A multi-component fuel self-ignition interval prediction method

Abstract

The application discloses a kind of multi-element fuel self-ignition interval prediction method, belong to fuel safety and intelligent prediction technical field.Acquire fuel component, equivalence ratio, initial temperature, initial pressure and self-ignition explosion state label, group according to same fuel component and equivalence ratio, construct initial temperature-logarithmic pressure two-dimensional target self-ignition explosion probability field;With fuel component and equivalence ratio constitute working condition characteristic vector, by full connection condition coding, two-dimensional feature mapping, bilinear upsampling and convolution decoding generate two-dimensional self-ignition explosion probability field, and adopt the model training of compound loss function based on physical constraint;Model is output complete probability field by inference, again by two-dimensional interpolation, threshold comparison and equiprobable contour line extraction, realize self-ignition explosion state determination, boundary extraction and interval prediction.The application improves the continuity, physical consistency and accuracy of prediction, with good engineering application value, especially suitable for explosion risk prediction and analysis in safety field.

Description

Technical Field

[0001] This invention belongs to the field of fuel combustion and explosion safety technology, specifically relating to a method for predicting the spontaneous combustion range of multi-fuel systems. Background Technology

[0002] Multi-fuel fuels are widely used in chemical production, energy utilization, and marine propulsion. Influenced by factors such as fuel composition, equivalence ratio, initial temperature, and initial pressure, their spontaneous combustion and explosion characteristics exhibit significant nonlinear features. Accurately determining the spontaneous combustion and explosion state and defining the spontaneous combustion and explosion range under different operating conditions is of great significance for fuel storage and transportation safety, combustion device design, and explosion risk prevention.

[0003] Existing spontaneous combustion ranges are mainly obtained through experimental measurements or reaction kinetic simulations. Experimental testing is costly and time-consuming, and it is difficult to comprehensively cover a wide range of operating conditions, including high temperature, high pressure, and mixed fuels. Reaction kinetic simulations, on the other hand, require solving rigid differential equations involving a large number of components and reactions, resulting in a large computational burden for multi-parameter simulations, which is difficult to meet the needs of rapid prediction and analysis.

[0004] In recent years, machine learning methods have been increasingly applied to predict combustion and explosion characteristics. However, existing methods mostly classify or regress at single operating points, making it difficult to directly obtain a continuous temperature-pressure two-dimensional spontaneous combustion and explosion probability field and the spontaneous combustion and explosion range. Furthermore, conventional models primarily rely on fitting sample data, lacking physical constraints on the smoothness of the probability field space and the monotonicity of the explosion boundary normal, leading to problems such as local oscillations, false boundaries, boundary fractures, and insufficient physical consistency in the prediction results. Therefore, it is necessary to propose a prediction method that integrates data features and physical constraints, enabling rapid generation of a two-dimensional spontaneous combustion and explosion probability field and accurate extraction of the explosion range. Summary of the Invention

[0005] The purpose of this invention is to address the above-mentioned problems by proposing a method for predicting the spontaneous combustion range of multi-fuel fuels.

[0006] The technical solution of the present invention: A method for predicting the spontaneous combustion range of multi-fuel fuels includes the following steps: Step (1): Obtain multiple sets of fuel composition, equivalence ratio, initial temperature, initial pressure and spontaneous combustion and explosion state labels. Divide the samples with the same fuel composition and equivalence ratio but different initial temperature and initial pressure into the same working condition sample group. Convert the initial pressure into logarithmic pressure, and construct a two-dimensional target spontaneous combustion and explosion probability field corresponding to each working condition sample group according to the preset initial temperature grid and logarithmic pressure grid and the spontaneous combustion and explosion state labels corresponding to different grid points in each working condition sample group. Step (2): Construct the working condition feature vector with the fuel composition and equivalence ratio corresponding to each working condition sample group, construct the condition probability field reconstruction model, and use the working condition feature vector as input and the corresponding two-dimensional target spontaneous combustion explosion probability field as supervision label. Use a composite loss function including binary cross-entropy data fitting loss, regional overlap loss, field smoothing loss and boundary physical sorting loss to train the condition probability field reconstruction model. Step (3): Obtain the fuel composition and equivalence ratio of the fuel to be predicted, input the trained conditional probability field reconstruction model, and generate a predicted two-dimensional spontaneous combustion explosion probability field through a single model inference; based on the initial temperature of the working condition to be judged and the logarithmic pressure converted from the initial pressure, perform two-dimensional interpolation in the predicted two-dimensional spontaneous combustion explosion probability field to determine the spontaneous combustion explosion state, and extract the equiprobability contour line from the predicted two-dimensional spontaneous combustion explosion probability field with the preset discrimination threshold as the equiprobability value; determine the equiprobability contour line as the spontaneous combustion explosion boundary, wherein the spontaneous combustion explosion boundary is used to characterize the boundary line between the spontaneous combustion explosion range and the non-spontaneous combustion explosion range.

