Method and system for dynamically monitoring combustion smoke density of rail transit material

Abstract

The invention provides a rail transit material combustion smoke density dynamic monitoring method and system, and relates to the technical field of rail transit, and the method comprises the steps: collecting the smoke concentration data of a plurality of sensors in a combustion test box, and carrying out the three-dimensional reconstruction through a self-adaptive grid refined three-dimensional space interpolation algorithm, the method comprises the following steps: constructing a combustion kinetic model by combining material combustion characteristic parameters, fusing a high-precision three-dimensional distribution diagram with the kinetic model, constructing a multi-scale diffusion prediction model based on a graph neural network, and generating smoke density distribution and diffusion path prediction data at future moments, thereby realizing accurate monitoring and early warning of the smoke density of a rail transit carriage.

Description

Technical Field

[0001] The present invention relates to rail transit technology, and in particular to a method and system for dynamically monitoring the density of combustion smoke of rail transit materials. Background Art

[0002] With the rapid development of urban rail transit, the safety of rail transit carriages has received increasing attention. In rail transit carriages, various decorative materials and equipment, once caught fire, will produce large amounts of toxic smoke, seriously threatening the lives of passengers. Monitoring the combustion smoke density of rail transit materials is an important indicator for evaluating the safety performance of materials, and is of great significance for preventing fire accidents and formulating effective evacuation strategies. Traditional rail transit material combustion smoke density detection mainly relies on static testing methods, using optical equipment in a standard test chamber to measure the smoke density generated by the combustion of the material and evaluate the safety performance of the material. With the development of computer technology and sensor technology, it has become possible to dynamically monitor the combustion smoke density of rail transit materials and predict its diffusion path, which is of great value to improving the level of rail transit safety.

[0003] First, traditional rail transit material combustion smoke density monitoring usually uses single-point or a small number of fixed-point sensors for data collection, which cannot fully reflect the three-dimensional distribution of smoke in the combustion test chamber or carriage, resulting in inaccurate monitoring results and difficulty in providing comprehensive data support for carriage safety assessment.

[0004] Secondly, existing monitoring methods are mostly limited to static data collection and simple data analysis. They lack the ability to predict the dynamic process of smoke diffusion and are unable to predict the smoke diffusion path and future concentration distribution in real time. This makes it difficult to make effective evacuation planning and ventilation control decisions when a fire occurs.

[0005] Finally, existing technologies generally fail to effectively combine the combustion characteristics of materials with smoke diffusion models, and lack comprehensive consideration of key parameters such as heat release rate and mass loss rate during material combustion. This leads to the inability to accurately simulate and predict smoke generation and diffusion behavior in complex environments, limiting the accuracy and practicality of fire warning systems. Summary of the Invention

[0006] The embodiments of the present invention provide a method and system for dynamically monitoring the density of combustion smoke of rail transit materials, which can solve the problems in the prior art.

[0007] A first aspect of an embodiment of the present invention provides a method for dynamically monitoring smoke density of rail transit material combustion, comprising:

[0008] Acquiring smoke concentration data collected by a plurality of smoke concentration sensors in a rail transit material combustion test chamber, wherein the plurality of smoke concentration sensors are distributed in a three-dimensional grid in the rail transit material combustion test chamber;

[0009] Based on the position coordinate information of the multiple smoke density sensors and the corresponding smoke concentration data, a three-dimensional spatial interpolation algorithm with adaptive grid refinement is used to perform three-dimensional reconstruction of the smoke density distribution in the rail transit material combustion test box. When the smoke concentration data is higher than a preset concentration threshold, the grid density is automatically increased, and when the smoke concentration data is lower than the preset concentration threshold, the grid density is automatically reduced to obtain a high-precision three-dimensional distribution map;

[0010] Collecting combustion characteristic parameters of materials in the vehicle compartment, including heat release rate, mass loss rate, and combustion product composition data of the materials; and establishing a material combustion kinetics model based on the material combustion characteristic parameters;

[0011] Deeply integrating the high-precision three-dimensional distribution map with the material combustion kinetics model to construct a multi-scale diffusion prediction model based on a graph neural network, wherein the multi-scale diffusion prediction model includes a macroscopic diffusion prediction layer and a microscopic diffusion prediction layer;

[0012] According to the output results of the multi-scale diffusion prediction model, smoke density distribution prediction data and diffusion path prediction data at future moments in the carriage are generated to guide fire evacuation and ventilation control in rail transit carriages.

[0013] The three-dimensional reconstruction of the smoke density distribution in the rail transit material combustion test chamber using a three-dimensional spatial interpolation algorithm with adaptive grid refinement includes:

[0014] Acquire smoke density data collected by multiple smoke concentration sensors in a rail transit material combustion test chamber, wherein the multiple smoke concentration sensors are three-dimensionally distributed in the rail transit material combustion test chamber, and map position information of the multiple smoke concentration sensors into a three-dimensional coordinate system;

[0015] Calculating a spatial gradient value of smoke density in the rail transit material combustion test box, multiplying a ratio of the spatial gradient value to the maximum smoke density at a current moment by an initial grid size to obtain a grid refinement criterion; and adaptively refining the spatial grid of the rail transit material combustion test box according to the grid refinement criterion;

[0016] Establishing a three-dimensional spatial interpolation model based on the adaptively refined grid structure, and determining an interpolation weight coefficient according to local grid characteristics within the rail transit material combustion test chamber, wherein the interpolation weight coefficient is composed of a credibility coefficient of a smoke concentration sensor and a distance attenuation function based on a local grid size;

[0017] The interpolation weight coefficient is input into the three-dimensional space interpolation model, the smoke density value of the grid node in the rail transit material combustion test box is calculated, and the three-dimensional reconstruction result of the smoke density distribution in the rail transit material combustion test box is obtained.

[0018] Establishing a material combustion kinetics model according to the material combustion characteristic parameters includes:

[0019] collecting thermogravimetric data of the material at different heating rates, calculating the mass loss rate at each temperature point based on the thermogravimetric data, determining a peak temperature point based on the mass loss rate, using the peak temperature point as a characteristic temperature for material decomposition, and dividing the combustion process of the material into multiple sub-reaction stages based on the characteristic temperature;

[0020] For each of the sub-reaction stages, selecting multiple characteristic conversion rate points and their corresponding temperature data, calculating the mass loss of each stage based on the characteristic conversion rate points, and substituting the mass loss, the temperature data, and the heating rate into the Flynn-Wall-Ozhawa equation;

[0021] Performing a logarithmic transformation on the Flynn-Wall-Ozawa equation to convert the nonlinear equation into a linear relationship between activation energy and frequency factor, establishing a parametric equation system, and solving the parametric equation system using the least squares method to obtain the initial kinetic parameters of each sub-reaction;

[0022] A parameter optimization space is constructed based on the initial kinetic parameters, and the parameter optimization space is input into a genetic algorithm optimization model, with the root mean square error between the experimental curve and the calculated curve as the optimization target; based on the optimization target, parameter optimization calculations are performed through selection, crossover and mutation operations of the genetic algorithm optimization model to obtain the optimal activation energy, frequency factor and reaction order, and a material combustion kinetics model is established based on the optimal activation energy, the frequency factor and the reaction order.

[0023] The Flynn-Wall-Ozawa equation is logarithmically transformed to convert the nonlinear equation into a linear relationship between activation energy and frequency factor, and a parametric equation system is established. The least squares method is used to solve the parametric equation system to obtain the initial kinetic parameters of each sub-reaction, including:

[0024] Obtaining a temperature reaction rate in a Flynn-Wall-Ozawa equation, wherein the Flynn-Wall-Ozawa equation includes a reaction conversion rate, a frequency factor, an activation energy, and a reaction order, wherein the Flynn-Wall-Ozawa equation represents an exponential functional relationship between the temperature reaction rate and the reaction conversion rate, the frequency factor, the activation energy, and the reaction order; performing a logarithmic transformation on the exponential functional relationship to obtain a linearized equation;

[0025] Establishing a set of parameter equations based on the linearized equations, constructing a target data sequence using multiple sets of reaction rate logarithmic values, and constructing a coefficient data sequence using corresponding temperature inverse values ​​and conversion rate logarithmic values, wherein the set of parameter equations characterizes a linear relationship between the target data sequence and the coefficient data sequence;

[0026] The least squares method is used to solve the parameter equation group, and the sum of the squares of the experimental values ​​and the fitted values ​​of the target data sequence is used as the optimization objective function. The parameter sequence that minimizes the optimization objective function is solved to obtain the optimal estimated value; the kinetic parameter recovery is performed on the optimal estimated value to obtain the initial kinetic parameters of each sub-reaction.

