Machine learning prediction method for permeability coefficient of geosynthetic clay liner under action of organic matters in vertical obstruction

Abstract

The invention discloses a machine learning prediction method for a permeability coefficient of a geosynthetic clay liner under the action of organic matters in vertical obstruction. The method comprises the following steps: S1, data collection; s2, data preprocessing and normalization are carried out; s3, algorithm selection; s4, optimizing parameters; and S5, model verification and evaluation. The method is based on influence factors such as bentonite characteristics (free expansion index and mass per unit area), prehydration conditions (effective confining pressure and pore liquid ion strength in backfill), organic matter pollution liquid chemical characteristics (concentration of water-soluble organic matter or dielectric constant of non-aqueous phase liquid organic matter), effective stress in a permeation process and the like. The permeability coefficient of the geosynthetic clay liner under the action of the organic matter can be effectively predicted through the machine learning prediction model, the accuracy of the prediction result is high, the time required for achieving chemical equilibrium in a traditional test can be greatly shortened, and manpower, material resources and financial resources are saved.

Description

Technical Field

[0001] The present invention relates to the field of environmental geotechnical engineering technology, and in particular to a machine learning prediction method for the permeability coefficient of a bentonite waterproof blanket under the action of organic matter in a vertical barrier. Background Art

[0002] The Yangtze River Delta region is home to numerous chemical parks. Due to leaks in processes such as material transportation, feeding, and sample preparation, as well as contamination from pipelines, valves, tanks, and other equipment, the soil and groundwater within these parks are extensively contaminated by organic matter. These organic pollutants migrate with groundwater, forming plumes of pollution, posing potential risks to human health and the ecological environment and necessitating urgent risk management. Furthermore, the Yangtze River Delta's coastal areas and shallow watersheds are complex, with groundwater levels fluctuating significantly with the seasons and experiencing alternating wet-dry cycles. The groundwater also contains high levels of salinity, and the complex connection between seawater and groundwater makes it corrosive, posing significant challenges to risk management.

[0003] In contaminated sites, vertical barriers are often used to block or adsorb pollutants to prevent them from spreading further to the surrounding areas. Commonly used vertical barrier barriers are concrete anti-seepage walls and soil-bentonite anti-seepage walls, but their hydration is inhibited in high-concentration organic polluted liquids, dry-wet cycles, and high-salt environments, and their anti-seepage performance is poor, which cannot meet the extremely low anti-seepage coefficient requirements in the field of environmental protection. Bentonite waterproof blanket (GCL) is a high-performance geosynthetic material in the anti-seepage system. It is a blanket structure formed by fixing bentonite between two layers of geotextiles through a needle-punching process. It has an extremely low hydraulic permeability coefficient and good chemical compatibility. It is widely used in anti-seepage systems in contaminated sites. The bentonite waterproof blanket vertical barrier composite barrier is developed on the basis of the conventional soil-bentonite anti-seepage wall: after the trench is excavated, the bentonite waterproof blanket composite component is vertically lowered in the middle of the trench width and along the length of the trench, and low-permeability materials are backfilled on both sides of the bentonite waterproof blanket to form a vertical barrier composite barrier with anti-seepage and anti-pollution functions. The permeability coefficient k under the action of pollutants is 2.5447 k / cm2. c ≤1×10 -9 cm / s.

[0004] The low permeability of bentonite waterproof blanket is mainly because the bentonite fills the voids inside the bentonite through osmotic expansion, making the liquid flow path more tortuous, thus making it have an extremely low hydraulic permeability coefficient (k w <5×10 -11 m / s). However, under the infiltration of high-concentration organic polluted liquid, the infiltration expansion of bentonite is suppressed. As the infiltration time increases, the permeability coefficient k of the bentonite waterproof blanket increases. c It gradually increases until it reaches the equilibrium stage, and the permeability coefficient tends to be stable. At this time, the permeability coefficient k of the organic contaminated liquid isc Compared with the permeability coefficient k of tap water w It may increase by several orders of magnitude, resulting in failure to meet risk control requirements (k c ≤1×10 -9 m / s). Therefore, it is necessary to study the permeability coefficient k of bentonite waterproof blanket under the action of high concentration organic polluted liquid. c Change rules to verify whether it can meet the risk control requirements.

