A method for predicting the content of coal soluble organic matter molecular classes based on solvent domains

Abstract

The application discloses a coal soluble organic matter molecular category content prediction method based on a solvent domain, relates to the field of coal chemical analysis, and comprises the following steps: S1, acquiring sample data and recording solvent domain identification; S2, performing component identification on each sample to obtain relative content; S3, merging component relative content into a molecular category content candidate set; S4, constructing a feature matrix with component relative content as input and performing data preprocessing; S5, performing unsupervised evaluation on the molecular category content candidate set; and S6, training a prediction model and outputting a content prediction result. The sample is grouped through sequential solvent sub-domain processing, component features are unified, and molecular categories are merged, unsupervised target category screening is performed in combination with principal component analysis and hierarchical cluster analysis, accurate prediction of the coal soluble organic matter molecular category content is realized, and data support is provided for clean and efficient utilization of coal and optimization of a conversion process.

Description

Technical Field

[0001] This invention relates to the field of coal chemical analysis technology, and in particular to a method for predicting the content of soluble organic matter molecular categories in coal based on solvent domain. Background Technology

[0002] Coal, as a crucial component of my country's energy structure, occupies a fundamental position in industrial production and energy supply. However, traditional coal combustion processes produce SO2 and NO. x Pollutants such as PM2.5 and VOCs pose potential threats to environmental quality and human health. Therefore, promoting the transformation of coal resources towards cleaner, more efficient, and higher value-added processes has become an important research topic in the energy and chemical industry. Coal organic matter is a complex organic macromolecular system formed by highly cross-linked polycyclic aromatic skeletons and heteroatom functional groups containing oxygen, nitrogen, and sulfur. The diversity of its molecular structure directly affects the reactivity and product selectivity of coal during pyrolysis, liquefaction, and gasification processes. A deep understanding of the molecular composition and distribution characteristics of soluble organic matter in coal has significant theoretical and engineering application value for optimizing coal conversion processes and developing high-value-added coal-based products.

[0003] Gas chromatography-mass spectrometry (GC / MS) is currently the mainstream method for analyzing the molecular composition of soluble organic matter in coal. This technique can separate and identify hundreds of organic compounds in coal thermal dissolution products, providing detailed information on molecular types and relative contents. With the continuous improvement of instrument precision, the amount of data obtained from GC / MS analysis has increased dramatically. Hundreds of chromatographic peaks can be detected in a single sample, corresponding to dozens of compound classes, resulting in high data dimensionality, complex structures, and numerous parameters. In current technologies, the processing of GC / MS data typically relies on manual experience for spectral interpretation and component classification, or on simple content summaries using conventional statistical methods. However, when processing multi-solvent domains and multiple batches of samples, it is difficult to address the inconsistency in detected components between different samples, and the class imbalance phenomenon in the data is not effectively handled, limiting the accuracy and stability of model predictions.

[0004] To address the aforementioned issues, existing research attempts to introduce chemometrics and machine learning methods for automatic feature extraction and modeling of GC / MS data. For example, some literature reports the use of principal component analysis to reduce the dimensionality of high-dimensional component data, or the use of cluster analysis for sample classification. However, existing methods still have the following technical shortcomings when processing coal soluble organic matter data: First, they do not fully consider the essential differences between extracts from different polar solvents, simply mixing data from all solvent domains for modeling, introducing cross-domain heterogeneity noise, which affects the chemical interpretation ability and prediction accuracy of the model; second, when constructing the prediction model, all detected molecular categories are usually used as prediction targets, while the clustering structure of some categories is unstable under resampling or different solvent domains, resulting in insufficient generalization ability of the model when faced with new samples; third, the content distribution of each category in coal soluble organic matter is extremely uneven, with some nitrogen-containing, sulfur-containing, and other heteroatom compounds having low content, making it difficult to accurately predict a few categories using conventional supervised learning methods, resulting in prediction bias caused by category imbalance. Summary of the Invention

[0005] This invention overcomes the shortcomings of existing technologies and provides a solvent domain-based method for predicting the content of soluble organic matter molecular categories in coal. It groups samples through sequential solvent domain processing, unifies component characteristics, and merges molecular categories. Principal component analysis and hierarchical clustering analysis are combined for unsupervised target category screening. Resampling technology is introduced to evaluate category stability, and an F1-Score-guided imbalanced learning strategy is employed to optimize the prediction model. This effectively suppresses cross-domain heterogeneity noise, improves the model's predictive ability for a few categories, and ultimately achieves accurate prediction of the content of soluble organic matter molecular categories in coal, providing data support for the clean and efficient utilization and optimization of coal conversion processes.

[0006] To achieve the above objectives, the technical solution adopted by this invention is as follows: This invention provides a method for predicting the content of soluble organic matter molecules in coal based on solvent domains, comprising the following steps:

[0007] S1. Obtain multiple sample data obtained through low-temperature thermal melting treatment, and record the corresponding solvent domain identifier for each sample;

[0008] S2. Gas chromatography-mass spectrometry was used to identify the components of each sample and obtain the relative content of each component.

[0009] S3. Based on the preset molecular category merging rules, the relative contents of the components are merged into a candidate set of molecular category contents;

[0010] S4. Construct a feature matrix with the relative content of components as input, and perform data preprocessing. The preprocessing includes at least cross-sample feature alignment of each sample based on the component identifier to form a unified input feature space.

