Full-scale experiments in a bifurcated tunnel in a hydropower station with a fire source at the junction area were conducted to investigate fire-induced smoke spread characteristics in connected regions. The spread and descent of smoke in each tunnel were studied by analyzing the temperature profiles and smoke layer heights coupled with on-site observations. The results show that the slope and ventilation strongly affect the smoke risks in each tunnel with the smoke temperature steadily decreasing with distinct stratification upwind of the fire source. The smoke and fresh air mixture flow disrupted the smoke stratification and caused the smoke level to lower in the tunnel with the vertical temperature distribution becoming more uniform. Moreover, the aggravation of smoke deposition in the uphill direction in tunnel 2# increased the smoke temperature in the longitudinal direction below 3 m. The different fire risks in different tunnel sections should be carefully considered in smoke control designs. For the present fire heat release rate, the smoke layer height remained at 4.2 m in the upstream tunnel, while the downstream tunnels faced higher risks that the smoke layer descending below 1.5 m. The uphill slope in tunnel 2# increased the smoke descent with the smoke height decreasing to 0.8 m after 2 min, which should be considered in smoke control designs and fire emergency.

Fires in very large oil tanks are suppressed by injecting foam into the annular seal ring around the top until the ring is completely filled with foam. However, most of the foam ends up on the unburned area which does not control the fire and results in insufficient supply to the fire region. This paper presents a partition fire extinguishing plan for very large oil tanks that is more efficient and reduces foam use without changing the existing oil tank structure or the fire control system. The scheme was evaluated using an annular oil pan and a small external floating roof oil pan for experimental research on the foam fluidity and partitioned fire control. The results compare the effective covering time, the amount of foam and the fire control time. The results show that the foam spillover into different regions is directly related to the gap width and is not related to the oil pan shape and size. The results also show that the fire control time is shorter when using partitioned areas rather than the unpartitioned design. The partitioned fire control method then reduces the foam use by more than 60%. Experiments with a 3 m diameter oil tank fire are consistent with the results of the annular oil pan experiment, which shows that the advantages of zonal fire control are not related to the tank size and that the partitioned fire control scheme can be applied to very large oil tanks.

The low-pressure environments during flights of civil aircraft complicate the fire control and safety in cargo compartments. This study analyzed the effectiveness of large aviation kerosene surface pool fire suppression by water mist in a full-scale simulated cargo compartmentwith a low-pressure environment. The results show that without the water mist, the mass loss rate, flame temperature and radiant heat flux of the pool fire decreasewith decreasing ambient pressure, while the flame height increases.At an initial ambient pressure of 76 kPa, the compartment pressure increases rapidly after ignition, reaching an overpressure of 16.7 kPa, which increases the risk of failure of the cabin structure and ventilation system. With a water mist, the mist droplet size and the burning rate decrease with decreasing ambient pressure and the fire extinguishing time is significantly shortened, but the large firesare difficult to extinguish quickly due to the limited spray flux. In addition, the spray reduces the compartment temperature and radiant heat flux,such as a maximum temperature drop of 150℃ near the ceiling above the flame for an initial ambient pressure of 76 kPa.

Substations are key power system components, so substation fires can have serious consequences on system operations. In particular, more research is needed on the pyrolysis mechanism of cables in substations. There is little research on the chemical reactions and the gas products produced during pyrolysis of substation cables. The current study analyzed the pyrolysis kinetics, solid-gas product characteristics and reaction mechanism of the sheath material of a typical flame-retardant low-voltage power cable (type ZRB-VV22-0.6/1.0 kV) in a substation in the absence of oxygen. Thermogravimetric (TG) experiments were used to investigate the pyrolysis of the sheath material in a nitrogen atmosphere. The experiments showed that the pyrolysis can be divided into two major stages with activation energies of 145.17 and 241.71 kJ/mol for the two stages based on the Friedman, FWO and KAS methods. A field scanning electron microscope and an X-ray energy spectrometer (EDS) were used to analyze the elemental mass fractions of the raw material before pyrolysis and the residue after pyrolysis. The carbon mass fraction decreased due to volatilization of the carbon in the sheath. Ca, Mg and Si were not volatilized but remained in the residue with significantly greater mass fractions. Cl and O were partially volatilized with somewhat higher mass fractions in the residue. The volatilization of Cl infers the existence of HCl in the gas products. The use of Py-GC/MS to quantify the volatile products showed that the products of the first pyrolysis stage (<623 K) are mainly HCl and benzene. When the pyrolysis temperature rises up to 773 K, the products contain polycyclic aromatic hydrocarbons and benzene series. The main pyrolysis process is attributed to the pyrolysis of the PVC in the sheath material. Therefore, the main products are similar to PVC pyrolysis. The HCl yield was 0.077 g/g and the benzene yield was 0.266 g/g while the benzene series yields were relatively low.

