A digital prediction and reverse causal chain diagnosis method for hydrogen refueling station fires

Abstract

This invention discloses a digital prediction and reverse causal chain diagnosis method for hydrogen refueling station fires, belonging to the field of fire prediction and diagnosis technology. Based on the environmental parameters and leakage source parameters of the hydrogen refueling station, a multivariate coupled operating condition is designed to simulate a hydrogen refueling station fire and construct a database. Based on the Transformer model, a learnable location encoding mechanism is designed to dynamically fuse spatiotemporal coordinate features, training a fire development prediction model. By fusing temperature sensor data through a spatiotemporally coupled Transformer model, a diffusion generation architecture with residual constraints is constructed, training a fire temperature visualization prediction model. Based on the temperature sensor data and environmental parameters of the hydrogen refueling station, combined with the fire development prediction model and the fire temperature visualization prediction model, digital prediction of hydrogen refueling station fires is achieved. A reverse causal chain diagnosis method based on fault trees and multi-source data is established to achieve second-level root cause diagnosis. This invention is used for more accurate prediction of hydrogen refueling station fires and reverse causal chain diagnosis.

Description

Technical Field

[0001] This invention relates to the field of fire prediction and diagnosis technology, and in particular to a digital prediction and reverse causal chain diagnosis method for hydrogen refueling station fires. Background Technology

[0002] Hydrogen refueling stations are a crucial part of hydrogen energy development. Their complex internal components make them prone to safety accidents, and they are often located near main roads with high traffic volume from vehicles and tourists. Hydrogen has a wide flammability range, low ignition energy, and rapid flame propagation; leaks can easily ignite jet fires, causing severe casualties and property damage. Therefore, fire monitoring is paramount. Promptly activating sprinklers or ventilation in the early stages of a jet fire can significantly reduce losses.

[0003] In recent years, traditional fire monitoring has relied on computer vision technology, using images and videos captured by surveillance cameras to monitor fires, enabling timely firefighting measures and alarms. However, hydrogen flames appear pale blue under natural light, making them extremely difficult to detect through images and videos. In practice, hydrogen concentration sensors and temperature sensors are commonly used for monitoring, but these are insufficient for pinpointing the location of the fire and predicting its size. Traditional algorithms also have limitations in detection accuracy and effectiveness, and cannot visualize the predicted temperature field data.

[0004] In addition, the current hydrogen fire accident analysis technology at hydrogen refueling stations has systemic defects. The core problems are insufficient fusion of multi-source data and lack of dynamic diagnostic capabilities; the lack of transparent logical connection between alarms and accident evolution paths leads to poor interpretability of emergency decisions; the lack of precision in fire fighting and rescue, and the extensive resource allocation model cannot lock key fire control nodes, which easily leads to low cost-effectiveness of safety investment and secondary disasters. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, this invention provides a digital prediction and reverse causal chain diagnosis method for hydrogen refueling station fires. This method enables spatiotemporal coupling prediction and temperature visualization of hydrogen refueling station fires based on Transformer and diffusion models, allowing for more accurate fire prediction. The reverse causal chain diagnosis method facilitates precise fire suppression and reduces secondary damage.

[0006] To achieve the above objectives, the present invention adopts the following technical solution, including:

[0007] A method for digital prediction and reverse causal chain diagnosis of hydrogen refueling station fires includes the following steps:

[0008] S1. Create a database: Based on the environmental parameters and leakage source parameters of the hydrogen refueling station, design multivariate coupled operating conditions, conduct hydrogen refueling station fire simulation, and construct a hydrogen fire dynamic characteristic parameter database covering multi-dimensional influencing factors, namely the hydrogen fire database.

[0009] S2, Construct a fire development prediction model for hydrogen refueling stations: Based on the Transformer model, a learnable location coding mechanism is designed to dynamically fuse spatiotemporal coordinate features, and the fire development prediction model is trained using a hydrogen fire database; the input of the fire development prediction model is the environmental parameters and temperature sensor data at the current moment, and the output is the flame length and temperature sensor data at each moment in the future period.

