A geotechnical body fracture evolution prediction method based on a physical information neural network and a slope monitoring system

Abstract

This invention discloses a method for predicting the evolution of cracks in soil and rock masses based on a physical information neural network and a slope monitoring system, belonging to the field of intelligent monitoring technology in geotechnical engineering. The method includes: acquiring multi-factor experimental data of expansive soil; performing normalization preprocessing on the data and dividing it into training and test sets; constructing a physical information neural network model containing an input layer, multiple hidden layers, and an output layer; constructing a composite loss function integrating monotonicity constraints, boundary constraints, and smoothness regularization, and training the model using an adaptive weight scheduling strategy; and using the trained model to predict the crack rate, number of cracks, and average crack width. The system includes a data acquisition module, a prediction module, and an early warning module. This invention achieves physically consistent multi-objective balanced prediction under small sample conditions by embedding the physical laws of crack development into the neural network training process, with a comprehensive determination coefficient reaching 0.834, which is superior to traditional machine learning methods.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for geotechnical engineering, and in particular to a method for predicting the evolution of cracks in soil and rock mass based on physical information neural networks and a slope monitoring system. Background Technology

[0002] Soil and rock masses refer to geological materials composed of rocks and soil, widely distributed in various engineering sites. They mainly include: rocks (such as granite, sandstone, shale, limestone, etc.), cohesive soils (such as clay, silty clay, silt, etc.), non-cohesive soils (such as sand, silt, gravel, etc.), and special soils (such as expansive soil, collapsible loess, frozen soil, saline soil, etc.). These soil and rock masses can develop fissures under natural environmental influences, affecting engineering stability. Expansive soil is a special type of cohesive soil mainly composed of hydrophilic minerals such as montmorillonite, exhibiting significant water absorption and expansion, and water loss and shrinkage characteristics. Under wet-dry cycles, fissures gradually develop on the surface of expansive soil slopes, extending deeper to form a complex fissure network, severely affecting the hydraulic properties and mechanical strength of the soil, ultimately leading to engineering disasters such as shallow landslides. Therefore, accurately predicting the evolution of fissures in expansive soil is of great significance for slope stability assessment and engineering safety.

[0003] Currently, research on the evolution of fractures in expansive soils mainly relies on laboratory experiments and empirical regression models. However, relying on empirical correlations or simplified analytical models has limitations in capturing the complex nonlinear relationships between multiple factors and fracture characteristics.

[0004] In recent years, machine learning methods have shown promising applications in geotechnical engineering. Among them, the Physics-Informed Neural Network (PINN) combines data-driven learning with physical constraints, enabling effective learning under limited data conditions while ensuring the physical consistency of prediction results. However, existing machine learning methods have the following shortcomings when applied to predicting the evolution of cracks in expansive soils:

[0005] (1) Traditional neural networks (such as multilayer perceptrons MLP) are prone to overfitting under small sample conditions, resulting in low prediction accuracy and the inability to guarantee that the prediction results conform to physical laws. They may also produce unreasonable predictions such as the gap rate decreasing with the number of cycles.

[0006] (2) Although traditional machine learning methods (such as random forest, support vector regression, etc.) may achieve good results on a single output variable, they are difficult to accurately predict multiple fracture feature parameters (fracture rate, number of fractures, average fracture width) at the same time, and the prediction accuracy among the outputs is uneven.

[0007] (3) Existing methods lack an effective mechanism to integrate the physical laws of crack development (such as monotonicity, boundary constraints, and smoothness) into the model training process, resulting in insufficient model generalization ability. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies, such as low accuracy in predicting the evolution of rock and soil fissures, poor physical consistency, and unbalanced predictions among multiple outputs, this invention provides a method for predicting the evolution of rock and soil fissures based on a physical information neural network and a slope monitoring system.