[0005] Bentonite blanket permeability testing refers to ASTM D6766, "Standard Test Method for Evaluation of Hydraulic Properties of Geosynthetic Clay Liners Permeated with Potentially Incompatible Aqueous Solutions." Using a flexible-wall permeameter, the permeability of bentonite blankets subjected to specific organic matter concentrations is tested to determine whether they can achieve the low permeability requirements of vertical barrier composite systems. However, meeting the test's termination criteria, particularly those corresponding to chemical equilibrium, such as a relative error of less than 10% between the influent and effluent pH, conductivity, contaminant concentration, or dielectric constant, and the absence of a significant monotonic increase or decrease in these parameters over time, can require testing times ranging from several months to several years, posing a significant challenge to project construction schedules. Therefore, it is necessary to develop a low-cost and more convenient alternative method for predicting the permeability of bentonite blankets subjected to organic matter. Summary of the Invention

[0006] In view of the above-mentioned deficiencies in the prior art, the present invention aims to provide a machine learning prediction method for the permeability coefficient of bentonite waterproof blanket under the action of organic matter in vertical barrier. This method can predict the permeability coefficient k of bentonite waterproof blanket under the action of organic matter based on factors such as bentonite properties (free expansion index, mass per unit area), prehydration conditions (effective confining pressure and pore fluid ion strength in backfill), chemical properties of contaminated liquid (concentration of water-soluble organic matter or dielectric constant of non-aqueous liquid organic matter), and effective stress of the infiltration process. c .

[0007] To achieve the above objectives, the present invention provides a machine learning prediction method for the permeability coefficient of a bentonite waterproof blanket under the action of organic matter in a vertical barrier, comprising the following steps:

[0008] A machine learning prediction method for the permeability coefficient of a bentonite waterproof blanket under the action of organic matter in a vertical barrier, characterized by comprising the following steps:

[0009] Step 1. Data collection: Through literature research, enterprise consultation and indoor experiments, a database of bentonite waterproofing blankets under the influence of water-soluble organic matter and non-aqueous liquid organic matter infiltration solutions is established. The database contains the following influencing factors: bentonite properties, prehydration conditions, chemical properties of organic contaminated liquids, effective stress during the infiltration process, and the corresponding permeability coefficients of bentonite waterproofing blankets under the combined effect of various influencing factors. The influencing factors and permeability coefficients are all measured values. One influencing factor and one corresponding permeability coefficient constitute a database sample. The bentonite waterproofing blanket database contains all database samples.

[0010] Step 2: Data preprocessing and normalization: (1) Process missing values ​​and outliers; For all the database samples collected in step 1, delete the samples with missing influencing factors and permeability coefficients, and only retain the samples with complete influencing factors and permeability coefficients; For samples with similar influencing factor data, exclude the corresponding samples with permeability coefficients that have a mutation of more than one order of magnitude; (2) Normalize the influencing factor data; Since the influence of influencing factor data of different orders of magnitude has different effects on the model learning process, the influencing factor data is normalized by the Euclidean norm, that is, the L2 norm, and the L2 norm is normalized to 0~1:

[0011]

[0012] Where m n is the original value of the nth impact factor;

[0013] The normalized value of the impact factor is the ratio of the original value to the L2 norm, that is, m n / L2 norm is obtained; the normalized values ​​of the influencing factors and their corresponding permeability coefficients are used as data sets, 75% of the data sets are used for subsequent model training, and 25% of the data sets are used for subsequent model verification;

[0014] Step 3. Algorithm selection: Use the multi-layer perceptron (MLP) neural network model in the Scikit-learn machine learning library in Python, with a hidden layer size of 100, the ReLU function as the activation function, and the mean square error (MSE) as the loss function. Update the weights using the stochastic gradient descent method to optimize the square error between the model prediction value and the measured value to predict the permeability coefficient of bentonite waterproofing blanket under the action of organic matter.

[0015] Step 4: Parameter optimization: The range of each parameter in the stochastic gradient descent method is set based on experience; the model parameters are taken from minimum to maximum values ​​in sequence, and the mean square error corresponding to the regression of different model parameters is calculated through hyperparameter tuning; when the mean square error is minimized, the corresponding model parameters are selected as the optimal parameters;

[0016] Step 5: Model validation and evaluation: Based on the influencing factors in the corresponding data set for model validation, the predicted value of the permeability coefficient is obtained through the optimal parameters of the multilayer perceptron neural network model; the predicted value of the permeability coefficient is compared with the measured value to evaluate the performance of the model.

[0017] As a further preferred solution, in step 1:

[0018] The bentonite properties include free expansion index and mass per unit area;

[0019] The prehydration conditions include effective confining pressure and ionic strength of pore fluid in backfill material;

[0020] The chemical properties of the organic contaminated liquid include the concentration of water-soluble organic matter or the dielectric constant of non-aqueous liquid organic matter.