[0007] Furthermore, step (1) specifically includes the following steps: Multiple conditional sampling points are generated within a preset range of fuel mole fraction and equivalence ratio using Sobol low-difference sequences. The mole fraction values ​​at each conditional sampling point are normalized. The sample is uniformly divided within a preset initial temperature and logarithmic pressure range to obtain multiple initial temperature grid points and multiple logarithmic pressure grid points. These initial temperature and logarithmic pressure grid points are paired to form a preset initial temperature-logarithmic pressure two-dimensional grid. For each conditional sampling point, it is combined with each grid point in the preset initial temperature-logarithmic pressure two-dimensional grid to form multiple sets of simulated operating condition samples containing fuel composition, equivalence ratio, initial temperature, and initial pressure. The initial pressure is determined based on the logarithmic pressure grid points. Reaction kinetics simulations are performed on each simulated operating condition sample. The results of the reaction kinetics simulations determine whether spontaneous combustion and explosion occur, obtaining the corresponding spontaneous combustion and explosion state label. The simulated operating condition samples and their corresponding spontaneous combustion and explosion state labels constitute a training sample set. The fuel composition and equivalence ratio are arranged in a preset order to generate the original operating condition feature vector. Where x1, x2, x3, x4, x5, and x6 represent the mole fractions of n-heptane, ammonia, hydrogen, nitrogen, water vapor, and carbon dioxide, respectively, and Φ represents the equivalence ratio; the original operating condition characteristic vector XThe features include the mole fraction and equivalence ratio of each fuel component; samples with the same fuel component and equivalence ratio are divided into the same working condition sample group; for each working condition sample group, the total number of training samples corresponding to each grid point and the number of spontaneous combustion and explosion samples labeled as explosion are counted according to the initial temperature grid and logarithmic pressure grid; the total number of training samples and the number of spontaneous combustion and explosion samples are smoothed; and the ratio of the smoothed number of spontaneous combustion and explosion samples to the smoothed total number of training samples is determined as the target spontaneous combustion and explosion probability of the grid point, thus constructing a two-dimensional target spontaneous combustion and explosion probability field corresponding to each working condition sample group.

[0008] Furthermore, step (2) specifically includes the following steps: Step 1: Arrange the fuel composition and equivalence ratio corresponding to each working condition sample group in a preset order, and standardize them based on the mean and standard deviation of the training samples to obtain the working condition feature vector. Save the mean and standard deviation of each feature for subsequent standardization processing of the working condition feature vector to be predicted. Initial temperature and initial pressure are not included as components of the working condition feature vector. The initial temperature is used to determine the temperature coordinates of the two-dimensional target spontaneous combustion and explosion probability field, and the logarithmic pressure obtained from the initial pressure transformation is used to determine the pressure coordinates of the two-dimensional target spontaneous combustion and explosion probability field. Input the working condition feature vector into a fully connected conditional coding network, and obtain the working condition hidden features through multi-layer linear transformation and ReLU activation processing. Z ; Step 2: Input the implicit feature Z of the operating condition into a fully connected mapping layer to map it into an initial two-dimensional feature map, and then reconstruct the initial two-dimensional feature map into a two-dimensional multi-channel feature map. Input the two-dimensional multi-channel feature map into a convolutional decoding network, and perform bilinear upsampling, convolution processing, and feature refinement in sequence to obtain a single-channel prediction logistic value matrix with the same size as the initial temperature grid points and logarithmic pressure grid points. Use the Sigmoid function to convert the single-channel prediction logistic value matrix into a two-dimensional spontaneous combustion and explosion probability matrix, where the rows and columns of the two-dimensional spontaneous combustion and explosion probability matrix correspond to the initial temperature grid points and logarithmic pressure grid points, respectively. Step 3: Fit the loss using data that includes binary cross-entropy. L BCE Regional overlap loss L Dice Field smoothing loss L TV and boundary physical sorting loss L phy Composite loss function L total The conditional probability field reconstruction model is trained; a sampling triplet for the spontaneous combustion and explosion boundary is selected along the normal direction of the boundary, including sampling points on the non-explosion side, sampling points in the boundary neighborhood, and sampling points on the explosion side; the composite loss function... Ltotal as follows: in, p i,j and y i,j These represent the grid positions ( i , j The predicted spontaneous combustion and explosion probability and the target spontaneous combustion and explosion probability at point () are given. In the region overlap loss, Ω represents the set of grid points labeled with the target spontaneous combustion and explosion probability and participating in the loss calculation. After flattening the initial temperature-log pressure two-dimensional grid, p i and y i They represent the first i The predicted spontaneous combustion explosion probability and the target spontaneous combustion explosion probability at each grid point; ε is a small constant used to avoid numerical singularities; K represents the number of boundary sampling triplets selected along the spontaneous combustion explosion boundary normal; p k non , p k boundary and p k exp They represent the first k The predicted probabilities of non-explosion side sampling points, boundary neighborhood sampling points, and explosion side sampling points in a triplet of spontaneous combustion explosion boundary normal sampling; λ Dice , λ TV and λ phy The weights are assigned to the region overlap loss, field smoothing loss, and boundary physical sorting loss, respectively, to balance data fitting, region overlap, probability field smoothness, and boundary normal monotonicity. The AdamW optimization algorithm is used to update the parameters of the conditional probability field reconstruction model, and the learning rate is dynamically adjusted according to the validation set loss. Training is stopped when the validation set evaluation index no longer improves or reaches the maximum number of training rounds, and the parameters of the conditional probability field reconstruction model with the best validation performance are saved. The trained conditional probability field reconstruction model is used to receive new working condition feature vectors and outputs the two-dimensional spontaneous combustion and explosion probability field under the corresponding initial temperature-log pressure two-dimensional grid.