[0010] S3, Construct a fire temperature visualization prediction model for hydrogen refueling stations: By fusing temperature sensor data through a spatiotemporally coupled Transformer model, a diffusion generation architecture with residual constraints is constructed and trained to obtain a fire temperature visualization prediction model; the input of the fire temperature visualization prediction model is temperature sensor data, and the output is a fire temperature distribution cloud map.

[0011] S4, based on temperature sensor data and environmental parameters of hydrogen refueling stations, combined with fire development prediction models and fire temperature visualization prediction models, realizes digital fire prediction of hydrogen refueling stations, locates fire locations, and identifies and predicts risk-sensitive areas.

[0012] S5 establishes a reverse causal chain diagnosis method based on fault tree and multi-source data to locate specific causal chain paths and achieve root cause diagnosis.

[0013] Preferably, in step S1, the specific implementation steps for constructing the hydrogen fire database are as follows:

[0014] S11. Establish a three-dimensional model of the hydrogen refueling station and use OpenFOAM software to simulate a fire at the hydrogen refueling station;

[0015] S12, considering the environmental parameters and leakage source parameters of the hydrogen refueling station, the temperature sensor dataset, flame length and fire temperature distribution cloud map under various working conditions are obtained periodically through numerical simulation;

[0016] Among them, environmental parameters include wind direction and wind speed; leakage source parameters include initial mass flow rate and leakage location;

[0017] S13. Perform data preprocessing on the dataset, including standardization, missing value filling, and noise reduction, to obtain the hydrogen fire database.

[0018] Preferably, the specific implementation steps of step S2 are as follows:

[0019] S21. The hydrogen fire database is divided into training set, validation set and test set. The training set is used to drive the training of the fire development prediction model, the validation set is used to support the hyperparameter tuning and architecture selection of the fire development prediction model, and the test set is used to evaluate and verify the reliability of the fire development prediction model.

[0020] S22, based on a learnable location coding mechanism, dynamically fuses spatiotemporal coordinate features to construct a high-dimensional joint representation space, and uses mean square error as the loss function to define the difference measure between the output value and the true value of the fire development prediction model;

[0021] S23, based on the Bayesian optimization algorithm, the feasible region of hyperparameters is sampled and evaluated, and the loss function value of the validation set is used as the performance index to select the model configuration with the best performance on the validation set, thereby achieving a balanced optimization of computational efficiency and generalization ability.

[0022] S24 combines the Adam optimizer to set the initial learning rate and momentum parameters, iteratively updates the hyperparameters through a mini-batch gradient descent strategy, and statistically analyzes the validation set loss function value after each training cycle to monitor overfitting trends and dynamically adjust the learning rate and implement an early stopping strategy.

[0023] Preferably, the specific implementation steps of step S3 are as follows:

[0024] S31, the features of the fire temperature distribution cloud map are extracted by the spatiotemporally coupled Transformer model, and the temperature sensor data is associated with the self-attention mechanism in the Transformer model for spatiotemporal encoding.

[0025] S32 embeds spatiotemporally encoded temperature sensor data as a conditional signal into the backbone network of the diffusion model. By adding physical constraints to the loss function and generation process, the temperature values ​​of the temperature sensor are mapped to a high-dimensional space through a fully connected layer, thereby realizing fire temperature visualization.

[0026] S33 uses the future temperature sensor data output by the fire development prediction model of S2 as a condition to drive the diffusion model of S32 to generate a predicted temperature cloud map, thereby constructing a fire temperature visualization prediction model; a probability decoder is integrated into the output layer of the diffusion model to quantify the uncertainty of spatial predicted temperature, and the prediction confidence is represented synchronously by color depth mapping in the visualization results.

[0027] Preferably, in step S32, the physical constraint includes an energy conservation constraint.

[0028] Preferably, the specific implementation steps of step S4 are as follows:

[0029] S41: Obtain temperature sensor data and real-time environmental parameters from the hydrogen refueling station, and input them into the fire development prediction model and the fire temperature visualization prediction model.

