A method, device and storage medium for evaluating temperature zoning of salinized permafrost

Abstract

The present application relates to the technical field of highway railway tunnel construction, and particularly relates to a salinization frozen soil temperature zoning evaluation method, equipment and storage medium. The method comprises: obtaining measured parameters of a salinization frozen soil site, wherein the measured parameters comprise salt index, ice content, particle size, lithology and salt uniformity; performing feature engineering processing on the measured parameters to obtain discretized input features; based on the input features, calling a classification model trained by supervised learning to obtain the ground temperature zoning result of the region. The present application replaces the traditional borehole temperature measurement with the on-site easy-to-measure parameters, and realizes the rapid and automatic identification of ground temperature zoning by combining feature engineering and classification model, without the need for on-site long-term monitoring holes or complex in-situ tests, thereby significantly reducing the survey cost and operation cycle, while avoiding the uncertainty of manual experience identification, and providing an efficient, reliable and reproducible ground temperature zoning evaluation method for cold region tunnels and railway and highway engineering.

Description

Technical Field

[0001] This invention relates to the field of highway and railway tunnel construction technology, and in particular to a method, equipment and storage medium for evaluating the temperature zones of saline-alkali permafrost. Background Technology

[0002] Saline-frozen soil refers to a special geological body widely distributed in high-altitude and high-latitude cold regions. Its core characteristic is that the content of soluble salts in the permafrost or seasonally frozen soil matrix exceeds 0.3%. Driven by both temperature and salinity, the physical and mechanical properties of this type of soil are extremely complex. Soluble salts (such as chlorides and sulfates) in saline soil have a significant freezing point depression effect. When the soil contains salt, its freezing temperature decreases significantly with increasing salt content, causing pore water in saline-frozen soil to remain unfrozen even in sub-zero temperatures where conventional permafrost would already be frozen. Simultaneously, sulfate-saline soil undergoes repeated crystallization-dissolution within a specific temperature range, resulting in severe salt swelling damage.

[0003] Ground temperature is the core indicator for evaluating the stability of permafrost. According to the national standard "Code for Engineering Geological Investigation of Permafrost" (GB50324-2014), permafrost is classified into four levels based on the annual average ground temperature (Tcp): stable (Tcp < -2.0℃), basically stable (-2.0℃ ≤ Tcp < -1.0℃), unstable (-1.0℃ ≤ Tcp < -0.5℃), and extremely unstable (Tcp ≥ -0.5℃).

[0004] For saline permafrost, geothermal zoning is not only a benchmark for judging the thermal stability of permafrost, but also a key basis for assessing its bearing capacity, deformation (thaw settlement and salt swelling), and long-term settlement risk. For example, saline permafrost in the high-temperature unstable zone (Tcp≥-0.5°C) is often in a critical state close to thawing and contains a large amount of unfrozen water, making it extremely easy for engineering disturbances to cause disasters.

[0005] Currently, the geothermal assessment of saline-frozen soil mainly relies on the following methods, but all of them have limitations:

[0006] Traditional geological exploration methods rely on borehole temperature measurement as the most direct means of obtaining ground temperature, but it is costly, time-consuming, and the point measurement points often fail to reflect the strong spatial variability of saline-frozen soil.

[0007] Empirical statistical models: Existing permafrost distribution models are mostly based on regression analysis of elevation and latitude. Although they can roughly classify permafrost types, they ignore the quantitative influence of salinity on phase change temperature, resulting in extremely large prediction errors in high salinity areas.

[0008] The disconnect between laboratory tests and numerical simulations: Although the mechanisms of laboratory tests (such as freezing temperature tests) and numerical simulations (such as water-heat-salt-force multi-field coupling) are clear, the parameters are complex to obtain and difficult to apply quickly in the field.

