Method and system for constructing a model for detecting unfrozen soil water content based on neural network

Abstract

The present application relates to the field of model construction technology, and discloses a method and system for constructing a model for detecting the unfrozen water content of frozen soil based on a neural network. The method comprises: performing segmented sampling tests on clay samples in the temperature range of 0°C to ‑20°C, obtaining a multidimensional original data set, standardizing the data set and performing principal component analysis to obtain a normalized feature vector. The feature vector is subjected to three-stage classification labeling to obtain a training data matrix. A deep neural network is used for five-fold cross-validation training to construct a prediction model with three hidden layers. The model output is corrected and calculated according to the mineral composition and porosity parameters of the soil body, and finally the outliers are identified by root mean square error analysis and the principle of three times the standard deviation to obtain a standardized prediction result. The present application can solve the technical problem of low prediction accuracy of unfrozen water content in the prior art by establishing a mapping relationship between thermal parameters and unfrozen water content, and correcting it in combination with soil characteristic parameters.

Description

Technical Field

[0001] The present application relates to the field of model construction, and in particular to a method and system for constructing a frozen soil unfrozen water content detection model based on a neural network. Background Art

[0002] The unfrozen water content of frozen soil is an important parameter to characterize the freezing characteristics of frozen soil, and has a significant impact on the thermodynamic and mechanical properties of frozen soil. At present, the test methods for the unfrozen water content of frozen soil mainly include gamma-ray method, nuclear magnetic resonance method, differential scanning calorimetry, temperature-controlled centrifuge method and temperature-controlled pressure plate method. These methods determine the unfrozen water content by measuring the energy change during the change of water state or phase change. At the same time, the thermal physical parameters such as thermal conductivity and specific heat of frozen soil are also closely related to the unfrozen water content, which can be measured by methods such as hot wire method and mixed calorimetry. Due to the influence of factors such as soil mineral composition and pore structure, there is a complex nonlinear relationship between the unfrozen water content and temperature, and it shows obvious stage characteristics during the freezing process.

[0003] However, the existing methods for testing the unfrozen water content in frozen soil have some limitations. First, most testing methods can only obtain the unfrozen water content values ​​at discrete temperature points, which makes it difficult to accurately describe the continuous change characteristics of the unfrozen water content during the freezing process. Secondly, affected by the soil structure and mineral composition, the test data often have certain discreteness and uncertainty, and need to be comprehensively analyzed and corrected in combination with multiple parameters. In addition, since frozen soil is a complex multiphase medium, the coupling relationship between its thermal parameters and unfrozen water content has not been fully understood, and there is a lack of effective prediction methods. Summary of the invention

[0004] The present application provides a method and system for constructing a frozen soil unfrozen water content detection model based on a neural network, which is used to solve the technical problem of low prediction accuracy of unfrozen water content in the prior art by establishing a mapping relationship between thermal parameters and unfrozen water content and making corrections in combination with soil characteristic parameters.

[0005] In the first aspect, the present application provides a method for constructing a detection model for unfrozen water content in frozen soil based on a neural network, and the method for constructing a detection model for unfrozen water content in frozen soil based on a neural network comprises: testing clay samples by segmented sampling in a temperature range of 0°C to -20°C, wherein the sampling interval of the first temperature range of 0°C to -5°C is 0.5°C, and the sampling interval of the second temperature range of -5°C to -20°C is 2°C, to obtain a multidimensional original data set; using Min-Max standardization and principal component analysis to process the multidimensional original data set to obtain a normalized feature vector with a numerical range of 0 to 1; based on a thermal parameter critical value, marking the normalized feature vector as an unfrozen stage The unfrozen water content prediction model is obtained by inputting the training data matrix into a deep neural network for five-fold cross-validation training, and the unfrozen water content prediction model has three hidden layers and the number of neurons in each layer is 2 times, 1.5 times and 1 times of the input dimension respectively. The output result of the unfrozen water content prediction model is calculated by linear weighted correction based on the soil mineral composition ratio coefficient and porosity characteristic parameters to obtain the unfrozen water content value after correction coefficient compensation. The abnormal values ​​in the unfrozen water content value after correction coefficient compensation are identified by root mean square error analysis and 3 times standard deviation principle to obtain standardized prediction results.

[0006] In a first implementation of the first aspect, the clay sample is tested by segmented sampling in a temperature range of 0°C to -20°C, wherein the sampling interval of the first temperature range of 0°C to -5°C is 0.5°C, and the sampling interval of the second temperature range of -5°C to -20°C is 2°C, to obtain a multidimensional original data set, including: setting the initial moisture content to 15%-35%, setting the density to 1.5-1.8g / cm³, processing the clay sample by density and moisture content control to obtain a clay sample to be tested, and obtaining basic physical parameter data through liquid limit, plastic limit, and particle grading tests; testing the unfrozen water content data of the clay sample to be tested in the range of 0°C to -5°C by a temperature-controlled centrifuge, and testing the unfrozen water content data of the clay sample to be tested in the range of -5°C to -20°C by a temperature-controlled pressure plate instrument; sampling The thermal conductivity data of the clay sample to be tested in the temperature range of 0°C to -20°C are tested by an improved hot wire method, and the specific heat data of the clay sample to be tested in the temperature range of 0°C to -20°C are tested by a mixed calorimetry method; the unfrozen water content test data, thermal conductivity data and specific heat data of each temperature point are tested three times to obtain average value data, and the average value data is combined and sorted with basic physical parameter data to obtain a multidimensional original data set; the Mahalanobis distance method is used to identify abnormal data points in the multidimensional original data set, and a data set after removing the abnormal values ​​is obtained, and the data set after removing the abnormal values ​​is smoothed by an interpolation method to obtain a continuously distributed temperature-thermal parameter data set; each parameter in the continuously distributed temperature-thermal parameter data set is normalized to obtain the multidimensional original data set.

[0007] In a second implementation of the first aspect, the multidimensional original data set is processed by Min-Max standardization and principal component analysis to obtain a normalized eigenvector with a numerical range of 0 to 1, including: performing zero mean processing on the temperature value, thermal conductivity, specific heat, and unfrozen water content data in the multidimensional original data set to obtain standardized data with a mean of 0, and processing the standardized data by variance normalization to obtain normalized data with a standard deviation of 1; analyzing the normalized data by feature correlation to obtain a feature correlation coefficient matrix, and decomposing the feature correlation coefficient matrix by eigenvalues ​​to obtain eigenvalues ​​and eigenvectors; arranging the eigenvalues ​​in descending order And calculate the cumulative contribution rate, determine the number of principal components, and sort the eigenvectors according to the corresponding eigenvalues ​​to obtain the principal component loading matrix; perform matrix multiplication on the normalized data and the principal component loading matrix to obtain the principal component score matrix, and use the Min-Max standardization method to process the principal component score matrix to obtain the initial eigenvector with a numerical range of 0 to 1; use the variance contribution rate to weightedly combine the initial eigenvectors to obtain a comprehensive characteristic index, and use the data distribution consistency test to obtain the calibrated characteristic index; use data dimensionality reduction mapping to perform coordinate transformation on the calibrated characteristic index to obtain a normalized eigenvector.

[0008] In a third implementation of the first aspect, the normalized feature vector is marked as a three-stage classification of an unfrozen stage, a frozen stage, and a frozen stage based on the critical value of the thermal parameter to obtain a training data matrix with stage attributes, including: using zero crossing points to analyze the normalized feature vector to obtain jump position data of the temperature-thermal parameter curve, and determining the thermal parameter critical value of the jump position data through cluster analysis to obtain the freezing stage boundary temperature data; performing piecewise linear fitting on the normalized feature vector to obtain slope change data in each temperature interval, and performing correspondence analysis between the slope change data and the freezing stage boundary temperature data to determine the stage feature discrimination criterion; grouping the normalized feature vector according to the stage feature discrimination criterion to obtain the unfrozen stage boundary temperature data. The method comprises the following steps: forming a stage data subset, a frozen stage data subset and a frozen stage data subset, assigning category identifiers to the three data subsets respectively, and establishing a labeled classification data group; analyzing the labeled classification data group by Fisher discriminant analysis to obtain the discrimination score of each stage, and performing fuzzy classification processing on the data points whose discrimination score is lower than a threshold value to generate stage classification correction data; processing the stage classification correction data based on the Sigmoid function to obtain stage boundary transition data, and analyzing the stage boundary transition data by probability distribution to calculate the membership value of each stage; performing matrix concatenation of the membership value and the normalized feature vector to construct an augmented feature matrix, and enhancing the features of the augmented feature matrix by covariance analysis to generate a training data matrix with stage attributes.

