Natural hydrogen cause identification and contribution rate quantification method based on multi-source data and machine learning

Abstract

The invention provides a natural hydrogen cause identification and contribution rate quantification method based on multi-source data and machine learning, and belongs to the technical field of natural hydrogen resource exploration and cause identification. A systematic data correction, standardization, air mixing correction and high-dimensional feature extraction method is constructed, and a comprehensive feature system covering isotope features, component ratio features, end member distance features and geological background parameters is formed; a machine learning technology is introduced, a multilayer neural network model oriented to natural hydrogen cause recognition is established, and automatic distinguishing and high-precision classification of different cause end members are achieved; according to the method, a cause contribution rate inversion network is provided, quantitative calculation of the contribution ratio of each cause in a multi-cause mixed sample is realized, the problem that quantitative splitting cannot be carried out on multi-cause superposition by a traditional method is effectively solved, and a scientific basis is provided for natural hydrogen resource quantity estimation, end member inversion and enrichment mode research.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural hydrogen resource exploration and genesis identification, and particularly relates to a natural hydrogen genesis identification and contribution rate quantification method based on multi-source data and machine learning. BACKGROUND

[0002] Natural hydrogen is an important clean energy with the advantages of clean combustion products, high energy density and great renewable potential. In the geological system, the genesis of natural hydrogen is complex and diverse, mainly including water-rock reaction, radiation decomposition, rock rupture, biological genesis and mantle source degassing processes. Different genesis often exists simultaneously and is superimposed on each other, making the genesis identification and quantitative analysis of natural hydrogen a key technical problem for natural hydrogen resource evaluation and exploration.

[0003] Currently, the common technical means for natural hydrogen genesis determination include:

[0004] Gas component analysis: determining H2, CH4, CO2 and other components by gas chromatography and other instruments to infer the gas source.

[0005] Stable isotope analysis: preliminary judgment of hydrogen genesis by hydrogen isotope (δD).

[0006] Geological-geochemical comparison: analyzing the gas generation environment in combination with lithology, structure and temperature-pressure conditions.

[0007] Although the above methods can provide preliminary genesis information, there are still the following limitations:

[0008] Strong index overlap: the H2 content and other parameters of different genesis types overlap, resulting in fuzzy classification boundaries.

[0009] Difficult to identify multiple genesis superposition: in complex geological environments, multiple genesis may act together, and traditional methods are difficult to quantitatively distinguish the contribution of each genesis.

[0010] Lack of quantitative ability: existing discrimination methods are mainly qualitative analysis, and cannot calculate the contribution proportion of each genesis in the mixed sample.

[0011] Strong subjectivity and poor repeatability: the discrimination method based on diagram and empirical threshold relies on artificial experience, lacks uniform standard and generalizability. SUMMARY

[0012] In view of the problems existing in the prior art, the present application designs an intelligent natural hydrogen genesis identification method integrating multi-source features, genesis classification, contribution rate quantification and end member reconstruction, aiming to realize intelligent identification of natural hydrogen samples with multiple genesis and quantitative calculation of contribution proportion, thereby solving the above problems in the prior art.

[0013] This invention provides a method for identifying the origin and quantifying the contribution rate of natural hydrogen based on multi-source data and machine learning, comprising:

[0014] Multi-source data acquisition: Gas composition analysis, hydrogen isotope measurement, helium isotope measurement, other associated gas measurement, and geological environment measurement were performed on natural hydrogen samples from different geological environments to obtain the gas composition and δD value of hydrogen in the natural hydrogen samples. 3 He and 4 He content, concentration of other associated gases, and geological parameters.

[0015] Data preprocessing: outlier detection and correction, noise correction, dimensional unification and standardization, air contamination correction, and missing value handling are performed on the collected multi-source data.

[0016] Feature construction: Isotope features, gas component ratio features, endmember distance and fractionation features, and geological background features were constructed using preprocessed data. Principal component analysis (PCA) and independent component analysis (ICA) were used to decouple the constructed high-dimensional features and extract the principal components that best distinguish the origin of natural hydrogen.

[0017] Multi-network construction: A natural hydrogen origin classification network C-Net, a natural hydrogen origin contribution rate inversion network R-Net, and an endmember reconstruction network E-Net were constructed to obtain a joint identification network. The genetic types include biogenetic, radiation decomposition, rock fracturing, water-rock interaction, and mantle degassing.