[0027] Deeply integrating the high-precision three-dimensional distribution map with the material combustion kinetics model to construct a multi-scale diffusion prediction model based on a graph neural network includes:

[0028] Dividing the high-precision three-dimensional distribution map into a grid structure, measuring physical parameters of nodes, substituting the physical parameters into the material combustion kinetics model to calculate node reaction rates, constructing node feature vectors, and constructing edge feature vectors based on calculated transmission parameters of adjacent nodes;

[0029] A dual feature fusion network is designed, and the node feature vector and the edge feature vector are mapped respectively using a first weight matrix and a second weight matrix. A micro-macro feature mapping matrix is ​​obtained through an activation function, and an evolution trend of the node state is calculated based on the micro-macro feature mapping matrix.

[0030] The evolution trend is input into a multi-level prediction network to calculate the prediction error of the local reaction rate and the overall transmission field, and when the prediction error is less than a preset error threshold, a material combustion diffusion prediction model is output.

[0031] Based on the output results of the multi-scale diffusion prediction model, the smoke density distribution prediction data and diffusion path prediction data at future moments in the carriage are generated to guide the fire evacuation and ventilation control of rail transit carriages, including:

[0032] Determining the temporal and spatial features of an output result of the multi-scale diffusion prediction model, and fusing the temporal and spatial features to obtain spatiotemporal enhancement features;

[0033] Inputting the spatiotemporal enhancement features into a causal convolutional network, extracting temporal features using a dilated convolution operation, transferring the temporal features between layers using a residual connection structure, calculating the temporal correlation coefficient of the transferred features, determining the optimal number of network layers and dilation rate based on the temporal correlation coefficient, and generating a feature sequence with long-range dependencies;

[0034] Constructing a dual-branch prediction network based on the feature sequence, generating smoke density distribution prediction data in the vehicle compartment at a future time using a density prediction branch, generating diffusion path prediction data at a future time using a path prediction branch, calculating a prediction error between the smoke density distribution prediction data and the diffusion path prediction data, and dynamically adjusting a weight coefficient of the prediction branch based on the prediction error;

[0035] Using a probabilistic prediction model to evaluate the reliability of the predicted data after weight adjustment, calculating the confidence interval and calibration error of the reliability of the predicted data, iteratively optimizing the parameters of the probabilistic prediction model based on the confidence interval and the calibration error, and outputting a prediction result with guaranteed reliability;

[0036] The safety level of each area in the carriage is calculated according to the prediction results, the optimal evacuation channel is planned based on the safety level, the smoke concentration deviation of each area is calculated using the smoke density distribution prediction data, and ventilation control instructions are generated according to the smoke concentration deviation.

[0037] Using the probabilistic prediction model to evaluate the reliability of the predicted data after weight adjustment, calculating the confidence interval and calibration error of the reliability of the predicted data, iteratively optimizing the parameters of the probabilistic prediction model based on the confidence interval and the calibration error, and outputting a prediction result with reliability assurance, including:

[0038] Constructing a weight matrix of a prediction parameter association network, performing association analysis on the prediction parameters based on the prediction parameter association network and the weight matrix, setting an error threshold between a predicted value and a true value as a reliability constraint, screening an optimal parameter combination based on the weight matrix and the reliability constraint, calculating a probability deviation between the predicted value and the true value based on the optimal parameter combination, and establishing a reliability probability distribution of the prediction result;

[0039] Constructing a dynamic confidence interval based on the reliability probability distribution and the weight matrix, constructing a calibration error calculation model using the dynamic confidence interval and the reliability probability distribution, inputting the uncertainty state of the prediction system into the calibration error calculation model, and obtaining a dynamic calibration error during the prediction process;

[0040] The weight matrix is ​​used to construct a knowledge modulation matrix, the dynamic calibration error is weighted by the knowledge modulation matrix, an adaptive learning rate is calculated based on the weighted dynamic calibration error, the parameters of the prediction model are iteratively optimized to obtain optimal prediction parameters; a prediction calculation is performed based on the optimal prediction parameters and the weight matrix, and a prediction result with reliability guarantee is output.

[0041] A second aspect of an embodiment of the present invention provides a dynamic monitoring system for smoke density of rail transit material combustion, comprising:

[0042] The first unit is used to obtain smoke concentration data collected by multiple smoke concentration sensors in a rail transit material combustion test box, wherein the multiple smoke concentration sensors are distributed in a three-dimensional grid in the rail transit material combustion test box;

[0043] The second unit is configured to perform three-dimensional reconstruction of the smoke density distribution in the rail transit material combustion test chamber using a three-dimensional spatial interpolation algorithm with adaptive grid refinement based on the position coordinate information of the multiple smoke density sensors and the corresponding smoke concentration data, and automatically increase the grid density when the smoke concentration data is higher than a preset concentration threshold, and automatically reduce the grid density when the smoke concentration data is lower than the preset concentration threshold, so as to obtain a high-precision three-dimensional distribution map;

[0044] The third unit is used to collect material combustion characteristic parameters in the vehicle compartment, including the material's heat release rate, mass loss rate, and combustion product composition data; and establish a material combustion kinetics model based on the material combustion characteristic parameters;

[0045] A fourth unit is configured to deeply integrate the high-precision three-dimensional distribution map with the material combustion kinetics model to construct a multi-scale diffusion prediction model based on a graph neural network, wherein the multi-scale diffusion prediction model includes a macroscopic diffusion prediction layer and a microscopic diffusion prediction layer;

[0046] The fifth unit is used to generate smoke density distribution prediction data and diffusion path prediction data in the car at future moments based on the output results of the multi-scale diffusion prediction model, which are used to guide fire evacuation and ventilation control in rail transit cars.

[0047] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0048] processor;

[0049] a memory for storing processor-executable instructions;

[0050] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0051] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0052] The beneficial effects of this application are as follows:

[0053] The present invention arranges multiple smoke concentration sensors in a three-dimensional grid inside a rail transit material combustion test chamber and uses a three-dimensional spatial interpolation algorithm with adaptive grid refinement to achieve high-precision three-dimensional reconstruction of smoke density distribution. This solves the technical problems of low spatial resolution and inability to reflect the actual uneven distribution of smoke in traditional smoke density measurement methods, and greatly improves the accuracy and spatial resolution of smoke density monitoring.

[0054] This invention deeply integrates high-precision three-dimensional distribution maps with material combustion dynamics models to construct a multi-scale diffusion prediction model based on graph neural networks. It can predict smoke diffusion behavior from both macro and micro levels, significantly improving the accuracy of smoke diffusion path and density prediction. Compared with traditional methods, the prediction accuracy is increased by 25%-35%, and the prediction time is advanced by 3-5 minutes.

[0055] The present invention realizes real-time monitoring of smoke density distribution in rail transit carriages and prediction of future moments, providing a scientific basis for fire evacuation and ventilation control when a carriage fire occurs. It can effectively reduce the risk of casualties in fire accidents and improve the safety level of rail transit. It has important social value and application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a flow chart of a method for dynamically monitoring smoke density of rail transit material combustion according to an embodiment of the present invention;

[0057] Figure 2 Schematic diagram of real-time reconstruction error under changing smoke density during combustion according to an embodiment of the present invention;

[0058] Figure 3 This is a flow chart of linearizing the Flynn-Wall-Ozhava equation to solve kinetic parameters according to an embodiment of the present invention;

[0059] Figure 4 This is a flow chart of material combustion and diffusion prediction based on dual feature fusion according to an embodiment of the present invention;

[0060] Figure 5 Schematic diagram of prediction accuracy comparison under different training sample sizes in an embodiment of the present invention. DETAILED DESCRIPTION

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0062] The technical solution of the present invention is described in detail below with reference to specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0063] Figure 1 FIG. 1 is a flow chart of a method for dynamically monitoring the density of combustion smoke of rail transit materials according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0064] Acquiring smoke concentration data collected by a plurality of smoke concentration sensors in a rail transit material combustion test chamber, wherein the plurality of smoke concentration sensors are distributed in a three-dimensional grid in the rail transit material combustion test chamber;

[0065] Based on the position coordinate information of the multiple smoke density sensors and the corresponding smoke concentration data, a three-dimensional spatial interpolation algorithm with adaptive grid refinement is used to perform three-dimensional reconstruction of the smoke density distribution in the rail transit material combustion test box. When the smoke concentration data is higher than a preset concentration threshold, the grid density is automatically increased, and when the smoke concentration data is lower than the preset concentration threshold, the grid density is automatically reduced to obtain a high-precision three-dimensional distribution map;

[0066] Collecting combustion characteristic parameters of materials in the vehicle compartment, including heat release rate, mass loss rate, and combustion product composition data of the materials; and establishing a material combustion kinetics model based on the material combustion characteristic parameters;

[0067] Deeply integrating the high-precision three-dimensional distribution map with the material combustion kinetics model to construct a multi-scale diffusion prediction model based on a graph neural network, wherein the multi-scale diffusion prediction model includes a macroscopic diffusion prediction layer and a microscopic diffusion prediction layer;

[0068] According to the output results of the multi-scale diffusion prediction model, smoke density distribution prediction data and diffusion path prediction data at future moments in the carriage are generated to guide fire evacuation and ventilation control in rail transit carriages.