[0009] Furthermore, the determination of the spontaneous combustion and explosion state in step (3) specifically includes: The fuel composition and equivalence ratio of the fuel to be predicted are obtained. The mean and standard deviation of each feature obtained during the training phase are used for standardization to obtain the feature vector of the operating condition to be predicted. The feature vector of the operating condition to be predicted is input into the conditional probability field reconstruction model after training. The spontaneous combustion and explosion probability of all grid points of the initial temperature-log pressure two-dimensional grid is obtained through one model inference to form the predicted two-dimensional spontaneous combustion and explosion probability field. The initial pressure of the operating condition to be judged is converted into log pressure. Based on the initial temperature and log pressure of the operating condition to be judged, adjacent temperature grid points and log pressure grid points are determined in the predicted two-dimensional spontaneous combustion and explosion probability field. The spontaneous combustion and explosion probability of the operating condition to be judged is calculated by bilinear interpolation. When the spontaneous combustion and explosion probability is greater than or equal to the preset judgment threshold, the operating condition to be judged is determined to be an explosion state; otherwise, it is determined to be a non-explosion state.

[0010] Furthermore, the extraction of the equal probability contour lines in step (3) specifically includes: Using a preset discrimination threshold as the equal probability value, contour lines are extracted from the predicted two-dimensional spontaneous combustion and explosion probability field. The logarithmic pressure coordinates of the extracted equal probability contour lines are converted into initial pressure coordinates to obtain the spontaneous combustion and explosion boundary and spontaneous combustion and explosion range under the initial temperature-initial pressure coordinates.

[0011] Furthermore, the prediction results are correlated, displayed, and interpreted, specifically including: The predicted two-dimensional spontaneous combustion and explosion probability field, spontaneous combustion and explosion boundary, and the location, spontaneous combustion and explosion probability, and explosion state of the working condition to be judged are displayed in association. The two-dimensional probability field generation process of the conditional probability field reconstruction model is combined with the two-dimensional interpolation process based on initial temperature and logarithmic pressure into an end-to-end calculation function. The Shapley additive interpretation analysis is performed on the end-to-end calculation function to determine the influence of fuel composition, equivalence ratio, initial temperature, and initial pressure on the spontaneous combustion and explosion probability of the working condition to be judged.

[0012] The beneficial effects of this invention are as follows: This invention uses fuel composition and equivalence ratio as input conditions to directly generate a two-dimensional spontaneous combustion and explosion probability field through a single model inference; it improves the continuity of the predicted probability field, the monotonicity of the boundary normal, and the physical consistency of the spontaneous combustion and explosion boundary through physical constraints; it realizes the state determination of the working condition to be judged through two-dimensional interpolation, and obtains the spontaneous combustion and explosion boundary and spontaneous combustion and explosion range through the extraction of equal probability contour lines, thereby reducing the computational cost of large-scale temperature and pressure working condition reaction kinetic simulation. Attached Figure Description

[0013] Figure 1 This is a technical roadmap for the multi-fuel auto-ignition range prediction and analysis method proposed in this invention.

