To solve the problems existing in the traditional NER methods in the domain of tunnel construction safety，such as fuzzy entity boundary，difficulty in small-sample learning，and insufficiently comprehensive extraction of feature information，an entity recognition method for tunnel construction accident text based on the BERT-BiLSTM-CRF model was proposed. Firstly，the BERT model was used to encode the tunnel construction accident text to obtain word vectors containing semantic features. Then，the word vectors output after the training of the BERT model were input into the BiLSTM model to further obtain the context feature of the tunnel construction accident text and conduct label probability prediction. Finally，by utilizing the constraints of the annotation rules of the CRF layer，the output result of the BiLSTM model was corrected，and the maximum probability sequence annotation result was obtained，so as to realize the intelligent classification of the labels of the tunnel construction accident texts. Comparative experiments were conducted between this model and other four commonly used traditional NER models on the tunnel construction safety accident corpus dataset. The results show that the recognition accuracy rate，recall rate and F₁ value of the BERT-BiLSTM-CRF model are 88%，89% and 88% respectively，and the entity recognition effect is better than other benchmark models. By using the established NER model to recognize the entities in the actual tunnel construction accident texts，its application effect in the domain of tunnel construction safety is verified.

To explore the approach for dealing with roof fall by the open TBM in coal mine excavation，the mechanism and pattern recognition of roof fall were investigated considering the unfavorable geological conditions such as abundant water，faults，joints and sandstone with fractured structure. Firstly，the roof fall mechanism was analyzed by utilizing the modified excavation compensation theory and the minimum support stress of surrounding rock which fully considered the intermediate principal stress. Based on the successful case of the main inclined shaft of Kekegai mine and TBM site construction data，the characteristics of roof fall were deeply analyzed. Then，in accordance with the collected on-site feedback monitoring information，the variations of excavation parameters before and after the roof fall were systematically examined，and the machine learning models of random forest (RF)，back propagation (BP) neural network，and Library for support vector machines (LIBSVM) were constructed to effectively identify the roof fall. The results demonstrate that the internal cause of roof fall is the deterioration of sandstone mechanical properties resulting from water-rock interlace in the cataclastic structure，the external cause is the energy release by mechanic-rock action，and the controlling cause is the excavation stress compensation and the timely application of steel anchor (cable) shotcrete + steel arch (steel plate belt). The sharp increase of penetration，and thrust of the hob，the torque of the cutter head and sharp decline of the cutter head speed are the characteristics of roof fall driving parameters. The RF model has the highest prediction accuracy for the classification of roof fall of surrounding rock，and its accuracy rate of identifying roof fall risk is 1.78% and 11.84% higher than that of BP and LIBSVM，respectively.

In order to effectively reduce the casualties and property losses caused by cabin turbulence，a decision analysis method based on a dynamic Bayesian network is proposed for cabin turbulence emergency response. Firstly，according to the relevant laws and regulations at domestic and international，combined with the emergency duties of key personnel on the ground and in the air，the turbulence emergency disposal process is analysed from pre-flight，in-flight and post-flight，and 24 key events are selected to construct a structured BT model. Secondly，the mapping conditions and transformation rules are established to form a DBN model. Then，the objective direct node a priori probability and the supplementary node fuzzy probability obtained by the triangular fuzzy probabilities of supplementary nodes obtained by the fuzzy number expert judgement method to obtain the a priori probabilities of all nodes. Finally，the time slice intervals of 1 and 2 min are selected to focus on the simulation inference of moderate and heavy turbulence，and to study the characteristics of the influence of each dynamic element on the failure of cabin turbulence event disposal. The results show that: the emergency response nodes are significantly affected by the degree of turbulence and time changes，and the optimal time for emergency response is within 5 min. Among them，the probability of failure for the failure of the crew fixation measures in place increases with the increase of the degree of turbulence，human factors such as the failure of the crew to fasten the seat belts and the over-servicing by the cabin crew are the key reasons for the failure of the response.