[0030] S42, the fire development prediction model and the fire temperature visualization prediction model complete data spatiotemporal alignment, outlier repair, and feature extraction at the local node to obtain the spatiotemporal evolution characteristics of the fire at the hydrogen refueling station and the fire temperature distribution cloud map; the spatiotemporal evolution characteristics of the fire include flame length and temperature sensor data.

[0031] S43, based on the spatiotemporal evolution characteristics of fires at hydrogen refueling stations and fire temperature distribution cloud maps, locates the fire location, identifies and predicts risk-sensitive areas, and triggers risk warnings; among them, the fire location is located based on the fire temperature distribution cloud map, and risk-sensitive areas are identified through different temperature ranges.

[0032] Preferably, the specific implementation steps of step S5 are as follows:

[0033] S51, analyze the causes and pathways of hydrogen jet fire incidents at hydrogen refueling stations, including hydrogen leakage from equipment and safety system failures, and establish corresponding accident trees;

[0034] S52, based on the location of the fire, extracts multi-source sensor data from a period of time before the fire, in order to activate and quantify the causal path in the fault tree, locate a specific causal chain path, perform reverse causal chain diagnosis, and realize root cause diagnosis.

[0035] S53 generates an analysis report that includes the root cause, a complete causal timeline, and supporting data.

[0036] The present invention also provides a readable storage medium having a computer program stored thereon, which, when executed, implements the aforementioned method for digital prediction and reverse causal chain diagnosis of hydrogen refueling station fires.

[0037] The present invention also provides an electronic device, which includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned method for digital prediction and reverse causal chain diagnosis of hydrogen refueling station fires.

[0038] The present invention also provides a computer program product, which includes a computer program / instruction that, when executed by a processor, implements the aforementioned method for digital prediction and reverse causal chain diagnosis of hydrogen refueling station fires.

[0039] The advantages of this invention are:

[0040] (1) This invention provides a method for spatiotemporal coupling prediction and temperature visualization of hydrogen refueling station fires based on Transformer model and diffusion model, which can be used to predict fires more accurately and facilitate timely alarm and fire rescue measures.

[0041] (2) The present invention combines the Transformer model with the diffusion model to realize fire temperature visualization. Compared with traditional methods, the core advantage is mainly reflected in the capture of the global spatiotemporal dependence of the temperature field by the self-attention mechanism of the Transformer model, combined with the probability generation and physical constraint capabilities of the diffusion model. It can accurately reconstruct the temperature field under sparse measurement points, retain local hot spots and gradient features, and realize real-time dynamic visualization, taking into account accuracy, efficiency and generalization.

[0042] (3) Based on the hydrogen refueling station fire database, Transformer model and diffusion model, this invention trains and constructs a fire development prediction model and a fire temperature visualization prediction model. Combined with the environmental information of the day and real-time sensor data, the fire development prediction model obtains key information such as fire location and leakage source parameters. The fire temperature visualization prediction model quickly predicts the flame temperature distribution cloud map within tens of seconds after the fire occurs, triggering an early warning signal, which facilitates emergency alarm and timely fire-fighting measures.

[0043] (4) The reverse causal chain diagnosis method based on fault tree has strong interpretability. Any alarm can be traced back to the specific path and real-time data in the fault tree. It responds quickly and the calculation speed of rule matching is extremely fast, meeting the real-time requirements. It is simple to deploy and can be directly built based on the existing data acquisition and monitoring control system and fault tree model. It facilitates precise fire extinguishing, optimizes resource allocation, improves the success rate of fire extinguishing, and reduces secondary damage. Attached Figure Description

[0044] Figure 1 This is a flowchart of a digital prediction and reverse causal chain diagnosis method for hydrogen refueling station fires proposed in this invention.

[0045] Figure 2 This is a schematic diagram of a jet fire simulation using OpenFOAM software.

[0046] Figure 3 This is a schematic diagram of the fire development prediction model of the present invention.