[0018] Multi-network joint training and overall optimization: The joint recognition network is jointly trained using the constructed features.

[0019] Application identification: The natural hydrogen sample to be analyzed is subjected to multi-source data acquisition, data preprocessing and feature construction in sequence, and the feature matrix is ​​input into the trained joint identification network to output the genetic type and genetic contribution rate of the natural hydrogen sample.

[0020] Optionally, the data preprocessing includes:

[0021] (1) Outlier detection and correction;

[0022] Outliers in multi-source data were identified using the interquartile range (IQR) and local outlier factor (LOF) methods. Data that may have involved gas leaks or abnormal peak shapes during the measurement process were removed by combining instrument log records. Large error data from repeated measurements were processed by weighted averaging or removal.

[0023] (2) Noise correction;

[0024] Drift correction is performed by repeating measurements with a standard gas to eliminate noise caused by temperature and pressure changes or instrument stability.

[0025] (3) Dimensional unification and standardization;

[0026] All data are standardized to a mean of 0 and a standard deviation of 1 using the Z-score method:

[0027] .

[0028] where, is the standardized data, x is the original data, is the sample mean on the training data set, is the sample standard deviation on the training data set.

[0029] (4) Air mixing correction;

[0030] The atmospheric correction is performed using the inert gas end-member method, and the true R 4 He / 20 Ne ratio of the sample is recovered by He :

[0031] .

[0032] .

[0033] where, R measured is the observed isotope ratio, R air is the isotope ratio of the atmospheric end-member, R Sample is the corrected true isotope ratio of the sample, f air is the proportion of air in the sample, which is calculated by the inert gas ratio.

[0034] (5) Missing value processing;

[0035] Bayesian multiple imputation (MICE) is used for missing data to ensure the integrity of the input matrix.

[0036] Optionally, the use of pre-processed data to construct isotope features, gas component ratio features, end-member distance and fractionation features, and geological background features, includes:

[0037] The δD value of hydrogen, 3 He and 4 He content to calculate the δD-H2 coupling feature, 3 He and 4 He ratio R He , to obtain the isotope feature.

[0038] The ratio of H2 to CH4, the ratio of H2 to CO2, and the ratio of H2 to He are calculated using H2 concentration and other associated gas concentrations to obtain the gas component ratio feature.

[0039] Compute the Mahalanobis distance between the sample and the known endmember, get the endmember distance:

[0040] .

[0041] where e is the feature vector of the sample, is the mean vector of the endmember, S is the covariance matrix of the endmember, is its inverse matrix.

[0042] Construct the coupling feature of δD-ln(H2) as the fractionation feature.

[0043] Use the geological parameters to construct the geological background feature composed of uranium thorium ratio, lithology parameters and porosity parameters.

[0044] Optionally, the constructing the natural hydrogen origin classification network C-Net, the natural hydrogen origin contribution rate inversion network R-Net and the endmember reconstruction network E-Net comprises:

[0045] (1) Constructing the natural hydrogen origin classification network C-Net;

[0046] A multi-layer fully connected neural network is adopted, which contains the following structures: an input layer that receives m-dimensional multi-source features; 128 nodes are set in a hidden layer 1 with ReLU activation; 64 nodes are set in a hidden layer 2 with ReLU activation; a Dropout is set with p=0.2 for preventing overfitting; an output layer is set with 5 nodes corresponding to five categories of origins, and the output layer adopts Softmax:

[0047] .

[0048] where P k is the probability that the sample belongs to the kth category of origin, z k is the logit value of the kth category output by the network.

[0049] The loss function of C-Net adopts a cross-entropy loss function:

[0050] .

[0051] where y k represents whether the kth category of origin is a true label, if the kth category of origin is true, y k =1, and the rest are 0.

[0052] The C-Net outputs the origin probability:

[0053] .

[0054] where P represents the probability distribution of five categories of origins, which is one of the inputs of R-Net.

[0055] (2) Constructing the R-Net network for inverting the contribution rate of natural hydrogen origin;

[0056] A regression-type network is adopted, which includes the following structure: input layer, which receives the feature vector Z after dimensionality reduction by PCA / ICA and the causal probability vector P output by C-Net; hidden layer 1 has 64 nodes and ReLU activation; hidden layer 2 has 32 nodes and ReLU activation; output layer has 5 nodes, corresponding to the contribution rates of the five causal categories.