[0011] S5. Perform unsupervised evaluation for label screening on the candidate set of molecular category content. The unsupervised evaluation includes at least principal component analysis and hierarchical cluster analysis, and combines resampling cluster stability index to screen out molecular categories that are indivisible or unpredictable, thereby determining the target set of molecular categories for supervised learning, and using the corresponding molecular category content as the supervised learning output label.

[0012] S6. Train at least two prediction models under a unified data interface and output the content prediction results of the target molecule category set; when the target molecule category set has class imbalance, introduce a class imbalance handling strategy and determine its key strategy parameters using the validation set F1-Score as the optimization criterion.

[0013] In a preferred embodiment of the present invention, the low-temperature thermal dissolution process is a sequential thermal dissolution process, in which samples of the cyclohexane domain, toluene domain, and ethanol domain are obtained in the order of primary solvent cyclohexane, secondary solvent toluene, and tertiary solvent ethanol; the solvent domain identifier is used as a grouping variable for domain training to suppress heterogeneity noise introduced by cross-domain mixing.

[0014] In a preferred embodiment of the present invention, the molecular category merging rule includes: classifying components based on their structure type, functional group type, skeleton type and / or spectral library category label, and summing the relative contents of components within the same molecular category to obtain the content of that molecular category.

[0015] In a preferred embodiment of the present invention, the unsupervised evaluation of cluster stability combined with resampling includes: resampling the samples and repeatedly performing principal component analysis and hierarchical cluster analysis to calculate cluster consistency or stability index; filtering out molecular categories whose clustering results are unstable under resampling based on the stability index, and jointly determining the target molecular category set by combining the separability evaluation results of principal component analysis and hierarchical cluster analysis.

[0016] In a preferred embodiment of the present invention, the data preprocessing further includes: establishing a global component feature index to map the detected components of each sample to a unified feature space; and filling undetected components with preset missing markers, so that samples from different batches or different solvent gradients can obtain comparable unified input feature expressions.

[0017] In a preferred embodiment of the present invention, the class imbalance handling strategy includes a resampling strategy or a cost-sensitive training strategy; wherein the resampling strategy includes synthetic minority class oversampling, and the cost-sensitive training strategy includes setting class weights; the parameters of the synthetic minority class oversampling technique or the class weight parameters are constrained by the class distribution of the training set and determined by the optimal F1-Score of the validation set.

[0018] In a preferred embodiment of the present invention, the prediction model includes at least two or more of the following: logistic regression, artificial neural network, support vector machine, random forest and XGBoost, and outputs the prediction result after training and comparison based on a unified data interface.

[0019] In a preferred embodiment of the present invention, the input of principal component analysis in the unsupervised evaluation is a normalized molecular category content matrix, which is used to reduce dimensionality and extract principal component score vectors; the principal component score vectors are used as input features for hierarchical clustering analysis to evaluate the clustering structure of samples in the principal component space.

[0020] In a preferred embodiment of the present invention, the hierarchical clustering analysis uses Euclidean distance and Ward's method to perform bottom-up clustering to obtain a clustering dendrogram of the samples; the clustering label of each sample is obtained according to the cutting height or number of clusters of the dendrogram, which is used for subsequent resampling consistency evaluation.

[0021] In a preferred embodiment of the present invention, the content prediction results of the target molecule category set are evaluated by model evaluation metrics, which include one or more of accuracy, Kappa score, Jaccard coefficient, macro-average F1-Score and / or Hamming loss, and the prediction performance of each category is displayed by confusion matrix or visualization chart.

[0022] This invention addresses the shortcomings of the prior art and has the following beneficial effects:

[0023] (1) This invention provides a method for predicting the content of soluble organic matter molecules in coal based on solvent domains. By introducing a solvent domain grouping mechanism, the soluble organic matter samples in coal are independently modeled. The samples are divided into cyclohexane domain, toluene domain and ethanol domain using the principle of progressive solvent polarity. This allows components of different polarities to be enriched in their respective solvent domains, thereby effectively suppressing the heterogeneity noise introduced by cross-solvent domain mixed modeling and improving the chemical consistency of data distribution. Compared with the existing technology that simply merges all solvent extracts, the domain modeling strategy of this invention enables the model to capture the unique component correlation rules in each solvent domain, thereby further improving the model's predictive stability for complex systems of soluble organic matter in coal.

[0024] (2) This invention uses an unsupervised learning method that combines principal component analysis and hierarchical clustering analysis to screen molecular categories, and introduces resampling technology to evaluate cluster stability. By calculating the cluster consistency of candidate molecular categories under resampling conditions, categories with unstable cluster structures are screened out, and only target categories with robust chemical distribution characteristics are retained for subsequent supervised learning. This avoids misjudgment caused by limited sample size or accidental clustering, and improves the reliability of target category screening. Compared with the existing method of directly modeling all detected categories, the screening strategy of this invention enables the model to maintain stable prediction performance when facing new samples or different experimental batches, thereby further enhancing the generalization ability of the model.