[0047] Figure 4 This is a schematic diagram of the fire temperature visualization prediction model of the present invention.

[0048] Figure 5 This is a fault tree diagram of hydrogen leakage in the main equipment of the hydrogen refueling station according to the present invention.

[0049] Figure 6 This is a fault tree diagram of the hydrogen refueling station safety system failure according to the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] A digital prediction and reverse causal chain diagnosis method for hydrogen refueling station fires, see reference. Figure 1 This includes the following steps:

[0052] S1. Create a hydrogen fire database covering multiple influencing factors: collect environmental parameters and leakage source parameters of hydrogen refueling stations, design multivariate coupled operating conditions, integrate a high-precision simulation platform for hydrogen refueling station fires and a standardized data processing flow, and construct a hydrogen fire dynamic characteristic parameter database covering multiple influencing factors, i.e., the hydrogen fire database.

[0053] Among them, multidimensional influencing factors include leakage source parameters (initial mass flow rate, leakage location, etc.), environmental parameters (wind speed, wind direction, etc.), and fire dynamic characteristic parameters including temperature, flame length, etc.

[0054] S2, Construct a fire development prediction model for hydrogen refueling stations: Divide the hydrogen fire database into training, validation, and test sets. Based on the Transformer model, design a learnable location encoding mechanism to dynamically fuse spatiotemporal coordinate features. Combine Bayesian hyperparameter tuning and validation feedback, dynamically adjust the learning rate based on validation loss and implement early shutdown. Use the hydrogen fire database to train the fire development prediction model.

[0055] The fire development prediction model takes as input environmental parameters (wind speed, wind direction, etc.) and temperature sensor data at the current moment, and outputs flame length and temperature sensor data for each moment in the future.

[0056] S3. Construct a fire temperature visualization prediction model for hydrogen refueling stations: By fusing temperature sensor data through a spatiotemporally coupled Transformer model, a diffusion generation architecture with residual constraints is constructed to train the fire temperature visualization prediction model, thereby achieving real-time mapping from low-resolution monitoring data to a high-precision temperature field.

[0057] The fire temperature visualization prediction model takes temperature sensor data as input and outputs a fire temperature distribution cloud map.

[0058] S4 enables digital fire prediction at hydrogen refueling stations: Based on temperature sensor data and environmental parameters of hydrogen refueling stations, combined with fire development prediction models and fire temperature visualization prediction models, it enables real-time identification and risk warning of key risk nodes.

[0059] S5 establishes a reverse causal chain diagnosis method based on fault tree and multi-source data to locate a specific causal chain path, achieve root cause diagnosis in seconds, and facilitate precise fire suppression.

[0060] The specific implementation steps of step S1 are as follows:

[0061] S11, by integrating historical accident cases and experimental data from hydrogen refueling stations, and after outlier removal and standardization preprocessing, optimizes the mesh generation method and simulation model settings by comparing numerical simulation results with experimental data, ensuring the reliability of the numerical simulation model and achieving high-precision reproduction of combustion dynamics. A 40m×27m×10m 3D model of a hydrogen refueling station was established, including a hydrogen tubing vehicle, hydrogen compressor, hydrogen storage tank, and hydrogen dispenser. OpenFOAM (computational fluid dynamics software) was used to simulate jet fire (immediate ignition of hydrogen leakage forming a jet fire). (See reference...) Figure 2 , Figure 2 In the text, T represents temperature.

[0062] S12, taking into account wind direction (east, west, south, north), wind speed (1m / s, 2m / s, 3m / s, 4m / s), leak location (hydrogen tubing vehicle, hydrogen compressor, hydrogen storage tank, hydrogen refueling machine), leak diameter (1mm, 5mm, 10mm, 20mm), and remaining hydrogen content in the equipment (25%, 50%, 75%, 100%), an OpenFOAM dataset with 1024 operating conditions and a simulation time of 60s was obtained. Temperature sensor data was collected at 10 locations every 0.5s. Through numerical simulation, temperature sensor simulation datasets and fire temperature distribution cloud maps under various operating conditions were periodically obtained.