[0009] In a fourth implementation of the first aspect, the training data matrix is ​​input into a deep neural network for five-fold cross-validation training to obtain an unfrozen water content prediction model containing three hidden layers and the number of neurons in each layer is 2 times, 1.5 times and 1 times the input dimension, respectively, including: dividing the training data matrix into a training set and a validation set in a ratio of 8:2, obtaining an initial training data subset, and dividing the initial training data subset into five equal parts by stratified sampling to obtain five groups of cross-validation data sets; processing the five groups of cross-validation data sets by data shuffling to obtain a randomly sorted training sample sequence, and batch-dividing the training sample sequence by a mini-batch generator to obtain batch training data blocks; setting the number of neurons in the first hidden layer of the deep neural network to twice the input dimension, establishing a dimensionality expansion feature extraction layer, and performing a nonlinear transformation on the output of the first hidden layer by a ReLU activation function to obtain an initial feature activation function. active value; setting the number of neurons in the second hidden layer to 1.5 times the input dimension, establishing a feature compression layer, and performing feature mapping on the output of the second hidden layer through the hyperbolic tangent function to obtain the intermediate layer feature representation; setting the number of neurons in the third hidden layer to 1 times the input dimension, establishing a feature reconstruction layer, and numerically compressing the output of the third hidden layer through the Sigmoid function to obtain normalized feature data; calculating the mean square error between the normalized feature data and the true value of the unfrozen water content to obtain the loss function value, and performing gradient back propagation on the loss function value through the Adam optimizer to calculate the network parameter update amount; performing weighted sum operation on the network parameter update amount and the original network parameter to generate a candidate network parameter group, and evaluating the performance of the candidate network parameter group through the validation set, and obtaining an unfrozen water content prediction model comprising three hidden layers and the number of neurons in each layer being 2 times, 1.5 times and 1 times the input dimension respectively.

[0010] In a fifth implementation of the first aspect, the output result of the unfrozen water content prediction model is calculated based on the soil mineral composition ratio coefficient and the porosity characteristic parameter through linear weighted correction to obtain the unfrozen water content value after the correction coefficient is compensated, including: performing deviation calculation on the output result of the unfrozen water content prediction model and the original unfrozen water content test value to obtain initial error data, and analyzing the initial error data through statistical distribution to obtain error distribution characteristic value; analyzing the soil mineral composition through X-ray diffraction to obtain mineral content percentage data, and weighting the mineral content percentage data to obtain the mineral composition ratio coefficient; measuring the porosity through mercury injection method to obtain the unfrozen water content value. A test soil sample is taken to obtain pore size distribution data, and the pore size distribution data is density-corrected to obtain porosity characteristic parameters; the mineral composition ratio coefficient is linearly combined with the porosity characteristic parameter to construct a soil property correction factor, and the soil property correction factor is processed by characteristic normalization to obtain a standardized correction coefficient; a correlation analysis is performed between the standardized correction coefficient and the error distribution characteristic value to obtain a weight adjustment coefficient, and the weight adjustment coefficient is fitted by parameters to obtain a correction function expression; the output result of the unfrozen water content prediction model is calculated by the correction function expression to obtain the unfrozen water content value after the correction coefficient is compensated.

[0011] In a sixth implementation of the first aspect, the abnormal value in the unfrozen water content value after the correction coefficient compensation is identified by root mean square error analysis and the principle of three times the standard deviation to obtain a standardized prediction result, including: performing a difference operation on the unfrozen water content value after the correction coefficient compensation and the measured value to obtain a prediction error sequence, and performing a square sum calculation on the prediction error sequence to obtain an error sum; performing a square root processing on the error sum to obtain a root mean square error value, and normalizing the root mean square error value to obtain a standardized error index; performing a statistical distribution analysis on the unfrozen water content value after the correction coefficient compensation The method comprises the following steps: analyzing the unfrozen water content after the correction coefficient is compensated, obtaining the mean and standard deviation data, and amplifying the standard deviation data by 3 times to obtain the outlier judgment threshold; performing difference calculation between the unfrozen water content value after the correction coefficient is compensated and the mean to obtain a deviation data sequence, and comparing and screening the deviation data sequence with the outlier judgment threshold to obtain an outlier marking result; removing the abnormal data points in the outlier marking result to obtain a data set after removing the outliers, and processing the data set after removing the outliers by data interpolation to obtain a continuous distribution sequence; processing the continuous distribution sequence by Min-Max standardization to obtain a standardized prediction result.

[0012] In a second aspect, the present application provides a system for constructing a frozen soil unfrozen water content detection model based on a neural network, and the system for constructing a frozen soil unfrozen water content detection model based on a neural network includes:

[0013] A test module is used to test clay samples in a temperature range of 0°C to -20°C by segmented sampling, wherein the sampling interval of the first temperature range of 0°C to -5°C is 0.5°C, and the sampling interval of the second temperature range of -5°C to -20°C is 2°C, to obtain a multidimensional original data set;

[0014] An analysis module, used for processing the multidimensional original data set by using Min-Max standardization and principal component analysis to obtain a normalized feature vector with a value range between 0 and 1;

[0015] A labeling module, used for labeling the normalized feature vector into three-stage classification of an unfrozen stage, a frozen stage, and a frozen stage based on a critical value of a thermal parameter, so as to obtain a training data matrix with stage attributes;

[0016] A training module is used to input the training data matrix into a deep neural network for five-fold cross-validation training to obtain an unfrozen water content prediction model with three hidden layers and the number of neurons in each layer being 2 times, 1.5 times and 1 times the input dimension respectively;

[0017] A correction module is used to calculate the output result of the unfrozen water content prediction model through linear weighted correction based on the soil mineral composition ratio coefficient and porosity characteristic parameters to obtain the unfrozen water content value after correction coefficient compensation;

[0018] The identification module is used to identify abnormal values ​​in the unfrozen water content values ​​after the correction coefficient compensation through root mean square error analysis and the principle of three times standard deviation, so as to obtain a standardized prediction result.

[0019] In the technical solution provided in this application, by adopting a segmented sampling test strategy in the temperature range of 0℃ to -20℃, using a small interval sampling of 0.5℃ in the range of 0℃ to -5℃, and a large interval sampling of 2℃ in the range of -5℃ to -20℃, it is ensured that more dense data points are obtained in the key temperature range of frozen soil phase change, and the accuracy of capturing the changing characteristics of unfrozen water content during the freezing process is improved, while the test workload is reasonably controlled. The multidimensional original data set is processed by Min-Max normalization and principal component analysis, and the numerical range is limited to the range of 0 to 1. It not only eliminates the influence between parameters of different dimensions, but also extracts the main features of the data through dimensionality reduction processing, reducing data redundancy. Based on the critical value of the thermal parameter, the normalized feature vector is classified and labeled into three stages: unfrozen stage, frozen stage, and frozen stage, which accurately reflects the physical state characteristics of frozen soil at different temperature stages and provides a more effective data structure for the training of neural networks. Five-fold cross-validation training was performed through a deep neural network, and a network structure design with three hidden layers was adopted, in which the number of neurons in each layer was 2 times, 1.5 times, and 1 times the input dimension, respectively. This not only ensured that the model had sufficient feature extraction capabilities, but also avoided the overfitting problem caused by the overly complex network structure. A linear weighted correction mechanism based on the soil mineral composition ratio coefficient and porosity characteristic parameters was introduced to effectively compensate for the prediction deviation caused by soil structure differences and improve the applicability of the model in different types of clay samples. Finally, outliers were identified through root mean square error analysis and the 3-times standard deviation principle to ensure the reliability and accuracy of the final prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0021] Figure 1 It is a schematic diagram of an embodiment of a method for constructing a frozen soil unfrozen water content detection model based on a neural network in an embodiment of the present application;

[0022] Figure 2 This is a schematic diagram of an embodiment of a system for constructing a frozen soil unfrozen water content detection model based on a neural network in an embodiment of the present application. DETAILED DESCRIPTION

[0023] The embodiment of the present application provides a method and system for constructing a frozen soil unfrozen water content detection model based on a neural network. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0024] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiment of the present application, an embodiment of a method for constructing a frozen soil unfrozen water content detection model based on a neural network includes:

[0025] Step S101, testing clay samples by segmented sampling in a temperature range of 0°C to -20°C, wherein the sampling interval of the first temperature range of 0°C to -5°C is 0.5°C, and the sampling interval of the second temperature range of -5°C to -20°C is 2°C, to obtain a multidimensional original data set;

[0026] Step S102, using Min-Max normalization and principal component analysis to process the multidimensional original data set to obtain a normalized feature vector with a value range between 0 and 1;

[0027] Step S103, based on the critical value of the thermal parameter, marking the normalized feature vector into three-stage classification of unfrozen stage, frozen stage, and frozen stage, to obtain a training data matrix with stage attributes;

[0028] Step S104, inputting the training data matrix into the deep neural network for five-fold cross validation training, and obtaining an unfrozen water content prediction model having three hidden layers and the number of neurons in each layer being 2 times, 1.5 times and 1 times of the input dimension respectively;

[0029] Step S105, based on the soil mineral composition ratio coefficient and porosity characteristic parameters, the output result of the unfrozen water content prediction model is calculated by linear weighted correction to obtain the unfrozen water content value after correction coefficient compensation;

[0030] Step S106, identifying abnormal values ​​in the unfrozen water content values ​​after correction coefficient compensation through root mean square error analysis and the principle of three times standard deviation, and obtaining a standardized prediction result.

[0031] It is understandable that the execution subject of the present application may be a neural network-based frozen soil unfrozen water content detection model construction system, or a terminal or a server, which is not specifically limited here. The present application embodiment is described by taking a server as the execution subject as an example.