[0009] The lack of intelligent applications: Currently, there is no technical solution that uses machine learning methods to establish a nonlinear mapping relationship between multi-dimensional measured features such as salinity, grain size, and ice content and refined geothermal zoning, which greatly limits the usability of exploration data. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of existing technologies, such as reliance on manual experience and complex temperature measurement for saline-thawed permafrost temperature identification, and lack of rapid on-site quantification methods, and to provide a method, equipment, and storage medium for evaluating saline-thawed permafrost temperature zones.

[0011] In a first aspect, the present invention provides a method for evaluating the temperature zoning of saline-alkali permafrost, comprising the following steps: Step 1: Obtain the measured parameters of the saline-frozen soil in the field. The measured parameters include salt content, ice content, particle size, lithology and salt uniformity. Step 2: Perform feature engineering on the measured parameters to obtain discretized input features; Step 3: Input the input features into the classification model trained by supervised learning to obtain the geothermal zoning evaluation results for the region.

[0012] Preferably, the feature engineering process includes: Step 21: Discretize the measured parameters of continuous numerical type. The discretization process includes one or more of the following methods: equal frequency discretization, equal width discretization, or adaptive segmentation based on the minimum description length principle. Step 22: Encode the discretized features and categorical features to generate input variables suitable for machine learning models.

[0013] Preferably, in the discretization process, for salinity parameters, an equal-frequency discretization combined with adaptive segmentation is used to automatically find the segmentation point that maximizes information gain for segmentation. For the ice content parameter, a segmentation method with equal width is used.

[0014] Preferably, in step 3, the training process of the classification model includes: Step 31: Obtain indoor test data and numerical simulation data of the reconstructed samples associated with the measured parameters; Step 32: Determine the freezing temperature under different salinity and moisture content conditions based on the indoor test data; Step 33: Simulate the variation law of the upper limit of permafrost under different salinity conditions based on the numerical simulation data, and determine the target variable of the geothermal zone based on the jump characteristics of the upper limit of permafrost. Step 34: Associate the input variables of the measured parameters after feature engineering with the target variables to construct a training sample set; Step 35: Use the gradient boosting algorithm to perform supervised learning on the training sample set and construct a classification model.

[0015] Preferably, the target variable is divided into multiple discrete partitions based on salinity level and geothermal range, including: Low-temperature stable zone of weakly saline soil, low-temperature basically stable zone of weakly saline soil, high-temperature unstable zone of weakly saline soil, high-temperature extremely unstable zone of weakly saline soil, low-temperature stable zone of moderately saline soil, low-temperature basically stable zone of moderately saline soil, high-temperature unstable zone of moderately saline soil, high-temperature extremely unstable zone of moderately saline soil.

[0016] Preferably, the classification model is an XGBoost classifier based on the gradient boosting algorithm.

[0017] Preferably, the method further includes step 4, model update, which specifically includes: Step 41: Obtain the measured data of the newly labeled data at a preset cycle; Step 42: Using the existing discretization method and encoding method, compile the newly labeled data into the training set format; Step 43: Freeze the bottom-level feature extractor of the classification model and fine-tune only the top-level classifier to achieve incremental learning.

[0018] Preferably, it also includes an interface-based service step: The classification model is integrated into an HTTP service, which receives on-site measured parameters and returns geothermal zoning prediction results via an API interface for use by mobile devices or browsers.

[0019] A temperature zoning evaluation device for saline permafrost includes a processor and a memory. When the processor executes a program stored in the memory, it implements the above-mentioned evaluation method.

[0020] A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the above-described evaluation method.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides a method for evaluating geothermal zoning in saline-frozen soil, significantly reducing the difficulty and cost of obtaining geothermal zoning data in the field. This method only requires obtaining conventional measured parameters from the saline-frozen soil field, eliminating the need for long-term geothermal monitoring wells or complex in-situ thermal response tests. This avoids the high costs and lengthy processes associated with borehole temperature measurement in traditional methods, enabling geothermal zoning determination to be completed quickly at the survey site.