[0025] (3) In the process of building the prediction model, the present invention introduces a class imbalance processing strategy guided by F1-Score. In view of the data distribution characteristics of the low content of a few categories such as nitrogen and sulfur in coal soluble organic matter, the present invention adopts synthetic minority class oversampling technology or cost-sensitive training method, and uses the validation set F1-Score as the optimization criterion to determine the relevant parameters, thereby improving the recall and precision of the model for the minority categories, effectively alleviating the prediction bias caused by class imbalance. Compared with the conventional modeling method that ignores the prediction accuracy of minority categories in the prior art, the imbalance processing strategy of the present invention enables the model to more comprehensively reflect the true composition characteristics of coal soluble organic matter, and further meet the actual needs of trace component analysis in industrial applications.

[0026] (4) This invention integrates multiple machine learning models for training and comparison under a unified data interface, enabling different algorithms such as logistic regression, artificial neural networks, support vector machines, random forests and XGBoost to be evaluated in the same feature space and label system. The optimal model can be selected or integrated output can be performed according to the prediction task, so that the complementary advantages of different algorithms in handling linear and nonlinear relationships can be fully utilized, thereby improving the accuracy and robustness of the prediction results. The multi-model integration strategy of this invention enables the prediction system to adapt to the data characteristics of different solvent domains and different molecular categories, thus providing a more reliable quantitative analysis tool for coal conversion process optimization. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a schematic diagram of the overall process of the prediction method of the present invention;

[0029] Figure 2 This is a schematic diagram of the sequential solvent thermal dissolution treatment and solvent domain modeling of the present invention;

[0030] Figure 3 This is a schematic diagram of the molecular category merging and molecular category content calculation process for GC / MS component identification in this invention;

[0031] Figure 4 This is a flowchart of the unified input feature space construction process for solvent domain feature alignment and missing filling in this invention;

[0032] Figure 5 This invention relates to the PCA score plot (PC1-PC2) projection and scree plot (principal component interpretation variance distribution plot) of the GC / MS molecular category content matrix based on solvent domain.

[0033] Figure 6 This is the HCA clustering diagram of the GC / MS molecular category content matrix based on solvent domain in this invention;

[0034] Figure 7 This is a schematic diagram of the cluster stability assessment and target molecule category screening process for resampling in this invention;

[0035] Figure 8 This is a schematic diagram of F1-Score-guided parameter optimization and multi-model integrated prediction under class imbalance conditions according to the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0038] Example 1:

[0039] Reference Figure 1 As shown, a GC / MS method for predicting soluble organic matter in coal using solvent-domain modeling and PCA / HCA screening includes the following steps:

[0040] S1. Reference Figure 2 As shown, coal sample preparation and sequential thermal melting treatment:

[0041] It should be noted that the purpose of sequential thermal dissolution in this embodiment is to utilize the principle of progressive solvent polarity to initially separate coal soluble organic matter of different polarities, thereby laying the foundation for subsequent solvent domain modeling and effectively suppressing cross-domain heterogeneity noise.

[0042] S11. Coal Sample Selection and Pretreatment: Nine coal samples, including lignite, sub-bituminous coal, and bituminous coal, were selected. After ball milling and pulverizing, the samples were passed through a 200-mesh sieve to obtain uniform coal powder. The coal powder was then vacuum dried at 80 ℃ for 48 h for later use.

[0043] S12. Sequential thermal melting process: Weigh 5-10 g of coal powder and 25-50 mL of solvent into a 250 mL high-pressure reactor and seal it. Pressurize with nitrogen and vent three times, maintaining a final pressure of approximately 1 MPa. Heat to 300 ℃, stir at 1000 r / min for 2 h, and cool to room temperature with forced air. After opening the reactor, transfer the mixture and rinse the reactor with the same solvent until the washings are clear. Load the extract into a pre-dried and weighed cellulose filter paper tube (lined with defatted cotton) and place it in a Soxhlet extractor. Add 150-200 mL of fresh solvent to the lower flask, and reflux for 24 h (6-8 times / hour) in an oil bath at 95 ℃ and a cold trap at 10 ℃. After extraction:

[0044] The filter cartridge was vacuum dried at 80 °C to constant weight to obtain the solvent-soluble residue ISP.

[0045] The solvent was recovered by rotary evaporation of the extract, and the residue was dried under vacuum at 65 °C to constant weight to obtain the solvent-soluble product SP.

[0046] The above steps were repeated sequentially with cyclohexane (CYH), toluene (TOL), and ethanol (ET) to obtain SP1 / ISP1, SP2 / ISP2, and SP3 / ISP3, respectively. Nine coal samples were treated using this method, resulting in a total of 27 heat-soluble samples.

[0047] The cyclohexane domain, toluene domain, and ethanol domain obtained by sequential extraction were modeled independently.

[0048] Solvent domain labels are retained in all stages, including modeling features, data distribution, and cluster screening, so that the heterogeneity of components is isolated by domain.

[0049] Each solvent domain can be trained independently with a personalized model, or feature vectors can be merged to build a global model for performance comparison within and across domains.

[0050] Furthermore, by accurately segmenting the solvent domains and using the progressively increasing solvent polarity to achieve grouped modeling, the mixed-model noise generated across solvent domains is avoided throughout the entire process from sample to feature. In the respective solvent domains of cyclohexane, toluene, and ethanol, the model performance shows significant differences, indicating that the chemical resolution of the domains is sufficient to improve prediction accuracy.

