Method for predicting fire source power and opening condition of subway carriage fire

Abstract

The invention discloses a subway carriage fire source power and opening condition prediction method. The method comprises the following steps: constructing a scaling-down train carriage experiment system; acquiring time sequence data in real time according to a temperature sensor in the train compartment experiment system; inputting the time sequence data into a CNN (Content Number)-Attention-LSTM (Long Short Term Memory) model, and optimizing hyper-parameters of the CNN-Attention-LSTM model by utilizing a dung beetle optimization algorithm, so as to obtain an optimized CNN-Attention-LSTM model; and obtaining data to be detected, and inputting the data to be detected into the optimized CNN-Attention-LSTM model to obtain a fire source power prediction value and a compartment opening state classification result. According to the method, the CNN-Attention-LSTM hybrid model is combined with the dung beetle optimization algorithm, so that the prediction performance of the fire parameters of the subway carriage is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent fire monitoring and prediction technology, and in particular relates to a method for predicting the fire source power and opening conditions in subway car fires. Background Technology

[0002] Laterally-opening train carriages typically have enclosed, elongated interiors. During a fire, heat, smoke, and flames accumulate rapidly within this confined space, accelerating the spread of the fire and increasing the difficulty of escape, posing a serious threat to human safety. Traditional fire monitoring methods rely primarily on physical models and empirical formulas (such as heat release rate calculations and temperature decay models), which have significant limitations in complex scenarios: physical models need to simplify the dynamic fire process and cannot simultaneously capture the evolution of fire power and changes in carriage opening status; empirical formulas are susceptible to environmental interference, resulting in insufficient prediction accuracy. While computational fluid dynamics (CFD) numerical simulation methods developed in recent years have improved accuracy, their high computational complexity and time consumption make them unsuitable for real-time monitoring.

[0003] With the development of artificial intelligence technology, deep learning models are gradually being applied to fire monitoring. Wu et al. (2021) trained an AI model based on a numerical database to detect fire sources in tunnels, but it relied too heavily on idealized simulation data and had poor adaptability to real-world scenarios. Kim et al. (2022) used image recognition to identify fires, but it was sensitive to camera location and lighting conditions and lacked robustness. Zhang et al. proposed a dual-agent deep learning model to improve real-time prediction, but the model complexity was too high and resource consumption was large. The digital twin system developed by Hong et al. (2024) is prone to failure when interrupted due to strong network dependency. Existing research also has common problems:

[0004] Data dependency limitations: Most models (such as the study by He et al. in 2023) are trained based on FDS simulation data, which makes it difficult to adapt to the dynamic changes of real fires;

[0005] Feature processing separation: Models such as Wu (2022) and Zhang (2024) failed to effectively integrate spatial distribution (temperature field), temporal evolution (fire spread) and key control parameters (carriage opening status);

[0006] Insufficient real-time performance and generalization ability: In complex carriage environments (closed at both ends with multiple openings and sensor failure), the prediction accuracy of existing AI models drops sharply and the computation delay is significant.

[0007] In summary, existing technologies have not yet solved three major shortcomings: poor adaptability to dynamic scenarios, insufficient fusion of multi-dimensional features, and high hardware dependence, which restrict the accurate and real-time monitoring of fires in train carriages. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention proposes a method for predicting the fire source power and opening conditions in subway car fires. By combining a CNN-Attention-LSTM hybrid model with a dung beetle optimization algorithm, the predictive performance of subway car fire parameters is significantly improved.

[0009] To achieve the above objectives, the present invention provides a method for predicting the fire source power and opening conditions in subway car fires, comprising:

[0010] Construct a scaled-down experimental system for train carriages;

[0011] The time-temperature sequence data is acquired in real time by the temperature sensor inside the train carriage experimental system.

[0012] The time-temperature series data is input into the CNN-Attention-LSTM model, and the hyperparameters of the CNN-Attention-LSTM model are optimized using the dung beetle optimization algorithm to obtain the optimized CNN-Attention-LSTM model.

[0013] The test data is obtained and input into the optimized CNN-Attention-LSTM model to obtain the predicted fire source power and the classification results of the carriage opening status.

[0014] Optionally, the train carriage experimental system is a 1 / 3 scaled-down structure, and the power scaling relationship of the fire source is established according to the Froude similarity scaling principle.