[0063] S13 involves data preprocessing, including standardization, missing value filling, and noise reduction, to ensure data quality and consistency, and to construct a hydrogen fire database.

[0064] In step S2, a fire development prediction model for hydrogen refueling stations is constructed. (See [reference needed]) Figure 3 The specific implementation steps are as follows:

[0065] S21, based on the principle of stratified random sampling, randomly divides the data into 70% training set, 15% validation set and 15% test set. The training set is used to drive the training of the prediction model and the optimization of spatiotemporal features, the validation set is used to support the hyperparameter tuning and architecture selection of the prediction model, and the test set is used to evaluate and verify the reliability of the prediction model.

[0066] S22 dynamically fuses spatiotemporal coordinate features based on a learnable location coding mechanism to construct a high-dimensional joint representation space, and uses mean square error as the loss function to define the difference measure between the output value (fire location, flame length, etc.) of the prediction model and the true value.

[0067] S23 uses Bayesian optimization algorithm to sample and evaluate the feasible region of hyperparameters, and uses the validation set loss function value as the performance index to select the model configuration with the best performance on the validation set, thereby achieving a balance between computational efficiency and generalization ability.

[0068] In step S24, the initial learning rate (determined by optimization in step S23) and momentum parameter (determined by optimization in step S23) are set using the Adam optimizer, and the network weights (hyperparameters) are iteratively updated using a mini-batch gradient descent strategy. After each training epoch, the validation set loss function value is statistically analyzed to monitor overfitting trends, and the learning rate is dynamically adjusted accordingly (e.g., learning rate decay is applied based on the validation loss plateau) and an early stopping strategy is implemented (e.g., when the validation loss continuously decreases). N (Stop training if there is no improvement after one cycle).

[0069] In step S3, a visual prediction model for fire temperature at hydrogen refueling stations is constructed. (See [link / reference]). Figure 4 The specific implementation steps are as follows:

[0070] S31 uses the ViT model (Vision Transformer) to extract features of the fire temperature distribution cloud map, and uses the self-attention mechanism in the Transformer model to associate temperature sensor data for spatiotemporal encoding.

[0071] ;

[0072] ;

[0073] in, t This refers to the time step number; x , y , z The three-dimensional coordinates of the temperature sensor; For encoding dimensions; i , j , k , m These are time-coded vectors ( PE vector), x Axis coordinates y Axis coordinates z Dimension grouping index of axis coordinates (range 0 to ); PE Indicates time encoding. SE Indicates spatial encoding; Indicates the first t The second time-encoded vector of the time step i Even-numbered dimensions; Indicates the first t The second time-encoded vector of the time step i +1 dimension (odd dimension); Indicates sensor x The second spatial encoding vector of coordinates j Even-numbered dimensions; Indicates sensor x The second spatial encoding vector of coordinates j +1 dimension (odd dimension); Indicates sensor y The second spatial encoding vector of coordinates k Even-numbered dimensions; Indicates sensor y The second spatial encoding vector of coordinates k +1 dimension (odd dimension); Indicates sensor z The second spatial encoding vector of coordinates m Even-numbered dimensions; Indicates sensor z The second spatial encoding vector of coordinates m +1 dimension (odd dimension).

[0074] S32 uses spatiotemporally encoded temperature sensor data as a conditional signal and embeds it into the backbone network of the diffusion model through feature splicing and cross-attention. By adding physical constraints to the loss function and generation process, the temperature values ​​of the temperature sensor are mapped to a high-dimensional space through a fully connected layer, thereby realizing fire temperature visualization.

[0075] Using a diffusion model as the core generation tool, the model is provided with a two-dimensional reference based on real-world scenarios by embedding temperature sensor data. At the same time, energy conservation constraints are added to the model training loss function (to correct errors by "increasing errors if physical rules are violated") and the inference generation process (to enforce compliance with fire physics laws such as temperature conduction and fuel ignition point) to avoid generating temperature distributions that do not conform to reality. Then, a fully connected layer is used to map the predicted temperature values ​​of discrete low-dimensional temperature sensors to a high-dimensional feature space that matches the visualization image.

