operational hazards on offshore oil platforms. When such accidents occur, the unique marine environment and complex platform structures often make firefighting efforts challenging to contain within short timeframes, potentially triggering chain reactions that result in severe casualties and significant property losses. To achieve effective prevention and rapid response to platform fires while providing scientific decision support for emergency rescue operations, this study develops an intelligent fire risk prediction model utilizing Support Vector Machine(SVM) algorithms. Using a deepwater drilling platform in the South China Sea as the research subject, we first conduct detailed numerical simulations through the Fire Dynamics Simulator(FDS) to analyze fire progression patterns under marine platform conditions. By strategically deploying monitoring points in simulated scenarios, we accurately capture temperature variations at critical locations and dynamic changes in carbon monoxide concentration over time. These time-series data serve as crucial input parameters for the predictive model. In constructing the model, this study employs SVM as the core framework while introducing Particle Swarm Optimization(PSO) to intelligently optimize key parameters of the predictive model, including kernel function parameters and penalty factors, thereby significantly enhancing the model's prediction accuracy and generalization ability. To systematically validate the reliability and advancement of the proposed model, the PSO-SVM model was compared with traditional SVM and GA-SVM models optimized using Genetic Algorithms(GA). Experimental results demonstrate that the traditional SVM model achieves average prediction accuracies of only 51.66% for temperature change curves and 50.5% for carbon monoxide volume fraction changes, indicating suboptimal performance. Although the GA-optimized GA-SVM model shows improved prediction performance with accuracies of 51.40% and 62.5%, respectively, there remains considerable room for further enhancement. Notably, the PSO-SVM model, optimized with the particle swarm optimization algorithm, demonstrated superior overall predictive performance among all comparison models. It achieved 83.26% accuracy in temperature prediction and 70.0% accuracy in carbon monoxide volume fraction prediction, significantly outperforming other models. This validation confirms the model's excellence and reliability in predicting the consequences of marine platform fire accidents, providing robust data support and technical assurance for on-site emergency rescue decision-making.
Landslide deformation analysis and prediction methods integrating InSAR with enhanced CNN-BiLSTM in Zhenxiong County
Hu Bo;Hou Yuelin;Li Wen;Li Yuebin;Zhang Runyuan;Yang Li;Wu Wenqing;Li Jia;Landslides are natural disasters that pose a significant threat to the safety of lives and property. Timely and accurate monitoring and prediction of deformation are crucial for effective disaster prevention and control. This paper focuses on the landslide-affected area in Zhenxiong County, Yunnan Province. It utilizes Small Baseline Subset Interferometric Synthetic Aperture Radar(SBAS-InSAR) technology to monitor landslide deformation and proposes an enhanced CNN-BiLSTM model, referred to as the Whale Migrating Algorithm-Convolutional Neural Network-Bi-directional Long Short-Term Memory-Attention(WMA-CNN-BiLSTM-Attention), for predicting landslide deformation. First, deformation information from the landslide area is obtained using SBAS-InSAR technology, with deformation features extracted through time-series data processing. Landslide deformation is often complex and nonlinear, making it difficult for traditional optimization methods to identify optimal solutions. To address this, the CNN-BiLSTM model is optimized using the Whale Migrating Algorithm(WMA) and integrated with a self-attention mechanism(Attention). This combination efficiently searches for the best hyperparameter configuration in a high-dimensional parameter space, significantly enhancing the model's performance in predicting landslide deformation. Experimental results show that the proposed model achieves an average Root Mean Square Error(RMSE) of just 1.69 mm in landslide deformation prediction. Compared to benchmark models, this represents a reduction in error ranging from 44% to 56%. Furthermore, the maximum and minimum absolute deviations are less than 4 mm and 0.1 mm, respectively, while the average coefficient of determination(R2) exceeds 0.94. These findings indicate that the model effectively captures the dynamic characteristics of landslide deformation, providing robust and reliable support for early warning and prevention efforts related to landslides. The integration of advanced remote sensing technology with optimized deep learning models enhances the accuracy of deformation monitoring and prediction while establishing a strong technical foundation for developing scientific and effective landslide prevention and control strategies. This approach holds significant practical importance in mitigating the losses associated with landslide disasters.
Risk assessment of enterprise water supply network leakage based on AHP-EWM improved logistic regression
Liu Ling;Wu Zhiwu;Zhang Kai;Lan Ling;Li Xilin;Yu Xiaoxi;Fu Meiru;This study addresses the challenge of limited data samples in leakage risk assessment for small-to medium-sized enterprise water supply networks by proposing an enhanced logistic regression method that integrates the Analytic Hierarchy Process(AHP) and the Entropy Weighting Method(EWM). The method systematically identifies six key static influencing factors—pipe material, diameter, age, connection type, burial depth, and road type—based on an analysis of pipeline operation and maintenance data from a steel enterprise in Liaoning Province collected between 2021 and 2024. These factors form a framework for assessing leakage risk indicators. Subjective weights are determined using AHP, which incorporates expert judgment, while objective weights are derived from EWM based on the characteristics of data distribution. These weights are then combined linearly through a least-squares deviation minimization method to produce comprehensive weights. The integrated weights are incorporated into a logistic regression model to develop a weighted logistic regression assessment model, with hyperparameters optimized via grid search and cross-validation. Risk probabilities are visualized in a Geographic Information System(GIS), generating a spatial distribution map of leakage risk across the enterprise's water supply network. The results demonstrate that the combined AHP-EWM weighting approach significantly enhances the rationality and stability of weight allocation, with burial depth(22.32%) and pipe material(21.08%) identified as the primary influencing factors. The weighted logistic regression model performs exceptionally well, achieving an accuracy of 91% and an AUC of 0.92, while surpassing both single-weighting models and traditional logistic regression in terms of precision, recall, and other evaluation metrics. Validation analysis reveals that 78.57% of the 14 actual leakage incidents in 2024 occurred in high-risk zones identified by the model, confirming its strong engineering relevance and early-warning capability. Based on the assessment results, a four-level risk management strategy—comprising “routine monitoring, targeted maintenance, emergency preparedness, and immediate repair”—is proposed to provide reliable data support and a decision-making foundation for transitioning the enterprise's pipeline operation and maintenance model from “reactive response” to “proactive prevention”.