[0015] Optionally, the CNN-Attention-LSTM model includes:

[0016] A convolutional neural network unit is used to extract the thermal distribution characteristics of the carriage space based on the time series data;

[0017] Attention mechanism unit, used to adjust feature weights at different time steps;

[0018] Long Short-Term Memory (LSTM) network units are used to capture long-term dependencies in temperature data.

[0019] Optionally, the process of optimizing the CNN-Attention-LSTM model using the dung beetle optimization algorithm includes:

[0020] Rolling dung balls enables global search; egg-rearing behavior performs local optimization; foraging behavior introduces random perturbations; stealing behavior accelerates convergence.

[0021] Optionally, the dung-rolling behavior for global search includes:

[0022]

[0023] in, Let pX be the new position of the i-th dung beetle in the (t+1)-th iteration; i x represents the best historical position of dung beetle i in the t-th iteration; w X represents the position of the worst-performing individual in the current population. i is the historical best position of dung beetle iii; a is the natural coefficient, which takes the value of 1 or -1, simulating the dung beetle moving forward or backward.

[0024] Optionally, local optimization of fertilization behavior includes:

[0025] x new1 =x g ×(1-R)x new2 =x g ×(1+R);

[0026] Where, x g The position is the global optimum in the current population; R is the reduction factor; x new1 The lower boundary of the spawning area; x new2 This marks the upper boundary of the spawning area.

[0027] Optionally, foraging behavior may introduce random perturbations, including:

[0028]

[0029] Where, x g The global optimal position in the current population; randn(1, dim) is the generated normally distributed random number vector; pX i This is the best historical position for dung beetle i; Let be the new position of the i-th dung beetle in the (t+1)-th iteration.

[0030] Optionally, theft behavior can accelerate convergence, including:

[0031]

[0032] in, Let be the new position of the i-th dung beetle in the (t+1)-th iteration; bestX is the optimal position of dung beetle iii in the current iteration; S is the scaling factor; g is a random vector that follows a normal distribution; pX i This is the best historical position for dung beetle i; x g This is the globally optimal position within the current population. A randomly selected reference individual.

[0033] Technical advantages of this invention: This invention discloses a method for predicting the fire source power and opening conditions in subway car fires. The mean absolute error (MAE) of the fire source power prediction test set is reduced to 0.75, the root mean square error (RMSE) is 3.39, and the coefficient of determination (R²) is [missing information]. 2 The F1 score reached 0.97, demonstrating the model's excellent ability to capture heat release rates. The overall accuracy of carriage opening status recognition reached 91.36%, with P2 (opening doors 2 and 3) and P3 (opening doors 2, 3, and 4) achieving accuracies of 96.4% and 91.8% respectively, effectively solving the misjudgment problem of traditional methods in dynamic scenes. In complex scenarios (such as partial sensor failure or multiple opening combinations), the model's F1 score remained stable at 0.91. The precision, recall, F1 score, and accuracy of the DBO-optimized model were approximately 9.04%, 9.01%, 9.75%, and 9.78% higher than the CNN-LSTM-Attention model, respectively. Regarding error control, the MAE and RMSE of the DBO-optimized model were reduced by approximately 69.39% and 14.54%, respectively. The training efficiency was high, requiring only 3 minutes to complete model optimization (based on an I5-12490F processor), meeting the requirements of real-time fire response. The system simultaneously processes spatial distribution (CNN extracts temperature field), temporal evolution (LSTM captures fire dynamics), and key parameters (attention mechanism assigns weights). The error fluctuation is less than 5% under different fire source power (11.8-70.8kW) and ventilation conditions (P1-P4 conditions), avoiding the data dependence limitations of existing technologies. Attached Figure Description

[0034] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0035] Figure 1 This is a flowchart illustrating a method for predicting the fire source power and opening conditions in a subway car fire according to an embodiment of the present invention.

[0036] Figure 2 This is a schematic diagram of the scaled-down train carriage experimental system according to an embodiment of the present invention;

[0037] Figure 3 This is a schematic diagram of a fire scene database according to an embodiment of the present invention;

[0038] Figure 4 This is a schematic diagram of the temperature change curves of four sensors under different fire source powers in an embodiment of the present invention;

[0039] Figure 5 This is a schematic diagram illustrating temperature changes under different operating conditions according to an embodiment of the present invention;

[0040] Figure 6This is a schematic diagram of the network model construction in an embodiment of the present invention;

[0041] Figure 7 This is a schematic diagram of the DBO optimization algorithm according to an embodiment of the present invention;

[0042] Figure 8 This is a schematic diagram of the fitness change curve of the dung beetle algorithm in an embodiment of the present invention. Detailed Implementation

[0043] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0044] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0045] like Figure 1 As shown, this embodiment provides a method for predicting the fire source power and opening conditions in a subway car fire, including:

[0046] Construct a scaled-down experimental system for train carriages;

[0047] Time-series data are acquired in real time from the temperature sensors inside the train carriage experimental system.