[0032] Specifically, the clay samples were sampled and tested in sections. Considering that the unfrozen water content of frozen soil changes dramatically in the temperature range of 0℃ to -5℃, a dense sampling interval of 0.5℃ was used in this range, while the change in the temperature range of -5℃ to -20℃ was relatively gentle, and a larger sampling interval of 2℃ was used. It is very important to control the initial parameters of the clay samples. The initial water content needs to be controlled between 15%-35%, and the density needs to be controlled within the range of 1.5-1.8g / cm³. The unfrozen water content in the range of 0℃ to -5℃ was tested by a temperature-controlled centrifuge, and the unfrozen water content in the range of -5℃ to -20℃ was tested by a temperature-controlled pressure plate instrument. At the same time, the thermal conductivity was tested by the improved hot wire method, and the specific heat was tested by the mixed calorimetry method. The average value was obtained by repeating the test three times for each temperature point, and a multidimensional original data set containing temperature, thermal conductivity, specific heat and unfrozen water content was obtained. The multidimensional original data set was then processed by Min-Max standardization and principal component analysis. Min-Max standardization is a linear transformation method that maps data to the interval [0,1]. For the original data x, it is normalized by the formula (x-min) / (max-min). Principal component analysis converts the original features into new linearly independent features through orthogonal transformation. By calculating the eigenvalues ​​and eigenvectors, the principal components with a cumulative contribution rate of more than 85% are selected as new feature representations. The standardized data is subjected to feature correlation analysis to obtain the feature correlation coefficient matrix. The eigenvalues ​​and eigenvectors are obtained through eigenvalue decomposition. The eigenvectors are sorted according to the corresponding eigenvalues ​​to obtain the principal component load matrix, and finally the normalized eigenvectors with a numerical range of 0 to 1 are obtained.

[0033] When the normalized feature vector is marked for three-stage classification, the freezing process of the frozen soil sample is divided into three stages: unfrozen, frozen, and frozen based on the critical value of the thermal parameter. The jump position of the temperature-thermal parameter curve is determined by zero crossing point analysis, and the critical value of the thermal parameter is determined by cluster analysis. The normalized feature vector is fitted with piecewise linear fitting, the slope change in each temperature interval is analyzed, and the stage feature discrimination criterion is established. Fisher discriminant analysis is used to evaluate the discrimination of each stage, and fuzzy classification is performed on the data points with a discrimination score lower than the threshold. The smooth transition of the stage boundary is achieved based on the Sigmoid function, and finally a training data matrix with stage attributes is obtained. Next, the model is trained by a deep neural network, and the training data is divided into a training set and a validation set in a ratio of 8:2, and a five-fold cross-validation method is used. The neural network structure contains three hidden layers. The number of neurons in the first layer is twice the input dimension, and the ReLU activation function is used for nonlinear transformation; the number of neurons in the second layer is 1.5 times the input dimension, and the hyperbolic tangent function is used for feature mapping; the number of neurons in the third layer is the same as the input dimension, and the Sigmoid function is used for numerical compression. The mean square error is used as the loss function, the Adam optimizer is used for parameter optimization, and the model performance is evaluated through the validation set.

[0034] Considering the influence of soil properties on the unfrozen water content, the model output results need to be corrected. X-ray diffraction analysis is used to obtain the percentage data of soil mineral content, and mercury injection testing is used to obtain the pore size distribution data, and the mineral composition ratio coefficient and porosity characteristic parameters are obtained respectively. The two parameters are linearly combined to obtain the soil property correction factor, and the standardized correction coefficient is obtained through characteristic normalization. The correction function expression is established to compensate the prediction results. Finally, the accuracy evaluation and outlier identification of the corrected unfrozen water content values ​​are carried out. The difference between the predicted value and the measured value is calculated to obtain the prediction error sequence, and the root mean square error is calculated and normalized. The 3 times standard deviation criterion is used to identify outliers, and the mean and standard deviation of the data are calculated. Data points that exceed the mean ± 3 times the standard deviation range are marked as outliers. The abnormal data points are eliminated and supplemented by interpolation, and finally the standardized prediction results are obtained.

[0035] Taking a clay sample as an example, the initial water content is 25% and the density is 1.65g / cm³. 11 tests were conducted at intervals of 0.5℃ in the range of 0℃ to -5℃, and 8 tests were conducted at intervals of 2℃ in the range of -5℃ to -20℃, and each temperature point was tested 3 times. The original unfrozen water content obtained after data processing was 25% at 0℃, dropped to 18.5% at -2℃, 12.3% at -5℃, 8.6% at -10℃, and stabilized at 6.2% at -20℃. The thermal conductivity data showed that it was 1.52W / (m·K) at 0℃, gradually increased with decreasing temperature, and reached 1.78W / (m·K) at -20℃. After principal component analysis, the first three principal components with a cumulative contribution rate of 87.3% were selected as features. The three-stage classification results showed that -0.5℃ and -4.5℃ were two key phase transition points, corresponding to the transition from unfrozen to frozen and frozen to frozen. The root mean square error after deep neural network training is 0.85%. After correction of soil properties, the final prediction accuracy is improved and the root mean square error is reduced to 0.62%.

[0036] In the embodiment of the present application, by adopting a segmented sampling test strategy in the temperature range of 0°C to -20°C, using a small interval sampling of 0.5°C in the range of 0°C to -5°C, and using a large interval sampling of 2°C in the range of -5°C to -20°C, it is ensured that more dense data points are obtained in the key temperature range of frozen soil phase change, and the accuracy of capturing the changing characteristics of unfrozen water content during the freezing process is improved, while the test workload is reasonably controlled. The multidimensional original data set is processed using Min-Max normalization and principal component analysis, and the numerical range is limited to the range of 0 to 1. It not only eliminates the influence between parameters of different dimensions, but also extracts the main features of the data through dimensionality reduction processing, reducing data redundancy. Based on the critical value of the thermal parameter, the normalized feature vector is classified and labeled in three stages: the unfrozen stage, the frozen stage, and the frozen stage, which accurately reflects the physical state characteristics of frozen soil at different temperature stages and provides a more effective data structure for the training of neural networks. Five-fold cross-validation training was performed through a deep neural network, and a network structure design with three hidden layers was adopted, in which the number of neurons in each layer was 2 times, 1.5 times, and 1 times the input dimension, respectively. This not only ensured that the model had sufficient feature extraction capabilities, but also avoided the overfitting problem caused by the overly complex network structure. A linear weighted correction mechanism based on the soil mineral composition ratio coefficient and porosity characteristic parameters was introduced to effectively compensate for the prediction deviation caused by soil structure differences and improve the applicability of the model in different types of clay samples. Finally, outliers were identified through root mean square error analysis and the 3-times standard deviation principle to ensure the reliability and accuracy of the final prediction results.

[0037] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0038] (1) The initial moisture content is set to 15%-35%, and the density is set to 1.5-1.8 g / cm³. The clay sample is processed by controlling the density and moisture content to obtain the clay sample to be tested. The basic physical parameter data are obtained through liquid limit, plastic limit, and particle grading tests.

[0039] (2) The unfrozen water content data of the clay sample to be tested in the range of 0°C to -5°C is tested by a temperature-controlled centrifuge, and the unfrozen water content data of the clay sample to be tested in the range of -5°C to -20°C is tested by a temperature-controlled pressure plate instrument;

[0040] (3) The thermal conductivity data of the clay sample to be tested in the temperature range of 0℃ to -20℃ are tested by the improved hot wire method, and the specific heat data of the clay sample to be tested in the temperature range of 0℃ to -20℃ are tested by the hybrid calorimetry method;

[0041] (4) Repeat the test data of unfrozen water content, thermal conductivity and specific heat at each temperature point three times to obtain the average value data, and then combine and organize the average value data with the basic physical parameter data to obtain a multidimensional original data set;

[0042] (5) The Mahalanobis distance method is used to identify abnormal data points in the multidimensional original data set, and the data set after removing the outliers is obtained. The data set after removing the outliers is smoothed by the interpolation method to obtain a continuously distributed temperature-thermal parameter data set;

[0043] (6) Perform normalization processing on each parameter in the continuously distributed temperature-thermal parameter data set to obtain a multidimensional original data set.

[0044] Specifically, the basic physical parameters of the clay samples were controlled and tested. The clay samples were modulated by artificial mixing and vibration table compaction to control their initial water content between 15% and 35%. This range is determined based on the basic properties of clay. Too low water content will lead to unstable soil structure, and too high water content will affect the freezing effect. At the same time, the sample density was controlled within the range of 1.5-1.8g / cm³, which is consistent with the common clay density in engineering practice. The modulated clay samples to be tested were tested for liquid limit, plastic limit and particle grading. The liquid limit test uses the multi-point method to determine the water content of the soil sample by the 82g cone drop method. The plastic limit uses the rubbing method to determine the water content of the soil sample when it is dried. The particle grading test uses screening and sedimentation analysis to obtain the weight percentage of soil particles of different particle sizes. For the unfrozen water content test in the temperature range of 0℃ to -5℃, a temperature-controlled centrifuge is used for measurement. The temperature-controlled centrifuge separates free water and bound water through a centrifugal field, and is equipped with a precision temperature control system with a temperature control accuracy of ±0.1°C. During the test, the clay sample was placed in a special centrifuge tube, and the sample temperature was gradually reduced from 0°C to -5°C through program control. Centrifugal separation was performed every 0.5°C drop in temperature, with a centrifugal force of 2000g and a time of 30 minutes. The separated moisture content was measured by weighing to obtain the unfrozen water content at that temperature point. For the temperature range of -5°C to -20°C, a temperature-controlled pressure plate instrument was used for testing. The temperature-controlled pressure plate instrument uses the principle of a semi-permeable membrane to control the water potential of the soil by applying different air pressures, and cooperates with the temperature control system to determine the unfrozen water content under low temperature conditions. During the test, sampling was carried out at intervals of 2°C, and each temperature point was measured after stabilization for 2 hours.