[0014] Figure 2The above are spatial sample distribution maps of the multi-component fuels constructed in this invention, wherein (a) is a sample distribution map of n-heptane / ammonia mixed fuel, and (b) is a sample distribution map of n-heptane / hydrogen mixed fuel.

[0015] Figure 3 The following are density distribution diagrams of the explosive and non-explosive samples used in this invention, wherein (a) is the density distribution diagram of the sample in the initial temperature dimension, and (b) is the density distribution diagram of the sample in the logarithmic dimension of the initial pressure.

[0016] Figure 4 This is a comparison and verification diagram of the simulation results and experimental data of the spontaneous combustion explosion limits of different fuels using the combustion mechanism adopted in this invention.

[0017] Figure 5 This is a structural diagram of the conditional probability field reconstruction model based on physical constraints constructed in this invention.

[0018] Figure 6 The following is a performance evaluation diagram of the conditional probability field reconstruction model constructed in this invention, wherein (a) is a distribution diagram of F1 value, crossover ratio and balance accuracy, and (b) is a cumulative distribution diagram of relative error of explosion probability.

[0019] Figure 7 The diagram shows a comparison between the spontaneous combustion explosion boundary predicted by the conditional probability field reconstruction model constructed in this invention and the spontaneous combustion explosion boundary simulated by reaction kinetics under different representative operating conditions. In this diagram, (a) represents the boundary comparison results under different fuel components and equivalence ratios of n-heptane / ammonia mixed fuel, and (b) represents the boundary comparison results under different fuel components and equivalence ratios of n-heptane / hydrogen mixed fuel.

[0020] Figure 8 This is a characteristic influence analysis diagram of the test set operating conditions of this invention. Detailed Implementation

[0021] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0022] like Figure 1 As shown, this invention provides a method for predicting the spontaneous combustion range of multiple fuels. Using the mole fractions and equivalence ratios of n-heptane, ammonia, hydrogen, nitrogen, water vapor, and carbon dioxide as operating parameters, it constructs a two-dimensional target spontaneous combustion and explosion probability field based on initial temperature-logarithmic pressure, trains a conditional probability field reconstruction model with physical constraints, and achieves the generation of the two-dimensional spontaneous combustion and explosion probability field, determination of the spontaneous combustion and explosion state under specified operating conditions, extraction of spontaneous combustion and explosion boundaries, and prediction of the spontaneous combustion and explosion range. Specifically, it includes the following steps: Step (1): Construct a training sample set. Use Sobol low-difference sequences to generate conditional sampling points for fuel composition and equivalence ratio, and combine them with a uniformly divided initial temperature-log pressure two-dimensional grid to form simulated operating condition samples; perform reaction kinetic simulation calculations on each simulated operating condition sample to obtain spontaneous combustion and explosion state labels, and construct a two-dimensional target spontaneous combustion and explosion probability field and operating condition feature vector based on the spontaneous combustion and explosion state labels.

[0023] Step 1: Determine the operating parameters of the multi-component fuel. The multi-component fuel is a fuel system containing combustible components and diluent components. The combustible components include n-heptane, ammonia, or hydrogen, and the diluent components include nitrogen, water vapor, and carbon dioxide. The operating parameters include the mole fraction and equivalence ratio of each component. The mole fraction of each component ranges from 0 to 1, and the equivalence ratio ranges from 0.5 to 3. Within the preset value range of each parameter, conditional sampling points are generated using a Sobol low-difference sequence, and these conditional sampling points are mapped to a range where the mole fraction of each component is not less than 0 and the sum of the mole fractions is 100%. For example... Figure 2 As shown, (a) is the conditional sampling point distribution of n-heptane / ammonia mixed fuel, and (b) is the conditional sampling point distribution of n-heptane / hydrogen mixed fuel. Step 2: Perform reaction kinetic simulations for different fuel compositions, stoichiometric ratios, initial temperatures, and initial pressures. Based on the simulation results, determine whether spontaneous combustion and explosion occur, and obtain the corresponding spontaneous combustion and explosion status labels. Samples that spontaneously combust and explode are labeled as 1, and samples that do not spontaneously combust and explode are labeled as 0. The initial temperature range is set to 400-1200 K, and the initial pressure range is set to 10⁻¹⁰ K. 7 Pa. For example. Figure 3 As shown, a distribution analysis was performed on spontaneous combustion explosion samples and non-spontaneous combustion explosion samples. (a) shows the sample density distribution along the initial temperature dimension, and (b) shows the sample density distribution along the logarithmic dimension of the initial pressure, used to examine the coverage of the training samples within the temperature and pressure ranges. Figure 4 As shown, the auto-ignition explosion limits of different fuels obtained from reaction kinetics simulations are compared with experimental data to verify the rationality of the sample data and the combustion mechanism adopted. Step 3: Samples with the same fuel composition and equivalence ratio are grouped into the same operating condition sample group. A two-dimensional temperature-logarithmic pressure grid is established by uniformly dividing the grid along the temperature direction (range 400-1200 K) and the logarithmic pressure direction (range 1-7). The number of spontaneous combustion explosion samples and the total number of samples corresponding to each grid point are counted, and Gaussian filtering is used to smooth the data. The ratio of the smoothed explosion sample count to the smoothed total number of samples is used as the target spontaneous combustion explosion probability for the corresponding grid point, thus establishing a two-dimensional target spontaneous combustion explosion probability field.