[0048] The time series data is input into the CNN-Attention-LSTM model, and the hyperparameters of the CNN-Attention-LSTM model are optimized using the dung beetle optimization algorithm to obtain the optimized CNN-Attention-LSTM model.

[0049] The test data is obtained and input into the optimized CNN-Attention-LSTM model to obtain the predicted fire source power and the classification results of the carriage opening status.

[0050] Specifically, the implementation process of this embodiment includes:

[0051] like Figure 2 As shown, to simulate a fire environment inside a train carriage, the experiment was conducted on a 1 / 3 scale train carriage test bench. Following the Froude similarity scaling principle, the variable relationships at various scales can be simplified as follows: In the formula Let T represent the heat release rate, m represent the temperature, and f represent the model-scale parameter and the full-scale parameter, respectively. The model is 6.6 meters long, 0.9 meters wide, and 0.76 meters high, constructed using 15.0 mm thick fireproof board to withstand the high temperatures during the experiment. Four identical transverse openings, each 0.43 meters wide and 0.63 meters high, are evenly distributed on one side of the model. A porous propane gas burner made of steel plate, 0.25 meters long, 0.25 meters wide, and 0.1 meters high, is used in the experiment to simulate a fire and generate a stable ignition source. The heat release power of the ignition source is calculated as follows: Where χ is the combustion efficiency, and ΔH c Heat of combustion of gas This represents the mass flow rate. Temperature sensors are located 2.0 cm below the top of the compartment, spaced 10 cm apart near the heat source and 30 cm apart at the far end, to measure temperature. To ensure the reliability and repeatability of the results, each experimental scenario is run at least twice.

[0052] like Figure 3 As shown, to verify the accuracy of the model algorithm, this embodiment constructs two types of experimental scenarios. Considering that in real-world scenarios, temperature sensors close to the fire source may be damaged and fail due to excessive fire power, and that sensor disconnection may occur in engineering projects, this experiment selects a temperature sensor located 2.3m directly above the fire source, and then selects one temperature sensor every six sensors, labeled A (230), B (300), C (450), and D (660), for a total of four temperature sensors to construct the database. The experiment records the entire process from ignition to fire stabilization, totaling 540 seconds, with slight variations in different experiments. Through preprocessing of the experimental data, including data cleaning, outlier handling, and dataset partitioning, a final database for training and validation is formed.

[0053] like Figure 4As shown, 11.8, 23.6, 35.4, 47.2, 59, and 70.8 kW were selected as fire sources, and all the carriage doors were opened. When all the carriage doors were open, the temperature change curves of the four sensors over time reflected the heat distribution characteristics under different fire source powers. Sensors A (approximately 2.3 m from the fire source) and B (approximately 3 m from the fire source) experienced rapid temperature increases and reached a high steady-state temperature in a short time. This is mainly due to the strong radiative heat transfer and convective heat transfer effect caused by the proximity of the fire source. According to Fourier's law of heat conduction: Q = -kδ(dT / dx), where Q is the convective heat flux, k is the thermal conductivity, h is the convective heat transfer coefficient, δ is the heat transfer area, and ΔT is the temperature difference. Sensors A and B, due to the high heat flux near the fire source, quickly reached a high-temperature steady state. The relatively farther sensors C (approximately 4.5 m from the fire source) and D (approximately 6 m from the fire source) experienced slower temperature increases and lower steady-state temperatures. This is because opening the carriage door significantly enhances ventilation, causing the heat generated by the fire source to be quickly dissipated to the external environment through convection and radiation, thus reducing the overall temperature gradient inside the carriage.

[0054] The reason why the temperature data from sensors C and D are similar is mainly because their locations are far from the fire source. With all the carriage doors open, a relatively uniform airflow is formed inside the carriage, and the effects of heat conduction and convection cooling tend to be consistent. According to the heat diffusion equation:

[0055]

[0056] in, This represents the rate of change of temperature over time, where α is the thermal diffusivity. It is the Laplace operator for temperature, representing the rate of temperature diffusion in space.