[0022] This invention achieves end-to-end automatic mapping of measured parameters to geothermal zoning. By performing feature engineering on the measured parameters, the original survey data is transformed into standardized, structured, discrete input features. Then, a pre-trained classification model is used to directly output the geothermal zoning results. The entire process does not require manual experience judgment or complex theoretical formula calculations, reducing the uncertainty in identification caused by human factors. Attached Figure Description

[0023] Figure 1 Flowchart of the evaluation method for temperature zones in saline-alkali permafrost regions.

[0024] Figure 2 Flowchart.

[0025] Figure 3 Chloride salt salinity characteristics after interval distribution histogram after equal frequency discretization and adaptive segmentation.

[0026] Figure 4 Histogram of the interval distribution of ice content characteristics after equal-width discretization.

[0027] Figure 5 : Numerical simulation results curve of freezing temperature-salinity-upper limit of permafrost.

[0028] Figure 6 : Graph showing the discrete results of the freezing temperature at the same frequency.

[0029] Figure 7 Flowchart of the XGBoost algorithm. Detailed Implementation

[0030] The present invention will now be described in further detail with reference to specific embodiments. However, this should not be construed as limiting the scope of the above-described subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0031] This invention provides a method and apparatus for evaluating geothermal zones in saline permafrost. The core idea of ​​this method is to utilize readily available conventional survey parameters from the field, enhance the data's expressiveness through feature engineering, and then rapidly map them onto geothermal zones that have been finely calibrated through indoor experiments and numerical simulations using machine learning models. This avoids the difficulties of conducting complex temperature measurements or mechanical tests directly in the field.

[0032] To facilitate understanding of the technical solution of this invention, the following explains the key technical terms involved in this application: (a) Gradient boosting: Gradient boosting is a machine learning technique that typically uses weak learners (such as shallow decision trees) to combine the predictions of multiple weak learners. Each learner attempts to correct the errors of the previous learner, thereby gradually improving the accuracy and generalization ability of the model and forming a powerful prediction model.

[0033] (ii) MDLP: Minimum Description Length Principle. Based on the "minimum description length" criterion, MDLP recursively searches for the bisection point on continuous attributes that minimizes the sum of "data encoding length + model encoding length". It automatically determines the number and boundaries of discrete intervals, suppressing overfitting while maximizing class information, and providing concise, efficient and interpretable discrete features for subsequent classifiers.

[0034] (iii) Feature discretization: Feature discretization is the process of converting continuous features into discrete features. By dividing the data into different intervals and mapping them to fixed values, the efficiency and stability of the model processing are improved.

[0035] First, before performing specific judgment steps, a reliable classification model needs to be built as the core discrimination tool. The model building process includes two main stages: training set preparation and supervised learning, such as... Figure 1-2 As shown.

[0036] During the training set preparation phase, two types of data need to be collected simultaneously: one is the conventional parameters measured in the field, which serve as the input variables of the model; the other is the geothermal zoning labels obtained through indoor experiments and numerical simulations, which serve as the target variables of the model.

[0037] Specifically, step 1 involves obtaining the measured parameters of the saline-frozen soil at the site. These parameters should be readily available indicators from the geological survey report, including salinity indicators such as chloride salinity, chloride salinity, sulfate salinity, and sulfite salinity, as well as eight characteristics including ice content, particle size, lithology, and salt uniformity. Among these, particle size, lithology, and salt uniformity are categorized data, while salinity and ice content are continuous numerical data.

[0038] Methods of obtaining information include on-site sampling, comparative analysis, or direct querying of existing survey reports.

[0039] Step 2: For these measured parameters, especially continuous numerical data, detailed feature engineering processing is required.

[0040] Step 21: Since different parameters follow different statistical distributions—for example, salinity distribution is mostly Gaussian, while freezing temperature is mostly exponential—a differentiated discretization strategy is needed. The discretization process includes one or more of the following: equal-frequency discretization, equal-width discretization, or adaptive segmentation based on the minimum description length principle.