[0024] Step (2): Development of the Conditional Probability Field Reconstruction Model. A conditional probability field reconstruction model is constructed, comprising a fully connected conditional encoding network, a fully connected mapping layer, and a convolutional decoding network. The model uses a feature vector of fuel composition and stoichiometry as conditional input and outputs a two-dimensional spontaneous combustion and explosion probability matrix on the corresponding initial temperature-logarithmic pressure grid. This two-dimensional spontaneous combustion and explosion probability matrix characterizes the two-dimensional spontaneous combustion and explosion probability field. The initial temperature and initial pressure are used to determine the temperature and pressure coordinates of the two-dimensional spontaneous combustion and explosion probability field, respectively, and the initial pressure is logarithmically transformed to determine the corresponding logarithmic pressure coordinates. Figure 5 As shown, the development of the conditional probability field reconstruction model includes the following steps: Step 1: Arrange the fuel components and equivalence ratios in a preset order to form a working condition feature vector. Then, standardize the working condition feature vector based on the feature mean and standard deviation of the training samples. Step 2: The fully connected conditional coding network consists of multiple fully connected neural networks with 256, 256, and 128 neurons in the hidden layers, respectively. The standardized operating condition feature vectors undergo linear transformation, ReLU activation, and random deactivation to obtain the operating condition latent features. These latent features are mapped and reconstructed into an initial two-dimensional feature map of size 128×16×16 through fully connected layers. The convolutional decoding network sequentially performs bilinear upsampling and convolution processing on the initial two-dimensional feature map, with the number of feature channels being 128, 64, and 32, respectively. Each convolutional module includes two 3×3 convolutional layers, a ReLU activation layer, and a two-dimensional random deactivation layer. After convolutional refinement, a 1×1 convolutional layer converts the multi-channel feature map into a single-channel output matrix, and bilinear interpolation is used to adjust it to the target initial temperature-log pressure grid size. The Sigmoid function is used to convert the model's output logic value into a spontaneous combustion explosion probability within the range of 0-1. Step 3: Fit the loss using data that includes binary cross-entropy. L BCE Regional overlap loss L Dice Field smoothing loss L TV and boundary physical sorting loss L phy Composite loss function L totalA conditional probability field reconstruction model is trained. The boundary physical ranking loss imposes a sequential, non-decreasing order constraint on the predicted probabilities of sampling points on the non-explosion side, boundary neighborhood, and explosion side selected along the normal to the spontaneous combustion explosion boundary. The Optuna method is used to optimize the learning rate, regularization parameters, and weights of each loss term. The data is divided according to the working condition sample groups, with the training set, validation set, and test set accounting for approximately 70%, 15%, and 15%, respectively, to avoid samples with the same fuel composition and equivalence ratio appearing in different datasets simultaneously. The AdamW optimization algorithm is used to update the model parameters, with a batch size of 128 and a maximum training epoch of 120. Training stops when the validation set metrics do not show effective improvement for 25 consecutive epochs, and the model parameters with the best validation performance are saved. Figure 6 As shown, the F1 score, crossover ratio (CRR), and balance accuracy were used to evaluate the model's classification performance, while the relative error of explosion probability was used to evaluate the model's probabilistic prediction performance. (a) shows the distribution of each classification performance index, and (b) shows the cumulative distribution of the relative error of explosion probability. The optimal CRR obtained during hyperparameter optimization was 0.9495, indicating a high degree of overlap between the predicted explosion area and the target explosion area.