[0009] To achieve the above-mentioned technical objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for predicting the evolution of fractures in soil and rock masses based on a physical information neural network, comprising: S1. Obtain multi-factor test data for soil and rock, including input parameters and fracture characteristic parameters; among which, the input parameters include soil type, initial moisture content, compaction degree, sample thickness and number of wet-dry cycles; fracture characteristic parameters include fracture ratio, number of fractures and average fracture width. S2. Normalize the experimental data by normalizing all input parameters and crack feature parameters to the [0,1] interval, and divide the preprocessed experimental data into training set and test set according to a preset ratio. S3. Construct a physical information neural network model; the model includes an input layer, multiple hidden layers, and an output layer; the input layer receives normalized input parameters, and the output layer outputs predicted values ​​of fracture rate, number of fractures, and average fracture width. S4. The physical information neural network model is trained based on a composite loss function that incorporates physical constraints. The composite loss function includes a data fitting loss term and a physical constraint loss term. The physical constraint loss term includes monotonicity constraint loss, boundary constraint loss, and smoothness regularization loss. An adaptive weight scheduling strategy is used to dynamically adjust the weights of each physical constraint loss term. S5. Input the input parameters of the soil and rock under the working condition to be predicted into the trained physical information neural network model, and output the corresponding predicted values ​​of the fracture characteristic parameters.

[0010] Optionally, in step S4, the monotonicity constraint loss is calculated as follows: ; in, For monotonicity-constrained loss, Let i be the predicted crack rate of the i-th sample. This refers to the number of wet and dry cycles. The sample size is represented by the number of samples. The monotonicity constraint is used to ensure that the crack rate increases monotonically with the number of wet-dry cycles.

[0011] Optionally, the adaptive weight scheduling strategy is as follows: ; ; in, Here, is a general weight scheduling function, and t is the current training epoch. For the number of rounds during the preheating period, As the initial weights, The final weights are determined by the data fitting process in the early stages of training, with the weights of physical constraints gradually increasing as training progresses. , , All adopt the aforementioned adaptive weight scheduling strategy, and each independently selects from... linear growth to warm-up period wheel.

[0012] Alternatively, the expression for the composite loss function is: ; ; ; ; in, For composite loss function, For the sample size, For data fitting loss, For the first Measured values ​​of each sample (after normalization). These are the normalized predicted values; The loss is due to monotonicity constraints. For boundary constraint loss, This is the tolerance margin; For smoothness regularization loss, Let x be the gradient vector of the input parameter vector. Let be the predicted crack rate of the i-th sample; , , These are the weighting coefficients for each constraint term.

[0013] Optionally, the input parameters include an initial moisture content ranging from 30% to 40%, a compaction degree ranging from 80% to 90%, a sample thickness ranging from 1 to 4 cm, and a wet-dry cycle number ranging from 1 to 5.

[0014] Optionally, the physical information neural network model includes 2-5 hidden layers, each containing 32-256 neurons, with the hyperbolic tangent function Tanh as the activation function; the model is trained using the Adam optimizer with an initial learning rate of It employs the ReduceLROnPlateau learning rate scheduling strategy.

[0015] In a second aspect, the present invention provides a slope monitoring system, the system comprising the steps of performing the method described above: The data acquisition module is used to collect information on soil type, moisture content, compaction degree, soil layer thickness, and number of wet-dry cycles of expansive soil on slopes. The prediction module is used to input the information collected by the data acquisition module into the physical information neural network model trained by the above method, and output the predicted values ​​of fracture rate, number of fractures and average fracture width. The early warning module is used to compare the predicted values ​​of the fracture characteristic parameters output by the prediction module with preset thresholds, and to issue an early warning signal when the fracture characteristic parameters reach the preset early warning threshold.

[0016] Optionally, in the early warning module, the preset early warning thresholds include a yellow early warning threshold and a red early warning threshold; a yellow early warning is triggered when the predicted value of the fracture characteristic parameter reaches 80% of the historical maximum value, and a red early warning is triggered when it reaches 100% of the historical maximum value.