[0069] In an optional embodiment, using a three-dimensional spatial interpolation algorithm with adaptive grid refinement to perform three-dimensional reconstruction of the smoke density distribution in the rail transit material combustion test chamber includes:

[0070] Acquire smoke density data collected by multiple smoke concentration sensors in a rail transit material combustion test chamber, wherein the multiple smoke concentration sensors are three-dimensionally distributed in the rail transit material combustion test chamber, and map position information of the multiple smoke concentration sensors into a three-dimensional coordinate system;

[0071] Calculating a spatial gradient value of smoke density in the rail transit material combustion test box, multiplying a ratio of the spatial gradient value to the maximum smoke density at a current moment by an initial grid size to obtain a grid refinement criterion; and adaptively refining the spatial grid of the rail transit material combustion test box according to the grid refinement criterion;

[0072] Establishing a three-dimensional spatial interpolation model based on the adaptively refined grid structure, and determining an interpolation weight coefficient according to local grid characteristics within the rail transit material combustion test chamber, wherein the interpolation weight coefficient is composed of a credibility coefficient of a smoke concentration sensor and a distance attenuation function based on a local grid size;

[0073] The interpolation weight coefficient is input into the three-dimensional space interpolation model, the smoke density value of the grid node in the rail transit material combustion test box is calculated, and the three-dimensional reconstruction result of the smoke density distribution in the rail transit material combustion test box is obtained.

[0074] Several smoke concentration sensors are installed in a three-dimensional distribution within a rail transit material combustion test chamber. The chamber measures 2 meters by 1.5 meters by 2.5 meters and houses 32 smoke concentration sensors. Each sensor has a unique three-dimensional coordinate identifier. For example, the first sensor's coordinates are (0.2, 0.3, 0.4) meters, the second's are (0.2, 0.3, 1.2) meters, and so on, until all 32 sensors are mapped. This coordinate data is stored in the system database and serves as the basis for subsequent calculations.

[0075] Obtain the smoke density readings from each smoke concentration sensor at a specific time t. For example, at t = 30 seconds, the first sensor reads 0.15 g / m³, the second sensor reads 0.21 g / m³, and so on for all 32 sensors. After collecting this data, the system associates it with the corresponding sensor location information to form the data point set {(xi, yi, zi, di)}, where (xi, yi, zi) represents the three-dimensional coordinates of the i-th sensor and di represents the smoke density value measured by that sensor.

[0076] The rail transit material combustion test chamber was initially meshed, with the initial grid size set to 0.2 m × 0.2 m × 0.2 m. This resulted in 1040 initial grid cells, 10 × 8 × 13. For each grid cell, the system estimated the smoke density gradient at the cell's center based on surrounding sensor data.

[0077] When calculating the spatial gradient of smoke density, the system considers the estimated smoke density values ​​at the center points of the six adjacent grid cells (top, bottom, left, right, front, and back) for each grid cell. Assuming the coordinates of a grid cell's center point are (x, y, z) and its estimated density is d, its x-direction gradient is calculated by dividing the density difference between that point and its adjacent points at x+0.2 and x-0.2 by the distance (0.4 meters). The same principle is used to calculate the gradients in the y and z directions. The square root of the sum of the squares of the gradient values ​​in the three directions is taken to obtain the spatial gradient amplitude at that point.

[0078] For each grid cell, calculate its mesh refinement criterion. Assuming the current maximum smoke density is dmax = 0.35 g / m³, and the spatial gradient value of a grid cell is g = 0.25 g / m³ / m, the mesh refinement criterion value for that cell is (g / dmax) × 0.2 = (0.25 / 0.35) × 0.2 ≈ 0.143. The system sets a threshold of 0.1; when the refinement criterion value exceeds this threshold, the grid cell is refined.

[0079] During the mesh refinement process, grid cells with criterion values ​​exceeding the threshold are divided into eight sub-grid cells (2 × 2 × 2). For example, if the original grid cell size is 0.2 m × 0.2 m × 0.2 m, the sub-grid cell size becomes 0.1 m × 0.1 m × 0.1 m after refinement. In this example, approximately 150 initial grid cells are refined, increasing the total number of grid cells in the system from 1040 to approximately 2240.

[0080] Based on the mesh refinement results, the system builds a three-dimensional spatial interpolation model. For each mesh node where smoke density needs to be estimated, the system first identifies the sensors that affect it. The influence radius is set to 0.5 meters, meaning that sensor data within a 0.5-meter radius of the target node will be included in the calculation. The selected sensor data is weighted, and the weight coefficient consists of two parts:

[0081] The sensor reliability factor is determined based on the sensor's historical stability and calibration accuracy. For example, the first sensor's reliability is 0.95, the second's is 0.92, and so on. The distance decay function uses an exponential decay function, with the effect decreasing with increasing distance. For example, if the sensor is r meters from the target node, the distance decay factor can be expressed as exp(-3r / h), where h is the local grid size. This decay factor differs for refined regions (h = 0.1 meter) and non-refined regions (h = 0.2 meter).

[0082] Multiply the sensor credibility factor by the distance attenuation factor to obtain the final interpolation weight coefficient. For example, if a sensor has a credibility of 0.95, is 0.15 meters away from the target node, and is located in an area with a grid size of 0.1 meters, its weight coefficient is 0.95×exp(-3×0.15 / 0.1)≈0.25.

[0083] Using the weighted coefficients above, we perform a weighted average of the smoke density data from each sensor to calculate the smoke density estimate for the target node. This calculation process is repeated for all grid nodes, ultimately resulting in a 3D reconstruction of the smoke density distribution for the entire test chamber.

[0084] Through the above implementation method, the system can generate high-resolution three-dimensional visualization results of smoke density distribution, especially providing more detailed reconstruction effects in areas with large smoke density gradients (such as around fire sources), thereby providing a more accurate analysis basis for the research on the combustion safety of rail transit materials.

[0085] Figure 2 This is a schematic diagram of real-time reconstruction error under changing smoke density during combustion according to an embodiment of the present invention:

[0086] This figure shows the change trend of the prediction error (RMSE) of three different grid methods (this technical solution, fixed grid method and standard adaptive method) over time during the material combustion process. The horizontal axis represents the combustion time, ranging from 0 to 60 seconds; the vertical axis represents the reconstruction error, ranging from 0 to 0.20. It can be observed from the figure that all three methods show a bimodal feature, but there are obvious differences in the error amplitude and degree of fluctuation. The fixed grid method (circular mark) has the largest error fluctuation, with two peaks at 20 seconds and 45 seconds respectively, and the highest error reaches about 0.19; the standard adaptive method (square mark) performs better, with two peaks appearing near 20 seconds and 45 seconds respectively, and the maximum error is about 0.13; the performance of this technical solution (triangle mark) is the best, and the overall error curve is the smoothest. The errors at the two peaks are only about 0.07 and 0.08 respectively, and the lowest prediction error is always maintained throughout the combustion process. Especially in the early stage (0-10 seconds) and late stage (50-60 seconds) of combustion, the error of this technical solution is maintained at a low level of 0.02-0.03, showing excellent prediction stability and accuracy.

[0087] In an optional embodiment, establishing a material combustion kinetics model according to the material combustion characteristic parameters includes:

[0088] collecting thermogravimetric data of the material at different heating rates, calculating the mass loss rate at each temperature point based on the thermogravimetric data, determining a peak temperature point based on the mass loss rate, using the peak temperature point as a characteristic temperature for material decomposition, and dividing the combustion process of the material into multiple sub-reaction stages based on the characteristic temperature;

[0089] For each of the sub-reaction stages, selecting multiple characteristic conversion rate points and their corresponding temperature data, calculating the mass loss of each stage based on the characteristic conversion rate points, and substituting the mass loss, the temperature data, and the heating rate into the Flynn-Wall-Ozhawa equation;

[0090] Performing a logarithmic transformation on the Flynn-Wall-Ozawa equation to convert the nonlinear equation into a linear relationship between activation energy and frequency factor, establishing a parametric equation system, and solving the parametric equation system using the least squares method to obtain the initial kinetic parameters of each sub-reaction;

[0091] A parameter optimization space is constructed based on the initial kinetic parameters, and the parameter optimization space is input into a genetic algorithm optimization model, with the root mean square error between the experimental curve and the calculated curve as the optimization target; based on the optimization target, parameter optimization calculations are performed through selection, crossover and mutation operations of the genetic algorithm optimization model to obtain the optimal activation energy, frequency factor and reaction order, and a material combustion kinetics model is established based on the optimal activation energy, the frequency factor and the reaction order.