[0041] In practice, for parameters such as chloride and sulfate salinity, which significantly impact later ensemble learning, an equal-frequency discretization combined with adaptive segmentation is employed. Adaptive segmentation, based on the minimum description length principle, recursively searches for the bisection point on continuous attributes that minimizes the sum of the data encoding length and the model encoding length, automatically determining the number and boundaries of discrete intervals, thereby suppressing overfitting while maximizing class information. For the ice content parameter, equal-width discretization is used for segmentation.

[0042] Table 1 summarizes the data from the discretization process.

[0043] Figure 3 The isofrequency discrete results of the salinity features in the training set are presented. Figure 4 The isofrequency discrete results of the ice content characteristics are shown, and the distribution density of samples in each interval can be clearly seen from the figure.

[0044] Step 22: After discretization, all features need to be encoded to prevent the machine learning model from becoming unbalanced due to dimensional issues. In this embodiment, one-hot encoding is used to encode the above eight features, completing the cleaning and transformation of the input variables to form standardized input variables X.

[0045] Simultaneously, it is also necessary to prepare geothermal zoning labels corresponding to these measured parameters. Specifically, this involves sampling soil and rock around the measured data and reconstructing the field environment in a laboratory. To ensure data traceability, both the measured and reconstructed data are bound together using a four-element tag: GNSS location-UTC time-depth-data collector, achieving a one-to-one association.

[0046] Step 3: Input the input features into the classification model trained through supervised learning to obtain the geothermal zoning evaluation results for the region. Specifically, this includes: Step 31: Obtain indoor test data and numerical simulation data of the reconstructed samples associated with the measured parameters.

[0047] Step 32: Subsequently, indoor experiments were conducted to determine the freezing temperature T under different salinity and moisture contents through multiple rounds of testing. See Table 2 for details. Table 2 summarizes the freezing temperatures T under different salinity and moisture contents.

[0048] Step 33: Based on the obtained freezing temperature data, further numerical simulations are conducted. This embodiment uses COMSOL multiphysics coupled numerical simulation software to simulate the migration pattern of the frozen soil upper limit under different salinity conditions as seasonal temperature changes occur. Through multiple simulation experiments, statistical data U of the frozen soil upper limit is generated. Based on the numerical simulation data, the variation pattern of the frozen soil upper limit under different salinity conditions is simulated, and the target variables for geothermal zoning are determined based on the abrupt change characteristics of the frozen soil upper limit.

[0049] Figure 5 The numerical simulation results of freezing temperature are presented. The figures show that when salinity is low, the upper limit of the frozen soil temperature decreases steadily with decreasing sub-zero temperatures; however, when salinity exceeds a certain threshold and ground temperature continues to increase, a significant jump occurs in the upper limit of the frozen soil temperature. This jump reflects the critical point at which saline-alkali permafrost transitions from a relatively stable state to an unstable state. Recording the "ground temperature-salinity" combination at the occurrence of these jump values ​​yields the interval data Y for the target variable.

[0050] This data is discrete data, and in this embodiment, it is divided into eight categories, specifically including: weakly saline soil low-temperature stable zone (ground temperature < -2.4℃), weakly saline soil low-temperature basically stable zone (ground temperature ≥ -2.4℃ and < -1.4℃), weakly saline soil high-temperature unstable zone (ground temperature ≥ -1.4℃ and < -0.9℃), weakly saline soil high-temperature extremely unstable zone (ground temperature ≥ -0.9℃), moderately saline soil low-temperature stable zone (ground temperature < -2.5℃), moderately saline soil low-temperature basically stable zone (ground temperature ≥ -2.5℃ and < -1.5℃), moderately saline soil high-temperature unstable zone (ground temperature ≥ -1.5℃ and < -1.0℃), and moderately saline soil high-temperature extremely unstable zone (ground temperature ≥ -1.0℃). That is, the target variable is divided into multiple discrete zones according to the salinity level and ground temperature range.

[0051] Table 3 shows the classification and coding of target variables for geothermal zones.

[0052] To facilitate machine learning processing, these classifications were one-hot encoded, mapped to eight-bit binary codes ranging from 10000000 to 00000001. Simultaneously, a comparative analysis was performed on the statistical data U of the permafrost upper limit using the equal-frequency discrete method, with results as follows: Figure 6 As shown, this temperature jump can also effectively segment the dataset, verifying the rationality of the above classification.