[0025] Step (3): Predict and determine the spontaneous combustion explosion range. Obtain the mole fraction and equivalence ratio of the fuel components of the fuel to be predicted, and input them into the trained conditional probability field reconstruction model after standardization. Generate a predicted two-dimensional spontaneous combustion explosion probability field through a single model inference. The predicted two-dimensional spontaneous combustion explosion probability field represents the spontaneous combustion explosion probability distribution of the fuel to be predicted on the initial temperature-logarithmic pressure grid. According to the initial temperature of the working condition to be judged and the logarithmic pressure converted from the initial pressure, bilinear interpolation is performed in the predicted two-dimensional spontaneous combustion explosion probability field to obtain the corresponding spontaneous combustion explosion probability. The spontaneous combustion explosion state is determined according to the preset discrimination threshold. The equiprobability contour line is extracted with the preset discrimination threshold as the equiprobability value and determined as the spontaneous combustion explosion boundary. The spontaneous combustion explosion range and the non-spontaneous combustion explosion range are divided accordingly.

[0026] Step 1: Obtain the initial temperature and initial pressure of the operating condition to be judged, and convert the initial pressure into logarithmic pressure. Based on the positions of the initial temperature and the converted logarithmic pressure in the predicted two-dimensional spontaneous combustion and explosion probability field, determine four grid points jointly determined by two adjacent initial temperature grid points and two logarithmic pressure grid points, and use bilinear interpolation to calculate the spontaneous combustion and explosion probability of the operating condition to be judged; Step 2: Compare the interpolated spontaneous combustion and explosion probability with a preset discrimination threshold. In this embodiment, the preset discrimination threshold is 0.50; when the spontaneous combustion and explosion probability is greater than or equal to 0.50, the condition to be judged is determined to be an explosion state; when the spontaneous combustion and explosion probability is less than 0.50, the condition to be judged is determined to be a non-explosion state. Step 3: Extract equal probability contour lines with a probability value of 0.50. Smooth the extracted equal probability contour lines and convert the logarithmic pressure coordinates to initial pressure coordinates to obtain the spontaneous combustion explosion boundary and spontaneous combustion explosion range under the initial temperature-initial pressure coordinates. Figure 7 As shown, the model-predicted spontaneous combustion explosion boundary is compared with the spontaneous combustion explosion boundary simulated by reaction kinetics. (a) and (b) represent the boundary comparison results under two different fuel composition and equivalence ratio conditions, respectively. The results show that the predicted boundary can reflect the spontaneous combustion explosion boundary migration caused by changes in fuel composition and equivalence ratio.

[0027] Optionally, the output includes fuel composition, equivalence ratio, initial temperature, initial pressure, discrimination threshold, spontaneous combustion / explosion probability, explosion state, and spontaneous combustion / explosion boundary. The two-dimensional spontaneous combustion / explosion probability field generation process is combined with a bilinear interpolation process based on initial temperature and logarithmic pressure into an end-to-end calculation function. A Shapley additive interpretation analysis is then performed on this end-to-end calculation function to determine the influence of fuel composition, equivalence ratio, initial temperature, and initial pressure on the spontaneous combustion / explosion probability of the condition to be discriminated. Figure 8 As shown, Figure 8 The left figure shows the ranking of the importance of the mean absolute SHAP values ​​of each input feature. Figure 8 The right figure shows the distribution of SHAP values ​​for each input feature.

[0028] This invention realizes a complete process from sample data construction, physical constraint-based conditional probability field reconstruction model training, two-dimensional spontaneous combustion and explosion probability field generation, specified working condition determination, spontaneous combustion and explosion boundary extraction, and spontaneous combustion and explosion interval prediction, which can improve the physical consistency, boundary continuity and prediction accuracy of the spontaneous combustion and explosion interval prediction results of multi-fuels.

[0029] The above-described embodiments are merely illustrative of the implementation methods of the present invention, but should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A method for predicting the spontaneous combustion range of multi-fuel fuels, characterized in that, Includes the following steps: Step (1): Obtain multiple sets of fuel composition, equivalence ratio, initial temperature, initial pressure and spontaneous combustion and explosion state labels. Divide the samples with the same fuel composition and equivalence ratio but different initial temperature and initial pressure into the same working condition sample group. Convert the initial pressure into logarithmic pressure, and construct a two-dimensional target spontaneous combustion and explosion probability field corresponding to each working condition sample group according to the preset initial temperature grid and logarithmic pressure grid and the spontaneous combustion and explosion state labels corresponding to different grid points in each working condition sample group. Step (2): Construct the working condition feature vector with the fuel composition and equivalence ratio corresponding to each working condition sample group, construct the condition probability field reconstruction model, and use the working condition feature vector as input and the corresponding two-dimensional target spontaneous combustion explosion probability field as supervision label. Use a composite loss function including binary cross-entropy data fitting loss, regional overlap loss, field smoothing loss and boundary physical sorting loss to train the condition probability field reconstruction model. Step (3): Obtain the fuel composition and equivalence ratio of the fuel to be predicted, input the trained conditional probability field reconstruction model, and generate a predicted two-dimensional spontaneous combustion explosion probability field through a single model inference; based on the initial temperature of the working condition to be judged and the logarithmic pressure converted from the initial pressure, perform two-dimensional interpolation in the predicted two-dimensional spontaneous combustion explosion probability field to determine the spontaneous combustion explosion state, and extract the equiprobability contour line from the predicted two-dimensional spontaneous combustion explosion probability field with the preset discrimination threshold as the equiprobability value; determine the equiprobability contour line as the spontaneous combustion explosion boundary, wherein the spontaneous combustion explosion boundary is used to characterize the boundary line between the spontaneous combustion explosion range and the non-spontaneous combustion explosion range.