[0092] To establish a combustion kinetics model based on the material's combustion characteristics, the first step is to collect thermogravimetric data. Using a thermogravimetric analyzer, the mass change of the target material at different heating rates (e.g., 5°C / min, 10°C / min, 15°C / min, and 20°C / min) is measured to generate a temperature-mass curve. By processing this collected thermogravimetric data, the mass loss rate at each temperature point—that is, the rate of mass change with temperature—is calculated.

[0093] By analyzing the mass loss rate curve, we can identify peak temperatures at various heating rates. For example, at a heating rate of 10°C / min, a polymer material might experience mass loss peaks at 350°C, 450°C, and 550°C. These peak temperatures are identified as characteristic temperatures of the material's thermal decomposition and used to divide the combustion process into multiple sub-reaction stages. For example, the combustion process can be divided into three stages: the first stage is from room temperature to the first characteristic temperature; the second stage is from the first characteristic temperature to the second characteristic temperature; and the third stage is from the second characteristic temperature to the end of combustion.

[0094] For each defined sub-reaction stage, multiple characteristic conversion rate points (e.g., 0.1, 0.2, 0.3, ..., 0.9) and their corresponding temperature data were selected. For a wood material, for example, in the first stage of the reaction, when the conversion rate was 0.2, the corresponding temperatures were 210°C, 225°C, and 240°C at heating rates of 5°C / min, 10°C / min, and 20°C / min, respectively.

[0095] Based on these characteristic conversion points, the mass loss rate for each stage was calculated. The mass loss rate is defined as the ratio of the mass lost at a certain point in the stage to the total mass lost during that stage. The mass loss rate, temperature data, and heating rate are substituted into the Flynn-Wall-Ozhawa equation, which describes the relationship between conversion rate, temperature, and time during the thermal decomposition of a material.

[0096] To simplify the solution process, the Flynn-Wall-Ozawa equation was logarithmically transformed, converting the nonlinear equation containing exponential relationships into a linear expression of the activation energy and frequency factor. This transformation simplifies the complex nonlinear problem into a linear one, facilitating subsequent solution. For this transformed linear equation, a parametric equation system was established, containing data from multiple characteristic conversion rate points.

[0097] The least-squares method was used to solve this parametric equation system, yielding the initial kinetic parameters for each sub-reaction stage, including activation energy, frequency factor, and reaction order. For a plastic material, the least-squares method calculated the initial activation energy of the first-stage reaction to be 120 kJ / mol, the frequency factor to be 10^13 s^-1, and the reaction order to be 1.2.

[0098] After obtaining the initial kinetic parameters, a parameter optimization space is constructed for further optimization. This parameter optimization space is a multidimensional search space centered on the initial kinetic parameters, with a range of fluctuations (e.g., activation energy ±20 kJ / mol, frequency factor ±1 order of magnitude, reaction order ±0.5). This parameter optimization space is then fed into the genetic algorithm optimization model, with the root mean square error (RMSE) between the experimentally measured thermogravimetric curve and the curve predicted by the computational model being used as the optimization target. The RMSE is calculated as the square root of the sum of the squares of the mass differences between the two curves at corresponding temperatures, divided by the number of data points.

[0099] During the genetic algorithm optimization process, parameter optimization is performed through selection, crossover, and mutation operations. Specifically, the selection operation selects high-quality individuals based on their fitness (the inverse of the error value) using a roulette wheel method; the crossover operation swaps the parameters of the selected individuals with a certain probability (such as 0.8); and the mutation operation randomly fine-tunes the individual parameters with a lower probability (such as 0.1).

[0100] After multiple generations of evolution (e.g., 100 generations), the algorithm converges to an optimal solution or a state close to it. For example, for a composite material, after genetic algorithm optimization, the optimal activation energy for the second-stage reaction was 155 kJ / mol, the optimal frequency factor was 10^14.5 s^-1, and the optimal reaction order was 1.5. At this point, the root mean square error between the calculated and experimental curves dropped to 0.003.

[0101] Based on the obtained optimal activation energy, frequency factor, and reaction order, a complete material combustion kinetic model was established. This model can describe the material's decomposition rate and mass loss under different temperature conditions, thereby predicting the material's combustion behavior in actual fire environments. Model validation shows that the established kinetic model can accurately predict the material's thermal decomposition process over a heating rate range of 5°C / min to 20°C / min, with a prediction error of less than 5%, meeting the requirements of engineering applications.

[0102] In an optional embodiment, the Flynn-Wall-Ozhawa equation is logarithmically transformed to convert the nonlinear equation into a linear relationship between activation energy and frequency factor, a parametric equation system is established, and the least squares method is used to solve the parametric equation system to obtain the initial kinetic parameters of each sub-reaction, including:

[0103] Obtaining a temperature reaction rate in a Flynn-Wall-Ozawa equation, wherein the Flynn-Wall-Ozawa equation includes a reaction conversion rate, a frequency factor, an activation energy, and a reaction order, wherein the Flynn-Wall-Ozawa equation represents an exponential functional relationship between the temperature reaction rate and the reaction conversion rate, the frequency factor, the activation energy, and the reaction order; performing a logarithmic transformation on the exponential functional relationship to obtain a linearized equation;

[0104] Establishing a set of parameter equations based on the linearized equations, constructing a target data sequence using multiple sets of reaction rate logarithmic values, and constructing a coefficient data sequence using corresponding temperature inverse values ​​and conversion rate logarithmic values, wherein the set of parameter equations characterizes a linear relationship between the target data sequence and the coefficient data sequence;

[0105] The least squares method is used to solve the parameter equation group, and the sum of the squares of the experimental values ​​and the fitted values ​​of the target data sequence is used as the optimization objective function. The parameter sequence that minimizes the optimization objective function is solved to obtain the optimal estimated value; the kinetic parameter recovery is performed on the optimal estimated value to obtain the initial kinetic parameters of each sub-reaction.

[0106] To obtain more accurate initial kinetic parameters during fuel kinetic analysis, this embodiment provides a parameter solution method based on the Flynn-Wall-Ozawa equation. The Flynn-Wall-Ozawa equation is a commonly used equation for describing the pyrolysis process of solid fuels. It expresses the exponential relationship between the temperature reaction rate and the reaction conversion rate, frequency factor, activation energy, and reaction order.

[0107] In this embodiment, the temperature-dependent reaction rate data from the Flynn-Wall-Ozhawa equation is first obtained. Through thermogravimetric analysis, the mass change data of the sample at different temperatures can be obtained. Based on the mass change curve, the reaction conversion rate α and the corresponding reaction rate dα / dt at different temperature points can be calculated. For example, if a coal sample is pyrolyzed at a heating rate of 10°C / min, mass change data can be obtained for the temperature range of 200°C to 900°C.

[0108] Based on the mass change data, we calculated that at 350°C, the reaction conversion α was 0.15, and the reaction rate dα / dt was 0.002 min-1; at 450°C, the reaction conversion α was 0.38, and the reaction rate dα / dt was 0.005 min-1; and at 550°C, the reaction conversion α was 0.72, and the reaction rate dα / dt was 0.009 min-1.

[0109] A logarithmic transformation is performed on the Flynn-Wall-Ozawa equation, converting the exponential relationship into a linear one. The Flynn-Wall-Ozawa equation describes the relationship between the reaction rate (dα / dt) and the reaction conversion (α), the frequency factor (A), the activation energy (E), and the reaction order (n). By performing a logarithmic transformation, a linearized equation is obtained, in which the logarithmic value of the reaction rate is linearly related to the inverse of the temperature and the logarithmic value of the reaction conversion.

[0110] For example, for the above experimental data, when the temperature is 350℃ (i.e. 623K), the logarithm of the reaction rate dα / dt is -6.2146, the reciprocal of the temperature 1 / T is 0.00161, and the logarithm of the reaction conversion rate α is -1.8971; when the temperature is 450℃ (i.e. 723K), the logarithm of the reaction rate is -5.2983, the reciprocal of the temperature is 0.00138, and the logarithm of the reaction conversion rate is -0.9676; when the temperature is 550℃ (i.e. 823K), the logarithm of the reaction rate is -4.7105, the reciprocal of the temperature is 0.00122, and the logarithm of the reaction conversion rate is -0.3285.

[0111] A parametric system of equations is established based on the linearized equations. Multiple sets of reaction rate logarithms are used to construct a target data sequence, and the corresponding temperature reciprocals and conversion logarithms are used to construct a coefficient data sequence. For example, for the three data sets mentioned above, the target data sequence is [-6.2146, -5.2983, -4.7105], and the coefficient data sequence includes the temperature reciprocals [0.00161, 0.00138, 0.00122] and the reaction conversion logarithms [-1.8971, -0.9676, -0.3285]. The parametric system of equations characterizes the linear relationship between the target data sequence and the coefficient data sequence. The parameters to be determined include coefficients related to activation energy, reaction order, and frequency factor.

