The situation of road traffic safety of non-motorized vehicles remains severe, and extensive research has been conducted in China and abroad. To comprehensively analyze the research status and prospect, 448 Chinese papers and 860 English papers from 2000 to 2025 were selected from China National Knowledge Infrastructure (CNKI) and Web of Science (WOS). Subsequently, the papers were analyzed for keyword co-occurrence and cluster by CiteSpace, and on this basis, the research hotspots and prospect were summarized. The results indicate that current research shows a significant growth trend, and primarily focuses on four areas: risk of riding behaviors, road traffic accidents, mixed traffic flow conflicts, and traffic management strategies. Future studies should expand into diverse scenarios and deepen the application of multimodal data, with methodological support of computer vision technologies and machine learning algorithms to enhance research capabilities.

To address the challenge of accurately predicting the fatigue crack initiation life caused by surface pitting corrosion in buried thermal pipelines during design, operation, and maintenance, this study employed the finite element method to investigate the influence of pit morphology, defect interaction, and axial loading on the maximum stress concentration factor K_t. An empirical formula for K_t was proposed, and a prediction method for fatigue crack initiation life under soil corrosion was developed. A pipeline in service in Beijing was used as a case study to verify the scientificity and effectiveness of this method. The results indicate that a 10-fold increase in pit depth leads to a 2.84-fold increase in K_t, while a 10-fold reduction in pit circumferential width results in a 4.75-fold increase. Deep and narrow defects characterized by a/c > 0.6 and b/c < 0.6 exert a stronger effect on increasing K_t and significantly shorten crack initiation life. When the defect spacing d=0, K_t reaches 1.03 times that of a single defect. The smaller the defect spacing, the stronger the interaction effects and the lower the crack initiation life. As defect spacing decreases, the fatigue crack initiation life of deep narrow pits is reduced to 0.12 times the original life, whereas shallow wide pits are more sensitive to spacing, with their crack initiation life reduced to 0.83 times the original life. Under conditions of low soil resistivity, low pH, and elevated temperature, crack initiation may occur within 20 years. The crack initiation life of shallow wide pits is more sensitive to soil parameters. Compared with internal pressure loading alone, an axial compressive load of 20 MPa significantly reduces crack initiation life by 0.74 times. Increasing axial tensile load from 20 MPa to 50 MPa results in a further life reduction of 0.82 times.

To address the current issues of low intelligence level in coal mine working faces and insufficient research on the performance monitoring of shaft structures, a digital twin-based performance monitoring method for vertical shafts is proposed. Firstly, a five-dimensional framework for the digital twin of vertical shafts is proposed based on the operational mechanism and performance monitoring requirements of the shafts. Secondly, a digital twin of the shaft is established by combining virtual-real mapping technology with a finite element surrogate model for grid dimensionality reduction. The structural performance of the vertical shaft is predicted online through artificial neural network technology, where the predicted data is the real-time prediction of shaft structure performance data obtained during the shaft operation process using a shaft structure performance prediction model. The prediction model for the structural performance of the vertical shaft adopts the RBF surrogate model, and the Unity3D virtual engine platform is built to integrate the above functions and achieve online prediction of the structure performance of the vertical shaft. The results indicate that during the operation, by simulating 120 sets of stress and strain data under different working conditions, the average coefficient of determination between predicted and simulated values is 0.995 5, indicating a high correlation between the predicted strain and simulated strain, thus verifying the feasibility of the digital twin framework for vertical shafts. This provides an effective reference for the digital improvement of vertical shafts.