[0076] The energy conservation constraint is shown below:

[0077] ;

[0078] in, Hydrogen concentration, kg / m³ 3 ; The specific heat capacity of hydrogen at constant pressure, J / (kg) K); The flame velocity is in m / s; The thermal conductivity of hydrogen is W / (m). K); For the heat source term of chemical reaction, MJ / m 3 ; Let L2 norm be the square of the residuals of the energy conservation equation; The gradient of temperature T; This refers to the nabla operator (gradient operator).

[0079] Specifically, the feature concatenation involves adjusting the dimensions of the spatiotemporally encoded temperature sensor data (conditional signal) to match the feature maps output by each layer of the diffusion model backbone network in both spatial and channel dimensions. Then, channel-dimensional feature concatenation is performed to achieve preliminary fusion of the conditional signal and backbone features, laying the foundation for subsequent precise feature interaction. The cross-attention mechanism involves constructing a cross-attention interaction module. The temperature features of the temperature sensor data mapped to a high-dimensional space via a fully connected layer are used as the query. The fused features obtained through the above feature concatenation are used as the key and value. Attention weights are obtained by calculating the similarity between the query and the key. Then, the values ​​are weighted and aggregated based on these attention weights to achieve precise interaction between the temperature features and the backbone features of the diffusion model, strengthening the guiding role of the conditional signal in the diffusion model generation process.

[0080] S33 uses the future temperature sensor data output by the fire development prediction model of S2 as a condition to drive the diffusion model of S32 to generate a predicted temperature cloud map, thereby constructing a fire temperature visualization prediction model; a probability decoder is integrated into the output layer of the diffusion model to quantify the uncertainty of spatial predicted temperature, and the prediction confidence is represented synchronously by color depth mapping in the visualization results.

[0081] The specific implementation steps of step S4 are as follows:

[0082] S41 integrates temperature sensor data from hydrogen refueling stations and real-time environmental parameters (wind speed, wind direction, etc.) and inputs them into the fire development prediction model and the fire temperature visualization prediction model.

[0083] S42, the fire development prediction model and the fire temperature visualization prediction model complete data spatiotemporal alignment, outlier repair, and feature extraction in the local node (hardware device in the local environment) to obtain the spatiotemporal evolution characteristics of the fire at the hydrogen refueling station and the fire temperature distribution cloud map. Among them, the spatiotemporal evolution characteristics of the fire include flame length, temperature, etc.

[0084] S43, based on the spatiotemporal evolution characteristics of fires at hydrogen refueling stations and fire temperature distribution cloud maps, can quickly locate fire locations, intelligently identify and predict risk-sensitive areas, trigger risk warnings, and minimize casualties and property losses. Specifically, it locates fire locations based on fire temperature distribution cloud maps and identifies risk-sensitive areas through different temperature ranges.

[0085] The specific implementation steps of step S5 are as follows:

[0086] S51 outlines the possible causes and pathways of hydrogen jet fires at hydrogen refueling stations. Possible causes include malfunctions in key equipment (tube bundle vehicle, hydrogen delivery pipelines, hydrogen storage cylinders, compressors, hydrogen dispensers), hydrogen leaks, and safety systems (leak detection, emergency shut-off, fire suppression systems). A corresponding fault tree is established. See details in the original document. Figure 5 and Figure 6 .

[0087] S52, based on the fire location identified in step S43, the system automatically extracts multi-source sensor data (time-series data, such as pressure, temperature, flow rate, hydrogen concentration, etc.), event log data (valve opening / closing, pump start / stop, maintenance operation records, etc.), and equipment status data (accumulated equipment runtime, last maintenance time, lifecycle stage, etc.) from a period prior to the fire. This data is used to activate and quantify causal paths in the fault tree, locate a specific causal chain path, and achieve second-level reverse causal chain diagnosis. Taking a jet fire caused by external leakage from the main shut-off valve of a hydrogen storage cylinder as an example, see Table 1 for details.