[0112] The least squares method is used to solve the parametric equations. The sum of the squared errors between the experimental and fitted values ​​of the target data sequence is used as the optimization objective function. The optimal estimate of the parameter sequence is determined by minimizing the optimization objective function. For example, for the above data, the least squares method yields a coefficient of -15,000 for the activation energy, 0.8 for the reaction order, and 16.2 for the frequency factor.

[0113] Kinetic parameter recovery was performed on the optimal estimates to obtain the initial kinetic parameters for each subreaction. Based on the coefficients related to the activation energy, the activation energy E was calculated to be 124.7 kJ / mol; based on the coefficients related to the reaction order, the reaction order n was calculated to be 0.8; and based on the coefficients related to the frequency factor, the frequency factor A was calculated to be 10^7 min^-1. These parameter values ​​served as initial values ​​for subsequent kinetic model optimization.

[0114] To verify the accuracy of the parameter solution, the obtained initial kinetic parameters were substituted into the Flynn-Wall-Ozhawa equation to calculate the theoretical reaction rate and compare it with the experimentally measured reaction rate. For example, for a temperature of 350°C and a reaction conversion of 0.15, the calculated theoretical reaction rate was 0.0019 min⁻¹, close to the experimental value of 0.002 min⁻¹. For a temperature of 450°C and a reaction conversion of 0.38, the calculated theoretical reaction rate was 0.0048 min⁻¹, close to the experimental value of 0.005 min⁻¹. For a temperature of 550°C and a reaction conversion of 0.72, the calculated theoretical reaction rate was 0.0087 min⁻¹, close to the experimental value of 0.009 min⁻¹. The calculation results demonstrate the good accuracy of the initial kinetic parameters obtained using this method.

[0115] Through the above method, the nonlinear Flynn-Wall-Ozhawa equation can be effectively converted into a linear relationship, and the least squares method can be used to solve the parametric equation group to obtain the initial kinetic parameters of each sub-reaction, providing basic data support for subsequent fuel pyrolysis kinetics analysis.

[0116] Figure 3 This is a flow chart of linearizing the Flynn-Wall-Ozhava equation to solve the kinetic parameters according to an embodiment of the present invention:

[0117] The figure details the kinetic parameter solution process based on the Flynn-Wall-Ozhawa equation, presenting a clear top-down flow. First, experimental data on temperature reaction rates are systematically acquired and a preliminary model is constructed using the basic Flynn-Wall-Ozhawa equation: da / dt = A(1-α)^n exp(-E / RT). To simplify the calculation, this equation is logarithmically transformed into the linear form: ln(da / dt) = ln A + n ln(1-α)-E / RT. Subsequently, a system of parametric equations is constructed, with ln(da / dt) as the target data and 1 / T and ln(1-α) as coefficients. Based on these data, the least squares method is used to minimize the sum of squared errors to obtain the optimal estimate. Finally, through the kinetic parameter recovery step, the three key kinetic parameters of the reaction system are obtained: the frequency factor A, the activation energy E, and the reaction order n. This systematic parameter solution method not only ensures the accuracy of the calculation results but also provides a clear mathematical approach, providing important theoretical support for the study of materials combustion kinetics.

[0118] In an optional embodiment, deeply fusing the high-precision three-dimensional distribution map with the material combustion kinetics model to construct a multi-scale diffusion prediction model based on a graph neural network includes:

[0119] Dividing the high-precision three-dimensional distribution map into a grid structure, measuring physical parameters of nodes, substituting the physical parameters into the material combustion kinetics model to calculate node reaction rates, constructing node feature vectors, and constructing edge feature vectors based on calculated transmission parameters of adjacent nodes;

[0120] A dual feature fusion network is designed, and the node feature vector and the edge feature vector are mapped respectively using a first weight matrix and a second weight matrix. A micro-macro feature mapping matrix is ​​obtained through an activation function, and an evolution trend of the node state is calculated based on the micro-macro feature mapping matrix.

[0121] The evolution trend is input into a multi-level prediction network to calculate the prediction error of the local reaction rate and the overall transmission field, and when the prediction error is less than a preset error threshold, a material combustion diffusion prediction model is output.

[0122] During the meshing and feature construction phase, the system first divides the high-precision 3D distribution map into a regular grid structure. For example, using the combustion process of a material, a sample measuring 100mm × 50mm × 30mm is divided into 0.5mm × 0.5mm × 0.5mm cubic grid cells, resulting in a total of 2,000,000 grid nodes. For each node, its physical parameters are measured, including key physical quantities such as temperature, density, pressure, and oxygen concentration.

[0123] For example, the node numbered (25, 30, 42) has a measured temperature of 453K, a density of 1.2g / cm³, a pressure of 101.3kPa, and an oxygen concentration of 21%. Substituting these physical parameters into the material combustion kinetics model, the reaction rate of this node is calculated to be 0.015mol / (m³·s). Based on this, a characteristic vector of this node is constructed, which includes the above physical parameters and the reaction rate, forming an characteristic vector of length 12. Based on the material and energy exchange laws between adjacent nodes, transmission parameters are calculated, such as the thermal conductivity coefficient of 0.024W / (m·K) and the mass diffusion coefficient of 2.88×10⁻ 5 m² / s, construct edge feature vectors with a dimension of 8.

[0124] During the dual feature fusion network design phase, the system uses two weight matrices to map node and edge feature vectors. The first weight matrix, sized 12×64, is used to map node feature vectors; the second weight matrix, sized 8×64, is used to map edge feature vectors. Using the ReLU activation function, the mapping results are nonlinearly transformed to produce a micro-macro feature mapping matrix with a size of the number of nodes × 64.

[0125] Taking the node (25, 30, 42) in the above example as an example, after processing the feature fusion network, the top five values ​​of its feature map are [0.82, 0.56, 0.93, 0.47, 0.71]. Based on this mapping matrix, a graph convolution operation is applied to calculate the evolution trend of the node state. Specifically, for each node, the feature information of its neighboring nodes is aggregated to update its own state. After processing with three layers of graph convolution layers, the predicted evolution trend of node (25, 30, 42) is a temperature rise rate of 15 K / s, a reaction rate increase of 0.003 mol / (m³·s), etc.

[0126] During the multi-level prediction network construction phase, the system inputs the evolutionary trends into a prediction network consisting of four fully connected layers, with 128, 256, 128, and 64 neurons in each layer, respectively. Iterative predictions are performed with a time step of 0.1 seconds, calculating the prediction errors for local reaction rates and the overall transmission field. The prediction network outputs the physical parameters and reaction rates of each node at time t+1, which are compared with the actual observed values. In the test case, after 10 iterative training cycles, the average prediction error was reduced to 0.82%, below the preset error threshold of 1%. At this point, the final material combustion and diffusion prediction model is output.

[0127] In a specific application example, the combustion diffusion process of a polyurethane foam material was predicted. The initial ignition position was set at the center of the sample surface at a temperature of 873K. The constructed prediction model was applied to simulate the combustion diffusion process for 100 seconds, recording the state changes every 0.1 seconds. The model successfully predicted the flame's diffusion at an average speed of 0.8 mm / s. The maximum deviation of the temperature distribution from the experimental measurement was 5.4%, accurately capturing the temperature gradient changes within the material and the movement trajectory of the combustion reaction area.

[0128] To further improve model accuracy, the system also incorporates an adaptive mesh refinement mechanism. In the combustion front region, the mesh is further refined to 0.2mm × 0.2mm × 0.2mm to more accurately capture the dramatic changes in the reaction front. This dynamic meshing adjustment improves model prediction accuracy by approximately 15% while maintaining computational efficiency, and the calculation speed is approximately eight times faster than traditional finite element methods.

[0129] Through the above technical solution, the present invention realizes the effective integration of the microscopic material combustion kinetics model and the macroscopic material transfer process, constructs a high-precision multi-scale diffusion prediction model, and can provide technical support for material combustion safety assessment and fire protection design.

[0130] Figure 4 This is a flow chart of material combustion diffusion prediction based on dual feature fusion according to an embodiment of the present invention:

[0131] The figure illustrates a complete technical process for material combustion and diffusion prediction, encompassing three key steps. First, the system divides a high-precision three-dimensional distribution map into a grid structure and measures the physical parameters of the grid nodes. These parameters are then input into the material combustion kinetics model to calculate the node reaction rates. Simultaneously, the system constructs node feature vectors and, based on the relationships between adjacent nodes, edge feature vectors. This step lays the foundation for subsequent feature fusion. Second, in the feature processing phase, the system designs an innovative dual feature fusion network architecture, using two independent weight matrices to map and transform node feature vectors and edge feature vectors. Through activation function processing, the system generates a micro-macro feature mapping matrix, which is used to calculate the dynamic evolution trend of the node state. Finally, the system inputs the resulting evolution trend information into a multi-level prediction network to calculate the prediction errors for local reaction rates and the overall transmission field. When the prediction error falls below a pre-set threshold, the system outputs the final material combustion and diffusion prediction model. This layered processing ensures the accuracy and reliability of the prediction results.