To address the challenge of locating the fire source in high-rise building fires, a full-scale indoor fire test platform was constructed and a series of tests were conducted to investigate the feasibility of inferring fire locations from the temperature field on the fire-unexposed surface of window glass in high-rise building fires. By varying the fire location and heat release rate, the temperature field on fire-unexposed surface and fire environment parameters were obtained, and the characteristics of fire-unexposed surface temperature field under different scenarios were analyzed. The results show that window glass regions at higher elevations and on the fire side exhibit significantly higher temperature rise rates under different fire location conditions. At 480 s after ignition, the temperature non-uniformity coefficient of fire-unexposed surface under different fire location conditions is not less than 33.52%, and reductions in the distance between fire location and window lead to a marked increase in the temperature non-uniformity coefficient on the fire-unexposed surface. With increasing heat release rate, the coefficient of variation of the increase in temperature rise rates across different glass regions on the fire-unexposed surface generally exceeds 10%, and this disparity becomes more pronounced with increasing heat release rate. When the normal distance between the fire location and the window decreases, window glass at higher elevations and on the fire side exhibits a greater increase in temperature rise rate. When the radial distance between the fire location and the window decreases, the increase in temperature rise rate at higher elevations is significantly greater than that on the fire side.

In order to improve the development of safety behavior of flight cadets and enhance the level of flight training, conduct an indepth exploration of the relationship and intrinsic mechanism between authentic leadership style and safety behavior of flight cadets, a theoretical model of the safety behavior of flight cadets based on the authentic leadership style theory was constructed. The questionnaire was developed by drawing on established scales and consulting experts to align it with the current training conditions of aviation schools. AMOS26.0 software was used to test the mediating effect of authentic follow and the moderating effect of basic psychological needs satisfaction, and verify the effect on all constructs, the applicability of the theoretical model of flight cadets safety behavior in different stages of learning to fly, flight level, instructor job groups and work environment. The results show that there is a positive correlation between the authentic leadership style of flight instructors and the safety behavior of flight cadets, and the authentic follow of flight cadets plays a mediating role between the two. The authentic leadership style has a positive impact on authentic follower, and the basic psychological needs satisfaction plays a moderating role in this process. Different groups between the four dimensions of stages of learning to fly, flight level, instructor job groups and work environment have no moderating effect on the model, further verifying the structural stability of the model.

To address the problem of reduced defect classification accuracy caused by noise contamination in the bend detection signals of oil and gas pipelines, this paper proposes an oil and gas elbow defect diagnosis model based on Welch power spectrum feature enhancement and multi-head attention improved dual-branch multi-scale-residual collaborative network. Firstly, the Welch method was used to convert the collected time domain signal into a feature-enhanced power spectrum, showing the energy distribution of the defect signals at different frequencies. Secondly, the multi-scale network branch composed of parallel stacked convolutional layers was responsible for extracting the multi-dimensional features of the signal power spectrum, and the multi-head attention mechanism was used to establish long-term associations between features. Simultaneously, the residual network branch captured the detail information of the signal power spectrum. Finally, the deep concatenation layer fused the features extracted by the dual-branch network to achieve defect classification. The experiment results show that in a high-noise environment, the test accuracy of the proposed model is 91.6%. Compared with the models based on Kaiser windows and flat-top windows, the classification accuracy is improved by 1%~7.9%; compared with convolutional neural network (CNN) and long short-term memory network (LSTM), the accuracy is improved by 36.9% and 10.3% respectively.

To obtain the powder suppressant with greater inhibition effect, NaHCO₃, Al(OH)₃, K₂CO₃ and NH₄H₂PO₄ were mixed by a compounding method, and a new composite powder was obtained. A visual spherical vessel was applied to study the effects of the composite powder with different mixing ratios on methane explosion overpressure, flame propagation and free radical production. The results indicate that as the mass proportion of NaHCO₃, Al(OH)₃, K₂CO₃ and NH₄H₂PO₄ is 1∶1∶2∶1, the inhibition effect of composite powder is significantly greater than that of the single powder. The synergistic mechanisms of the composite powder are as follows: Before 250 ℃, the composite powder undergoes a metathesis reaction, a large number of gases such as H₂O, CO₂ and NH₃ are released. In the early stage of methane explosion, the energy is absorbed and methane is diluted by the gases, so the rapid flame propagation is inhibited. NH₄H₂PO₄ plays a role in absorbing C-containing free radicals during methane combustion. This effect is enhanced when powder is mixed, leading to a further inhibition for methane combustion reaction.