2. The method for predicting the spontaneous combustion range of multiple fuels according to claim 1, characterized in that, Step (1) specifically includes the following steps: Multiple conditional sampling points are generated within a preset range of mole fraction and equivalence ratio of fuel using the Sobol low-difference sequence, and the sampled values ​​of mole fraction in each conditional sampling point are normalized. The grid is uniformly divided within a preset initial temperature range and a logarithmic pressure range to obtain multiple initial temperature grid points and multiple logarithmic pressure grid points. Multiple initial temperature grid points and multiple logarithmic pressure grid points are paired to form a preset initial temperature-logarithmic pressure two-dimensional grid. For each conditional sampling point, it is combined with each grid point in the preset initial temperature-logarithmic pressure two-dimensional grid to form multiple sets of simulated operating condition samples containing fuel composition, equivalence ratio, initial temperature, and initial pressure. The corresponding initial pressure is determined based on the logarithmic pressure grid points. Reaction kinetics simulation calculations are performed on each simulated operating condition sample, and the results are used to determine whether spontaneous combustion and explosion occur, obtaining the corresponding spontaneous combustion and explosion state label. The simulated operating condition samples and their corresponding spontaneous combustion and explosion state labels constitute a training sample set. The fuel composition and equivalence ratio are arranged in a preset order to generate the original operating condition feature vector. Where x1, x2, x3, x4, x5, and x6 represent the mole fractions of n-heptane, ammonia, hydrogen, nitrogen, water vapor, and carbon dioxide, respectively, and Φ represents the equivalence ratio; the original operating condition characteristic vector X The features include the mole fraction and equivalence ratio of each fuel component; samples with the same fuel component and equivalence ratio are divided into the same working condition sample group; for each working condition sample group, the total number of training samples corresponding to each grid point and the number of spontaneous combustion and explosion samples labeled as explosion are counted according to the initial temperature grid and logarithmic pressure grid; the total number of training samples and the number of spontaneous combustion and explosion samples are smoothed; and the ratio of the smoothed number of spontaneous combustion and explosion samples to the smoothed total number of training samples is determined as the target spontaneous combustion and explosion probability of the grid point, thus constructing a two-dimensional target spontaneous combustion and explosion probability field corresponding to each working condition sample group.