[0057] Heat diffuses rapidly in an open system, causing temperature variations to become more uniform in areas far from the heat source. Furthermore, the radiative heat flux is lower in areas far from the heat source, as shown in the following formula:

[0058]

[0059] Where q is the radiative heat flux density, σ is the Stefan-Boltzmann constant, ∈ is the surface emissivity, T is the surface temperature of the object, and T∞ is the ambient temperature.

[0060] The opening of the carriage door increased the air exchange rate, resulting in more even heat dissipation from areas away from the fire source, thus making the temperature curves of sensors C and D similar. This uniform air convection and heat dissipation mechanism plays a crucial role in fire scenarios, explaining the similarity in temperature data between sensors C and D.

[0061] like Figure 5 As shown, under different fire scales, the higher the heat release rate (HRR), the higher the smoke temperature, and the more significant the temperature difference at locations farther from the fire source. To verify the effectiveness of the proposed algorithm, this embodiment selected a relatively small fire source power, namely 11.8kW, for the experiment. In the experiment, the four openings of the carriage were arranged in different combinations, specifically: doors 1 and 4 open (condition P1); doors 2 and 3 open (condition P2); doors 2, 3, and 4 open (condition P3); and all doors fully open (condition P4). These opening combinations aim to explore the model's predictive ability for the internal temperature distribution of the carriage under different opening configurations and evaluate its applicability and accuracy in complex fire scenarios. Through this experimental setup, this embodiment can analyze in detail the model's performance in processing temperature distribution data under different opening conditions and verify its predictive ability for the carriage door opening status in fire scenarios. According to the temperature change curves, the temperature changes over time of the four sensors (A, B, C, D) under different opening combinations (P1, P2, P3, P4) showed significant similarities. Despite the different opening combinations, the temperature curves of all sensors exhibited similar characteristics in terms of both the upward trend and the steady-state temperature. Specifically, whether it was P1 (doors 1 and 4 open), P2 (doors 2 and 3 open), P3 (doors 2, 3, and 4 open), or P4 (fully open), the temperatures of all four sensors rose rapidly in the initial stage and remained constant after reaching a steady state. This phenomenon may be attributed to the uniformity of heat distribution and the similarity of ventilation conditions inside the carriage, leading to a consistent trend in the temperature data recorded by the sensors under different opening combinations. This data similarity indicates that relying solely on traditional physical models and empirical formulas is insufficient to effectively distinguish between different opening combinations, thus challenging existing methods for predicting fire heat release rates and temperature distribution.

[0062] In this context, the introduction of a deep learning-based CNN-Attention-LSTM model is particularly important. Convolutional Neural Networks (CNNs) can efficiently extract spatial features from sensor temperature data, the attention mechanism can enhance the model's focus on key features, and Long Short-Term Memory (LSTM) networks excel at capturing long-term dependencies in time-series data. Combining the advantages of these three models, the CNN-Attention-LSTM model demonstrates superior performance in processing complex multidimensional data and automatic feature extraction. Through training on a large amount of experimental data, this model can capture subtle differences in sensor temperature variations, thereby accurately identifying different combinations of openings. This ensemble model not only improves the accuracy and robustness of predictions but also efficiently processes and analyzes complex data in fire scenarios without relying on traditional empirical formulas, providing more scientific and reliable decision support for fire prevention and control.

[0063] like Figure 6 As shown, this embodiment designs and constructs a CNN-Attention-LSTM model that combines a convolutional neural network (CNN), an attention mechanism, and a long short-term memory network (LSTM) to process complex multidimensional temperature data obtained from a train carriage fire experiment. This model can handle fire scenarios with different opening combinations, accurately predict the fire source power and opening combinations, thereby providing effective decision support.

[0064] The model takes time-series data from four temperature sensors as input, which is first fed into the model through a sequence input layer. To process the sequence data, this embodiment folds the data into a form suitable for convolutional operations. Then, a convolutional neural network (CNN) extracts local spatial features from the time-series data through multiple convolutional layers.