[0045] The thermal conductivity test adopts the improved hot wire method, which determines the thermal conductivity by measuring the relationship between the hot wire power input and the temperature response. The test device includes a heating wire, a temperature sensor and a data acquisition system. The heating wire adopts a nickel-chromium wire with a diameter of 0.2 mm, and provides stable heat through a constant current source. The temperature sensor adopts a PT100 platinum resistor with a temperature measurement accuracy of 0.01°C. The heating wire is buried in the center of the clay sample, and a constant heat power is applied at different temperature points. The temperature change curve with time is recorded, and the thermal conductivity is calculated according to Fourier's thermal conductivity law. The specific heat test adopts the mixed calorimetry method. The soil sample with a known temperature is put into a calorimeter with different temperatures, and the specific heat value is calculated by measuring the equilibrium temperature after mixing. The calorimeter uses a 500ml copper Dewar flask, and the thermometer accuracy is 0.01°C. To ensure the reliability of the data, the unfrozen water content, thermal conductivity and specific heat at each temperature point are tested three times. Taking the unfrozen water content test as an example, the centrifugal separation or pressure plate test is repeated three times at each temperature point, and the arithmetic mean of the three test results is taken as the representative value of the temperature point. All test data and basic physical parameters (liquid limit, plastic limit, particle size distribution, etc.) are organized into a data table in a standard format to form a multidimensional original data set.

[0046] The Mahalanobis distance method is used to identify outliers in the original data set. The Mahalanobis distance takes into account the correlation between the dimensions of the data and identifies outliers by calculating the standardized distance from the sample point to the data center. The Mahalanobis distance of each data point is calculated, and the data points whose distance value exceeds the threshold (usually the 95% quantile of the chi-square distribution) are marked as outliers. The cubic spline interpolation method is used to smooth the data set after removing the outliers. This method fits the data points by a piecewise cubic polynomial function to ensure the continuity and smoothness of the interpolation curve.

[0047] Take the actual test as an example: after modulation, the moisture content of a clay sample is 28%, the density is 1.65g / cm³, the liquid limit is 42%, the plastic limit is 21%, and the content of particles with a particle size less than 0.075mm is 85%. At 0℃, the unfrozen water content was measured to be 28.0%, 27.8%, and 28.2%, respectively, with an average value of 28.0%; the thermal conductivity test values ​​were 1.42, 1.45, and 1.43 W / (m·K), with an average value of 1.43 W / (m·K); the specific heat test values ​​were 1680, 1695, and 1688 J / (kg·K), with an average value of 1688 J / (kg·K). Through Mahalanobis distance analysis, a set of abnormal data was found at the temperature point of -3℃: the unfrozen water content was 22.5% (the other two test values ​​were 19.8% and 20.1%), and this data point was eliminated. The unfrozen water content at -3°C was corrected by cubic spline interpolation, and the interpolation result was 20.0%, and finally a continuously distributed temperature-thermal parameter data set was obtained. All parameters were normalized so that the data distribution was within a reasonable numerical range, which was convenient for subsequent model training.

[0048] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0049] (1) Perform zero mean processing on the temperature value, thermal conductivity, specific heat, and unfrozen water content data in the multidimensional original data set to obtain standardized data with a mean of 0, and then perform variance normalization on the standardized data to obtain normalized data with a standard deviation of 1;

[0050] (2) Normalize the data through feature correlation analysis to obtain the feature correlation coefficient matrix, and decompose the feature correlation coefficient matrix through eigenvalue decomposition to obtain eigenvalues ​​and eigenvectors;

[0051] (3) Arrange the eigenvalues ​​in descending order and calculate the cumulative contribution rate to determine the number of principal components. Then, sort the eigenvectors according to the corresponding eigenvalues ​​to obtain the principal component loading matrix.

[0052] (4) Perform matrix multiplication on the normalized data and the principal component loading matrix to obtain the principal component score matrix, and use the Min-Max normalization method to process the principal component score matrix to obtain the initial eigenvector with a value range of 0 to 1;

[0053] (5) The initial eigenvectors are weighted and combined using the variance contribution rate to obtain a comprehensive characteristic index, and the comprehensive characteristic index is tested through data distribution consistency to obtain a calibrated characteristic index;

[0054] (6) Use data dimension reduction mapping to transform the coordinates of the calibrated feature indicators to obtain a normalized feature vector.

[0055] Specifically, the multidimensional original data set is zero-mean processed, and the parameters such as temperature, thermal conductivity, specific heat and unfrozen water content are subtracted from their respective means so that the data center point is located at the origin. For each parameter, its sample mean is calculated and the original data is subtracted from the mean to obtain standardized data with a mean of 0. Subsequently, variance normalization is performed, and the standardized data is divided by its standard deviation so that the processed data has a unit standard deviation, so that parameters of different dimensions are comparable. In order to analyze the correlation between the parameters, it is necessary to calculate the characteristic correlation coefficient matrix. By calculating the Pearson correlation coefficient between each two parameters, a symmetric correlation coefficient matrix is ​​obtained. The correlation coefficient range is between [-1,1], and the larger the absolute value, the stronger the correlation. For n parameters, the correlation coefficient matrix is ​​an n×n symmetric matrix. The matrix is ​​subjected to eigenvalue decomposition to obtain n eigenvalues ​​and corresponding eigenvectors.

[0056] In the result of eigenvalue decomposition, the eigenvalues ​​are arranged in order from large to small, and the cumulative contribution rate is calculated. The calculation formula for the cumulative contribution rate is:

[0057]

[0058] in, represents the cumulative contribution rate, represents the i-th eigenvalue, m represents the number of selected eigenvalues, and n represents the total number of eigenvalues. represents the sum of the first m eigenvalues, Represents the sum of all eigenvalues. When the cumulative contribution rate reaches the preset threshold (usually 85% or 90%), the corresponding number of eigenvalues ​​is the number of principal components. Based on the determined number of principal components, the corresponding eigenvectors are sorted by eigenvalue size to form a principal component loading matrix. The normalized raw data is multiplied by the principal component loading matrix to obtain the principal component score matrix. Each column of the matrix represents a principal component, and each row represents the score of a sample on each principal component. The principal component score matrix is ​​subjected to Min-Max normalization, and the data is mapped to the [0,1] interval to obtain the initial eigenvector.

[0059] The initial feature vector is weighted by using the variance contribution rate as the weight. The principal component with a larger variance contribution rate has a larger weight. The comprehensive feature index obtained by weighted combination needs to be tested for data distribution consistency. The KS test (Kolmogorov-Smirnov test) method is used to verify whether the data obeys the normal distribution to ensure that the statistical characteristics of the feature index meet the requirements. Finally, the calibrated feature index is mapped for data dimensionality reduction, and the t-SNE (t-distributed Stochastic Neighbor Embedding) algorithm is used for nonlinear dimensionality reduction to obtain a normalized feature vector.

[0060] For example, a set of clay sample test data contains measurement values ​​at 19 temperature points, including four parameters: temperature (T), thermal conductivity (λ), specific heat (c), and unfrozen water content (w). The original data is zero-mean processed. Taking thermal conductivity as an example, the original value range is 1.42-1.78 W / (m·K), and the mean is 1.60 W / (m·K). The processed data range becomes -0.18 to 0.18. After variance normalization, the standard deviation of all parameters is 1. The characteristic correlation coefficient matrix is ​​calculated to obtain a 4×4 symmetric matrix, in which the correlation coefficient between temperature and unfrozen water content is 0.92, indicating that the two have a strong positive correlation. Four eigenvalues ​​are obtained by eigenvalue decomposition: 2.85, 0.82, 0.21, and 0.12. Calculate the cumulative contribution rate: the contribution rate of the first principal component is 71.25%, and the cumulative contribution rate of the first two principal components is 91.75%, which exceeds the threshold of 90%, so the first two principal components are selected. According to the number of principal components selected, a 2×4 principal component loading matrix was constructed and multiplied with the 19×4 normalized data matrix to obtain a 19×2 principal component score matrix. The score matrix was Min-Max standardized so that the data was distributed in the interval [0,1]. The variance contribution rates of the two principal components were 0.7125 and 0.2050, respectively, and a weighted combination was performed to obtain the comprehensive feature index. The normality of the comprehensive feature index was verified by the KS test, and finally the t-SNE algorithm was used to map the high-dimensional features to the low-dimensional space.