In order to solve the problems of incomplete index system, low accuracy of calculation results and weak applicability of control measures when evaluating the safety of shield tunneling underpasses, a method for evaluating the safety status of shield tunneling underpasses was studied based on entropy weight cloud model. By analyzing various factors affecting the safety status of the project, it was determined that overall safety level was jointly determined by the original safety level of building and disturbance effect of underpass construction process. An original safety status evaluation index system for buildings was established, which consisted of 13 indicators in three categories: the existing deformation resistance of the building, the healthy and intact state of the building, and the importance of building. A safety evaluation index system for underpass construction process was established, which consisted of 22 indicators in five categories: excavation face instability, displacement of soil on side of shield machine ring, subsequent soil consolidation, construction management level, and relationship between building and tunnel. The weights of each index were determined by entropy weight method, the digital characteristics of the index were determined based on the cloud model, safety level and development trend were determined by FSM algorithm, and the weighted fitting weight value was used to judge the primary and secondary risk control points. Taking a subway project under two buildings as an example, the index system and evaluation method were verified. The results show that original safety status of two buildings and the construction process under two buildings are generally safe, and overall risk is within an acceptable range. For secondary indicators that are not inherent attributes of the project, the factors with weighted fitting weight values exceeding the average level (0.04) are used as key control points, those between 0.02 and 0.04 are used as secondary control points, and those between 0 and 0.02 are used as general control points.

To enhance China's work safety governance level, policy texts on work safety issued by the central government and 30 provincial-level governments from 2014 to 2023, as well as safety accident data from the same period, were used as the research sample. By integrating the Levenshtein distance algorithm, the Jaccard similarity algorithm, and other similarity measures, a "theme-content" two-stage policy synergy analysis model was proposed. Central-local policy synergy degree was quantitatively measured through a weighted evaluation of policy theme matching degree and content similarity. A complex synergy network was then constructed. Using social network analysis and modularity algorithms, the structural characteristics of the central-local safety policy synergy network were analyzed in depth, and its spatiotemporal evolution was revealed. Based on panel data, the impact of central-local synergy outcomes of safety policies on accident incidence were further investigated. The results show that China's central-local safety policy synergy network is exhibiting an increasingly integrated development trend over time. The responsiveness of provincial governments to central policies is continuously improving, evolving from early regional differentiation toward comprehensive nationwide coordination, and an initial "central planning-local response" national work safety policy system has been initially established. Moreover, the central-local policy synergy degree is significantly negatively correlated with the frequency of safety accidents, and improving the level of coordination can effectively reduce the risk of work safety accidents.

In order to address the constraints of limited corpus resources, restricted input capacity, and data privacy in applying LLMs to the field of safety engineering, a localized accident question-answering model integrating the DeepSeek with a RAG mechanism was constructed to enable intelligent parsing and knowledge services for complex texts, thereby supporting safety management decision-making. A semantic-feature corpus was built based on accident investigation reports and laws and regulations released by government emergency management systems, and technologies such as PaddleOCR, LayoutLMv3, and YOLOv8 were incorporated to accomplish document structure reconstruction and semantic modeling. The model encompassed four stages—document parsing, semantic alignment, knowledge-base construction, and hybrid retrieval—and was designed with capabilities for causal-chain extraction, regulation matching, and semantic mapping. The results indicated that, compared with the Deepseek-r1:32b model without the RAG mechanism, the enhanced model achieved improvements of 7.7% in automated scoring and 17.6% in human evaluation, and the response-speed and stability metrics presented higher numerical performance than those of the baseline model. The model performance was still influenced by the local parameter scale and the knowledge-updating mechanism, yet the experimental findings demonstrate that it is capable of fulfilling the intended functions in the present study.