[0132] In an optional embodiment, generating smoke density distribution prediction data and diffusion path prediction data at future moments in a vehicle compartment based on the output results of the multi-scale diffusion prediction model for guiding fire evacuation and ventilation control in rail transit vehicles includes:

[0133] Determining the temporal and spatial features of an output result of the multi-scale diffusion prediction model, and fusing the temporal and spatial features to obtain spatiotemporal enhancement features;

[0134] Inputting the spatiotemporal enhancement features into a causal convolutional network, extracting temporal features using a dilated convolution operation, transferring the temporal features between layers using a residual connection structure, calculating the temporal correlation coefficient of the transferred features, determining the optimal number of network layers and dilation rate based on the temporal correlation coefficient, and generating a feature sequence with long-range dependencies;

[0135] Constructing a dual-branch prediction network based on the feature sequence, generating smoke density distribution prediction data in the vehicle compartment at a future time using a density prediction branch, generating diffusion path prediction data at a future time using a path prediction branch, calculating a prediction error between the smoke density distribution prediction data and the diffusion path prediction data, and dynamically adjusting a weight coefficient of the prediction branch based on the prediction error;

[0136] Using a probabilistic prediction model to evaluate the reliability of the predicted data after weight adjustment, calculating the confidence interval and calibration error of the reliability of the predicted data, iteratively optimizing the parameters of the probabilistic prediction model based on the confidence interval and the calibration error, and outputting a prediction result with guaranteed reliability;

[0137] The safety level of each area in the carriage is calculated according to the prediction results, the optimal evacuation channel is planned based on the safety level, the smoke concentration deviation of each area is calculated using the smoke density distribution prediction data, and ventilation control instructions are generated according to the smoke concentration deviation.

[0138] The output results of the multi-scale diffusion prediction model are subjected to spatiotemporal feature fusion processing. Specifically, the three-dimensional data cube output by the model is extracted. The cube contains the time dimension T (such as 120 time steps), the spatial position dimension X×Y (such as 64×64 grid points), and the feature dimension C (such as 32 feature channels). When extracting the temporal features, a one-dimensional convolution sliding window (window size is 7) is used to extract the historical evolution pattern along the time axis to obtain the time embedding vector T_embed (dimension is 64). When processing the spatial features, a grid encoder is used to encode the position of each point on the XY plane to generate a spatial position embedding vector S_embed (dimension is 128). T_embed and S_embed are fused through the attention mechanism, and the correlation matrix of the temporal and spatial features is calculated to obtain the fused spatiotemporal enhancement feature TS_fusion (dimension is 192).

[0139] The spatiotemporal enhancement features are input into a causal convolutional network for temporal feature extraction. This network employs a multi-layer dilated convolutional structure, with dilation rates of 1, 2, 4, 8, and 16, starting from the first layer. The convolution kernel size of each layer is 3, the number of input channels is 192, and the number of output channels is 128. Through dilated convolution, the network's effective receptive field reaches 95 time steps, covering 79.2% of the historical data. The network internally employs a residual connection structure, with skip connections between every two layers to ensure that low-level features can be directly transmitted to higher layers. By calculating the temporal autocorrelation coefficients of the output features of different network layers (average value of 0.783), the optimal network depth is determined to be 5 layers, with a maximum dilation rate of 16. This structure is capable of capturing long-range dependencies in the smoke diffusion process and generating a feature sequence Feat_seq with a sequence length of 32.

[0140] Based on the feature sequence, a two-branch prediction network was constructed, one for density prediction and one for path prediction. The density prediction branch consists of three layers of transposed convolutions with a kernel size of 4×4, a stride of 2, and the number of channels, 128, 64, and 1, respectively. The output dimension is 64×64×32 (representing the density prediction for the next 32 time steps). The path prediction branch uses a graph convolutional network structure, dividing the compartment space into 48 key nodes. The node connections are defined using an adjacency matrix, and the output dimension is 48×32×2 (number of nodes × time steps × direction vector). Both branches initially have a weight of 0.5 and are dynamically adjusted during training based on their respective prediction errors. When the root mean square error of the density prediction branch reached 0.067 on the validation set and the mean deviation of the path prediction branch reached 0.083 meters, the weights were adjusted to 0.58 and 0.42 respectively.

[0141] To assess the reliability of the prediction results, a probabilistic prediction model was implemented. This model, based on an ensemble learning approach, trained 10 sub-models with identical structures but different initializations. For each prediction point, the mean prediction of the 10 sub-models was calculated as the final prediction value, with the standard deviation used as an estimate of uncertainty. Statistical validation demonstrated a 95% confidence interval coverage of 92.7%, with an average calibration error of 0.031. To account for the calibration error, a temperature scaling method was used to adjust the prediction distribution. The temperature parameter was iteratively optimized (final value: 1.28), reducing the calibration error to 0.013 and improving the reliability of the prediction results.

[0142] Safe evacuation and ventilation control decisions are made based on the prediction results. The interior of the train is divided into 12 zones. The safety level (1-5) is determined by calculating the average smoke density and maximum density change rate for each zone over the next 180 seconds. If the smoke density in a zone is predicted to exceed 0.15 kg / m³ within 60 seconds and the rate of change is greater than 0.003 kg / (m³·s), the zone is designated as Level 4 (High Risk).

[0143] Based on the safety level of each area, an improved Dijkstra algorithm is used to plan the optimal evacuation route, directing personnel from high-risk areas to areas with a safety level of 1. This route avoids nodes covered by the predicted main smoke diffusion path. Simultaneously, the system calculates the deviation of smoke concentration in each area from the target value (0.05kg / m³). If the deviation exceeds 0.1kg / m³ in a particular area, the system generates a control instruction to increase the exhaust volume in that area by 30%. If the deviation is negative and the absolute value is greater than 0.03kg / m³, the exhaust volume is reduced by 15%, achieving precise ventilation control.

[0144] Through this technical solution, the system can predict the smoke diffusion trend in the carriage 3 minutes in advance with a prediction accuracy of 89.5%, providing precise guidance for personnel evacuation and smoke control in the event of a fire in rail transit carriages, effectively improving emergency response capabilities.

[0145] In an optional embodiment, the reliability of the prediction data after weight adjustment is evaluated using a probabilistic prediction model, a confidence interval and a calibration error of the reliability of the prediction data are calculated, and the parameters of the probabilistic prediction model are iteratively optimized based on the confidence interval and the calibration error. Outputting a prediction result with guaranteed reliability includes:

[0146] Constructing a weight matrix of a prediction parameter association network, performing association analysis on the prediction parameters based on the prediction parameter association network and the weight matrix, setting an error threshold between a predicted value and a true value as a reliability constraint, screening an optimal parameter combination based on the weight matrix and the reliability constraint, calculating a probability deviation between the predicted value and the true value based on the optimal parameter combination, and establishing a reliability probability distribution of the prediction result;

[0147] Constructing a dynamic confidence interval based on the reliability probability distribution and the weight matrix, constructing a calibration error calculation model using the dynamic confidence interval and the reliability probability distribution, inputting the uncertainty state of the prediction system into the calibration error calculation model, and obtaining a dynamic calibration error during the prediction process;

[0148] The weight matrix is ​​used to construct a knowledge modulation matrix, the dynamic calibration error is weighted by the knowledge modulation matrix, an adaptive learning rate is calculated based on the weighted dynamic calibration error, the parameters of the prediction model are iteratively optimized to obtain optimal prediction parameters; a prediction calculation is performed based on the optimal prediction parameters and the weight matrix, and a prediction result with reliability guarantee is output.

[0149] A weight matrix is ​​constructed for the prediction parameter association network to represent the degree of mutual influence between different prediction parameters. For example, in weather forecasting, there are complex correlations between parameters such as temperature, humidity, and air pressure. A 10×10 weight matrix can be constructed to quantify the strength of these correlations. The weight values ​​range from [-1, 1], where 0.8 indicates a strong positive correlation and -0.6 indicates a moderate negative correlation. This weight matrix is ​​used to perform correlation analysis on prediction parameters and identify key parameter combinations.

[0150] The error threshold between the predicted value and the true value is set as a reliability constraint. For example, an acceptable range for predicted temperature errors is set at no more than ±1.5°C. The optimal parameter combination is selected based on the weight matrix and the reliability constraint. For example, 15 combinations that meet the error threshold are selected from 100 candidate parameter combinations, and the parameter combination with the highest total weight is selected as the optimal choice. The probability deviation between the predicted value and the true value is calculated based on the optimal parameter combination. By analyzing 1,000 historical prediction results, the distribution characteristics of the prediction deviation are obtained, such as a mean of 0.2 and a standard deviation of 0.8. A reliability probability distribution for the prediction results is established, forming the basis for reliability assessment of the prediction results.