[0051] S2, GC / MS component identification and relative content determination:

[0052] S21, GC / MS Detection Protocol: Gas chromatography-mass spectrometry (GC-MS) was used. Extracts from each solvent domain were dissolved and diluted to a concentration of 1 mg / mL, with 2 μL injected. The split ratio was 10:1, and the injection port temperature was 300 ℃. An HP-5MS capillary column (60.0 m × 0.25 mm × 0.25 μm) was used, with nitrogen as the carrier gas at a flow rate of 1.0 mL / min and a solvent delay time of 5 min. The column oven temperature program was: initial temperature 80 ℃ for 3 min, then increased to 260 ℃ at a rate of 3 ℃ / min and held for 3 min. Mass spectrometry conditions: EI ionization, ion source temperature 280 ℃, electron energy 70 eV, mass scan range 1.6–1050 amu, and a quadrupole mass analyzer.

[0053] S22. Spectral Interpretation and Molecular Assignment: The total ion chromatogram (TIC) obtained from GC / MS detection of samples in each solvent domain is used for peak identification. The corresponding mass spectrometry information for each chromatographic peak is extracted and compared with a standard mass spectrometry library, such as the NIST library. The matching factor obtained from the search is used as the component assignment index. When the matching factor is not lower than 70 and the search result meets the optimal matching conditions, the compound corresponding to the peak is determined as a valid named component. Peaks with matching factors below the threshold, non-optimal matching, and obvious impurity or low-noise peaks are removed.

[0054] Furthermore, for each component determined to be validly named, its characteristic parameters are extracted and recorded, including at least: retention time (RT), compound name, matching factor (MF), molecular formula, CAS number, base peak area (PA), and relative content (MJ), for peak quality control or overlapping peak identification, thereby forming structured data for subsequent component statistics, comparative analysis, and process evaluation.

[0055] Furthermore, the relative content is calculated using the peak area normalization method: the sum of all effective peak areas is taken as 100%, and the relative content of each component is the percentage of its peak area.

[0056] S3. Reference Figure 3 As shown, molecular categories are grouped and content is calculated:

[0057] It should be noted that GC / MS detects a wide variety of individual compounds, and directly using them as prediction targets would lead to complex and difficult-to-interpret models. Therefore, it is necessary to perform merging based on chemical structure to enhance the chemical interpretability of the model and reduce data sparsity.

[0058] S31. Classification Rules: For all detected molecules, they are grouped according to their structural framework, functional group type, and standard spectral library affiliation, resulting in complex components being aggregated into the following 20 categories: Alkanes, Cyclanes, Alkenes, Arenes, Alcohols, Phenols, Esters, Ketones, Aldehydes, Acids, Ethers, Furans, Other oxygens (two or more oxygen-containing functional groups), Nitrogen-containing compounds (NCOCs), Sulfur-containing compounds (SCOCs), Oxygen-and-nitrogen-containing compounds (ONCOCs), Oxygen-and-sulfur-containing compounds (OSCOCs), Nitrogen-and-sulfur-containing compounds (NSCOCs), and Oxygen-, nitrogen-, and... sulfur-containing compounds and other compounds (others, halogen-containing compounds).

[0059] S32. Content Calculation: For each category, sum the relative contents of all components belonging to that category within the sample:

[0060] ;

[0061] in, This indicates the total content of the molecule category cat; This represents the total number of molecules of this category detected in the sample. denoted as the relative percentage of the i-th molecule.

[0062] Understandably, merging molecular categories reduces feature dimensionality, improves the chemical interpretability of the model, and reduces the impact of data sparsity.

[0063] S4, Reference Figure 4 As shown, feature matrix construction and missing data alignment:

[0064] To address the issue of inconsistent component detection across different samples and solvent domains, this embodiment establishes a global feature index and performs missing data filling, ensuring comparability of all samples within the same feature space.

[0065] S41. Global Feature Index: The total feature index is defined by the total number of molecular categories detected in all samples and all solvent domains.

[0066] S42. Input Feature Construction: For any given sample (a coal sample in a certain solvent domain), according to the position mapping of the global feature index, the molecular categories detected in this sample are marked with the corresponding content values. If no molecular category is detected, the position is filled with 0, or missing symbols such as NAN are used, depending on model compatibility. In this way, each sample obtains a feature vector input of equal length, comparable, and stackable.

[0067] Specifically, based on the principles of feature engineering and information alignment: global feature indexing and data alignment enable each sample to be uniformly represented in a complete feature space, effectively preventing information gaps caused by factors such as fluctuations in the sensitivity of detection instruments and background values ​​between batches; global feature indexing and zero-filling scheme enable multiple components, different solvent domains, and different detection data to fall into an equal-dimensional space, and each component can track its predicted performance throughout its entire life cycle, thereby greatly enhancing the horizontal comparability between different samples.

[0068] S5. Combining PCA and HCA for resampling to filter target categories:

[0069] In order to select those categories with stable chemical spatial distribution and reliable predictability from the 20 merged molecular categories as targets for subsequent supervised learning, this embodiment introduces an unsupervised screening strategy that combines PCA, HCA and resampling techniques. This strategy can effectively improve the robustness and generalization ability of the model and avoid misjudgments caused by accidental clustering or data fluctuations.