This study searched Microblog data related to Typhoon "Mangkhut" in 2018 and "Lekima" in 2019 and then used the Bayesian sentiment analysis model to analyze the public opinions related to these typhoons. The results show two different typhoon disaster public opinion temporal and spatial evolution laws and an emotional evolution law. Then, the urban typhoon disaster loss was estimated for coastal and inland cities based on the temporal and spatial evolution and emotional evolution laws. The data includes the city's geographic location, economy, population and typhoon disaster damage as well as sentiments. The disaster damage assessment model is consistent with the disaster assessment. The research results and methods provide references for disaster research and disaster relief demand analyses and guidelines for supply allocation in cities during initial emergency responses during typhoons.

Rainstorm induced flooding greatly impacts the environment which leads to secondary events and complicates emergency responses. This study analyzes evaluation methods for rainstorm disasters in emerging disaster environments. An emergency decision-making method is developed for rainstorm disasters based on a Markov decision making model that considers the temporal and spatial characteristics of the rainstorm effects. The optimal decision-making strategy that gives the minimum disaster loss is identified by analyzing the rainstorm effects and environmental conditions in a single space along with the impact of various emergency responses. The results show that the spatiotemporal analysis of the rainstorm disaster scenarios based on the Markov decision making process can help decision makers adjust emergency response strategies as the conditions develop to minimize the losses through improved emergency decision-making.

Pluvial floods can inundate urban road networks which then disrupts traffic flows and public services; thereby, increasing emergency response times. This study used the TELEMAC-2D model to simulate flooding of the Qianshanhe catchment for 50 and 100 year design rainstorms to analyze the response times of hospital ambulances, fire station emergency vehicles and police vehicles in the Qianshanhe catchment for various flooding scenarios. The results show that the average ambulance response time without flooding is 19 min, the average emergency vehicle response time is 24 min and the average police vehicle response time is 15.8 min. The 50 year flood scenario will flood some of the roads which will reduce the average ambulance response time to 133.7 min, the average emergency vehicle response time to 241.8 min and the average police vehicle response time to 201 min, which are much longer than the response times without flooding. The 100 year scenario will flood most roads which will reduce the average ambulance response time to 220.1 min, the average emergency vehicle response time to 366 min and the average police vehicle response time to 304 min. Only areas near the hospitals, fire stations or police stations will get rapid responses, while other areas will get very slow responses. These results show that flooding will significantly affect emergency response times and that roads need to be improved to avoid greater losses. The TELEMAC-2D model is very useful for analyzing the effects of flooding and the emergency response capabilities for urban flooding for urban emergency management.

This study analyzed the planning of emergency evacuation routes in densely populated urban areas after earthquakes using an urban university campus as an example. The available campus refuge places were determined from campus literature and in-person investigations. The analysis then considered the influence of a falling outer building wall onto the evacuation route for various earthquake acceleration rates using the concept of an "evacuation virtual wall". Then, the Pathfinder software was used to develop an optimized evacuation plan for various earthquake acceleration rates, crowd distributions, personal coping behavior characteristics and exit arch phenomena that may occur during the evacuation. Finally, escape and evacuation simulation results are compared to determine the best evacuation plan to improve campus safety management.

Inspired by user cognition research, this study explores the image position and layout effects on the transformation process from image content to user engagement behavior in multi-image tweets on the Sina Weibo platform. The XGBoost model trained on the single-image tweet data was used to predict each image's "user engagement potential" in multi-image tweets. Correlation analysis, Z-test, and ordinary least squares (OLS) regression analysis were used to verify the relationship between image position, layout, and user engagement. The results show that in multi-image tweets on the Sina Weibo platform, the image position and layout factors can affect user engagement behavior to a certain extent. That is, (1) for the image position effect, tweets containing 2, 4, 5, and 8 images had the right position effect. Tweets containing 6 and 8 images had the bottom position effect. Tweets containing 3 and 8 images had the edge and middle effects, respectively. Others, in most cases, had symmetric effects. (2) For the image layout effect, Layouts 2, 3, 4, 5, 6, and 8 can positively promote the user engagement potential transfer of images compared with the single image so that the image tweet may achieve user engagement beyond the average level of its potential. However, Layout 7 had a negative effect, and Layout 9 had no significant difference from single-image tweets. The results of this study can provide references for the release of social media image information.