[0057] The output layer uses Softmax normalization to ensure that the contribution rate satisfies:

[0058] ;

[0059] Among them, C k The contribution rate to the k-th type of cause.

[0060] R-Net loss function uses Kullback-Leibler divergence and L2 regularization:

[0061] .

[0062] Among them, D KL Used to measure the true contribution rate C true Contribution rate of model prediction C pred Differences λ is the L2 regularization term, used to prevent model overfitting and control the weight magnitude, W is the set of model weight matrices, and λ is the regularization strength hyperparameter.

[0063] R-Net outputs the contribution rates of the five causes:

[0064] .

[0065] Here, C represents the probability distribution of the five causes, which serves as the input to E-Net.

[0066] (3) Construct the end-member reconstruction network E-Net;

[0067] Constructing an endmember linear superposition model:

[0068] .

[0069] .

[0070] Wherein, δD k R He,k This represents the value of the k-th type of causal endmember. , The hydrogen δD value obtained from E-Net reconstruction, 3 He and4 He ratio R He predicted value.

[0071] E-Net parameterizes end members in a learnable manner, and comprises the following structure: an input layer receives the cause contribution rate C output by R-Net; a hidden layer 1 is provided with 32 nodes and ReLU activation; a hidden layer 2 is provided with 16 nodes and ReLU activation; and an output layer is provided with 2 nodes corresponding to delta D pred , .

[0072] The E-Net loss function adopts an isotope reconstruction error:

[0073] .

[0074] wherein, , the measured value of the hydrogen delta D value, 3 He ratio R 4 He and He .

[0075] An end member updating mechanism is designed, and the end member is taken as a trainable parameter:

[0076] ;

[0077] .

[0078] After the above technical solution is adopted, the present application at least has the following beneficial effects:

[0079] 1. Improve the natural hydrogen origin recognition accuracy: the present application can accurately identify the multi-genetic types of natural hydrogen in a complex geological environment by combining helium, hydrogen isotope ratio, multi-dimensional gas composition (H2 content, CH4 content, H2 / CH4 ratio, etc.), geological parameters and machine learning algorithm, significantly improving the recognition accuracy and reducing the error of human experience interpretation.

[0080] 2. Quantify the natural hydrogen genetic contribution ratio: the present application fuses the C-Net output probability and multi-dimensional features through the R-Net network, outputs five types of genetic contribution rates, and realizes quantitative analysis of multi-genetic samples through normalization processing, overcoming the limitation of traditional methods which can only be qualitatively determined, and providing reliable data for resource quantity estimation and enrichment rule research.

[0081] 3. Enhance data utilization efficiency and intelligent level: the preprocessed high-dimensional features are reduced in dimension through PCA / ICA and network training, which can automatically identify the origin and contribution rate of new samples, realize intelligent and automatic data analysis, and improve the sample processing efficiency.

[0082] 4. Improve applicability and repeatability: the present application can be applied in different geological environments and sampling areas through standardized data preprocessing, feature selection and model training, and the determination results are repeatable and stable, reducing the dependence on artificial experience.

[0083] 5. Optimize resource exploration and evaluation efficiency: the present application provides a complete process from data collection to C-Net, R-Net and E-Net joint training and cause output, which can quickly process a large number of samples, improve the efficiency of natural hydrogen resource exploration, cause analysis and evaluation, save experimental time and operation cost, and support enrichment area optimization. BRIEF DESCRIPTION OF DRAWINGS

[0084] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0085] Figure 1 A flowchart of a natural hydrogen cause identification and contribution rate quantification method based on multi-source data and machine learning provided by the embodiments of the present disclosure.

[0086] Figure 2 A structural diagram of a natural hydrogen cause classification network C-Net.

[0087] Figure 3 A structural diagram of a natural hydrogen cause contribution rate inversion network R-Net.

[0088] Figure 4 A structural diagram of an end member reconstruction network E-Net.

[0089] Figure 5 A natural hydrogen cause identification and contribution rate quantification result of a certain underground mine. DETAILED DESCRIPTION

[0090] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0091] The technical solution closest to the present application at present is a discrimination method based on hydrogen isotope and H2 / CH4 ratio, which mainly divides samples into shell source or mantle source degassing types through scatter plot or experience interval. However, this scheme still has significant deficiencies, which are as follows:

[0092] 1. Classification boundary is not uniform, different research areas have inconsistent standards, and the classification results lack universality.