[0070] S51. Principal Component Analysis (PCA): Input the normalized molecular category content matrix X (samples as rows, categories as columns) into PCA to obtain the principal component score matrix and loading matrix. The score of the j-th principal component can be expressed as:

[0071] ;

[0072] in, Let the j-th principal component be the feature weight of the cat category; This represents the normalized content of the cat category.

[0073] Furthermore, to characterize the degree of influence of each molecular category on the principal components, a category cat is defined in the principal components. The contribution value (%) is the result of load normalization, for example:

[0074] .

[0075] Reference Figure 5 As shown, PC1 and PC2 contribute 50.3% and 31.5% respectively, with a cumulative contribution of approximately 81.8%. PC1-PC2 is primarily used for two-dimensional visualization projection and preliminary separability demonstration; the projection of the sample onto the PC1-PC2 plane is referenced. Figure 6 As shown, samples from different solvent domains exhibit differential distribution and clustering trends in the principal component space. To ensure information retention in subsequent cluster analysis, the PCA input dimensions used for HCA are selected as PC1 to PCK, where K satisfies that the cumulative contribution rate is not lower than a preset threshold. ,like or That is, choose the smallest K such that:

[0076] ;

[0077] in, Let be the variance contribution rate of the j-th principal component. The resulting PC1~PCK score vectors are used as input features for HCA, and are employed for clustering and stability assessment.

[0078] Specifically, PCA works by maximizing the variance between feature vectors to reduce high-dimensional features to a few dimensions (principal components) that express the main differences, thus achieving feature compression and information denoising. Its formula is as follows:

[0079] ;

[0080] Where X is the original data matrix; W is the principal component eigenvector matrix after dimensionality reduction.

[0081] Furthermore, based on the contribution results (see Table 1), a set of candidate target molecule categories was constructed. The candidate set is determined according to the following rules:

[0082] Top-N rule: For PC1 to PCK respectively, ... Sort the data from largest to smallest, then select the top N categories by contribution for each principal component and set them as a subset. Table 1, based on the Top-N rule, identifies several key categories that contribute significantly to PC1 to PC2. This indicates that the direction of principal components is primarily driven by changes in the content of 11 key categories, such as alkanes, aromatics, alcohols, phenols, esters, ketones, ethers, other oxygen-containing compounds, N-containing compounds, S-containing compounds, and other halogen-containing components.

[0083] Table 1: Contribution values ​​of various thermally soluble components of coal to PC1 and PC2

[0084] Family components (Chinese / English) PC1 contribution (%) PC2 contribution value (%) Alkanes 25.95 6.00 Cycloalkanes 2.82 3.31 Alkenes 6.49 3.47 Aromatic hydrocarbons 22.20 9.64 Alcohols 2.87 3.39 Phenols 1.24 10.74 Esters 1.85 27.86 Ketones 2.34 0.13 Aldehydes 3.60 2.51 Acids 3.60 1.92 Ethers 2.71 0.32 Furans 3.45 3.20 Other oxygen compounds 0.86 8.78 Nitrogen-containing organic compounds (NCOCs) 2.04 3.46 Sulfur-containing organic compounds (SCOCs) 3.74 2.82 Oxygen-nitrogen-containing organic compounds (ONCOCs) 2.58 2.43 Oxygen-sulfur organic compounds (OSCOCs) 3.11 1.79 Nitrogen-sulfur organic compounds (NSCOCs) 3.80 3.25 Oxygen-nitrogen-sulfur organic compounds (ONSCOCs) 3.72 3.18 Other categories 1.03 1.14

[0085] It should be noted that the contribution values ​​in the table are used to characterize the degree of influence of each molecular category on the corresponding principal components (PC1, PC2), and are obtained by normalizing the square of the loading coefficient of that category on the principal component.

[0086] S52. Hierarchical Cluster Analysis (HCA): To evaluate the clusterability of samples in the reduced-dimensional space and provide clustering labels for assessing resampling stability, the principal component score vectors obtained from PCA are used as input features for HCA. Specifically, a score vector is constructed for each sample m:

[0087] ;

[0088] Where k is the threshold based on the cumulative contribution rate in the PCA stage. The number of principal components selected.

[0089] Furthermore, the distance metric is to calculate the Euclidean distance between any two samples m and n in the principal component space:

[0090] .

[0091] It should be noted that when it is necessary to emphasize shape similarity, relevant distance / cosine distance can also be used. In this embodiment, Euclidean distance is preferred for matching with Ward's method.

[0092] Furthermore, the hierarchical clustering output is clustered bottom-up using the Ward method (or average linking method) to obtain a dendrogram of the samples. The cluster label for each sample is obtained based on the dendrogram's cutting height h or the number of clusters Q (e.g., determined by the maximum gap / profile coefficient). This is used for subsequent resampling consistency evaluation.

[0093] Specifically, HCA works by recursively merging nearest neighbor samples using a distance function to synthesize a tree-like hierarchical relationship, thereby effectively exposing the natural aggregation structure of the data.

[0094] To visually demonstrate the correspondence between sample clustering results and molecular category contribution patterns, based on the candidate target molecular category set... Constructing submatrices (Sample × Candidate Category), for Hierarchical clustering is performed on rows (samples) and columns (categories) respectively, and the results are output as heatmaps (see...). Figure 6 (As shown) The heatmap colors represent the normalized content level, and the dendrograms on both sides of the rows / columns reflect the clustering structure of the samples and categories, respectively.