With the advancement of research and development in recommendation systems, more attention has been paid to the precise recommendation of personalized information. The traditional method of collaborative filtering cannot meet the demand due to the scarcity of data in recommendation; thus, contextual information has been introduced to recommendation systems. Review text information containing user preferences is also widely used to alleviate data sparseness and cold start problems. As an unsupervised learning method, the autoencoder performs well in anomaly detection, face recognition, data augmentation, and data generation. The variational autoencoder can learn the distribution of latent vectors of users and items via neural networks. At present, only a few researchers are working for review-aware recommendations using variational autoencoder. This paper proposes a review-aware heterogeneous variational autoencoder recommendation model that introduces comment context information into the variational autoencoder through attention mechanism and neural network. The learning about latent feature distribution of the rating information by the variational autoencoder is retained, and feature fusion is performed in the early and late stages to construct a multimodal heterogeneous variational autoencoder model. Besides, the compound prior term and the balance factor calculation term for multimodal model training are further optimized. The experimental results showed that the proposed model outperforms the state-of-the-art other baseline models in recall and the normalized cumulative gain.

Currently, more people are becoming interested in the field of complex networks. Link prediction is a popular subdiscipline in complex networks and is used to predict missing links and identify false links. The traditional similarity-based complex network link prediction focuses on a particular similarity index of each node. This paper proposes the link prediction algorithm based on clustering coefficient and node centrality (CCNC), which combines the degree index, clustering coefficient index, and proximity centrality index into the link prediction of a complex network. This algorithm considers local information using clustering coefficient and degree by introducing proximity centrality to consider the importance of nodes in the network. Finally, using six real networks as examples, the feasibility and effectiveness of the CCNC algorithm are verified by comparing the AUC and the precision values.

Shallow algorithms of mining and learning such as statistics, clustering, and association relationships are generally used in current research on the prediction and mining of college students' daily behaviors on campus, and there is a lack of in-depth and high-level analysis of the applications in time series, spatial location, and correlation of students' behaviors on campus. Based on the network structure of spatio-temporal graphs, this paper proposes a multifragment semantic spatiotemporal graph convolutional network (MFSTGCN) model that considers the time series of campus activities and the correlation between hierarchy and spatial semantic features. By constructing a data set of college students' campus behaviors and conducting experiments, the model in this paper achieves 90.4% behavior prediction accuracy, which is better than typical prediction models. Finally, we provide students with early warning of abnormal daily behaviors to monitor students' individual growth and excavate high-level information such as student's behaviors and habits to provide a meaningful reference for the construction of personalized education.

Residual welding stresses affect the strength of drive axle housings. The residual welding stresses in axle housings are difficult and expensive to measure, especially the residual stress distribution over the entire structure. This study simulated the residual welding stresses in a drive axle housing of a commercial vehicle using a finite element analysis, including the effects of the strength and volume changes caused by the phase change during welding. The predicted stresses near the weld compared well with experimental data from neutron diffraction measurements, which verified the accuracy of the simulated results. This method can be used to determine the residual welding stress distribution throughout the axle housing and provide guidance for the optimal design of axle housing.

Vehicle mass and the road grade are key input parameters for automated vehicle decision and control algorithms. These two parameters are difficult to know accurately, but are tightly coupled for small acceleration driving scenarios with small inertias. The accuracy and convergence speed of existing estimation algorithms need to be further improved. A vehicle dynamics model is used here with an adaptive unscented Kalman filter (AUKF) algorithm for rapid, accurate joint estimates of the vehicle mass and road grade during small acceleration scenarios. The prediction model has a large initial mass noise with an adaptive shrink coefficient designed to adjust the prediction error covariance matrices. The adaptive shrink coefficient is related to the estimation result and the prediction error covariance matrices. Simulations and vehicle tests for various initial conditions show that the AUKF algorithm can accurately estimate the vehicle mass and road grade with an estimated vehicle mass error of less than 3% and a root-mean-square error in the road grade of less than 0.4°. In addition, the AUKF algorithm gives better convergence speeds for the vehicle mass and road grade estimates with longitudinal accelerations less than 0.3 m/s² than the traditional UKF algorithm. When the initial mass error is within 40%, the estimated mass error converges to less than 3% within 10 seconds.

Accurate motion models for tracked vehicles are difficult to build due to the complex interactions between the tracks and the terrain. A dual-layer unscented Kalman filter (DUKF) is developed to estimate the slip parameters in real time for tracked vehicles. An upper-layer unscented Kalman filter (UKF) is used first to estimate the slip parameters based on the historical trajectory information, which are then imported into the vehicle model based on the instantaneous centers of rotation (ICRs) to predict the forward trajectory. A lower-layer UKF is then used to correct the preliminarily estimated slip parameters based on the residual position of the trajectory prediction. The effectiveness of the DUKF is verified by simulations on RecurDyn and MATLAB/Simulink. The simulations show that the DUKF improves the accuracy of the slip parameter estimation and reduces the trajectory prediction errors with curvatures compared with predictions using the UKF and the extended Kalman filter (EKF).