[0053] Step 34: Subsequently, based on the four-element labels, the input variable X of the measured data is bound and associated with the target variable Y, and a data table is created in the relational database to store these {X: Y} relation pairs, thereby completing the preparation of the training set.

[0054] Step 35: Use the gradient boosting algorithm to perform supervised learning on the training sample set and construct a classification model.

[0055] Next, we proceed to the supervised learning phase to build the classification model. This embodiment employs the gradient boosting algorithm, specifically the XGBoost classifier, which is particularly suitable for low-cardinality machine learning tasks with moderate amounts of data, as discussed in this invention. First, the training set is divided into a training subset and a validation subset in an 8:2 ratio, namely X_train, Y_train and X_valid, Y_valid. The specific algorithm flow of XGBoost is as follows... Figure 7 As shown, its core idea is to use weak learners (such as shallow decision trees) as base models, and gradually combine the prediction results of multiple weak learners. Each learner attempts to correct the errors of the previous learner, thereby gradually improving the accuracy and generalization ability of the model. Through multiple rounds of iterative training, an ensemble learning algorithm F is finally formed, which is the required classification model.

[0056] Once the classification model is trained, it can be used for actual geothermal zoning prediction. In field applications, the first step is to obtain measured parameters from the saline-alkali permafrost field. These parameters are identical in type to those used in the training phase, including eight features: chloride salinity, chloride salinity, sulfate salinity, sulfite salinity, ice content, particle size, lithology, and salinity uniformity. Subsequently, these measured parameters undergo the same feature engineering processing as in the training phase. Specifically, for continuous numerical salinity and ice content data, they are segmented according to a predetermined discretization strategy and converted into standardized input feature vectors using one-heat encoding. Finally, the classification model trained through supervised learning (i.e., the aforementioned XGBoost classifier F) is called, and the processed input features are fed into the model's prediction function. The model then outputs the geothermal zoning results for the region. For example, the prediction result might be "weakly saline soil, low-temperature basically stable zone" or "medium-saline soil, high-temperature unstable zone," etc.

[0057] To facilitate quick access for on-site geologists without relying on computers, this invention can also summarize reference intervals for geothermal zoning parameters based on a large amount of training data. After training and statistical analysis on nearly ten thousand medium-sized datasets, the reference thresholds shown in Table 4 below can be obtained: Table 4 shows the reference thresholds for temperature zoning parameters of saline-frozen soil.

[0058] Step 4. In addition, considering that geological exploration work is usually carried out in stages, and new measured data will be obtained continuously as the exploration progresses, this invention also designed a model update mechanism, namely incremental learning.

[0059] Step 41: Obtain newly labeled test data at a preset cycle. Specifically, establish a user feedback and incremental learning process on a weekly basis.

[0060] Step 42: Based on the annotation method of the old model, using the existing discretization method, segmentation boundary and one-hot encoding system, recompile the newly annotated data into a dataset consistent with the original training set format.

[0061] Step 43: Freeze the bottom-level feature extractor of the classification model and fine-tune only the top-level classifier to achieve incremental learning. The purpose of this is to adapt to new data distribution changes (i.e., concept drift) while retaining historical knowledge, thereby effectively avoiding the problem of catastrophic forgetting and enabling the model to continuously optimize as data accumulates without forgetting the stable rules learned in the early stages.

[0062] The invention also includes an interface-based service step. To enhance the convenience of field applications, the aforementioned algorithm is integrated into an HTTP service, enabling geothermal zoning prediction via a standard API interface. Field personnel can use mobile devices (such as Android phones or HarmonyOS phones), tablets, or regular browsers to submit eight measured parameters via a simple HTTP request. The server-side backend calls the classification model for calculation and immediately returns the geothermal zoning results. This method greatly facilitates field personnel in quickly obtaining accurate geothermal zoning information when writing survey or test reports, significantly improving work efficiency.