[0095] S53. Reference Figure 7As shown, resampling stability assessment: To avoid accidental clustering results caused by limited sample size, class imbalance, or solvent domain differences, this embodiment resamples the candidate target molecule category set. Based on this, Bootstrap resampling is used to repeatedly sample the sample set, and PCA and HCA are repeatedly executed in each resampling to calculate the cluster consistency index and the class contribution stability index, thereby screening a stable set of target molecular classes.

[0096] Understandably, the working principle of resampling stability assessment is to automatically mask the detected drift categories by using whether the Bootstrapping or K-fold clustering labels are stable in different sample subsets.

[0097] Furthermore, Bootstrap resampling generates a sample set assuming the original sample set is... Repeat sampling B = 100 times, and the b-th resampling starts from... To ensure comparability of different resampling results, the fixed number of clusters Q is preferably the number of clusters determined by dendrite cutting / profile coefficients on all samples, and the same Q is used for cutting in each resampling.

[0098] Furthermore, the PCA+HCA is calculated for each resampling. For each :

[0099] 1. Based on candidate category set Constructing sub-moments (Rows: samples; Columns: candidate categories), and then normalize them;

[0100] 2. Regarding Perform PCA based on cumulative contribution rate threshold. Select the number of principal components The sample score vector is obtained as follows:

[0101] ;

[0102] 3. Based on Calculate the Euclidean distance between samples and perform HCA using the Ward method (or average linking method). Cut the dendrogram according to a fixed number of clusters Q to obtain cluster labels. .

[0103] In this embodiment, the clustering stability index (sample clustering consistency) is based on the consistency of all samples in clustering. The reference clustering labels obtained above As a baseline, for each resample b, it is based on the set of samples in which it appears. Calculate cluster consistency:

[0104] .

[0105] Simultaneously calculate the average contour coefficient of this resampling. Calculate the mean and standard deviation of 100 resampling operations: , , , . and When determining the stability of the cluster structure (preferred), , ).

[0106] Category contribution stability (used to filter stable target categories): Calculate the category cat under each resampling b. Maximum contribution:

[0107] .

[0108] Furthermore, statistics were compiled on each category after 100 resampling operations. With coefficient of variation:

[0109] .

[0110] And count the frequency of each category being selected by the Top-N rule. (Optional). When and (and (If enabled) retain this category in the stable target set. Typical threshold: , , .

[0111] Furthermore, the output is: a set of stable target molecule categories. This is used for subsequent supervised modeling / target category prediction; at the same time, the resampling stability statistics are output as evidence for the implementation examples, as shown in Tables 2-4.

[0112] Table 2: Parameter Settings for Resampling Stability Assessment

[0113] Parameters Value description Resampling method Bootstrap (sampling with replacement) Number of resampling times 100 Number of samples per sampling Same as the original sample size PCA cumulative contribution rate threshold η 95% or 98% HCA Distance Metric Euclidean distance HCA Linking Method Ward method or average link method Number of clustersQ The number of clusters for the reference cluster is fixed at 3. Stability threshold , (Adjustable) Category Filtering Threshold , , (Optional)

[0114] Table 3: Statistical results of cluster stability index under Bootstrap (B=100)

[0115] Dataset / Solvent Domain The average value of K Number of clustersQ All samples 3.0 3 0.86 0.04 0.54 0.07 CYH domain 3.0 3 0.89 0.05 0.57 0.08 TOL Domain 3.0 3 0.84 0.05 0.51 0.08 ET domain 3.0 3 0.82 0.03 0.49 0.06

[0116] Table 4: Statistics on the contribution stability of candidate molecule categories and screening results ( )

[0117] Categories (Chinese / English) Average maximum contribution (%) Top-N Selection Frequency (Optional) Does it meet the threshold? Conclusion (Retain / Reject) Alkanes 26.40 1.90 0.07 0.98 yes reserve Cycloalkanes 3.10 0.90 0.29 0.22 no Eliminate Alkenes 4.60 1.30 0.28 0.35 no Eliminate Aromatic hydrocarbons 22.80 1.80 0.08 0.97 yes reserve Alcohols 6.20 1.10 0.18 0.74 yes reserve Phenols 11.20 1.70 0.15 0.90 yes reserve Esters 28.50 2.40 0.08 0.99 yes reserve Ketones 5.60 0.90 0.16 0.72 yes reserve Aldehydes 3.80 1.00 0.26 0.40 no Eliminate Acids 3.90 1.20 0.31 0.38 no Eliminate Ethers 5.30 0.80 0.15 0.71 yes reserve Furans 4.10 1.10 0.27 0.33 no Eliminate Other oxygen compounds 9.00 1.60 0.18 0.88 yes reserve

[0118] As shown in Table 2, this embodiment uses Bootstrap (B=100) for resampling stability evaluation, and the PCA cumulative contribution rate threshold is set to... The optimal cluster size is 98%. HCA uses Euclidean distance and Ward's method with a fixed cluster number Q=3, and a stability threshold is set. , and category filtering thresholds , , .