[0017] The beneficial effects of this invention include: 1) By embedding the physical laws of crack development (monotonicity constraints, boundary constraints, smoothness regularization) into the neural network loss function, the physical consistency of the prediction results is guaranteed, avoiding unreasonable predictions that may be generated by traditional neural networks (such as crack rate decreasing with the number of cycles). 2) A multi-output network architecture is used to simultaneously predict three fracture characteristic parameters: fracture rate, fracture number, and average fracture width. This achieves balanced multi-objective prediction and a comprehensive determination coefficient. It achieves a score of 0.834, outperforming traditional machine learning methods such as Random Forest (0.793) and XGBoost (0.758); 3) By using an adaptive weight scheduling strategy, the training focuses on data fitting in the early stage and gradually strengthens physical constraints in the later stage, which effectively solves the overfitting problem under small sample conditions and requires only a small number of experimental samples to achieve effective prediction. 4) Combined with the slope monitoring system, the entire process from data acquisition and crack prediction to early warning output is automated, providing intelligent technical means for the stability assessment of expansive soil slopes. Attached Figure Description

[0018] Figure 1A flowchart illustrating the overall process of predicting the evolution of fractures in soil and rock based on a physical information neural network, as provided in this embodiment of the invention.

[0019] Figure 2 This is a schematic diagram of the dry-wet cycle test process and crack development provided in an embodiment of the present invention; wherein, Figure 2 (a) is a schematic diagram of the wet-dry cycle test process; Figure 2 (b) is a schematic diagram of fracture development; Figure 3 This is a schematic diagram of the physical information neural network model provided in an embodiment of the present invention; Figure 4 The loss function variation curve during model training provided in this embodiment of the invention; wherein, Figure 4 (a) is a schematic diagram of training and validation loss; Figure 4 (b) represents data loss and physical constraint loss; Figure 4 (c) is a schematic diagram of learning rate scheduling; Figure 4 (d) is a schematic diagram of training loss and validation results in the final training phase; Figure 5 This is a comparison chart of predicted and measured values ​​for the test set provided in an embodiment of the present invention; wherein, Figure 5 (a) is a schematic diagram showing the predicted and measured values ​​of the fracture rate. Figure 5 (b) is a schematic diagram showing the predicted and measured values ​​of the number of fractures predicted by the model; Figure 5 (c) is a schematic diagram showing the predicted and measured values ​​of the average crack width predicted by the model; Figure 6 The model provided in this embodiment of the invention provides a prediction curve of the evolution of porosity with the number of wet-dry cycles under different initial moisture contents; wherein, Figure 6 (a) is moderately expansive soil Prediction curve of the evolution of porosity with the number of wet-dry cycles. Figure 6 (b) is a highly expansive soil Prediction curve of the evolution of time-dependent porosity with the number of wet-dry cycles; Figure 7 This is a schematic diagram of the slope monitoring system provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the on-site deployment of the slope monitoring system provided in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0022] The method for predicting the evolution of cracks in soil and rock masses based on physical information neural networks provided by this invention is applicable to various types of soil and rock masses, including but not limited to rocks, cohesive soils, non-cohesive soils, and special soil types. The following embodiments use expansive soil as a typical example for detailed explanation; however, those skilled in the art will understand that the technical principles and implementation steps of this method are also applicable to the prediction of crack evolution in other types of soil and rock masses, requiring only adjustments to the corresponding input parameters and physical constraints based on the specific physical and mechanical properties of the soil and rock mass.

[0023] Example 1

[0024] like Figure 1 As shown, this embodiment provides a method for predicting the evolution of fractures in soil and rock masses based on physical information neural networks, including: S1: Obtain multi-factor test data for expansive soil. The test data includes input parameters and crack characteristic parameters. The input parameters include soil type, initial moisture content, compaction degree, sample thickness, and number of wet-dry cycles. The crack characteristic parameters include crack ratio, number of cracks, and average crack width.

[0025] Specifically, among the input parameters, the initial moisture content ranges from 30% to 40%, the compaction degree ranges from 80% to 90%, the sample thickness ranges from 1 to 4 cm, and the number of wet-dry cycles ranges from 1 to 5.