[0093] 2. Unable to quantify the contribution ratio, only able to determine the dominant cause type, unable to give the proportion of each cause in the mixed sample.

[0094] 3. Large manual interpretation error, lack of intelligent and automated recognition ability.

[0095] 4. Unable to fully utilize multi-dimensional gas information, only relying on isotope indicators and H2, CH4 content, ignoring other gas concentrations and ratio characteristics.

[0096] Therefore, the prior art still lacks a method capable of realizing intelligent identification and contribution rate quantification of natural hydrogen multi-genesis, which is difficult to meet the needs of resource evaluation and exploration under complex geological background.

[0097] The main technical problems to be solved by the present application include:

[0098] 1. Low accuracy of cause identification: the existing method only relies on empirical graphic or single isotope indicator, and cannot accurately distinguish samples with multiple causes; limited by index overlap and geological complexity, the discrimination result is highly subjective and difficult to be applied stably in different regions.

[0099] 2. Lack of quantitative calculation method of contribution ratio: the traditional scheme can only qualitatively determine the dominant cause type, and cannot output the relative contribution of each cause in the mixed sample.

[0100] It cannot meet the needs of natural hydrogen resource estimation, potential area analysis and enrichment rule research.

[0101] 3. Low data utilization efficiency and difficulty in intelligent analysis: multi-dimensional gas information (H2, CH4 content, H2 / CH4 ratio) is not fully utilized; lack of learnable and generalizable model, unable to automatically identify and predict new collected samples. 3 He / 4 He, δD, H2 content, CH4 content, H2 / CH4 ratio) is not fully utilized; lack of learnable and generalizable model, unable to automatically identify and predict new collected samples.

[0102] 4. Poor repeatability and generalization ability: manual experience threshold and graphic method rely on researchers' experience, lack of standardized process; unable to realize unified determination across regions or geological environments, reducing the generalizability of the technology.

[0103] The difficulties in solving these technical problems are:

[0104] 1. Multi-genesis mixing effect leads to the inability of traditional single indicator or two-dimensional discrimination diagram to achieve clear distinction.

[0105] 2. Hydrogen genesis is coupled with the source of the earth's sphere, and the sample distribution has a nonlinear relationship, increasing the difficulty of model establishment.

[0106] 3. The amount of data is limited, and there are outliers and dimension differences, which require strict preprocessing and feature extraction.

[0107] The inventors have researched and solved the above technical problems by adopting the following technical solutions:

[0108] 1) Multi-source data fusion and feature extraction: Helium and hydrogen isotope ratios are combined with multi-dimensional gas components (H2, CH4, CO2, etc.) to collect and build a systematic data correction, standardization, air mixing correction and high-dimensional feature extraction method, forming a comprehensive feature system covering isotope features, component ratio features, end-member distance features and geological background parameters, breaking through the technical bottleneck of traditional reliance on single indicators and ambiguous discrimination boundaries.

[0109] 2) Intelligent genesis identification model: Machine learning technology is introduced to establish a multi-layer neural network model for natural hydrogen genesis identification, achieving automatic differentiation and high-precision classification of different genesis end-members, thereby significantly reducing the dependence on artificial experience in traditional methods and improving the reliable identification capability in complex geological environments.

[0110] 3) Quantitative contribution ratio calculation: A genesis contribution rate inversion network is proposed to realize quantitative calculation of the contribution ratio of each genesis in multi-genesis mixed samples, effectively solving the problem of traditional methods that cannot quantitatively split multi-genesis superposition, and providing scientific basis for natural hydrogen resource estimation, end-member inversion and enrichment mode research.

[0111] 4) Application in multi-genesis interwoven environment: It can effectively handle the natural hydrogen genesis identification problem in complex geological environments with multi-genesis superposition.

[0112] The technical solutions of the present application will be described in detail below with specific examples. The following specific examples can be combined with each other, and the same or similar concepts or processes may not be described in detail in some examples.

[0113] As shown in Figure 1 , the present disclosure provides a natural hydrogen genesis identification and contribution rate quantification method based on multi-source data and machine learning, comprising:

[0114] 1. Multi-source data acquisition.