[0088] Table 1. Diagnostic Table of Reverse Causal Chain for External Leakage and Spray Fire from Hydrogen Storage Cylinders at Hydrogen Refueling Stations

[0089]

[0090] S53 automatically generates an analysis report containing the root cause, a complete cause-and-effect timeline, and key data support, facilitating precise fire suppression, improving the success rate of fire extinguishing, and reducing secondary damage. Taking a jet fire caused by external leakage from the main shut-off valve of a hydrogen storage cylinder as an example, see Table 2 for details.

[0091] Table 2 Accident Analysis Table of Hydrogen Storage Cylinder Leakage and Spray Fire at Hydrogen Refueling Stations

[0092]

[0093] In addition, if there is a critical path in the cause-effect graph that cannot be matched with the existing fault tree, it is automatically saved as a "candidate new fault chain" and submitted to professionals for in-depth analysis, which will be used to expand the fault tree in the future.

[0094] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for digital prediction and reverse causal chain diagnosis of hydrogen refueling station fires, characterized in that, Includes the following steps: S1. Create a database: Based on the environmental parameters and leakage source parameters of the hydrogen refueling station, design multivariate coupled operating conditions, conduct hydrogen refueling station fire simulation, and construct a hydrogen fire dynamic characteristic parameter database covering multi-dimensional influencing factors, namely the hydrogen fire database. S2, Construct a fire development prediction model for hydrogen refueling stations: Based on the Transformer model, a learnable location coding mechanism is designed to dynamically fuse spatiotemporal coordinate features, and the fire development prediction model is trained using a hydrogen fire database; the input of the fire development prediction model is the environmental parameters and temperature sensor data at the current moment, and the output is the flame length and temperature sensor data at each moment in the future period. S3, Construct a fire temperature visualization prediction model for hydrogen refueling stations: By fusing temperature sensor data through a spatiotemporally coupled Transformer model, a diffusion generation architecture with residual constraints is constructed and trained to obtain a fire temperature visualization prediction model; the input of the fire temperature visualization prediction model is temperature sensor data, and the output is a fire temperature distribution cloud map. S4, based on temperature sensor data and environmental parameters of hydrogen refueling stations, combined with fire development prediction models and fire temperature visualization prediction models, realizes digital fire prediction of hydrogen refueling stations, locates fire locations, and identifies and predicts risk-sensitive areas. S5. Establish a reverse causal chain diagnosis method based on fault tree and multi-source data to locate specific causal chain paths and achieve root cause diagnosis. The specific implementation steps of step S3 are as follows: S31, the features of the fire temperature distribution cloud map are extracted by the spatiotemporally coupled Transformer model, and the temperature sensor data is associated with the self-attention mechanism in the Transformer model for spatiotemporal encoding. S32 embeds spatiotemporally encoded temperature sensor data as a conditional signal into the backbone network of the diffusion model. By adding physical constraints to the loss function and generation process, the temperature values ​​of the temperature sensor are mapped to a high-dimensional space through a fully connected layer, thereby realizing fire temperature visualization. S33 uses the future temperature sensor data output by the fire development prediction model of S2 as a condition to drive the diffusion model of S32 to generate a predicted temperature cloud map, thereby constructing a fire temperature visualization prediction model; a probability decoder is integrated into the output layer of the diffusion model to quantify the uncertainty of spatial predicted temperature, and the prediction confidence is represented synchronously by color depth mapping in the visualization results.

2. The method for digital prediction and reverse causal chain diagnosis of hydrogen refueling station fires according to claim 1, characterized in that, In step S1, the specific implementation steps for constructing the hydrogen fire database are as follows: S11. Establish a three-dimensional model of the hydrogen refueling station and use OpenFOAM software to simulate a fire at the hydrogen refueling station; S12, considering the environmental parameters and leakage source parameters of the hydrogen refueling station, the temperature sensor dataset, flame length and fire temperature distribution cloud map under various working conditions are obtained periodically through numerical simulation; Among them, environmental parameters include wind direction and wind speed; leakage source parameters include initial mass flow rate and leakage location; S13. Perform data preprocessing on the dataset, including standardization, missing value filling, and noise reduction, to obtain the hydrogen fire database.