[0065] Conv1=ReLU(Conv2D(X;W1,b1));

[0066] Here, Conv1 is the output of the convolutional layer, typically a feature map after convolution and activation functions. X represents the input data from the four temperature sensors, W1 and b1 are the convolutional kernel weights and biases, respectively, and ReLU is the activation function. Subsequently, a second convolutional layer (Conv2st) performs deeper feature extraction and reduces data dimensionality through a global average pooling layer. Then, by introducing an attention mechanism, the model can dynamically adjust the feature weights at different time steps. The calculation formula for the attention mechanism is as follows:

[0067] e t =tanh(W e h t +b e );

[0068]

[0069] Among them, e t It is the score at time step t, α t It is the normalized attention weight, W e h is a learnable weight matrix. t b is the hidden state at the current time step. e This is the bias vector. Next, attention weights are applied to the convolutional feature map through a multiplication layer:

[0070] h′ t =α t h t ;

[0071] Where, h′ t Let α be the convolutional feature map vector at the t-th position (or channel, region). t h is the attention weight at position t. t This represents the new feature after weighting.

[0072] Finally, a Long Short-Term Memory (LSTM) network is used to capture long-term dependencies in the temperature data. The state update of the LSTM can be represented as:

[0073]

[0074] Among them, i t f is the input gate vector. t Forget gate vector, o t C is the output gate vector. t For the new cell state, h t-1 For the hidden state (short-term memory), σ is the sigmoid function, and W... i Let b be the weight matrix. i W is the bias vector. fThe weight matrix determines how the "forget gate" views the concatenated vector, x. t Let b be the input vector. f W is the bias vector. o The weight matrix is ​​the output gate parameter, b. o The bias vector is used for the output gate, and f is the default bias vector. t C is the forget gate vector. t-1 Given the old cell state, tanh is a function that compresses any real number to the interval (-1, 1), W C Let b be the weight matrix. C For the bias vector, the candidate memory default bias is tanh(C). t ) represents compressed memory, C t This is a new cell state.

[0075] Using multiple LSTM units and fully connected layers, the model can output predicted results for the combination of fire source power and carriage openings. The output of the LSTM layers is mapped to the predicted results of fire source power and opening combinations through fully connected layers, and finally classified through softmax layers and classification layers.

[0076] Through model training, the model parameters are optimized using the cross-entropy loss function:

[0077]

[0078] Among them, y i For real labels, To predict probabilities for the model, Let N be the loss function, N be the total number of samples, and i be the i-th sample.

[0079] The experiment used the Adam optimizer for training, setting the maximum number of training epochs to 2000 and the initial learning rate to 0.05. A segmented learning rate adjustment strategy was employed, with a learning rate decrease factor of 0.1 and a decrease period of 1000 epochs. The mini-batch size was set to 512, and the data was shuffled at the end of each training epoch. Long training time and segmented learning rate adjustments were used to improve the model's convergence performance and final accuracy. The network was trained using a 6-core I5-12490F processor, taking approximately 3 minutes in total. Random shuffling of the dataset during each training iteration ensured the model's generalization ability. After multiple training iterations, the model's performance showed minimal differences.

[0080] like Figure 7As shown, the DBO optimization algorithm addresses the following: Although the CNN-Attention-LSTM model exhibits strong feature extraction and temporal dependency capture capabilities when processing multidimensional time-series data, its performance still largely depends on the selection of hyperparameters, including kernel size, learning rate, and the number of neurons in the LSTM layer. These hyperparameter settings not only affect the model's convergence speed but also directly relate to its prediction accuracy and generalization ability. Traditional hyperparameter tuning methods, such as grid search or random search, are often inefficient and prone to getting trapped in local optima when dealing with complex deep learning models. Therefore, this embodiment introduces the Dung Beetle Optimization (DBO) algorithm, which simulates the foraging and egg-rearing behaviors of dung beetles in nature to achieve global optimization of the CNN-Attention-LSTM model's hyperparameters. DBO optimization achieves a good balance between global exploration and local exploitation, thereby further improving the model's performance.

[0081] This algorithm simulates four main behaviors of dung beetles: rolling dung balls, laying eggs, foraging, and stealing. Through the organic combination of these behaviors, the DBO algorithm achieves a good balance between global exploration and local exploitation. This embodiment details each step of the algorithm, explains its core principles through mathematical formulas, and applies it to the hyperparameter optimization of the CNN-LSTM-Attention model.

[0082] Initialize the population:

[0083] The algorithm begins by initializing the population, where the position of each dung beetle (i.e., the hyperparameter combination) is randomly generated within a given upper and lower bound. These initial positions provide a starting point for the subsequent optimization process. The parameters are set as follows:

[0084] The kernel size is between 1 and 5.