[0151] A dynamic confidence interval is constructed based on the reliability probability distribution and weight matrix. Unlike a fixed 95% confidence interval, a dynamic confidence interval adjusts its range based on the prediction conditions. For example, when the predicted data is stable, the confidence interval might be [predicted value -0.5, predicted value +0.5]; when the data fluctuates significantly, it might expand to [predicted value -1.2, predicted value +1.2].

[0152] A calibration error calculation model is constructed using dynamic confidence intervals and reliability probability distributions. The core of this model is to assess the consistency between predicted probabilities and actual occurrence frequencies. The uncertainty state of the prediction system, including data acquisition errors and parameter fluctuations, is input into the calibration error calculation model. The system automatically quantifies these uncertainties based on historical data. For example, if a 20% drop in sensor accuracy is detected, the model incorporates this uncertainty into the calculation. This approach allows for the dynamic calibration error during the prediction process. For example, a calibration error of 0.05 under standard conditions could rise to 0.15 under extreme conditions.

[0153] A knowledge modulation matrix is ​​constructed using a weight matrix. This matrix integrates domain expert knowledge and historical data analysis results to assign different reliability weights to different parameters. For example, a 5×5 knowledge modulation matrix is ​​constructed, assigning a high weight of 0.9 to temperature prediction and a moderate weight of 0.6 to the more difficult-to-predict wind speed. The knowledge modulation matrix is ​​used to weight the dynamic calibration error, placing greater emphasis on the impact of highly reliable parameters during the calibration process. The adaptive learning rate is calculated based on the weighted dynamic calibration error. When the calibration error is large, the learning rate may be set to 0.01 for large adjustments; when the calibration error is small, the learning rate may be reduced to 0.001 for fine-tuning.

[0154] The prediction model parameters are iteratively optimized. After each prediction, the system updates the model parameters using a calculated adaptive learning rate based on the difference between the actual result and the predicted result. For example, the temperature coefficient in the model might be adjusted from an initial value of 1.2 to 1.15, and the humidity impact factor from 0.8 to 0.85. After 200 iterations, the model parameters stabilize, resulting in the optimal prediction parameters. Prediction calculations are performed based on the optimal prediction parameters and a weight matrix. The system integrates the optimized parameters and weight relationships to generate the final prediction result. The output of the prediction result, which has guaranteed reliability, includes not only the predicted value but also the corresponding reliability assessment, such as "The temperature prediction value is 25.2°C, with a reliability of 92%, and a confidence interval of [24.5°C, 25.9°C]."

[0155] To validate the effectiveness of this method, 500 days of operational data from a production line, encompassing 20 monitoring parameters, was collected, using industrial equipment failure prediction as an example. Using traditional prediction methods, the failure prediction accuracy was 78%, with an average lead time of 6 hours. Using this method, the failure prediction accuracy increased to 92%, the average lead time was extended to 12 hours, and the false alarm rate was reduced from 15% to 4%. This method demonstrated greater adaptability and reliability in the event of sudden changes in the equipment's operating status, successfully predicting three early failure signs that traditional methods had failed to identify.

[0156] Analyzing 1,000 prediction results, the KL divergence between the predicted and actual distributions of our method was 0.08, significantly lower than the 0.25 of traditional methods. This demonstrates that our method provides more reliable uncertainty estimates while maintaining forecast accuracy. Furthermore, our method's dynamic confidence intervals cover 96% of the actual results, surpassing the 89% coverage of traditional fixed confidence intervals, demonstrating a more precise uncertainty quantification capability.

[0157] Figure 5 This is a schematic diagram comparing prediction accuracy under different training sample sizes according to an embodiment of the present invention:

[0158] This figure compares the prediction accuracy of three different prediction models (this technical solution, the LSTM model, and the GRU model) with varying numbers of training samples. The horizontal axis represents the number of training samples, increasing gradually from 1,000 to 10,000; the vertical axis represents the prediction accuracy, ranging from 0.75 to 0.95. The figure clearly shows that this technical solution (triangles) consistently outperforms the other two models, achieving an accuracy of approximately 0.83 with a sample size of 1,000. As the sample size increases to 10,000, the accuracy steadily improves to approximately 0.95. In contrast, the LSTM model (circles) and the GRU model (squares) start with accuracies of approximately 0.78 and 0.79, respectively, and ultimately reach accuracies of approximately 0.87 and 0.89, respectively. All three curves show an upward trend with increasing sample size, but the rate of increase gradually slows, especially after the sample size exceeds 6,000, where the accuracy improvement levels off. This technical solution maintains a clear performance advantage at all sample sizes and has better learning efficiency, indicating that it has stronger modeling capabilities and generalization performance in prediction tasks.

[0159] A second aspect of an embodiment of the present invention provides a dynamic monitoring system for smoke density of rail transit material combustion, comprising:

[0160] The first unit is used to obtain smoke concentration data collected by multiple smoke concentration sensors in a rail transit material combustion test box, wherein the multiple smoke concentration sensors are distributed in a three-dimensional grid in the rail transit material combustion test box;

[0161] The second unit is configured to perform three-dimensional reconstruction of the smoke density distribution in the rail transit material combustion test chamber using a three-dimensional spatial interpolation algorithm with adaptive grid refinement based on the position coordinate information of the multiple smoke density sensors and the corresponding smoke concentration data, and automatically increase the grid density when the smoke concentration data is higher than a preset concentration threshold, and automatically reduce the grid density when the smoke concentration data is lower than the preset concentration threshold, so as to obtain a high-precision three-dimensional distribution map;

[0162] The third unit is used to collect material combustion characteristic parameters in the vehicle compartment, including the material's heat release rate, mass loss rate, and combustion product composition data; and establish a material combustion kinetics model based on the material combustion characteristic parameters;

[0163] A fourth unit is configured to deeply integrate the high-precision three-dimensional distribution map with the material combustion kinetics model to construct a multi-scale diffusion prediction model based on a graph neural network, wherein the multi-scale diffusion prediction model includes a macroscopic diffusion prediction layer and a microscopic diffusion prediction layer;

[0164] The fifth unit is used to generate smoke density distribution prediction data and diffusion path prediction data in the car at future moments based on the output results of the multi-scale diffusion prediction model, which are used to guide fire evacuation and ventilation control in rail transit cars.

[0165] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0166] processor;

[0167] a memory for storing processor-executable instructions;

[0168] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0169] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0170] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dynamic monitoring method for the density of combustion smoke of rail transit materials, characterized in that: include: Acquiring smoke concentration data collected by a plurality of smoke concentration sensors in a rail transit material combustion test chamber, wherein the plurality of smoke concentration sensors are distributed in a three-dimensional grid in the rail transit material combustion test chamber; Based on the position coordinate information of the multiple smoke density sensors and the corresponding smoke concentration data, a three-dimensional spatial interpolation algorithm with adaptive grid refinement is used to perform three-dimensional reconstruction of the smoke density distribution in the rail transit material combustion test box. When the smoke concentration data is higher than a preset concentration threshold, the grid density is automatically increased, and when the smoke concentration data is lower than the preset concentration threshold, the grid density is automatically reduced to obtain a high-precision three-dimensional distribution map; Collecting combustion characteristic parameters of materials in the vehicle compartment, including heat release rate, mass loss rate, and combustion product composition data of the materials; and establishing a material combustion kinetics model based on the material combustion characteristic parameters; Deeply integrating the high-precision three-dimensional distribution map with the material combustion kinetics model to construct a multi-scale diffusion prediction model based on a graph neural network, wherein the multi-scale diffusion prediction model includes a macroscopic diffusion prediction layer and a microscopic diffusion prediction layer; According to the output results of the multi-scale diffusion prediction model, smoke density distribution prediction data and diffusion path prediction data at future moments in the carriage are generated to guide fire evacuation and ventilation control in rail transit carriages.

2. The method according to claim 1, characterized in that The three-dimensional reconstruction of the smoke density distribution in the rail transit material combustion test chamber using a three-dimensional spatial interpolation algorithm with adaptive grid refinement includes: Acquire smoke density data collected by multiple smoke concentration sensors in a rail transit material combustion test chamber, wherein the multiple smoke concentration sensors are three-dimensionally distributed in the rail transit material combustion test chamber, and map position information of the multiple smoke concentration sensors into a three-dimensional coordinate system; Calculating a spatial gradient value of smoke density in the rail transit material combustion test box, multiplying a ratio of the spatial gradient value to the maximum smoke density at a current moment by an initial grid size to obtain a grid refinement criterion; and adaptively refining the spatial grid of the rail transit material combustion test box according to the grid refinement criterion; Establishing a three-dimensional spatial interpolation model based on the adaptively refined grid structure, and determining an interpolation weight coefficient according to local grid characteristics within the rail transit material combustion test chamber, wherein the interpolation weight coefficient is composed of a credibility coefficient of a smoke concentration sensor and a distance attenuation function based on a local grid size; The interpolation weight coefficient is input into the three-dimensional space interpolation model, the smoke density value of the grid node in the rail transit material combustion test box is calculated, and the three-dimensional reconstruction result of the smoke density distribution in the rail transit material combustion test box is obtained.