[0061] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0062] (1) The zero-crossing point analysis is used to normalize the feature vector to obtain the jump position data of the temperature-thermal parameter curve, and the critical value of the thermal parameter of the jump position data is determined by cluster analysis to obtain the boundary temperature data of the freezing stage;

[0063] (2) Perform piecewise linear fitting on the normalized feature vector to obtain the slope change data in each temperature range, and analyze the corresponding relationship between the slope change data and the freezing stage boundary temperature data to determine the stage feature discrimination criterion;

[0064] (3) According to the stage feature discrimination criteria, the normalized feature vectors are grouped to obtain the unfrozen stage data subset, the frozen stage data subset, and the frozen stage data subset, and the three data subsets are assigned category identifiers to establish a labeled classification data group;

[0065] (4) Analyze the labeled classification data set through Fisher discriminant analysis to obtain the discrimination scores of each stage, and perform fuzzy classification processing on the data points with discrimination scores below the threshold to generate stage classification correction data;

[0066] (5) Process the stage classification correction data based on the Sigmoid function, obtain the stage boundary transition data, and analyze the stage boundary transition data through probability distribution to calculate the membership value of each stage;

[0067] (6) Concatenate the membership value and the normalized feature vector to construct an augmented feature matrix, and enhance the features of the augmented feature matrix through covariance analysis to generate a training data matrix with stage attributes.

[0068] Specifically, the zero crossing point analysis is first performed. By identifying the sign change point of the first-order derivative of the temperature-thermal parameter curve, the position where the slope of the curve changes significantly is determined. The zero crossing point corresponds to the extreme point of the thermal parameter change rate, reflecting the transformation of the physical state of frozen soil. The identified jump position data is analyzed using the K-means clustering algorithm, and the temperature points are clustered into several categories. The cluster center is the critical value of the thermal parameter, thereby obtaining the boundary temperature data of the freezing stage. The piecewise linear fitting method is used for the normalized feature vector, and the temperature interval is divided into multiple sub-intervals. The least squares method is used for linear fitting in each sub-interval to obtain the slope data of each interval. For each sub-interval, the slope of the fitting line is calculated, which reflects the rate at which the thermal parameter changes with temperature in the interval. The obtained slope change data is analyzed for correspondence with the aforementioned freezing stage boundary temperature data to construct the stage feature discrimination criterion. The discrimination criterion is based on the absolute value and change trend of the slope, and is used to determine the freezing stage to which the data point belongs.

[0069] According to the established stage feature discrimination criteria, the normalized feature vectors were grouped. The unfrozen stage was characterized by small changes in thermal parameters and a slope close to zero; the frozen stage was characterized by rapid changes in thermal parameters and a large absolute slope; the frozen stage was characterized by stable thermal parameters and a smaller slope. The three grouped data subsets were assigned category labels: the unfrozen stage was marked as 1, the frozen stage was marked as 2, and the frozen stage was marked as 3, and the labeled classification data groups were obtained. Fisher discriminant analysis was performed on the labeled classification data groups. This method evaluates the classification effect by maximizing the ratio of the between-group variance to the within-group variance. The Fisher discriminant coefficients between the stages were calculated to obtain the discrimination scores. For data points with a discrimination score lower than the preset threshold, it indicates that there is uncertainty in their stage attribution and fuzzy classification processing is required. Fuzzy classification uses the distance weight method to determine the attribution probability of the data point according to the distance from the data point to the center of each category, and obtains the stage classification correction data.

[0070] The Sigmoid function is used to perform smooth transition processing on the stage classification correction data, making the transition between stages more continuous. The calculation formula of the membership value is:

[0071]

[0072] in, represents the membership of the i-th data point to the j-th stage on the k-th feature dimension, represents the distance from data point i to the center of stage j, represents the threshold parameter of the kth feature dimension, and To adjust the parameters, represents the feature vector of data point i, represents the center vector of the jth stage, l The target stage is labeled. The probability distribution analysis is performed on the processed stage boundary transition data to verify the continuity and smoothness of the transformation process.

[0073] The calculated membership values ​​and the normalized feature vectors are concatenated to obtain an augmented feature matrix. Covariance analysis is performed on the augmented feature matrix to calculate the correlation between the features. The expressiveness of the data is enhanced by combining the features, and finally a training data matrix with stage attributes is obtained.

[0074] Taking a set of frozen soil sample data as an example: the normalized feature vector was analyzed for zero crossing points, and two significant jump points were found near -0.5℃ and -4.5℃. The two temperature points were determined as the boundary temperatures of the freezing stage by K-means clustering. The piecewise linear fitting results showed that the slope of the interval from 0℃ to -0.5℃ was 0.05, the slope of the interval from -0.5℃ to -4.5℃ was 0.42, and the slope of the interval from -4.5℃ to -20℃ was 0.08. Based on this, the discrimination criteria were established. Among the discrimination scores obtained by Fisher discriminant analysis, the scores in the two transition intervals from -0.3℃ to -0.7℃ and -4.3℃ to -4.7℃ were low, and fuzzy classification processing was required. After Sigmoid function processing, the data point at -0.5℃ had a membership of 0.45 to the unfrozen stage and a membership of 0.55 to the frozen stage, reflecting the gradual characteristics of state transition. The training data matrix finally obtained contains the original features and stage attribute information.

[0075] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0076] (1) Divide the training data matrix into a training set and a validation set in a ratio of 8:2 to obtain the initial training data subset, and then divide the initial training data subset into five equal parts through stratified sampling to obtain five groups of cross-validation data sets;

[0077] (2) Five groups of cross-validation data sets are processed by shuffling the data to obtain randomly ordered training sample sequences, and the training sample sequences are divided into batches through a mini-batch generator to obtain batch training data blocks;

[0078] (3) In the first hidden layer of the deep neural network, the number of neurons is set to twice the input dimension, an expanded-dimensional feature extraction layer is established, and the output of the first hidden layer is nonlinearly transformed through the ReLU activation function to obtain the initial feature activation value;

[0079] (4) In the second hidden layer, the number of neurons is set to 1.5 times the input dimension, a feature compression layer is established, and the output of the second hidden layer is feature mapped using a hyperbolic tangent function to obtain the feature representation of the intermediate layer;

[0080] (5) In the third hidden layer, the number of neurons is set to 1 times the input dimension, a feature reconstruction layer is established, and the output of the third hidden layer is numerically compressed using the Sigmoid function to obtain normalized feature data;

[0081] (6) Calculate the mean square error between the normalized feature data and the true value of the unfrozen water content to obtain the loss function value, and then perform gradient backpropagation on the loss function value through the Adam optimizer to calculate the network parameter update amount;

[0082] (7) The network parameter update amount is weighted and calculated with the original network parameters to generate a candidate network parameter group. The performance of the candidate network parameter group is evaluated through a validation set, and a prediction model for unfrozen water content is obtained, which contains three hidden layers and the number of neurons in each layer is 2 times, 1.5 times, and 1 times the input dimension, respectively.

[0083] Specifically, the training data matrix is ​​first divided reasonably, and the data set is divided into a training set and a validation set in a ratio of 8:2. This ratio ensures that there is enough data for model training, while leaving an appropriate amount of data for validation. In the division process, in order to maintain the consistency of data distribution, the initial training data subset is divided into five equal-sized subsets using a stratified sampling method as a cross-validation data set. Stratified sampling ensures that the proportion of samples in the three stages of unfrozen, frozen, and frozen in each subset is consistent with the original data set. The five groups of cross-validation data sets are shuffled to eliminate the order correlation between samples and improve the generalization ability of the model. The Fisher-Yates shuffle algorithm is used for data shuffling. The algorithm traverses the array from back to front and exchanges the current element with the element at the previous random position to ensure that the probability of each permutation is equal. The shuffled training sample sequence is divided into small batches, and the batch size is set to 32 samples. This size strikes a balance between computational efficiency and training stability. The small batch generator intercepts samples in a fixed size order, and the last batch with less than 32 samples is still retained to ensure that all samples participate in training.

[0084] The first hidden layer of the neural network is used as a dimension-expanding feature extraction layer, and the number of neurons is set to twice the input dimension. For 4-dimensional input data including temperature, thermal conductivity, specific heat, and unfrozen water content, the first hidden layer is set to 8 neurons. This layer uses the ReLU activation function for nonlinear transformation. The form of the ReLU function is f(x)=max(0,x). When the input is positive, it directly outputs, and when it is negative, it outputs zero. This simple nonlinear transformation not only speeds up the training speed, but also avoids the gradient vanishing problem. The second hidden layer is used as a feature compression layer, and the number of neurons is 1.5 times the input dimension, that is, 6 neurons. This layer uses the hyperbolic tangent function (tanh) as the activation function. The tanh function maps the input to the (-1,1) interval. Its output is centered at zero and is symmetrical, which is helpful for feature extraction and compression. Compared with the ReLU function, the hyperbolic tangent function has a larger gradient near the zero point, making the network more sensitive to learning in this area.

[0085] The third hidden layer is used as a feature reconstruction layer. The number of neurons is the same as the input dimension, and 4 neurons are set. This layer uses the Sigmoid activation function for numerical compression. The Sigmoid function maps the input to the (0,1) interval, which is consistent with the physical meaning of the unfrozen water content because the unfrozen water content is a percentage value. Through the sequential feature transformation of the three hidden layers, the network can effectively learn the inherent laws of the data.