[0085] The learning rate is between 0.001 and 0.01.

[0086] The number of neurons is between 1,200.

[0087] The location of each dung beetle (hyperparameter combination) x i The formula for generating it is:

[0088] x i =lb+(ub-lb)×rand(1,dim);

[0089] Where: x iLet represent the position vector of the i-th dung beetle, containing three hyperparameters: kernel size, learning rate, and number of neurons. lb and ub are the lower and upper bounds of the hyperparameters, respectively, lb = [1, 0.001, 1] and ub = [5, 0.01, 200]. rand(1, dim) is a vector of random numbers generated between (0, 10) and (10, 1) to ensure that the dung beetles are uniformly distributed in the search space. The dung beetle colony is initially randomly distributed in a vast "foraging" area, and they begin to explore this area, hoping to find fertile land (i.e., the optimal solution).

[0090] Fitness assessment:

[0091] After initialization, the fitness of each dung beetle's position needs to be evaluated. The fitness function (fobj) is typically used to measure the performance of each hyperparameter combination on the target task. For example, in neural network hyperparameter optimization, the fitness function could be the model's classification accuracy on the validation set.

[0092]

[0093] Where, f(x) i ) indicates that dung beetle i is at position x i fitness value on This indicates that the model is at position x i The predicted value y for the j-th sample j Let represent the true value of the j-th sample, and n be the number of samples in the test set.

[0094] Rolling a dung ball (global exploration):

[0095] The dung ball rolling behavior is the first action executed in the DBO algorithm, primarily used for global exploration. In this phase, some dung beetles adjust their positions to explore new regions of the solution space, avoiding getting trapped in local optima.

[0096]

[0097] in: Let pX be the new position of the i-th dung beetle in the (t+1)-th iteration. i Let x be the historical best position of dung beetle i in the t-th iteration. w X represents the position of the worst-performing individual in the current population. i This represents the historical best position of dung beetle iii. 'a' is a random variable with a value of 1 or -1, used to determine the direction of position adjustment. Through this behavior, the dung beetle can explore new search space regions, effectively avoiding early local optima traps.

[0098] Ovulation behavior (partial development):

[0099] After the dung-rolling behavior ends, the algorithm enters the egg-rearing behavior phase, focusing on the vicinity of the current optimal solution. It further optimizes the solution quality by fine-tuning hyperparameters through local search. The update formula for the egg-rearing behavior is as follows:

[0100] x new1 =x g ×(1-R)x new2 =x g ×(1+R);

[0101] Where: x new1 x is the lower boundary of the spawning area. new2 x is the upper boundary of the spawning area. g This is the globally optimal position in the current population. It is a reduction factor that gradually decreases with the number of iterations. Through the breeding behavior, the algorithm can perform a detailed search around the optimal solution, thereby improving the accuracy of the solution.

[0102] Foraging behavior (a supplement to global exploration):

[0103] Foraging behavior complements global exploration, aiming to introduce randomness and continue exploring underdeveloped regions of the solution space. This behavior prevents premature population convergence and helps discover new potential optimal solutions. The update formula for foraging behavior is as follows:

[0104]

[0105] Where, x g The global optimal position in the current population; randn(1,dim) is the generated normally distributed random number vector; pX i This is the best historical position for dung beetle i; Let be the new position of the i-th dung beetle in the (t+1)-th iteration.

[0106] Theft (Competitive Optimization):

[0107] The stealing behavior is the last action executed in the DBO algorithm, simulating competition among dung beetles in nature. By "stealing" high-quality solutions from other dung beetles' positions, dung beetles with suboptimal solutions can quickly improve their fitness. The update formula for the stealing behavior is as follows:

[0108]

[0109] in, Let be the new position of the i-th dung beetle in the (t+1)-th iteration; bestX is the optimal position of dung beetle iii in the current iteration; S is the scaling factor; g is a random vector that follows a normal distribution; pX i This is the best historical position for dung beetle i; x gThis is the globally optimal position within the current population. Let be the new position of the i-th dung beetle in the (t+1)-th iteration.

[0110] Update population position and global optimum:

[0111] After performing the four actions described above, the algorithm needs to update the historical best position of each dung beetle and the global best position of the population. First, for each dung beetle, if the fitness of the new position in the current iteration is better than its historical best position, then its historical best position is updated.