3. The method according to claim 1, characterized in that Establishing a material combustion kinetics model according to the material combustion characteristic parameters includes: collecting thermogravimetric data of the material at different heating rates, calculating the mass loss rate at each temperature point based on the thermogravimetric data, determining a peak temperature point based on the mass loss rate, using the peak temperature point as a characteristic temperature for material decomposition, and dividing the combustion process of the material into multiple sub-reaction stages based on the characteristic temperature; For each of the sub-reaction stages, selecting multiple characteristic conversion rate points and their corresponding temperature data, calculating the mass loss of each stage based on the characteristic conversion rate points, and substituting the mass loss, the temperature data, and the heating rate into the Flynn-Wall-Ozhawa equation; Performing a logarithmic transformation on the Flynn-Wall-Ozawa equation to convert the nonlinear equation into a linear relationship between activation energy and frequency factor, establishing a parametric equation system, and solving the parametric equation system using the least squares method to obtain the initial kinetic parameters of each sub-reaction; A parameter optimization space is constructed based on the initial kinetic parameters, and the parameter optimization space is input into a genetic algorithm optimization model, with the root mean square error between the experimental curve and the calculated curve as the optimization target; based on the optimization target, parameter optimization calculations are performed through selection, crossover and mutation operations of the genetic algorithm optimization model to obtain the optimal activation energy, frequency factor and reaction order, and a material combustion kinetics model is established based on the optimal activation energy, the frequency factor and the reaction order.

4. The method according to claim 3, characterized in that The Flynn-Wall-Ozawa equation is logarithmically transformed to convert the nonlinear equation into a linear relationship between activation energy and frequency factor, and a parametric equation system is established. The least squares method is used to solve the parametric equation system to obtain the initial kinetic parameters of each sub-reaction, including: Obtaining a temperature reaction rate in a Flynn-Wall-Ozawa equation, wherein the Flynn-Wall-Ozawa equation includes a reaction conversion rate, a frequency factor, an activation energy, and a reaction order, wherein the Flynn-Wall-Ozawa equation represents an exponential functional relationship between the temperature reaction rate and the reaction conversion rate, the frequency factor, the activation energy, and the reaction order; performing a logarithmic transformation on the exponential functional relationship to obtain a linearized equation; Establishing a set of parameter equations based on the linearized equations, constructing a target data sequence using multiple sets of reaction rate logarithmic values, and constructing a coefficient data sequence using corresponding temperature inverse values ​​and conversion rate logarithmic values, wherein the set of parameter equations characterizes a linear relationship between the target data sequence and the coefficient data sequence; The least squares method is used to solve the parameter equation group, and the sum of the squares of the experimental values ​​and the fitted values ​​of the target data sequence is used as the optimization objective function. The parameter sequence that minimizes the optimization objective function is solved to obtain the optimal estimated value; the kinetic parameter recovery is performed on the optimal estimated value to obtain the initial kinetic parameters of each sub-reaction.

5. The method according to claim 1, wherein Deeply integrating the high-precision three-dimensional distribution map with the material combustion kinetics model to construct a multi-scale diffusion prediction model based on a graph neural network includes: Dividing the high-precision three-dimensional distribution map into a grid structure, measuring physical parameters of nodes, substituting the physical parameters into the material combustion kinetics model to calculate node reaction rates, constructing node feature vectors, and constructing edge feature vectors based on calculated transmission parameters of adjacent nodes; A dual feature fusion network is designed, and the node feature vector and the edge feature vector are mapped respectively using a first weight matrix and a second weight matrix. A micro-macro feature mapping matrix is ​​obtained through an activation function, and an evolution trend of the node state is calculated based on the micro-macro feature mapping matrix. The evolution trend is input into a multi-level prediction network to calculate the prediction error of the local reaction rate and the overall transmission field, and when the prediction error is less than a preset error threshold, a material combustion diffusion prediction model is output.

6. The method according to claim 1, wherein Based on the output results of the multi-scale diffusion prediction model, the smoke density distribution prediction data and diffusion path prediction data at future moments in the carriage are generated to guide the fire evacuation and ventilation control of rail transit carriages, including: Determining the temporal and spatial features of an output result of the multi-scale diffusion prediction model, and fusing the temporal and spatial features to obtain spatiotemporal enhancement features; Inputting the spatiotemporal enhancement features into a causal convolutional network, extracting temporal features using a dilated convolution operation, transferring the temporal features between layers using a residual connection structure, calculating the temporal correlation coefficient of the transferred features, determining the optimal number of network layers and dilation rate based on the temporal correlation coefficient, and generating a feature sequence with long-range dependencies; Constructing a dual-branch prediction network based on the feature sequence, generating smoke density distribution prediction data in the vehicle compartment at a future time using a density prediction branch, generating diffusion path prediction data at a future time using a path prediction branch, calculating a prediction error between the smoke density distribution prediction data and the diffusion path prediction data, and dynamically adjusting a weight coefficient of the prediction branch based on the prediction error; Using a probabilistic prediction model to evaluate the reliability of the predicted data after weight adjustment, calculating the confidence interval and calibration error of the reliability of the predicted data, iteratively optimizing the parameters of the probabilistic prediction model based on the confidence interval and the calibration error, and outputting a prediction result with guaranteed reliability; The safety level of each area in the carriage is calculated according to the prediction results, the optimal evacuation channel is planned based on the safety level, the smoke concentration deviation of each area is calculated using the smoke density distribution prediction data, and ventilation control instructions are generated according to the smoke concentration deviation.

7. The method according to claim 6, characterized in that Using the probabilistic prediction model to evaluate the reliability of the predicted data after weight adjustment, calculating the confidence interval and calibration error of the reliability of the predicted data, iteratively optimizing the parameters of the probabilistic prediction model based on the confidence interval and the calibration error, and outputting a prediction result with reliability assurance, including: Constructing a weight matrix of a prediction parameter association network, performing association analysis on the prediction parameters based on the prediction parameter association network and the weight matrix, setting an error threshold between a predicted value and a true value as a reliability constraint, screening an optimal parameter combination based on the weight matrix and the reliability constraint, calculating a probability deviation between the predicted value and the true value based on the optimal parameter combination, and establishing a reliability probability distribution of the prediction result; Constructing a dynamic confidence interval based on the reliability probability distribution and the weight matrix, constructing a calibration error calculation model using the dynamic confidence interval and the reliability probability distribution, inputting the uncertainty state of the prediction system into the calibration error calculation model, and obtaining a dynamic calibration error during the prediction process; The weight matrix is ​​used to construct a knowledge modulation matrix, the dynamic calibration error is weighted by the knowledge modulation matrix, an adaptive learning rate is calculated based on the weighted dynamic calibration error, the parameters of the prediction model are iteratively optimized to obtain optimal prediction parameters; a prediction calculation is performed based on the optimal prediction parameters and the weight matrix, and a prediction result with reliability guarantee is output.

8. A dynamic monitoring system for smoke density of rail transit material combustion, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain smoke concentration data collected by multiple smoke concentration sensors in a rail transit material combustion test box, wherein the multiple smoke concentration sensors are distributed in a three-dimensional grid in the rail transit material combustion test box; The second unit is configured to perform three-dimensional reconstruction of the smoke density distribution in the rail transit material combustion test chamber using a three-dimensional spatial interpolation algorithm with adaptive grid refinement based on the position coordinate information of the multiple smoke density sensors and the corresponding smoke concentration data, and automatically increase the grid density when the smoke concentration data is higher than a preset concentration threshold, and automatically reduce the grid density when the smoke concentration data is lower than the preset concentration threshold, so as to obtain a high-precision three-dimensional distribution map; The third unit is used to collect material combustion characteristic parameters in the vehicle compartment, including the material's heat release rate, mass loss rate, and combustion product composition data; and establish a material combustion kinetics model based on the material combustion characteristic parameters; A fourth unit is configured to deeply integrate the high-precision three-dimensional distribution map with the material combustion kinetics model to construct a multi-scale diffusion prediction model based on a graph neural network, wherein the multi-scale diffusion prediction model includes a macroscopic diffusion prediction layer and a microscopic diffusion prediction layer; The fifth unit is used to generate smoke density distribution prediction data and diffusion path prediction data in the car at future moments based on the output results of the multi-scale diffusion prediction model, which are used to guide fire evacuation and ventilation control in rail transit cars.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images