[0021] As a further preferred solution, the larger the free expansion index and the larger the mass per unit area in the bentonite properties, the lower the permeability coefficient of the bentonite waterproofing blanket, and the influencing factor and the permeability coefficient are negatively correlated;

[0022] The greater the effective confining pressure of the prehydration condition and the greater the effective stress of the infiltration process, the lower the permeability coefficient of the bentonite waterproof blanket, and the influencing factor and the permeability coefficient are negatively correlated;

[0023] The greater the ionic strength of the pore fluid in the pre-hydrated backfill material, the higher the permeability coefficient of the bentonite waterproof blanket, and the influencing factor and the permeability coefficient are positively correlated;

[0024] When the organic contaminated liquid is a water-soluble liquid, the greater the concentration, the higher the permeability coefficient; when the organic contaminated liquid is a non-aqueous phase liquid, the greater the dielectric constant, the higher the permeability coefficient;

[0025] The impact factor and permeability coefficient are positively correlated.

[0026] As a further preferred solution, in step three, the bentonite properties, prehydration conditions, chemical properties of the organic contaminated liquid, and effective stress of the infiltration process are normalized to 0-1 and used as model input data; and a predicted value of the permeability coefficient is given based on the multi-layer perceptron neural network model.

[0027] As a further preferred solution, in step 4, during the optimization process, the model training dataset is randomly divided into five consecutive subsets, each of which can be used for model validation, while the other subsets are used for iterative model training; this method enhances data reusability and avoids overfitting.

[0028] As a further preferred solution, in step five, when evaluating the performance of the model: when the predicted value is greater than the measured value, the ratio is greater than 1, which is a conservative prediction, and the predicted value is safe in actual engineering applications; when the predicted value is less than the measured value, the ratio is less than 1, which is a non-conservative prediction, and the predicted value is dangerous in actual engineering applications, that is, outside the value range of the non-conservative prediction, the prediction model should be used with caution within this value range.

[0029] Beneficial effects: The permeability coefficient test of bentonite waterproof blanket under the action of organic matter may require several months to several years of testing time to reach the termination standard corresponding to the chemical equilibrium, posing a severe challenge to the construction period of the project; the present invention is based on influencing factors such as bentonite properties, prehydration conditions, chemical properties of organic contaminated liquid, and effective stress of the infiltration process. The permeability coefficient of bentonite waterproof blanket under the action of organic matter can be effectively predicted through the "multi-layer perceptron (MLP)" neural network model. The prediction results are highly accurate, which can greatly reduce the time required for traditional tests to reach chemical equilibrium and save manpower, material and financial resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 Schematic diagram of neural network algorithm;

[0031] Figure 2 The results of model verification of permeability coefficient of bentonite waterproof blanket under the action of organic matter are presented;

[0032] Figure 3 It is the ratio of the predicted value to the measured value of the permeability coefficient of bentonite waterproof blanket under the action of organic matter. DETAILED DESCRIPTION

[0033] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0034] The present invention adopts the following technical solution: a machine learning prediction model is constructed using a multi-layer perceptron (MLP) neural network model. The neural network imitates the way the human brain creates interconnected neurons, such as Figure 1 As shown. Taking the impact factor as the input neuron (a group of neurons {x i |x1,x2,…x m}), the permeability coefficient of organic contaminated liquid of bentonite waterproof blanket (f(x)) is used as the output neuron. Between the input and output neurons, one or more layers of neurons, namely hidden layers, are inserted to transmit information. Different weights are applied to the connection between each neuron to indicate the importance of the previous layer of neurons to the next layer of receiving neurons. For example, neuron a1 is the value of the input layer plus the weight (w) of each neuron (i.e. w1x1+w2x2+…w m x m ). The values ​​of the neurons in the last hidden layer are transmitted to the output layer and converted into output values. By comparing the output values ​​with the corresponding measured permeability coefficients, the square error (R 2 The Multilayer Perceptron (MLP) neural network model uses the ReLU function as the activation function and the mean squared error (MSE) as the loss function. Stochastic gradient descent (Adam optimizer, with the default adaptive learning rate) is used to update weights to optimize the squared error between the model's predicted and measured values. Iterations terminate when the squared error falls below a specified value or when the maximum number of iterations is reached.

[0035] The present invention provides a machine learning prediction method for the permeability coefficient of a bentonite waterproof blanket under the action of organic matter in a vertical barrier, comprising the following steps:

[0036] S1. Data Collection: Through literature research, enterprise consultation and laboratory experiments, a database of bentonite waterproofing blankets under the action of penetrating solutions of water-soluble organic matter (methanol, ethanol, formaldehyde, glucose, etc.) and non-aqueous liquid organic matter (petroleum hydrocarbons, benzene series, chlorinated organic solvents, coal tar) was established;

[0037] The database includes the following influencing factors: bentonite properties, prehydration conditions, chemical properties of organic contaminated liquid, effective stress during the infiltration process, and the corresponding permeability coefficient of the bentonite waterproofing blanket under the combined effect of each influencing factor. The influencing factors and permeability coefficient are all measured values.