[0063] In summary, the method for evaluating temperature zones in saline permafrost provided by this invention establishes a nonlinear mapping relationship between readily measurable field parameters and the results of detailed indoor experiments and numerical simulations using machine learning algorithms. After hundreds of field applications and verifications, the method achieved an accuracy rate of over 90%, effectively saving manual labor and providing reliable technical support for engineering geological exploration in cold regions.

[0064] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for evaluating the temperature zones of saline-frozen ground, characterized in that, Includes the following steps: Step 1: Obtain the measured parameters of the saline-frozen soil in the field. The measured parameters include salt content, ice content, particle size, lithology and salt uniformity. Step 2: Perform feature engineering on the measured parameters to obtain discretized input features; Step 3: Input the input features into the classification model trained by supervised learning to obtain the geothermal zoning evaluation results of the target area.

2. The evaluation method for temperature zoning of saline-frozen ground according to claim 1, characterized in that, In step 2, the feature engineering process includes: Step 21: Discretize the measured parameters of continuous numerical type. The discretization process includes one or more of the following methods: equal frequency discretization, equal width discretization, or adaptive segmentation based on the minimum description length principle. Step 22: Encode the discretized features and categorical features to generate input variables suitable for machine learning models.

3. The evaluation method for temperature zoning of saline-frozen ground according to claim 2, characterized in that, In the discretization process, for saltiness parameters, an equal-frequency discretization combined with adaptive segmentation is used to automatically find the segmentation point that maximizes information gain for segmentation. For the ice content parameter, a segmentation method with equal width is used.

4. The evaluation method for temperature zoning of saline-frozen ground according to claim 1, characterized in that, Step 3, the training process of the classification model includes: Step 31: Obtain indoor test data and numerical simulation data of the reconstructed samples associated with the measured parameters; Step 32: Determine the freezing temperature under different salinity and moisture content conditions based on the indoor test data; Step 33: Simulate the variation law of the upper limit of permafrost under different salinity conditions based on the numerical simulation data, and determine the target variable of the geothermal zone based on the jump characteristics of the upper limit of permafrost. Step 34: Associate the input variables of the measured parameters after feature engineering with the target variables to construct a training sample set; Step 35: Use the gradient boosting algorithm to perform supervised learning on the training sample set and construct a classification model.

5. The evaluation method for temperature zoning of saline-frozen ground according to claim 4, characterized in that, The target variable is divided into multiple discrete partitions based on salinity level and geothermal range, including: Low-temperature stable zone of weakly saline soil, low-temperature basically stable zone of weakly saline soil, high-temperature unstable zone of weakly saline soil, high-temperature extremely unstable zone of weakly saline soil, low-temperature stable zone of moderately saline soil, low-temperature basically stable zone of moderately saline soil, high-temperature unstable zone of moderately saline soil, and high-temperature extremely unstable zone of moderately saline soil.

6. The evaluation method for temperature zoning of saline-frozen ground according to claim 1, characterized in that, The classification model is the XGBoost classifier based on the gradient boosting algorithm.

7. The evaluation method for temperature zoning of saline-frozen ground according to claim 1, characterized in that, It also includes step 4, model update, which specifically includes: Step 41: Obtain the measured data of the newly labeled data at a preset cycle; Step 42: Using the existing discretization method and encoding method, compile the newly labeled data into the training set format; Step 43: Freeze the bottom-level feature extractor of the classification model and fine-tune only the top-level classifier to achieve incremental learning.

8. The evaluation method for temperature zoning of saline-frozen ground according to claim 1, characterized in that, It also includes the interface service steps: The classification model is integrated into an HTTP service, which receives on-site measured parameters and returns geothermal zoning prediction results via an API interface for use by mobile devices or browsers.

9. A temperature zoning evaluation device for saline permafrost, characterized in that, It includes a processor and a memory, wherein when the processor executes a program stored in the memory, it implements the evaluation method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the evaluation method according to any one of claims 1 to 8.

Images