[0086] The mean square error between the normalized feature data and the true value of the unfrozen water content was calculated as the loss function of the network. The Adam optimizer was used for parameter optimization. The Adam optimizer combined the momentum method and the adaptive learning rate. The initial learning rate was set to 0.001, β1=0.9, and β2=0.999. In each training batch, the gradient of the loss function to the parameters of each layer was calculated by the back propagation algorithm, and the parameters were adjusted according to the update rule of the Adam optimizer. The network parameter update amount was weighted and calculated with the original network parameters. The weight coefficient used an exponential moving average. The weight of the historical parameters was 0.9, and the weight of the update amount was 0.1. This smooth update method improved the stability of the training. The performance of each set of candidate network parameters was evaluated on the validation set, and the parameter group with the smallest validation error was selected as the final model parameter.

[0087] Take a set of experimental data as an example: the training data matrix contains 200 samples, which are divided into 160 training samples and 40 validation samples in a ratio of 8:2. Five groups of cross-validation data sets are obtained by stratified sampling, each containing 32 samples. After the output of the 8 neurons in the first hidden layer is activated by the ReLU function, about 40% of the neurons are activated. The output values ​​of the tanh function of the 6 neurons in the second hidden layer are distributed in the range of (-0.8, 0.8). The output of the 4 neurons in the third hidden layer after compression by the Sigmoid function is highly consistent with the actual unfrozen water content. The mean square error of the model on the training set gradually decreases from the initial 0.152 to 0.008, and the error on the validation set is 0.012, indicating that the model has good generalization ability. The average error of the five-fold cross validation is 0.010, and the standard deviation is 0.002, indicating that the model performance is stable.

[0088] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0089] (1) Calculate the deviation between the output result of the unfrozen water content prediction model and the original unfrozen water content test value to obtain the initial error data, and analyze the initial error data through statistical distribution to obtain the error distribution characteristic value;

[0090] (2) Analyze the mineral composition of the soil by X-ray diffraction to obtain the mineral content percentage data, and then weight the mineral content percentage data to obtain the mineral composition ratio coefficient;

[0091] (3) Using mercury intrusion porosity testing on soil samples to obtain pore size distribution data, and then performing density correction on the pore size distribution data to obtain porosity characteristic parameters;

[0092] (4) Linearly combine the mineral composition ratio coefficient and the porosity characteristic parameter to construct the soil property correction factor, and then process the soil property correction factor through characteristic normalization to obtain the standardized correction coefficient;

[0093] (5) Perform correlation analysis on the standardized correction coefficient and the characteristic value of the error distribution to obtain the weight adjustment coefficient, and obtain the correction function expression by parameter fitting the weight adjustment coefficient;

[0094] (6) The output result of the unfrozen water content prediction model is calculated through the correction function expression to obtain the unfrozen water content value after correction coefficient compensation.

[0095] Specifically, the output results of the prediction model are first compared with the original unfrozen water content test value, and the deviation value of each test point is calculated to obtain the initial error data. The statistical distribution analysis of these error data includes the calculation of statistical characteristics such as mean error, standard deviation, skewness and kurtosis, and the analysis of the distribution law of errors. The standard deviation reflects the degree of dispersion of the predicted value, the skewness characterizes the asymmetry of the error distribution, and the kurtosis reflects the peak or flat characteristics of the error distribution. The analysis of the soil mineral composition adopts X-ray diffraction technology (XRD), and the lattice structure and mineral composition of crystalline materials are determined by the Bragg diffraction law. During the test, the soil sample is ground to less than 200 mesh, and Cu target X-ray is used, the tube voltage is 40kV, the tube current is 30mA, the scanning step is 0.02°, and the time for each step is 0.5s. The content percentage of each mineral is obtained by analyzing the peak position and peak intensity of the diffraction spectrum. According to the influence of minerals on the unfrozen water content, weight coefficients are set for different minerals. Clay minerals such as montmorillonite and illite have higher weights, while inert minerals such as quartz have lower weights. The mineral composition ratio coefficient is thus obtained.

[0096] The pore characteristics of soil samples were tested by mercury intrusion method, which is based on the capillary principle and allows mercury to enter pores of different sizes by gradually increasing pressure. A mercury intrusion instrument with a pore size test range of 3nm-100μm was used for the test. The sample pretreatment temperature was 105℃ and the equilibrium time was 30s. The volume of mercury injected under different pressures was recorded, and the pore size distribution was calculated by the Washburn equation. Considering that the actual density of the soil may be different from that of the test sample, it is necessary to correct the pore distribution data according to the original soil density to obtain accurate porosity characteristic parameters. The mineral composition ratio coefficient and the porosity characteristic parameter were linearly combined, and the influence mechanism of the two on the unfrozen water content was considered during the combination. The mineral composition mainly affects the amount of adsorbed water, while the pore characteristics affect the existence form of capillary water. The comprehensive soil property correction factor was obtained by weighted summation, and then the correction factor was normalized so that its numerical distribution was within a reasonable range to obtain the standardized correction factor.

[0097] The correlation analysis of the standardized correction coefficient and the previously obtained error distribution characteristic value is carried out, and the correlation coefficient matrix is ​​calculated to determine the correlation strength between the parameters. According to the results of the correlation analysis, the mapping relationship between the correction coefficient and the error is established to obtain the weight adjustment coefficient. Through methods such as polynomial fitting or piecewise function fitting, a specific correction function expression is established, which will be used as a compensation term for the prediction results. Finally, the original output result of the prediction model is substituted into the correction function expression for calculation to obtain the unfrozen water content value after soil property compensation. The correction process takes into account the influence of the physical properties of the soil on the unfrozen water content, which improves the accuracy of the prediction results.

[0098] Take a clay sample as an example: the mineral composition obtained by XRD analysis is 35% montmorillonite, 25% illite, 15% kaolinite, 20% quartz, and 5% feldspar. According to the influence of minerals on the unfrozen water content, weight coefficients of 1.5, 1.2, 1.0, 0.5, and 0.5 are assigned respectively, and the calculated mineral composition ratio coefficient is 1.08. The mercury injection test shows that the total porosity of the soil sample is 42%, of which micropores (<0.1μm) account for 35%, mesopores (0.1-1μm) account for 45%, and macropores (>1μm) account for 20%. Considering that the density of the original soil is 1.65g / cm³ and the density of the test sample is 1.55g / cm³, the pore distribution data is density-corrected, and the corrected porosity characteristic parameter is 0.92. The mineral composition ratio coefficient and porosity characteristic parameter are linearly combined in a ratio of 6:4, and the soil property correction factor is 1.016. After normalization, the standardized correction coefficient is 0.85. The original prediction result is 18.5% of unfrozen water content at -2℃, which is 1.7% different from the measured value of 20.2%. The compensation value calculated by the correction function expression is 1.5%. The final corrected prediction result is 20.0%, and the prediction accuracy is significantly improved.

[0099] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0100] (1) Perform a difference operation on the unfrozen water content value after correction coefficient compensation and the measured value to obtain a prediction error sequence, and then perform a square sum calculation on the prediction error sequence to obtain the error sum of squares;

[0101] (2) Taking the square root of the error sum to obtain the root mean square error value, and normalizing the root mean square error value to obtain the standardized error index;

[0102] (3) Perform statistical distribution analysis on the unfrozen water content after correction coefficient compensation to obtain mean and standard deviation data, and amplify the standard deviation data by 3 times to obtain the outlier judgment threshold;

[0103] (4) Calculate the difference between the unfrozen water content value after correction coefficient compensation and the mean value to obtain the deviation data sequence, and compare and screen the deviation data sequence with the outlier judgment threshold to obtain the outlier marking result;

[0104] (5) Eliminate the abnormal data points in the outlier marking results to obtain the data set after eliminating the outliers, and process the data set after eliminating the outliers by data interpolation to obtain a continuous distribution sequence;

[0105] (6) The continuous distribution sequence is processed through Min-Max normalization to obtain the standardized prediction results.

[0106] Specifically, the prediction error sequence is obtained by performing a difference operation between the unfrozen water content value after correction coefficient compensation and the actual measured value of the unfrozen water content. For each temperature point, the difference between the predicted value and the measured value is calculated and recorded in the prediction error sequence. Then, each error value in the prediction error sequence is squared and summed to obtain the error sum of squares. In practical applications, taking the temperature point of -2°C as an example, the unfrozen water content value of a clay sample after correction coefficient compensation predicted by the neural network model is 26.8%, while the measured value is 27.2%. The prediction error of this temperature point is -0.4%. The prediction errors of all temperature points are similarly calculated and squared and summed. Assuming that there are 40 temperature measurement points, the final error sum of squares is 8.92.