3. The method for digital prediction and reverse causal chain diagnosis of hydrogen refueling station fires according to claim 2, characterized in that, The specific implementation steps of step S2 are as follows: S21. The hydrogen fire database is divided into training set, validation set and test set. The training set is used to drive the training of the fire development prediction model, the validation set is used to support the hyperparameter tuning and architecture selection of the fire development prediction model, and the test set is used to evaluate and verify the reliability of the fire development prediction model. S22, based on a learnable location coding mechanism, dynamically fuses spatiotemporal coordinate features to construct a high-dimensional joint representation space, and uses mean square error as the loss function to define the difference measure between the output value and the true value of the fire development prediction model; S23, based on the Bayesian optimization algorithm, the feasible region of hyperparameters is sampled and evaluated, and the loss function value of the validation set is used as the performance index to select the model configuration with the best performance on the validation set, thereby achieving a balanced optimization of computational efficiency and generalization ability. S24 combines the Adam optimizer to set the initial learning rate and momentum parameters, iteratively updates the hyperparameters through a mini-batch gradient descent strategy, and statistically analyzes the validation set loss function value after each training cycle to monitor overfitting trends and dynamically adjust the learning rate and implement an early stopping strategy.

4. The method for digital prediction and reverse causal chain diagnosis of hydrogen refueling station fires according to claim 1, characterized in that, In step S32, the physical constraints include energy conservation constraints.

5. The method for digital prediction and reverse causal chain diagnosis of hydrogen refueling station fires according to claim 2, characterized in that, The specific implementation steps of step S4 are as follows: S41: Acquire temperature sensor data and real-time environmental parameters from the hydrogen refueling station and input them into the fire development prediction model and the fire temperature visualization prediction model. S42, the fire development prediction model and the fire temperature visualization prediction model complete data spatiotemporal alignment, outlier repair, and feature extraction at the local node to obtain the spatiotemporal evolution characteristics of the fire at the hydrogen refueling station and the fire temperature distribution cloud map; the spatiotemporal evolution characteristics of the fire include flame length and temperature sensor data. S43, based on the spatiotemporal evolution characteristics of fires at hydrogen refueling stations and fire temperature distribution cloud maps, locates the fire location, identifies and predicts risk-sensitive areas, and triggers risk warnings; among them, the fire location is located based on the fire temperature distribution cloud map, and risk-sensitive areas are identified through different temperature ranges.

6. The method for digital prediction and reverse causal chain diagnosis of hydrogen refueling station fires according to claim 1, characterized in that, The specific implementation steps of step S5 are as follows: S51, analyze the causes and pathways of hydrogen jet fire incidents at hydrogen refueling stations, including hydrogen leakage from equipment and safety system failures, and establish corresponding accident trees; S52, based on the location of the fire, extracts multi-source sensor data from a period of time before the fire, in order to activate and quantify the causal path in the fault tree, locate a specific causal chain path, perform reverse causal chain diagnosis, and realize root cause diagnosis. S53 generates an analysis report that includes the root cause, a complete causal timeline, and supporting data.

7. A readable storage medium, characterized in that, It stores a computer program, which, when executed, implements the method for digital prediction and reverse causal chain diagnosis of hydrogen refueling station fires as described in any one of claims 1 to 6.

8. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method for digital prediction and reverse causal chain diagnosis of hydrogen refueling station fires as described in any one of claims 1 to 6.

9. A computer program product, characterized in that, It includes a computer program / instruction that, when executed by a processor, implements the method for digital prediction and reverse causal chain diagnosis of hydrogen refueling station fires as described in any one of claims 1 to 6.

Images