[0115] Gas component analysis, hydrogen isotope measurement, helium isotope measurement, other associated gas measurement and geological environment measurement are performed on natural hydrogen samples from different geological environments to obtain the gas composition, δD value of hydrogen, He and 3 He and 4He content, other associated gas concentration and geological parameters.

[0116] (1) Gas component analysis.

[0117] Quantitative analysis of H2, CH4, CO2, N2, He and other gas components in natural hydrogen samples was performed by gas chromatograph (GC). Repeated measurements can be used to calculate measurement errors and for subsequent data cleaning.

[0118] (2) Hydrogen isotope measurement (δD).

[0119] The δD value of hydrogen was measured by stable isotope mass spectrometer (IRMS) to distinguish different genetic types. During the measurement process, standard gas was used for correction to reduce the influence of instrument drift.

[0120] (3) Helium isotope measurement (δ3He). 3 He, 4 He).

[0121] Noble gas mass spectrometer (Noble Gas MS) was used to measure 3 He / 4 He ratio (R He ) and other indicators to distinguish between mantle degassing and rock radiation decomposition. Through multi-channel simultaneous measurement, the accuracy and repeatability of the data were ensured.

[0122] (4) Measurement of other associated gases.

[0123] CH4, CO2, N2, Ar and other gas concentrations were measured to distinguish biological action, pyrolysis and water-rock reaction background processes.

[0124] (5) Geological environment measurement.

[0125] Geological parameters of the geological environment of natural hydrogen were measured.

[0126] 2. Data preprocessing.

[0127] (1) Outlier detection and correction.

[0128] The interquartile range method (IQR) and local outlier factor (LOF) method were used to identify outliers in multi-source data. Combined with instrument log records, data with possible gas leakage or peak shape abnormalities during measurement were excluded. Weighted average or rejection processing was used for large error data of repeated measurements.

[0129] (2) Noise correction.

[0130] Standard gas was used for repeated measurements to correct drift and eliminate noise caused by temperature, pressure changes or instrument stability.

[0131] (3) Dimension unification and standardization.

[0132] Because of the large difference in data dimension, for example 4 He is in the order of 10 -6 and δD is in ‰, we use Z-score standardization method to unify all data to the scale of mean 0 and standard deviation 1:

[0133] .

[0134] where, is the standardized data, x is the original data, is the sample mean on the training data set, is the sample standard deviation on the training data set.

[0135] (4) Air mixing correction.

[0136] Air mixing will affect the He isotope value. We use the inert gas end-member method to correct the atmosphere. Through 4 He / 20 Ne ratio to restore the true R He value of the sample:

[0137] .

[0138] .

[0139] where, R measured is the observed isotope ratio, R air is the isotope ratio of the atmospheric end-member, R Sample is the corrected true isotope ratio of the sample, f air is the proportion of air in the sample, which is calculated by the inert gas ratio.

[0140] (5) Missing value processing.

[0141] For missing data, we use Bayesian multiple imputation (MICE) to ensure the integrity of the input matrix.

[0142] 3. Feature construction.

[0143] On the basis of preprocessed data, multi-source features are constructed to enhance the model's ability to distinguish the origin of natural hydrogen.

[0144] (1) Isotope features.

[0145] δD-H2 coupling feature: indicates water-rock reaction and biogenic origin.

[0146] R He : to distinguish between mantle-derived degassing and deep hydrogen sources.

[0147] 4 He: Reflects the degree of radiation accumulation.

[0148] (2) Gas component ratio features.

[0149] H2 / CH4: Biological and non-biological hydrogen differentiation.

[0150] H2 / CO2: Auxiliary identification of water-rock reaction.

[0151] H2 / He: Indicates rock fracture or deep recharge.

[0152] (3) End-member distance and fractionation features.

[0153] Calculate the Mahalanobis distance of the sample and the known genetic end-member, and obtain the end-member distance:

[0154] .

[0155] where e is the feature vector of the sample, is the end-member mean vector, S is the covariance matrix of the end-member, is its inverse matrix.

[0156] Coupling features of δD-ln(H2) are constructed as fractionation features, reflecting whether hydrogen has undergone water-rock reaction or microbial action.

[0157] (4) Geological background features.

[0158] U / Th content: Characterizes the intensity of radiation genesis.

[0159] Lithology parameters: Characterize the possibility of fracture genesis.

[0160] Pore parameters: Characterize the hydrogen migration ability.

[0161] (5) High-dimensional feature fusion.