[0107] Next, the calculated sum of squared errors is squared to obtain the root mean square error value reflecting the prediction accuracy. In the above example, the square root of the sum of squared errors 8.92 is approximately 2.987, indicating that the root mean square error of the prediction result is 2.987%. In order to facilitate the comparative analysis of prediction results under different conditions, the root mean square error value needs to be normalized to obtain a standardized error index. The normalization process uses the maximum and minimum normalization method, that is, the root mean square error value is mapped to the interval of 0-1 to obtain a dimensionless standardized error index. When performing statistical distribution analysis on the unfrozen water content values ​​after correction coefficient compensation, the arithmetic mean of all predicted values ​​is first calculated as the mean feature of the data set. At the same time, the degree of dispersion of the data is calculated to obtain the standard deviation data. In actual operation, for the above prediction results containing 40 temperature measurement points, the average value of the unfrozen water content prediction value is calculated to be 23.5%, and the standard deviation is 2.1%. According to the 3σ principle, 3 times the standard deviation is used as the threshold for outlier judgment, that is, in this example, the threshold value is 6.3%.

[0108] Subsequently, the difference between the unfrozen water content value after correction coefficient compensation and the calculated mean is calculated to obtain the degree of deviation of each data point relative to the mean, and the deviation data sequence is obtained. Each value in the deviation data sequence is compared with the previously determined outlier judgment threshold. When the absolute value of the deviation of a data point exceeds the threshold, it is marked as an outlier. For example, the predicted unfrozen water content at the -1°C temperature point is 32.4%, and the difference with the mean of 23.5% is 8.9%, which exceeds the threshold of 6.3%, so the point is marked as an outlier. For all marked abnormal data points, the data removal method is used to temporarily remove them from the data set to obtain a data set after the outliers are removed. In order to maintain the continuity and integrity of the data, the interpolation method is used to interpolate the positions of the removed abnormal data points. Specifically, the cubic spline interpolation method is used to fit a smooth curve based on the normal data points on both sides of the abnormal point, so as to obtain the correction value of the abnormal point position. In the example of the -1°C temperature point, the correction value obtained by interpolation calculation is 25.8%, which is more coordinated with the overall data distribution. By processing all outliers in this way, we can finally get a continuously distributed data sequence.

[0109] The obtained continuous distribution sequence is standardized, and the Min-Max standardization method is used to map the data to the range of 0-1. The purpose of this processing is to eliminate the influence of dimension and facilitate the comparison and analysis of prediction results under different conditions. The standardization process uses the normalization formula, that is, the original value minus the minimum value and then divided by the difference between the maximum and minimum values. In this way, the final standardized prediction result is obtained.

[0110] For example, a clay sample was selected in a research project to predict the unfrozen water content in the temperature range of 0℃ to -20℃. First, the predicted values ​​of the unfrozen water content at 40 temperature points after correction coefficient compensation were obtained. The root mean square error calculated by comparing with the measured value was 2.987%, and the standardized error index was 0.348 after normalization. Statistical analysis showed that the average value of the predicted value was 23.5% and the standard deviation was 2.1%. Based on this, the threshold for judging the abnormal value was determined to be 6.3%. Through deviation analysis, it was found that the predicted values ​​at the three temperature points of -1℃, -1.5℃ and -2.5℃ may be abnormal, which were 32.4%, 31.2% and 15.6% respectively. Cubic spline interpolation was performed on these abnormal points, and the corrected values ​​were 25.8%, 24.9% and 22.1% respectively. Finally, the entire data sequence was Min-Max standardized to obtain the standardized prediction results in the range of 0-1.

[0111] The above describes the method for constructing a frozen soil unfrozen water content detection model based on a neural network in the embodiment of the present application. The following describes the system for constructing a frozen soil unfrozen water content detection model based on a neural network in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of a system for constructing a model for detecting unfrozen water content in frozen soil based on a neural network includes:

[0112] A test module is used to test clay samples in a temperature range of 0°C to -20°C by segmented sampling, wherein the sampling interval of the first temperature range of 0°C to -5°C is 0.5°C, and the sampling interval of the second temperature range of -5°C to -20°C is 2°C, to obtain a multidimensional original data set;

[0113] An analysis module, used for processing the multidimensional original data set by using Min-Max standardization and principal component analysis to obtain a normalized feature vector with a value range between 0 and 1;

[0114] A labeling module, used for labeling the normalized feature vector into three-stage classification of an unfrozen stage, a frozen stage, and a frozen stage based on a critical value of a thermal parameter, so as to obtain a training data matrix with stage attributes;

[0115] A training module is used to input the training data matrix into a deep neural network for five-fold cross-validation training to obtain an unfrozen water content prediction model with three hidden layers and the number of neurons in each layer being 2 times, 1.5 times and 1 times the input dimension respectively;

[0116] A correction module is used to calculate the output result of the unfrozen water content prediction model through linear weighted correction based on the soil mineral composition ratio coefficient and porosity characteristic parameters to obtain the unfrozen water content value after correction coefficient compensation;

[0117] The identification module is used to identify abnormal values ​​in the unfrozen water content values ​​after the correction coefficient compensation through root mean square error analysis and the principle of three times standard deviation, so as to obtain a standardized prediction result.

[0118] Through the cooperation of the above components, a segmented sampling test strategy is adopted in the temperature range of 0℃ to -20℃, a small interval sampling of 0.5℃ is used in the range of 0℃ to -5℃, and a large interval sampling of 2℃ is used in the range of -5℃ to -20℃, which ensures that more dense data points are obtained in the key temperature range of frozen soil phase change, improves the accuracy of capturing the change characteristics of unfrozen water content during freezing, and reasonably controls the test workload. The multidimensional original data set is processed by Min-Max normalization and principal component analysis, and the numerical range is limited to the range of 0 to 1. It not only eliminates the influence of parameters of different dimensions, but also extracts the main features of the data through dimensionality reduction processing, reducing data redundancy. Based on the critical value of thermal parameters, the normalized feature vector is classified and labeled into three stages: unfrozen stage, frozen stage and frozen stage, which accurately reflects the physical state characteristics of frozen soil at different temperature stages and provides a more effective data structure for neural network training. Five-fold cross-validation training was performed through a deep neural network, and a network structure design with three hidden layers was adopted, in which the number of neurons in each layer was 2 times, 1.5 times, and 1 times the input dimension, respectively. This not only ensured that the model had sufficient feature extraction capabilities, but also avoided the overfitting problem caused by the overly complex network structure. A linear weighted correction mechanism based on the soil mineral composition ratio coefficient and porosity characteristic parameters was introduced to effectively compensate for the prediction deviation caused by soil structure differences and improve the applicability of the model in different types of clay samples. Finally, outliers were identified through root mean square error analysis and the 3-times standard deviation principle to ensure the reliability and accuracy of the final prediction results.

[0119] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for constructing a frozen soil unfrozen water content detection model based on a neural network, characterized in that: The method for constructing a frozen soil unfrozen water content detection model based on a neural network includes: In the temperature range of 0°C to -20°C, clay samples were tested by segmented sampling, wherein the sampling interval of the first temperature range of 0°C to -5°C was 0.5°C, and the sampling interval of the second temperature range of -5°C to -20°C was 2°C, to obtain a multidimensional original data set; The multidimensional original data set is processed by Min-Max standardization and principal component analysis to obtain a normalized feature vector with a value range between 0 and 1; Based on the critical value of the thermal parameter, the normalized feature vector is marked as a three-stage classification of an unfrozen stage, a frozen stage, and a frozen stage, so as to obtain a training data matrix with stage attributes; The training data matrix is ​​input into a deep neural network for five-fold cross validation training to obtain an unfrozen water content prediction model having three hidden layers and the number of neurons in each layer is 2 times, 1.5 times and 1 times the input dimension respectively; Based on the soil mineral composition ratio coefficient and porosity characteristic parameters, the output result of the unfrozen water content prediction model is calculated by linear weighted correction to obtain the unfrozen water content value after correction coefficient compensation; The abnormal values ​​in the unfrozen water content values ​​after the correction coefficient compensation are identified by root mean square error analysis and the three-times standard deviation principle, and the standardized prediction results are obtained.

2. The method for constructing a frozen soil unfrozen water content detection model based on a neural network according to claim 1, characterized in that: The clay sample is tested by segmented sampling in the temperature range of 0°C to -20°C, wherein the sampling interval of the first temperature range of 0°C to -5°C is 0.5°C, and the sampling interval of the second temperature range of -5°C to -20°C is 2°C, to obtain a multidimensional original data set, including: The initial moisture content is set to 15%-35%, and the density is set to 1.5-1.8g / cm³. The clay sample is processed by controlling the density and moisture content to obtain the clay sample to be tested. The basic physical parameter data are obtained through liquid limit, plastic limit, and particle grading tests. The unfrozen water content data of the clay sample to be tested in the range of 0°C to -5°C is tested by a temperature-controlled centrifuge, and the unfrozen water content data of the clay sample to be tested in the range of -5°C to -20°C is tested by a temperature-controlled pressure plate instrument; The thermal conductivity data of the clay sample to be tested in the temperature range of 0°C to -20°C are tested by using an improved hot wire method, and the specific heat data of the clay sample to be tested in the temperature range of 0°C to -20°C are tested by using a mixed calorimetry method; The unfrozen water content test data, thermal conductivity data and specific heat data at each temperature point are tested three times repeatedly to obtain average data, and the average data is combined and sorted with the basic physical parameter data to obtain a multidimensional original data set; The Mahalanobis distance method is used to identify abnormal data points in the multidimensional original data set, obtain a data set after removing the abnormal values, and smooth the data set after removing the abnormal values ​​by an interpolation method to obtain a temperature-thermal parameter data set with continuous distribution; A normalization process is performed on each parameter in the continuously distributed temperature-thermal parameter data set to obtain the multi-dimensional original data set.