[0112] Termination condition check and next iteration:

[0113] Finally, the algorithm checks if the termination condition is met, such as reaching the maximum number of iterations or finding a satisfactory solution. If the condition is met, the globally optimal solution is output; otherwise, it returns to the next iteration and starts again from the rolling ball behavior.

[0114] like Figure 8 The figure shows the fitness change curve of the dung beetle algorithm, where the distributed data points represent different dung beetle individuals. A total of six iterations were performed, with each iteration containing 25 dung beetle individuals. The fitness change curve of the dung beetle optimization algorithm (DBO) shows the gradual decrease in fitness during the optimization process. A total of seven iterations were performed, with each iteration containing 25 dung beetle individuals. The fitness rapidly decreased from an initial 0.133 to 0.006, indicating that the optimization process successfully found a better solution. In the first two iterations, dung beetle individuals conducted global exploration through the "dung ball rolling" behavior, randomly distributing within the search space and quickly finding solutions with lower fitness, resulting in a significant decrease in fitness values. As iterations progressed, the dung beetle's "egg-rearing" behavior further explored the vicinity of high-quality solutions locally, optimizing the individual's fitness value. Subsequent "foraging" behavior expanded the individual's search range, preventing them from getting trapped in local optima, while continuing to optimize the quality of solutions. Finally, through "stealing" behavior, individuals competed to share high-quality solutions, further improving the overall fitness. After the sixth iteration, the fitness value stabilized, indicating that the algorithm had converged to the optimal solution. Finally, through DBO optimization, the optimal hyperparameters of the network were determined, including an initial learning rate of 0.01, a kernel size of 3, and 167 LSTM neurons. These parameters will be used to build the network model to ensure its optimal performance in fire monitoring tasks.

[0115] This embodiment proposes a deep learning model based on convolutional neural networks (CNN), long short-term memory networks (LSTM), and attention mechanisms. Hyperparameter optimization is performed using the dung beetle optimization algorithm (DBO) for predicting fire source power in train carriages and identifying carriage opening status. Experimental results show that the model has extremely high accuracy in fire source power prediction, with a mean absolute error (MAE) of 0.75, a root mean square error (RMSE) of 3.39, and a coefficient of determination (R²) of [missing value]. 2 The accuracy reached 0.97. In terms of recognizing the opening status of the carriage, the overall accuracy of the model reached 91.36%, with recognition accuracies of 96.4% and 91.8% for condition P2 (opening doors 2 and 3) and condition P3 (opening doors 2, 3, and 4), respectively. These results demonstrate that the DBO-optimized model can effectively extract multi-dimensional spatial and temporal features in complex fire scenarios, exhibiting excellent predictive ability and robustness.

[0116] Furthermore, this embodiment systematically compares the performance of different deep learning models. The results show that the DBO-optimized CNN-LSTM-Attention model proposed in this embodiment significantly outperforms single CNN or LSTM models in terms of accuracy and stability in fire source power prediction and carriage opening status recognition. Traditional CNN models perform poorly when processing time-series data, while LSTM models are insufficient in spatial feature extraction. In contrast, the model in this embodiment, by integrating the spatial feature extraction capability of CNN, the time-series processing capability of LSTM, and the feature weighting capability of the attention mechanism, demonstrates a stronger ability to capture spatiotemporal dynamic changes in complex fire scenarios, especially maintaining a low prediction error under high fire source power (59kW and 70.8kW). In addition, the DBO optimization algorithm further improves the convergence speed and the rationality of hyperparameter selection, enabling the model to achieve F1 scores of 0.95 and 0.91 on the training and test sets, respectively, further verifying its excellent generalization ability and application potential.

[0117] In summary, the DBO-optimized CNN-LSTM-Attention model proposed in this embodiment demonstrates significant advantages in fire source power prediction and carriage opening status recognition. It not only significantly improves prediction accuracy and robustness but also exhibits superior adaptability in complex scenarios. Performance comparisons with other deep learning models further validate the model's effectiveness and practicality. This research provides crucial technical support for real-time monitoring and emergency response to train fires and lays a solid theoretical foundation for the development and optimization of intelligent fire early warning systems, demonstrating broad application prospects.