[0162] Decouple high-dimensional features using principal component analysis (PCA) and independent component analysis (ICA) to extract the principal components that best distinguish the genesis of natural hydrogen.

[0163] 4. Multi-network construction.

[0164] Construct natural hydrogen genesis classification network C-Net, natural hydrogen genesis contribution rate inversion network R-Net, and end-member reconstruction network E-Net to obtain a joint identification network, where the genesis types include biological genesis, radiation decomposition, rock fracture, water-rock interaction, and mantle source degassing.

[0165] Natural hydrogen genesis classification network C-Net.

[0166] C-Net is used to input multi-source features and output the probability of the origin type of natural hydrogen, to solve the problem that traditional methods cannot accurately distinguish multi-origin superposition and quantitatively determine the contribution.

[0167] (1) Network input.

[0168] Input matrix , where n is the number of samples; m is the dimension of multi-source features, including isotope features, gas component ratio features, endmember distance, geological background features, etc., and the feature vector after standardization is represented as .

[0169] (2) Network structure.

[0170] C-Net adopts a multi-layer fully connected neural network, as shown in Figure 2 , which includes the following structures: input layer: receives m-dimensional multi-source features; hidden layer 1 sets 128 nodes, ReLU activation; hidden layer 2 sets 64 nodes, ReLU activation; Dropout sets p=0.2, used to prevent overfitting; output layer sets 5 nodes, corresponding to five types of origins, and the output layer uses Softmax:

[0171] .

[0172] where P k is the probability that the sample belongs to the kth origin, and z k is the logit value of the kth class output by the network.

[0173] (3) Loss function.

[0174] The loss function of C-Net adopts the cross-entropy loss function:

[0175] .

[0176] where y k represents whether the kth origin is the true label, if the kth origin is true, then y k =1, and the rest are 0.

[0177] (4) Network output.

[0178] C-Net outputs the origin probability:

[0179] .

[0180] where P represents the probability distribution of five types of origins, which is one of the inputs of R-Net.

[0181] Natural hydrogen origin contribution rate inversion network R-Net.

[0182] R-Net is used to quantitatively calculate the contribution ratio of each type of origin, realizing the "quantitative superposition of multiple origins" which cannot be achieved by traditional methods.

[0183] (1) Network input.

[0184] Two inputs are adopted, including the feature vector Z after PCA / ICA dimensionality reduction and the origin probability vector P output by C-Net.

[0185] (2) Network structure.

[0186] R-Net adopts a regression network, as shown in Figure 3 , which includes the following structures: an input layer that receives the feature vector Z after PCA / ICA dimensionality reduction and the origin probability vector P output by C-Net; 64 nodes in hidden layer 1 with ReLU activation; 32 nodes in hidden layer 2 with ReLU activation; and 5 nodes in the output layer corresponding to the contribution rates of the five types of origins.

[0187] The output layer adopts Softmax normalization, so that the contribution rate satisfies:

[0188] .

[0189] Where C k is the contribution rate of the kth type of origin.

[0190] (3) Loss function.

[0191] The loss function of R-Net adopts Kullback-Leibler divergence and L2 regularization:

[0192] .

[0193] Where D KL is used to measure the difference between the true contribution rate C true and the model-predicted contribution rate C pred , is the L2 regularization term, which is used to prevent model overfitting and control the weight amplitude, W is the model weight matrix set, and λ is the regularization strength hyperparameter.

[0194] (4) Network output.

[0195] R-Net outputs the contribution rates of the five types of origins:

[0196] .

[0197] Where C represents the probability distribution of the five types of origins, which is used as the input of E-Net.

[0198] Endmember reconstruction network E-Net.

[0199] E-Net is used to deduce the difference between theoretical isotopic endmembers and observed values ​​from the causal contribution rate, for model calibration, and also for generating endmembers δD and R. He The fitted value.

[0200] (1) Endmember linear superposition model.

[0201] The isotopic values ​​of natural hydrogen samples can generally be considered as a weighted sum of endmembers from different origins:

[0202] .

[0203] .

[0204] Wherein, δD k R He,k This represents the value of the k-th type of causal endmember. , The hydrogen δD value obtained from E-Net reconstruction, 3 He and 4 He's ratio R He The predicted value.

[0205] (2) Network structure.