Parallel cable-driven exoskeletons for motion assistance have the advantages of good compatibility with the human body, large range of motion, and many degrees of freedom. However, it often requires multiple motors to drive multiple cables, increasing the weight and energy consumption of the actuator. In the previous design, it was impossible to reduce the number of motors while realizing the multi-cable force control and multi-degree-of-freedom assistance. This paper proposes an exoskeleton for shoulder joint assistance based on an underactuated parallel cable mechanism, utilizing cams and springs to realize the three-degree-of-freedom assistance of the shoulder joint. According to a given action of the arm, this research conducts kinematics modeling, static modeling, cable arrangement optimization and cam profile designing. As a result, only one motor is used to achieve the different output speeds required by multiple cables. The analysis and simulation of the relationship between the rotation angle of the motor and the force of each cable in accordance with the three-degree-of-freedom assisting relationship of the shoulder joint were done. Results show that the corresponding relationship between the motor rotation angle and the cable force is accurate, verifying the feasibility of the under-driven mechanism for multi-cable and multi-degree-of-freedom force control.

The accuracy of the normal direction during hole drilling greatly affects the assembly accuracy and the service life and safety of aircraft. Therefore, better methods are needed for pose correction based on advanced visual detection methods. A reference normal direction detection method was developed for robotic hole drilling using a robotic visual inspection system that was tested using simulations and a prototype. The systems used a 3-D measurement model of the normal vector detection module to generate a point cloud on the workpiece surface. The point cloud in the camera coordinate system was then transformed into a point cloud in the hole drilling tool coordinate system by an online hand-eye calibration method based on a untraceless Kalman filter. After the normal vector to the surface is identified, a robotic hole drilling attitude correction algorithm is used to guide the drill. 3-D measurements show that this normal vector detection method is accurate and meets the hole specification requirements for the aviation industry.

Visual sensing is an effective means for obtaining arc-welding characteristic information, such as the position and pose of the welding torch, the shape and size of the welding groove, for closed-loop feedback control in intelligent arc welding. The paper describes a visual sensing image processing and feature extraction method for arc welding. A multi-source sensor was developed based on the fusion of visual information with the effect of gravity. The hardware and image preprocessing algorithm are optimized to reduce the interference of the strong arc light, spatter, and other effects on the CCD image. The algorithm then uses the edge extraction based on a Canny operator or the skeleton thinning algorithm based on iterative erosion. The two algorithms separately process the CCD image of the welding groove collected by the multi-source sensor to extract the laser lines, the laser line intersection coordinates and the laser line bending points coordinates caused by the welding groove. Comparison of the feature information extraction speeds and recognition accuracies of the two algorithms shows that the edge extraction algorithm based on the Canny operator can provide real-time weld seam tracking during arc welding.

Vector tracking technology is one of the key technologies of satellite navigation. As the threat of spoofing interference to satellite navigation security is increasing enormously, the influence of spoofing interference on vector receivers must be better understood to improve the anti-spoofing capability of vector tracking technology. This paper focuses on the influence of spoofing interference on three typical vector tracking loops. This study derives the mean tracking error caused by spoofing interference and the loss of lock conditions in the tracking loops. The analysis results were verified using a signal source simulator and a software receiver. The results show that spoofing interference can lead to non-zero mean tracking errors, reduce loop tracking accuracy, and even cause loop lock loses. Vector receivers can use this feature to improve anti-spoofing abilities.

EPFD (equivalence power flux density) is an important indicator for evaluating the interference of a NGSO (non-geostationary orbit) constellation system to a GSO (geostationary orbit) system. The EPFD probability distribution is usually obtained by extrapolating the satellite orbit position and compile statistics the occurrence time proportion of various EPFD values. For mega-constellation systems like Starlink, the traditional extrapolation method to calculate EPFD will involve long, inefficient simulations. This paper presents a simplified method to calculate the EPFD probability distribution. Constellation snapshots are generated by placing reference satellites at various locations in a region to obtain all possible satellite distributions. The snapshot probabilities are then characterized by the occurrence probabilities of the reference satellites, and the EPFD generated for various snapshots used to obtain the EPFD probability distribution. Simulations show that this method can be applied to both traditional constellations with fewer satellites and to mega-constellations. These simulations have the same accuracy as the traditional method, but with higher computational efficiencies.

A position tracking control method was designed with an attitude extraction algorithm for a tailsitter vertical take-off and landing (VTOL) unmanned aerial vehicle (UAV) subjected to time-varying crosswind disturbances. A mathematical model was constructed to predict the relationship between the crosswind disturbance and the aircraft's attitude. An attitude extraction algorithm was then developed using the aircraft's roll angle which significantly reduces the effect of the disturbance. A backstepping control method was used with a saturated function technique to account for the thrust boundedness with a PD plus feedforward control method for the attitude tracking control. The unknown derivatives of the desired attitude caused by the time-varying crosswind disturbance were resolved by a command filter. Tests show that the tracking error can be made arbitrarily small as long as the command filter frequency is sufficiently large.