[0119] As shown in Table 3, under Bootstrap (B=100), the average ARI of all samples and the solvent domains (CYH / TOL / ET) were 0.86 / 0.89 / 0.84 / 0.82, with standard deviations of 0.04 / 0.05 / 0.05 / 0.03, and average silhouette coefficients were 0.54 / 0.57 / 0.51 / 0.49, with standard deviations of 0.07 / 0.08 / 0.08 / 0.06. This indicates that the cluster structure has high consistency and small fluctuations under different resampling conditions, and the clustering results have good stability and repeatability.

[0120] Furthermore, based on the statistical results of the contribution stability of candidate molecule categories in Table 4, a set of stable target molecule categories was obtained. Its members satisfy and (and (If enabled) and other conditions. The final result is: S stable ={alkane, aromatic, ester, alcohol, phenol, ketone, ether, other oxygen, nitrogen-containing (NCOC), sulfur-containing (SCOC), other halogens}, a total of 11 categories of stable target molecules.

[0121] By using resampling stability screening, cluster label drift and unstable categories are masked through resampling and consistency indices during the unsupervised clustering stage, while retaining high-confidence molecular set modeling; thereby improving the generalization ability of the final model under extreme sample conditions or new data scenarios.

[0122] S6. Constructing a multi-model integrated prediction system for the solvent domain:

[0123] After determining a stable set of target molecule categories Subsequently, this embodiment constructs a multi-model integrated prediction system within each solvent domain. By comparing the performance of different models, the optimal model is selected or integrated to improve the accuracy and generalization ability of the prediction.

[0124] Specifically, within each domain, for the filtered set of categories... Multi-model ensemble training was employed. The ensemble models included: Linear Regression (LR), Multilayer Perceptron (ANN), Support Vector Machine (RBF kernel) (SVM), Random Forest (RF), and Gradient Boosting Tree (XGBoost).

[0125] All models use a unified data interface to input the sample feature matrix. (n is the number of samples, m is the number of features), output predicted labels. The training and validation sets are typically split in an 8:2 ratio, and cross-validation (such as k-fold CV) is used to evaluate model generalization.

[0126] The core algorithm flow is referenced in the following mathematical expression:

[0127] Normalization: x represents the original eigenvalues; This is the mean of the feature across all training samples; This represents the standard deviation of the feature across all training samples.

[0128] Principal component transformation: Z = XW; W is the characteristic loading matrix;

[0129] Random Forest Output: ;majorityvote is the majority voting function, which takes the mode (for classification tasks) or the average (for regression tasks) of the prediction results of all single trees as the final output; This represents the single-tree prediction result of the i-th decision tree in the random forest for the input feature x.

[0130] XGBoost output: M represents the number of frame trees; Let be the prediction residual of the m-th boosting tree for the input feature x.

[0131] Specifically, supervised models automatically capture subtle intra-sample co-variations by learning the relationship between input features (molecular category merge content) and output labels (target category / content); nonlinear models such as RF / XGBoost can fit complex interactions between categories and feature threshold relationships, thereby overcoming the strong nonlinearity between spectra and true content.

[0132] S7, Reference Figure 8 As shown, class imbalance handling and F1-Score-oriented parameter optimization:

[0133] To address the data imbalance problem caused by the low content of certain categories of soluble organic matter in coal, such as nitrogen-containing and sulfur-containing compounds, this embodiment introduces an imbalance handling strategy and uses F1-Score as the optimization criterion to ensure the model's predictive ability for a few categories.

[0134] Specifically, for training data in a few categories (such as Nitrogen / Sulfur-containing compounds, Others, etc.), two imbalanced processing strategies are employed:

[0135] 1. Resampling (SMOTE, synthesizing minority class samples).

[0136] 2. Cost-sensitive training (class weight adjustment).

[0137] The SMOTE parameter k is set to 1%-5% of the sample size, and the class weights are... Distribution by Category Automatic allocation, with the optimal F1 score on the validation set as the final criterion.

[0138] Understandably, the class imbalance resolution mechanism in this step fundamentally stems from the information theory principle of adjusting the sensitivity of the loss function to rare events, achieved by optimizing the following objective function:

[0139] ;

[0140] in For categories The weighting factor, This is the loss function.

[0141] The introduction of F1-Score essentially balances precision and recall, especially when there is a severe imbalance between positive and negative samples, it can avoid misjudgments caused by relying solely on accuracy:

[0142] .

[0143] Furthermore, multi-model ensemble and F1-driven optimization demonstrate the global optimality of tree-based ensemble models such as XGBoost / RF when facing typical scenarios with small sample sizes, extreme imbalance, and overlapping chemical categories. The F1 score and Jaccard coefficient are significantly better than linear / neural network methods. The F1 optimization scheme increases the recall of a few categories, making the model more adaptable to actual industrial needs.

[0144] S8. Model Performance Evaluation and Visualization Output:

[0145] All sub-model evaluation metrics include, but are not limited to: accuracy, Kappa score (measures prediction consistency), Jaccard (intersection over union of prediction and true sets), macro average F1-score (statistics by category), and Hamming loss (refinement of error rate).

[0146] Furthermore, the evaluation results are visually compared using F1-score, confusion matrix, and various other scores. Finally, each sub-model outputs the predicted content of the target molecule category. All evaluation reports and model parameters can be exported, providing support for real-time analysis and decision-making in production and research.