[0118] This invention discloses a method for predicting the fire source power and opening conditions in subway car fires. The mean absolute error (MAE) of the fire source power prediction test set is reduced to 0.75, the root mean square error (RMSE) is 3.39, and the coefficient of determination (R²) is [missing value]. 2 The F1 score reached 0.97, demonstrating the model's excellent ability to capture heat release rates. The overall accuracy of identifying carriage opening states reached 91.36%, with accuracies of 96.4% and 91.8% for conditions P2 (doors 2 and 3 open) and P3 (doors 2, 3, and 4 open), effectively solving the misjudgment problem of traditional methods in dynamic scenarios. In complex scenarios (such as partial sensor failure or multiple opening combinations), the model's F1 score remained stable at 0.91, with generalization ability improved by more than 23% compared to a single CNN / LSTM model. Training efficiency was high, requiring only 3 minutes to complete model optimization (based on an I5-12490F processor), meeting the requirements of real-time fire response. Simultaneous processing of spatial distribution (CNN extracts the temperature field), temporal evolution (LSTM captures fire dynamics), and key parameters (attention mechanism weighted opening states) resulted in error fluctuations of less than 5% across fire source power (11.8-70.8kW) and ventilation conditions (conditions P1-P4), avoiding the data dependence limitations of existing technologies.

[0119] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method of predicting a fire source power and an opening condition in a subway car, characterized by, The application relates to a method for predicting fire source power and opening state of a train carriage. The method comprises the following steps: constructing a train carriage experimental system with a reduced scale; acquiring time-temperature sequence data in real time according to temperature sensors in the train carriage experimental system; inputting the time-temperature sequence data into a CNN-Attention-LSTM model, optimizing hyperparameters of the CNN-Attention-LSTM model by using a scarab optimization algorithm, and obtaining an optimized CNN-Attention-LSTM model; 2. The method of predicting a fire source power and an opening condition in a subway car according to claim 1, wherein acquiring to-be-measured data, inputting the to-be-measured data into the optimized CNN-Attention-LSTM model, and obtaining a fire source power prediction value and a carriage opening state classification result.

3. The method of predicting a fire source power and an opening condition in a subway car according to claim 1, wherein The train carriage experimental system is a 1 / 3 scale structure, and a fire source power scaling relationship is established according to the Froude similarity scaling principle. The CNN-Attention-LSTM model comprises: a convolutional neural network unit for extracting a carriage space heat distribution feature according to the time sequence data; an attention mechanism unit for adjusting feature weights of different time steps; 4. The method of predicting a fire source power and an opening condition in a subway car according to claim 1, wherein a long short-term memory network unit for capturing long-term dependence in temperature data. The process of optimizing the CNN-Attention-LSTM model by using the scarab optimization algorithm comprises:

5. The method of predicting a fire source power and an opening condition in a subway car according to claim 4, wherein a rolling dung ball behavior for realizing global search, an egg laying behavior for performing local optimization, a foraging behavior for introducing random disturbance, and a stealing behavior for accelerating convergence. wherein, is the new position of the ith dung beetle in the (t+1)th iteration; pX i is the historical best position of the dung beetle i in the tth iteration; x w is the position of the worst individual in the current population; X i is the historical best position of the dung beetle iii; a is a natural coefficient taking the value 1 or -1, simulating the advance or the retreat of the dung beetle.

6. The method of predicting a fire source power and an opening condition in a subway car according to claim 4, wherein The rolling dung ball behavior for realizing global search comprises: x new1 = x g x (1 - R) x new2 = x g x (1 + R); where x g is the global optimum position in the current population; R is the reduction factor; x new1 is the lower bound of the egg-laying region; x new2 is the upper bound of the egg-laying region.

7. The method of predicting a fire source power and an opening condition in a subway car according to claim 4, wherein The egg laying behavior for performing local optimization comprises: where x g is the global best position in the current population; randn(l,dim) is a generated normal distribution random number vector; pX i is the historical best position of the ith dung beetle; is the new position of the ith dung beetle in the (t+1)th iteration.

8. The method of predicting a fire source power and an opening condition in a subway car according to claim 4, wherein The foraging behavior for introducing random disturbance comprises: The stealing behavior for accelerating convergence comprises: wherein, is the new position of the ith dung beetle in the (t+1)th iteration; bestX is the best position of the dung beetle iii in the current iteration; S is a scaling factor; g is a random vector following a normal distribution; pX i is the historical best position of the dung beetle i; x g is the global best position in the current population; is a randomly selected reference individual.

Images