[0206] E-Net performs learnable parameterization of endmembers, such as Figure 4 As shown, it includes the following structure: Input layer: receives the causal contribution rate C from the R-Net output; Hidden layer 1 has 32 nodes, ReLU activated; Hidden layer 2 has 16 nodes, ReLU activated; Output layer has 2 nodes, corresponding to δD pred , .

[0207] (3) Loss function.

[0208] E-Net loss function uses isotopic reconstruction error:

[0209] .

[0210] in, , For hydrogen δD value, 3 He and 4 He's ratio R He The measured value.

[0211] (4) End-member update mechanism.

[0212] Endmembers are not fixed values; they are treated as trainable parameters.

[0213] ;

[0214] .

[0215] 5. Multi-network joint training and overall optimization.

[0216] The constructed features are used for joint training of the joint identification network, and the process is as follows:

[0217] ① Input the feature matrix X;

[0218] ② C-Net forward output cause probability P;

[0219] ③ R-Net forward output contribution rate C;

[0220] ④ E-Net forward reconstruction δD, R He and calculate the error;

[0221] ⑤ Construct joint loss ;

[0222] ⑥ Back propagation and update for all networks (including end member parameters);

[0223] ⑦ Repeat iteration until the loss converges.

[0224] 6. Application identification.

[0225] A natural hydrogen sample is collected from a certain underground mine, and multi-source data collection, data preprocessing and feature construction are sequentially performed, and the feature matrix is input into the trained joint identification network, and the cause type and cause contribution rate of the natural hydrogen sample are output, as shown in Figure 5 .

[0226] The technical solutions of the present application have been described in conjunction with the preferred embodiments shown in the accompanying drawings, but those skilled in the art can easily understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to related technical features without deviating from the principles of the present application, and the technical solutions after these changes or replacements will fall within the protection scope of the present application.

Claims

1. A method for natural hydrogen genesis identification and contribution rate quantification based on multi-source data and machine learning, characterized in that, The application relates to a method for identifying the genesis of natural hydrogen, and belongs to the technical field of natural hydrogen identification. Multi-source data collection: gas component analysis, hydrogen isotope measurement, helium isotope measurement, other associated gas measurement and geological environment measurement are performed on natural hydrogen samples from different geological environments to obtain the gas composition, δD value of hydrogen, He content, other associated gas concentration and geological parameters of the natural hydrogen samples, 3 He and 4 He content, other associated gas concentration and geological parameters; The method comprises the following steps: Data preprocessing: detecting and correcting outliers, noise correction, dimension unification and standardization, air mixing correction and missing value processing on the collected multi-source data; Feature construction: constructing isotope features, gas component ratio features, end-member distance and fractionation features and geological background features by using the preprocessed data, and decoupling the constructed high-dimensional features by using principal component analysis (PCA) and independent component analysis (ICA) to extract the main components that can best distinguish the genesis of natural hydrogen; Multi-network construction: constructing a natural hydrogen genesis classification network C-Net, a natural hydrogen genesis contribution rate inversion network R-Net and an end-member reconstruction network E-Net to obtain a joint identification network, wherein the genesis types include biological genesis, radiation decomposition, rock rupture, water-rock interaction and mantle source degassing; Multi-network joint training and overall optimization: jointly training the joint identification network by using the constructed features; 2. The method of claim 1, wherein the method is characterized by, Application identification: sequentially performing multi-source data collection, data preprocessing and feature construction on a natural hydrogen sample to be analyzed, inputting the feature matrix into the trained joint identification network, and outputting the genesis type and genesis contribution rate of the natural hydrogen sample. The data preprocessing comprises the following steps: (1) Outlier detection and correction: the four-quartile range method (IQR) and the local outlier factor (LOF) method are used to identify outliers of the multi-source data, and data that may exist gas leakage or peak abnormality in the measurement process is removed according to the instrument log record; the weighted average or removal processing is performed on the large error data of repeated measurement; (2) Noise correction: drift correction is performed by using standard gas repeated measurement to eliminate noise caused by temperature, pressure change or instrument stability; (3) Dimension unification and standardization: the Z-score standardization method is used to unify all data to the scale with the mean value of 0 and the standard deviation of 1; (4) Air mixing correction; (5) Missing value processing: the Bayesian multiple imputation (MICE) is used for the missing data to ensure the integrity of the input matrix. ； wherein, is the standardized data, x is the original data, is the sample mean on the training data set, is the sample standard deviation on the training data set; The isotope features, gas component ratio features, end-member distance and fractionation features and geological background features are constructed by using the preprocessed data, and the construction comprises the following steps: The inert gas end member method is used for atmospheric correction, and the real R 4 He / 20 Ne ratio restores the real R He value of the sample: ； ； where R measured is the observed isotope ratio, R air is the isotope ratio of the atmospheric end member, R Sample is the corrected true isotope ratio of the sample, f air is the proportion of air in the sample, calculated from the noble gas ratios. The H2 / CH4 ratio, H2 / CO2 ratio and H2 / He ratio are calculated by using the H2 concentration and other associated gas concentrations to obtain the gas component ratio features; The Mahalanobis distance of the sample and the known genesis end-member is calculated to obtain the end-member distance; 3.The method of claim 1, wherein, The delta D-ln (H2) coupling feature is constructed as the fractionation feature; The isotope characteristics are obtained by using the δD value of hydrogen, 3 He and 4 He content calculation δD-H2 coupling characteristics, 3 He and 4 He ratio R He , and obtaining isotope characteristics; The geological background features composed of the uranium / thorium ratio, lithology parameters and pore parameters are constructed by using the geological parameters. The natural hydrogen genesis classification network C-Net, the natural hydrogen genesis contribution rate inversion network R-Net and the end-member reconstruction network E-Net are constructed, and the construction comprises the following steps: ； wherein e is the eigenvector of the sample, is the endmember mean vector, S is the covariance matrix of the endmembers, is the inverse matrix thereof; (1) Constructing the natural hydrogen genesis classification network C-Net; A multi-layer fully connected neural network is adopted, and the following structure is included: an input layer that receives m-dimensional multi-source features; 128 nodes in a hidden layer 1, and ReLU activation; 64 nodes in a hidden layer 2, and ReLU activation; a Dropout setting p=0.2 for preventing overfitting; and 5 nodes in an output layer corresponding to five types of genesis, and the output layer adopts Softmax:

4. The method according to claim 3, wherein, ​ ​ ​ ； where P k is the probability that the sample belongs to the kth class of origin, z k is the logit value of the kth class output by the network. The C-Net loss function adopts the cross-entropy loss function: ； wherein y k represents whether the kth cause is a true label, if the kth cause is true, y k = 1, and the rest is 0; The C-Net output is the probability of the cause: ； Where P represents the probability distribution of the five types of causes, which is one of the inputs of R-Net; (2) Construct the natural hydrogen origin contribution rate inversion network R-Net; A regression network is adopted, which includes the following structures: input layer, receiving the feature vector Z after PCA / ICA dimensionality reduction and the cause probability vector P output by C-Net; 64 nodes are set in hidden layer 1, with ReLU activation; 32 nodes are set in hidden layer 2, with ReLU activation; 5 nodes are set in the output layer, corresponding to the contribution rate of the five types of causes; The output layer adopts Softmax normalization, so that the contribution rate satisfies: ； wherein C k is the contribution rate of the kth cause; The R-Net loss function adopts Kullback-Leibler divergence and L2 regularization: ； wherein D KL for measuring the real contribution rate C true the difference between the model prediction contribution rate C pred is an L2 regularization term for preventing model overfitting and controlling weight amplitude, W is a model weight matrix set, and λ is a regularization strength hyperparameter;​ The R-Net outputs the contribution rate of the five types of causes: ； Where C represents the probability distribution of the five types of causes, which is the input of E-Net; (3) Construct the endmember reconstruction network E-Net; Construct the endmember linear superposition model: ； ； where δD k , R He,k is the representative value of the kth end-member, , is the hydrogen δD value reconstructed by E-Net, 3 He and 4 the ratio R He of the predicted value of He; E-Net parameterizes endmembers with learnable parameters, containing the following structure: input layer: receives the causal contribution rate C output by R-Net; hidden layer 1: 32 nodes are set, and ReLU activation is performed; hidden layer 2: 16 nodes are set, and ReLU activation is performed; output layer: 2 nodes are set, corresponding to δD pred 、 ; The E-Net loss function adopts the isotope reconstruction error: ； wherein , the δD value of hydrogen, 3 the ratio R of He to 4 the measured value of the ratio R of He to He the measured value of the ratio R of He to Design an endmember update mechanism to regard the endmember as a trainable parameter: ； 。

Images