3. The method for constructing a frozen soil unfrozen water content detection model based on a neural network according to claim 1, characterized in that: The method of processing the multidimensional original data set by Min-Max normalization and principal component analysis to obtain a normalized feature vector with a value range between 0 and 1 includes: Performing zero mean processing on the temperature value, thermal conductivity, specific heat, and unfrozen water content data in the multidimensional original data set to obtain standardized data with a mean of 0, and performing variance normalization processing on the standardized data to obtain normalized data with a standard deviation of 1; Analyzing the normalized data by feature correlation to obtain a feature correlation coefficient matrix, and decomposing the feature correlation coefficient matrix by eigenvalue to obtain eigenvalues ​​and eigenvectors; The eigenvalues ​​are arranged in descending order and the cumulative contribution rate is calculated to determine the number of principal components, and the eigenvectors are sorted according to the corresponding eigenvalues ​​to obtain a principal component loading matrix; Performing a matrix multiplication operation on the normalized data and the principal component loading matrix to obtain a principal component score matrix, and processing the principal component score matrix using a Min-Max normalization method to obtain an initial eigenvector with a value range between 0 and 1; The initial feature vectors are weightedly combined using variance contribution rate to obtain a comprehensive feature index, and the comprehensive feature index is tested by data distribution consistency to obtain a calibrated feature index; The calibrated feature index is subjected to coordinate transformation by using data dimension reduction mapping to obtain a normalized feature vector.

4. The method for constructing a frozen soil unfrozen water content detection model based on a neural network according to claim 1, characterized in that: Based on the thermal parameter critical value, the normalized feature vector is marked as a three-stage classification of an unfrozen stage, a frozen stage, and a frozen stage, and a training data matrix with stage attributes is obtained, including: The normalized feature vector is analyzed by zero crossing point to obtain the jump position data of the temperature-thermal parameter curve, and the thermal parameter critical value of the jump position data is determined by cluster analysis to obtain the freezing stage boundary temperature data; Performing piecewise linear fitting on the normalized feature vector to obtain slope change data in each temperature interval, and performing corresponding relationship analysis on the slope change data and the freezing stage boundary temperature data to determine the stage feature discrimination criterion; According to the stage feature discrimination criterion, the normalized feature vector is grouped to obtain an unfrozen stage data subset, a frozen stage data subset and a frozen stage data subset, and category identifiers are respectively assigned to the three data subsets to establish a labeled classification data group; Analyzing the labeled classification data set by Fisher discriminant analysis to obtain discrimination scores of each stage, and performing fuzzy classification processing on data points whose discrimination scores are lower than a threshold value to generate stage classification correction data; Processing the stage classification correction data based on the Sigmoid function, obtaining stage boundary transition data, and analyzing the stage boundary transition data through probability distribution to calculate the membership value of each stage; The membership value and the normalized feature vector are concatenated in a matrix to construct an augmented feature matrix, and the features of the augmented feature matrix are enhanced through covariance analysis to generate a training data matrix with stage attributes.

5. The method for constructing a frozen soil unfrozen water content detection model based on a neural network according to claim 1, characterized in that: The training data matrix is ​​input into a deep neural network for five-fold cross validation training, and an unfrozen water content prediction model is obtained, which has three hidden layers and the number of neurons in each layer is 2 times, 1.5 times and 1 times of the input dimension, respectively, including: The training data matrix is ​​divided into a training set and a validation set in a ratio of 8:2 to obtain an initial training data subset, and the initial training data subset is divided into five equal parts by stratified sampling to obtain five groups of cross-validation data sets; Processing the five groups of cross-validation data sets by data shuffling to obtain randomly ordered training sample sequences, and dividing the training sample sequences into batches by a mini-batch generator to obtain batch training data blocks; Setting the number of neurons in the first hidden layer of the deep neural network to be twice the input dimension, establishing a dimension-expanding feature extraction layer, and performing a nonlinear transformation on the output of the first hidden layer through a ReLU activation function to obtain an initial feature activation value; In the second hidden layer, the number of neurons is set to be 1.5 times of the input dimension, a feature compression layer is established, and the output of the second hidden layer is feature mapped by a hyperbolic tangent function to obtain an intermediate layer feature representation; In the third hidden layer, the number of neurons is set to be 1 times the input dimension, a feature reconstruction layer is established, and the output of the third hidden layer is numerically compressed by a Sigmoid function to obtain normalized feature data; The normalized characteristic data and the true value of the unfrozen water content are subjected to mean square error calculation to obtain a loss function value, and the loss function value is subjected to gradient backpropagation through an Adam optimizer to calculate the network parameter update amount; The network parameter update amount and the original network parameter are weighted and calculated to generate a candidate network parameter group, and the performance of the candidate network parameter group is evaluated through a validation set, thereby obtaining an unfrozen water content prediction model comprising three hidden layers, in which the number of neurons in each layer is 2 times, 1.5 times and 1 times the input dimension respectively.

6. The method for constructing a frozen soil unfrozen water content detection model based on a neural network according to claim 1, characterized in that: The output result of the unfrozen water content prediction model is calculated by linear weighted correction based on the soil mineral composition ratio coefficient and porosity characteristic parameters to obtain the unfrozen water content value after correction coefficient compensation, including: Calculate the deviation between the output result of the unfrozen water content prediction model and the original unfrozen water content test value to obtain initial error data, and analyze the initial error data through statistical distribution to obtain error distribution characteristic values; Analyze the mineral composition of the soil by X-ray diffraction to obtain mineral content percentage data, and distribute the mineral content percentage data by weight to obtain the mineral composition ratio coefficient; Testing the porosity of soil samples by mercury intrusion porosity method to obtain pore size distribution data, and performing density correction on the pore size distribution data to obtain porosity characteristic parameters; Linearly combining the mineral composition ratio coefficient and the porosity characteristic parameter to construct a soil property correction factor, and processing the soil property correction factor by characteristic normalization to obtain a standardized correction coefficient; Performing correlation analysis on the standardized correction coefficient and the error distribution characteristic value to obtain a weight adjustment coefficient, and obtaining a correction function expression by parameter fitting the weight adjustment coefficient; The output result of the unfrozen water content prediction model is calculated by the correction function expression to obtain the unfrozen water content value after correction coefficient compensation.

7. The method for constructing a frozen soil unfrozen water content detection model based on a neural network according to claim 1, characterized in that: The method of identifying abnormal values ​​in the unfrozen water content value after the correction coefficient compensation by using the root mean square error analysis and the three-times standard deviation principle to obtain a standardized prediction result includes: Performing a difference operation on the unfrozen water content value after the correction coefficient compensation and the measured value to obtain a prediction error sequence, and performing a square sum calculation on the prediction error sequence to obtain an error square sum; Performing a square root processing on the error sum of squares to obtain a root mean square error value, and normalizing the root mean square error value to obtain a standardized error index; Performing statistical distribution analysis on the unfrozen water content value after the correction coefficient compensation to obtain mean and standard deviation data, and amplifying the standard deviation data by 3 times to obtain an abnormal value judgment threshold; The difference between the unfrozen water content value after the correction coefficient compensation and the mean value is calculated to obtain a deviation data sequence, and the deviation data sequence is compared and screened with the abnormal value judgment threshold to obtain an abnormal value marking result; The abnormal data points in the abnormal value marking result are removed to obtain a data set after the abnormal value is removed, and the data set after the abnormal value is removed is processed by data interpolation to obtain a continuous distribution sequence; The continuous distribution sequence is processed by Min-Max normalization to obtain a standardized prediction result.

8. A system for constructing a frozen soil unfrozen water content detection model based on a neural network, used to implement the method for constructing a frozen soil unfrozen water content detection model based on a neural network as described in any one of claims 1 to 7, characterized in that: The neural network-based frozen soil unfrozen water content detection model construction system includes: A test module is used to test clay samples in a temperature range of 0°C to -20°C by segmented sampling, wherein the sampling interval of the first temperature range of 0°C to -5°C is 0.5°C, and the sampling interval of the second temperature range of -5°C to -20°C is 2°C, to obtain a multidimensional original data set; An analysis module, used for processing the multidimensional original data set by using Min-Max standardization and principal component analysis to obtain a normalized feature vector with a value range between 0 and 1; A labeling module, used for labeling the normalized feature vector into three-stage classification of an unfrozen stage, a frozen stage, and a frozen stage based on a critical value of a thermal parameter, so as to obtain a training data matrix with stage attributes; A training module is used to input the training data matrix into a deep neural network for five-fold cross-validation training to obtain an unfrozen water content prediction model with three hidden layers and the number of neurons in each layer being 2 times, 1.5 times and 1 times the input dimension respectively; A correction module is used to calculate the output result of the unfrozen water content prediction model through linear weighted correction based on the soil mineral composition ratio coefficient and porosity characteristic parameters to obtain the unfrozen water content value after correction coefficient compensation; The identification module is used to identify abnormal values ​​in the unfrozen water content values ​​after the correction coefficient compensation through root mean square error analysis and the principle of three times standard deviation, so as to obtain a standardized prediction result.

Images