3. The method for predicting the spontaneous combustion range of multiple fuels according to claim 2, characterized in that, Step (2) specifically includes the following steps: Step 1: Arrange the fuel composition and equivalence ratio corresponding to each working condition sample group in a preset order, and standardize them based on the mean and standard deviation of the training samples to obtain the working condition feature vector. Save the mean and standard deviation of each feature for subsequent standardization processing of the working condition feature vector to be predicted. Initial temperature and initial pressure are not included as components of the working condition feature vector. The initial temperature is used to determine the temperature coordinates of the two-dimensional target spontaneous combustion and explosion probability field, and the logarithmic pressure obtained from the initial pressure transformation is used to determine the pressure coordinates of the two-dimensional target spontaneous combustion and explosion probability field. Input the working condition feature vector into a fully connected conditional coding network, and obtain the working condition hidden features through multi-layer linear transformation and ReLU activation processing. Z ; Step 2: Input the implicit feature Z of the operating condition into a fully connected mapping layer to map it into an initial two-dimensional feature map, and then reconstruct the initial two-dimensional feature map into a two-dimensional multi-channel feature map. Input the two-dimensional multi-channel feature map into a convolutional decoding network, and perform bilinear upsampling, convolution processing, and feature refinement in sequence to obtain a single-channel prediction logistic value matrix with the same size as the initial temperature grid points and logarithmic pressure grid points. Use the Sigmoid function to convert the single-channel prediction logistic value matrix into a two-dimensional spontaneous combustion and explosion probability matrix, where the rows and columns of the two-dimensional spontaneous combustion and explosion probability matrix correspond to the initial temperature grid points and logarithmic pressure grid points, respectively. Step 3: Fit the loss using data that includes binary cross-entropy. L BCE Regional overlap loss L Dice Field smoothing loss L TV and boundary physical sorting loss L phy Composite loss function L total The conditional probability field reconstruction model is trained; a sampling triplet for the spontaneous combustion and explosion boundary is selected along the normal direction of the boundary, including sampling points on the non-explosion side, sampling points in the boundary neighborhood, and sampling points on the explosion side; the composite loss function... L total as follows: in, p i,j and y i,j These represent the grid positions ( i , j The predicted spontaneous combustion and explosion probability and the target spontaneous combustion and explosion probability at point () are given. In the region overlap loss, Ω represents the set of grid points labeled with the target spontaneous combustion and explosion probability and participating in the loss calculation. After flattening the initial temperature-log pressure two-dimensional grid, p i and y i They represent the first i The predicted spontaneous combustion explosion probability and the target spontaneous combustion explosion probability at each grid point; ε is a small constant used to avoid numerical singularities; K represents the number of boundary sampling triplets selected along the spontaneous combustion explosion boundary normal; p k non , p k boundary and p k exp They represent the first k The predicted probabilities of non-explosion side sampling points, boundary neighborhood sampling points, and explosion side sampling points in a triplet of spontaneous combustion explosion boundary normal sampling; λ Dice , λ TV and λ phy The weights are assigned to the region overlap loss, field smoothing loss, and boundary physical sorting loss, respectively, to balance data fitting, region overlap, probability field smoothness, and boundary normal monotonicity. The AdamW optimization algorithm is used to update the parameters of the conditional probability field reconstruction model, and the learning rate is dynamically adjusted according to the validation set loss. Training is stopped when the validation set evaluation index no longer improves or reaches the maximum number of training rounds, and the parameters of the conditional probability field reconstruction model with the best validation performance are saved. The trained conditional probability field reconstruction model is used to receive new working condition feature vectors and outputs the two-dimensional spontaneous combustion and explosion probability field under the corresponding initial temperature-log pressure two-dimensional grid.

4. The method for predicting the spontaneous combustion range of multiple fuels according to claim 3, characterized in that, The determination of spontaneous combustion and explosion status in step (3) specifically includes: The fuel composition and equivalence ratio of the fuel to be predicted are obtained. The mean and standard deviation of each feature obtained during the training phase are used for standardization to obtain the feature vector of the operating condition to be predicted. The feature vector of the operating condition to be predicted is input into the conditional probability field reconstruction model after training. The spontaneous combustion and explosion probability of all grid points of the initial temperature-log pressure two-dimensional grid is obtained through one model inference to form the predicted two-dimensional spontaneous combustion and explosion probability field. The initial pressure of the operating condition to be judged is converted into log pressure. Based on the initial temperature and log pressure of the operating condition to be judged, adjacent temperature grid points and log pressure grid points are determined in the predicted two-dimensional spontaneous combustion and explosion probability field. The spontaneous combustion and explosion probability of the operating condition to be judged is calculated by bilinear interpolation. When the spontaneous combustion and explosion probability is greater than or equal to the preset judgment threshold, the operating condition to be judged is determined to be an explosion state; otherwise, it is determined to be a non-explosion state.

5. The method for predicting the spontaneous combustion range of multiple fuels according to claim 4, characterized in that, Step (3) of extracting the contour lines of equal probability specifically includes: Using a preset discrimination threshold as the equal probability value, contour lines are extracted from the predicted two-dimensional spontaneous combustion and explosion probability field. The logarithmic pressure coordinates of the extracted equal probability contour lines are converted into initial pressure coordinates to obtain the spontaneous combustion and explosion boundary and spontaneous combustion and explosion range under the initial temperature-initial pressure coordinates.

6. The method for predicting the spontaneous combustion range of multiple fuels according to claim 5, characterized in that, The prediction results are presented and interpreted in a correlated manner, including: The predicted two-dimensional spontaneous combustion and explosion probability field, spontaneous combustion and explosion boundary, and the location, spontaneous combustion and explosion probability, and explosion state of the working condition to be judged are displayed in association. The two-dimensional probability field generation process of the conditional probability field reconstruction model is combined with the two-dimensional interpolation process based on initial temperature and logarithmic pressure into an end-to-end calculation function. The Shapley additive interpretation analysis is performed on the end-to-end calculation function to determine the influence of fuel composition, equivalence ratio, initial temperature, and initial pressure on the spontaneous combustion and explosion probability of the working condition to be judged.

Images