[0038] One influencing factor and a corresponding permeability coefficient constitute a database sample, forming multiple database samples; some data from this test are excerpted, as shown in Table 1:

[0039] Table 1 Summary of influencing factors and permeability coefficient parameters

[0040]

[0041]

[0042] S2. Data preprocessing and normalization: (1) Process missing values ​​and outliers; For all database samples collected in step 1, delete the database samples with missing influencing factors and permeability coefficients (database samples only have influencing factors or permeability coefficients), and only retain the samples with complete influencing factors and permeability coefficients; For samples with similar influencing factor data, filter the corresponding samples with permeability coefficients that produce mutations of more than one order of magnitude. Generally, when the influencing factors are close, the permeability coefficients will not differ by more than one order of magnitude; (2) Normalize the influencing factor data; Since the influencing factor data of different orders of magnitude have different effects on the model learning process, the influencing factor data are normalized by the Euclidean norm, that is, the L2 norm, and the L2 norm is normalized to 0~1:

[0043]

[0044] Where m n is the original value of the nth impact factor;

[0045] The normalized value of the impact factor is the ratio of the original value to the L2 norm, that is, m n / L2 norm is obtained; the normalized values ​​of the influencing factors and their corresponding permeability coefficients are used as data sets, 75% of the data sets are used for subsequent model training, and 25% of the data sets are used for subsequent model verification.

[0046] S3. Using the "Multi-layer Perceptron (MLP)" neural network model in the Scikit-learn machine learning library in Python, with a hidden layer size of 100, the ReLU function as the activation function, and the mean square error (MSE) as the loss function, the weights were updated by stochastic gradient descent (Adam optimizer, default adaptive learning rate) to predict the permeability coefficient of bentonite waterproofing blanket under the action of organic contaminated liquid. Among them, the bentonite properties (free swelling index, unit area mass), prehydration conditions (effective confining pressure, pore fluid ion strength in backfill), chemical properties of organic contaminated liquid (concentration of water-soluble organic matter or dielectric constant of non-aqueous phase liquid organic matter), and effective stress of the infiltration process were normalized to 0-1 and used as model input data. The permeability coefficient prediction value was given based on the multi-layer perceptron neural network model. S4. Parameter optimization: The value range of each model parameter (learning rate, number of iterations, etc.) is set based on experience; the model parameters are taken from minimum to maximum values ​​in turn, and the mean square error (MSE) corresponding to the regression of different model parameters is calculated through hyperparameter tuning (grid search) to optimize the square error between the model prediction value and the measured value; when the mean square error is minimized, the corresponding model parameter is selected as the optimal parameter. During the optimization process, the model training data set is randomly divided into 5 consecutive subsets, each of which can be used for model verification, while the other subsets are used for iterative model training; this method enhances the reusability of data and avoids overfitting;

[0047] S5. Model validation and evaluation: The predicted and measured values ​​of the model validation dataset are as follows: Figure 2 As shown, the values ​​on the black dashed line represent a 10-fold ratio of predicted value to measured value or measured value to predicted value, while the values ​​on the double black line represent a 100-fold ratio of predicted value to measured value or measured value to predicted value. In the validation dataset (58 measurements in total), 87% of the predicted values ​​were within the range of 0.1 to 10 times the measured value, and 97% were within the range of 0.01 to 100 times the measured value. Only two predicted values ​​fell below 0.01 times or exceeded 100 times the measured value. There was no significant bias in the data.

[0048] Comparison of chemical compatibility prediction and measured data of bentonite waterproof blanket Figure 3 As shown, when the measured value of bentonite waterproof blanket is 10 -8 m / s~10 -11 When the speed changes between m / s, the ratio of the predicted value to the measured value changes between 0.1 and 100, and the predicted value is basically within the acceptable error range; when the measured value of the bentonite waterproof blanket is greater than 10 -8 m / s, the ratio of the measured value to the predicted value may be greater than 100, and the error of the predicted value is relatively large. It is a non-conservative prediction and the prediction model should be used with caution. At this time, the measured value may still need to be used as the basis.