In order to promote the precise governance goals of accident risks，a research method combining process-tracking based on actor network theory and case comparative analysis was adopted to study the weak prevention generation mechanism in the process of accident risk governance. Firstly，the technical environment for analyzing the evolution process of accident risk was clarified，including prepositive contexts，structural scenarios，and developmental circumstances，as well as the analysis logic based on Lens Model and dimensions of time，space，and structure，to provide a research foundation for the integrity analysis of accident risk production process. Secondly，based on the case comparative analysis of the tracking of the interaction process between different types of actors and risks，the role change picture of the core actors，homeowners and rainstorm，in the process of accident/disaster risk governance and accident/disaster generation was presented integrally. Finally，a comparative analysis and reflection of case studies based on tracking the interaction between different types of actors and risks were summarized，and the essence and generation mechanism of weak prevention in the process of accident risk governance were proposed. The results indicate that the mixed interaction between human and non-human actors can continuously activate the emergence and mutual construction process of new actors，new intersectionality，and new vulnerabilities in the system，subsequently resulting in the non-stationary evolution of accident risk production environment and risk structure，which leads to intervention failure and recurrence of weak prevention in the process of accident risk governance.

To protect wooden buildings from fire threat，the influence mechanism of aging on wood burning characteristics was discussed. Firstly，the combustion behavior of wood was systematically discussed in four aspects: pyrolysis，flame combustion，smoldering combustion and flame spread. Secondly，the combustion characteristics of naturally aged wood and artificially accelerated aged wood were compared and analyzed. Finally，the effects of aging on wood fire risk were analyzed based on the smoke generation characteristics of wood combustion and the flame-spreading behavior of wood structure buildings. The results show that the mechanical properties of wood are significantly reduced by changing the internal composition and carbonization degree of wood，thus weakening the fire resistance and affecting the smoke generation. At the same time，the changes in physical and chemical properties and structural characteristics of aged wood promote the fire spread of ancient wooden buildings in the initial stage of fire. However，the study on the mechanism of wood carbonization by aging is still insufficient，and the influence of aging mode and environmental conditions on the dynamic characteristics of fire at the later stage of combustion has not been clarified. To evaluate the impact of aging on the fire risk of ancient wooden buildings，it is essential to integrate material science and structural mechanics for effective fire safety measures.

In order to investigate the thermal runaway characteristics of lithium-ion batteries following short-term exposure to high or low temperatures shocks during transport and usage，thermal runaway tests were conducted on fully charged lithium-ion batteries. These batteries were subjected to temperature shocks at -40 to 60 ℃ for 24，48，and 96 hours before external heating was applied. Analyses were performed on thermal runaway characteristic parameters to explore change rules in thermal runaway characteristics after short-term high- or low-temperature shocks. The results indicate that the overall severity of thermal runaway and the intensity of ejected flames were attenuated by the shock factor. With rising shock temperature，the time difference (Δt₁) between the onset of thermal runaway and the rupture of the battery safety valve shows slight fluctuations overall，while both the time difference (Δt₂) between peak temperature and the onset of thermal runaway and the initial thermal runaway temperature tends to increase. In contrast，peak thermal runaway temperature and flame heat flux decrease relatively. When the impact temperature decreases，Δt₁ gradually shortens，while Δt₂ initially increases and then decreases with lower temperatures. Both the onset and peak temperatures of thermal runaway drop significantly，with reductions of 15.2 and 175.4 ℃，respectively，observed at -40 ℃，along with a reduction in flame heat flux. Additionally，with extended shock durations，Δt₁，initial and peak temperatures of thermal runaway，and flame heat flux all decrease.