[0147] It should be noted that this invention, through techniques such as unsupervised-supervised concatenation, domain-specific modeling, stability screening, feature alignment, and data resampling, not only improves robustness to heterogeneity within and outside the overall sample but also facilitates rapid model migration to new sample data. The synergistic mechanism between solvent domain, chemical category, and model performance ultimately determines the high performance and high reliability of the prediction model.

[0148] The above description is based on the preferred embodiments of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered exemplary and non-limiting in all respects. The scope of the invention is defined by the appended claims rather than the foregoing description, and all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0149] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for predicting the content of soluble organic matter molecules in coal based on solvent domain, characterized in that, Includes the following steps: S1. Obtain multiple sample data obtained through low-temperature thermal melting treatment, and record the corresponding solvent domain identifier for each sample; S2. Gas chromatography-mass spectrometry was used to identify the components of each sample and obtain the relative content of each component. S3. Based on the preset molecular category merging rules, the relative contents of the components are merged into a candidate set of molecular category contents; S4. Construct a feature matrix with the relative content of components as input, and perform data preprocessing. The preprocessing includes at least cross-sample feature alignment of each sample based on the component identifier to form a unified input feature space. S5. Perform unsupervised evaluation for label screening on the candidate set of molecular category content. The unsupervised evaluation includes at least principal component analysis and hierarchical cluster analysis, and combines resampling cluster stability index to screen out molecular categories that are indivisible or unpredictable, thereby determining the target set of molecular categories for supervised learning, and using the corresponding molecular category content as the supervised learning output label. S6. Train at least two prediction models under a unified data interface and output the content prediction results of the target molecule category set; when the target molecule category set has class imbalance, introduce a class imbalance handling strategy and determine its key strategy parameters using the validation set F1-Score as the optimization criterion.

2. The method for predicting the content of soluble organic matter molecules in coal based on solvent domain according to claim 1, characterized in that: The low-temperature thermal dissolution process is a sequential thermal dissolution process, in which samples in the cyclohexane domain, toluene domain, and ethanol domain are obtained in the order of primary solvent cyclohexane, secondary solvent toluene, and tertiary solvent ethanol. The solvent domain identifier is used as a grouping variable for domain-specific training to suppress heterogeneous noise introduced by cross-domain model mixing.

3. The method for predicting the content of soluble organic matter molecules in coal based on solvent domain according to claim 1, characterized in that: The molecular category merging rules include: classifying components based on their structure type, functional group type, skeleton type, and / or spectral library category labels, and summing the relative contents of components within the same molecular category to obtain the content of that molecular category.

4. The method for predicting the content of soluble organic matter molecules in coal based on solvent domain according to claim 1, characterized in that: The unsupervised evaluation of cluster stability combined with resampling includes: resampling the samples and repeatedly performing principal component analysis and hierarchical cluster analysis to calculate cluster consistency or stability indices; filtering out molecular categories with unstable clustering results under resampling based on the stability indices, and jointly determining the target molecular category set by combining the separability evaluation results of principal component analysis and hierarchical cluster analysis.

5. The method for predicting the content of soluble organic matter molecules in coal based on solvent domain according to claim 1, characterized in that: The data preprocessing also includes: establishing a global component feature index to map the detected components of each sample to a unified feature space; filling undetected components with preset missing markers, so that samples from different batches or different solvent gradients can obtain comparable unified input feature expressions.

6. The method for predicting the content of soluble organic matter molecules in coal based on solvent domain according to claim 1, characterized in that: The class imbalance handling strategy includes a resampling strategy or a cost-sensitive training strategy; wherein the resampling strategy includes synthetic minority class oversampling technique, and the cost-sensitive training strategy includes setting class weights; the parameters of the synthetic minority class oversampling technique or the class weight parameters are constrained by the class distribution of the training set and determined by the criterion of optimal F1-Score on the validation set.

7. The method for predicting the content of soluble organic matter molecules in coal based on solvent domain according to claim 1, characterized in that: The prediction model includes at least two or more of the following: logistic regression, artificial neural network, support vector machine, random forest and XGBoost. The prediction results are output after training and comparison based on a unified data interface.

8. The method for predicting the content of soluble organic matter molecules in coal based on solvent domain according to claim 1, characterized in that: In the unsupervised evaluation, the input to principal component analysis is the normalized molecular category content matrix, which is used to reduce dimensionality and extract principal component score vectors. The principal component score vectors are used as input features for hierarchical cluster analysis to evaluate the clustering structure of samples in the principal component space.

9. The method for predicting the content of soluble organic matter molecules in coal based on solvent domain according to claim 8, characterized in that: The hierarchical clustering analysis uses Euclidean distance and Ward's method to perform bottom-up clustering, resulting in a cluster dendrogram of the samples. The cluster label for each sample is obtained based on the cutting height or number of clusters in the dendrogram, which is used for subsequent resampling consistency evaluation.

10. The method for predicting the content of soluble organic matter molecules in coal based on solvent domain according to claim 1, characterized in that: The predicted content of the target molecule category set is evaluated using model evaluation metrics, which include one or more of the following: accuracy, Kappa score, Jaccard coefficient, macro-average F1-Score, and / or Hamming loss. The prediction performance of each category is displayed through a confusion matrix or a visualization chart.

Images