[0049] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A machine learning prediction method for the permeability coefficient of a bentonite waterproof blanket under the action of organic matter in a vertical barrier, characterized by: The following steps are involved: Step 1. Data Collection: Through research, consultation, and testing, the following influencing factors were obtained: bentonite properties, prehydration conditions, chemical properties of organic contaminated liquid, effective stress during the infiltration process, and the corresponding permeability coefficient of the bentonite waterproofing blanket under the combined effect of each influencing factor. The influencing factors and permeability coefficients are measured values, and the influencing factors and corresponding permeability coefficients constitute the database sample; Step 2: Data preprocessing and normalization: (1) Process missing values ​​and outliers; for all database samples collected in step 1, delete the samples with missing impact factors and permeability coefficients; (2) Normalize the impact factor data; normalize the impact factor data using the Euclidean norm, and normalize the L2 norm to 0-1: Where m n is the original value of the nth impact factor; The normalized value of the impact factor is the ratio of the original value to the L2 norm, that is, m n / L2 norm is obtained; the normalized values ​​of the influencing factors and their corresponding permeability coefficients are used as data sets for subsequent model training and model verification respectively; Step 3: Algorithm selection: Use the multilayer perceptron neural network model in the Scikit-learn machine learning library in Python to optimize the squared error between the model prediction value and the measured value to predict the permeability coefficient of bentonite waterproofing blanket under the action of organic matter; Step 4. Parameter optimization: The value range of each parameter in the stochastic gradient descent method is based on experience. The model parameters are taken from minimum to maximum values ​​in turn, and the mean square error corresponding to the regression of different model parameters is calculated through hyperparameter tuning; when the mean square error is the smallest, the corresponding model parameters are selected as the optimal parameters; Step 5: Model validation and evaluation: Based on the influencing factors in the corresponding data set for model validation, the predicted value of the permeability coefficient is obtained through the optimal parameters of the multilayer perceptron neural network model; The predicted values ​​of permeability coefficients were compared with the measured values ​​to evaluate the performance of the model.

2. The machine learning prediction method for the permeability coefficient of a bentonite waterproof blanket under the action of organic matter in a vertical barrier according to claim 1 is characterized in that: In step one: The bentonite properties include free expansion index and mass per unit area; The prehydration conditions include effective confining pressure and ionic strength of pore fluid in backfill material; The chemical properties of the organic contaminated liquid include the concentration of water-soluble organic matter or the dielectric constant of non-aqueous liquid organic matter.

3. The machine learning prediction method for the permeability coefficient of a bentonite waterproof blanket under the action of organic matter in a vertical barrier according to claim 2, characterized in that: The larger the free expansion index and the greater the mass per unit area in the bentonite properties, the lower the permeability coefficient of the bentonite waterproofing blanket, and the influencing factor and the permeability coefficient are negatively correlated; The greater the effective confining pressure of the prehydration condition and the greater the effective stress of the infiltration process, the lower the permeability coefficient of the bentonite waterproof blanket, and the influencing factor and the permeability coefficient are negatively correlated; The greater the ionic strength of the pore fluid in the pre-hydrated backfill material, the higher the permeability coefficient of the bentonite waterproof blanket, and the influencing factor and the permeability coefficient are positively correlated; When the organic contaminated liquid is a water-soluble liquid, the greater the concentration, the higher the permeability coefficient; when the organic contaminated liquid is a non-aqueous phase liquid, the greater the dielectric constant, the higher the permeability coefficient; The impact factor and permeability coefficient are positively correlated.

4. The machine learning prediction method for the permeability coefficient of a bentonite waterproof blanket under the action of organic matter in a vertical barrier according to claim 2 is characterized by: In step three, the bentonite properties, prehydration conditions, chemical properties of the organic contaminated liquid, and effective stress of the infiltration process are normalized to 0-1 and used as model input data; the permeability coefficient prediction value is given based on the multi-layer perceptron neural network model.

5. The machine learning prediction method for the permeability coefficient of a bentonite waterproof blanket under the action of organic matter in a vertical barrier according to claim 1 is characterized by: In step 4, during the optimization process, the model training dataset is randomly divided into five consecutive subsets, each of which can be used for model validation, while the other subsets are used for iterative model training; this approach enhances data reusability and avoids overfitting.

6. The machine learning prediction method for the permeability coefficient of a bentonite waterproof blanket under the action of organic matter in a vertical barrier according to claim 1, characterized in that: In step five, when evaluating the performance of the model: when the predicted value is greater than the measured value, the ratio is greater than 1, which is a conservative prediction, and the predicted value is safe in actual engineering applications; when the predicted value is less than the measured value, the ratio is less than 1, which is a non-conservative prediction, and the predicted value is dangerous in actual engineering applications, that is, it is outside the value range of the non-conservative prediction, and the prediction model should be used with caution within this value range.

Images