To reveal the influence of asymmetric load on gas seepage and extraction radius，a multi-physics coal and gas fluid-solid coupling model was proposed to analyze the gas seepage characteristics of coal seams. Matrix-adsorbed gas was used as the mass source in the proposed model introducing asymmetric loads into the boundary conditions. Furthermore，segmented drilling was used under asymmetric load conditions to optimize the gas extraction radius and improve extraction efficiency. The results indicated that greater stress compressed the cracks inside the concentrated stress zone，making it more difficult for gas to flow and to be extracted more challenging. The gas pressure in the concentrated stress zone decreased by approximately 2% less than that in the original stress zone，and the permeability decreased by about 9%. Asymmetric load had different degrees of influence on the diffusion and seepage processes. Within 180 days，the mass of diffused gas of the original stress area decreased by 19% and the seepage mass decreased by 20.5%，while these values in the concentrated stress zone decreased by 16.9% and 17.9%，respectively. Asymmetric loads had adverse effects on gas extraction，increasing extraction time under uniform load conditions. By adjusting the extraction radius under asymmetric load conditions，not only can the extraction efficiency be improved by approximately 3%，but it can also ensure that the extraction standards are met within 180 days，thereby effectively improving the overall performance of gas extraction.

This study investigated the protective effect of an active steering control strategy when VRU start from the outside or inside of a curve and cross the road with uniform acceleration，deceleration，and uniform velocity. Firstly，the spatial positional relationship models for vehicles and VRU，the safety assessment models，and the active steering safety distance models were established to propose an active steering control strategy. Then，a lateral collision avoidance controller was designed using the quintic polynomial lane-change method，the Frenet coordinate transformation method，and the model predictive control method. Finally，with electric bicycle riders as the collision avoidance targets，18 mixed conditions were constructed based on the state of the target lane status，the movement direction，and the speed of the electric bicycle to verify the collision avoidance effect of the active steering control strategy. The results indicated the active steering control strategy avoided collisions between vehicles and electric bicycles under mixed conditions. The preceding，ego，and following vehicles in the target lane can drive normally during lane-changing，and the vehicle ride comfort was satisfactory.

In order to assist in the development of safety hazard management measures for hydropower project construction，the safety hazard texts accumulated during the construction inspection of hydropower projects were collected. Entities and relationships from the semi-structured safety hazard texts were extracted using Python. A knowledge graph of safety hazards was constructed and imported into the neo4j graph database for storage. A Sentence-Bidirectional Encoder Representations from Transformer (BERT) model based on bidirectional coding was built for the semantic matching of construction hazards in hydropower projects. The deep semantic features of target hazards and historical hazards were learned，and the historical safety hazards most similar to target hazards were recommended. Using the Cypher query statement，the governance measures corresponding to the historical security risk were searched. The results show that the Sentence-BERT model has an accuracy of 96.48% in identifying architecturally and historically similar safety hazards，which is significantly better than BERT，Word2vec-Deep Semantic Similarity Model (Word2vec-DSSM)，and BERT-DSSM models. Among 150 randomly selected target safety hazard data，the accuracy rate of testing historical similar safety hazard suggestions reaches 92%，and the retrieval effect of hazard management measures is demonstrated through the hazard knowledge graph，which verifies the applicability and effectiveness of the method.

To predict unsafe events for pilots in real time，a LSTM neural network was used to assess pilot maneuver stability and pilot maneuvering quality was improved by optimizing indicators. Firstly，a set of human-machine maneuvering factors presenting the pilot's maneuvering behavior characteristics was proposed by analyzing the pilot's stability maneuvering QAR data in flight. Secondly，the factors affecting the stability maneuvering of the aircraft were analyzed，and a gray correlation analysis method was used to determine the 15 characteristic parameters of associated risks from the 37 monitoring parameters closely related to the stability of the aircraft. Then，the LSTM model was used to train and test the data to predict the pilot's maneuvering stability，and indicators were proposed to evaluate safety stability quality. Finally，ML was used to rank the importance of relevant influencing factors to improve model validity. The results indicated that the time series model effectively eliminated the interference of parameters with little or no correlation with the prediction results in the original parameters. The stability model can predict risks with high accuracy and provide pilots with a 3-4 s time margin to take preventive measures and reduce unsafe incident occurrence during flight.