[0026] like Figure 2 As shown in (a), in the specific implementation process of this embodiment, the test soil sample was taken from the K23+800 section of a certain expressway. According to the free swelling rate test, the soil sample was divided into medium swelling soil (free swelling rate...). ) and highly expansive soil (free expansion rate) There are two types of soil samples. The basic physical properties of the two types of soil samples are shown in Table 1 below:

[0027] Table 1. Basic physical properties of the two types of soil samples

[0028] The sample preparation process is as follows: Dry soil is crushed and passed through a 1mm sieve, then uniformly sprayed with distilled water to the target moisture content, sealed for 48 hours to ensure uniform moisture distribution, and then statically compacted to the specified compaction degree. A diameter of [missing information - likely a diameter measurement] is used. =61.8mm ring cutter sampling.

[0029] The wet-dry cycle test design is as follows: a. Initial moisture content: three levels: 30%, 35%, and 40%; b. Compaction degree: two levels, 80% and 90%; c. Sample thickness: 1cm, 2cm, 3cm, and 4cm (four levels); c. Number of dry-wet cycles: 1-5 times.

[0030] In each cycle, the sample was dried in a 35°C oven until the moisture content dropped to approximately 12% (shrinkage limit), and then saturated using a vacuum saturation device for 24 hours. Surface crack images were captured after each drying stage, such as... Figure 2 As shown in (b).

[0031] The characteristic parameters of the fractures were extracted using the PCAS image processing software for porosity and fracture analysis, including: fracture ratio (i.e., the percentage of fracture area to the total surface area of ​​the sample, %), number of fractures (i.e., the total number of fractures, in mm), and average fracture width (i.e., the ratio of the total fracture area to the total fracture length, in mm). In this specific implementation, a total of 80 test data points were obtained, 40 for moderately expansive soil and 40 for highly expansive soil. Since the wet-dry cycle test of expansive soil involves a multi-factor orthogonal design and has a long experimental cycle and high cost, 80 samples constitute a small dataset. Traditional pure data-driven neural networks usually require hundreds of samples for effective training. This invention employs the Physical Information Neural Network (PINN) method, embedding the physical constraints of crack development into the neural network loss function to achieve a hybrid modeling approach of "physical prior knowledge + data-driven," overcoming the limitations of traditional empirical methods in capturing complex nonlinear relationships among multiple factors. Under small sample conditions, the prediction determination coefficients R² for crack rate, crack number, and average crack width reach 0.75, 0.74, and 0.80, respectively, significantly outperforming traditional methods, and the prediction results conform to physical laws.

[0032] This leads to the construction of the input parameter vector and the output parameter vector: The input parameter vector is represented as ,in The soil type is indicated by the code 0 for moderately expansive soil and 1 for highly expansive soil. Initial moisture content (%) Compaction degree (%) The thickness of the sample is (cm). This represents the number of wet and dry cycles.

[0033] The output parameter vector is represented as ,in The crack ratio is expressed as a percentage. This represents the number of cracks (cracks). The average crack width is (mm).

[0034] S2: Normalize the experimental data and divide the preprocessed experimental data into training and test sets according to a preset ratio.

[0035] In this specific implementation, all input and output parameters are subjected to Min-Max normalization, mapping the data to the [0,1] interval: ; in, and These are the minimum and maximum values ​​of the parameter in the training set, respectively.

[0036] The 80 preprocessed data points were randomly divided into a training set (64 samples) and a test set (16 samples) at a ratio of 80% / 20%.

[0037] S3: Construct a physical information neural network model; the model includes an input layer, multiple hidden layers and an output layer; the input layer receives normalized input parameters, and the output layer outputs the predicted values ​​of fracture rate, number of fractures and average fracture width.

[0038] like Figure 3 As shown, the physical information neural network model constructed in this embodiment adopts a fully connected neural network architecture, with the following specific configuration: Input layer: 5 neurons, corresponding to 5 input parameters; Hidden layer: 3 layers, 64 neurons per layer; Output layer: 3 neurons, corresponding to the gap ratio, gap number and average gap width; Activation function: hyperbolic tangent function Tanh; Total number of model parameters: 8899.

[0039] The reason for choosing Tanh as the activation function is that the Tanh function has an output range of (-1,1) and has zero-centeredness, which is beneficial for gradient propagation; at the same time, the smoothness of the Tanh function ensures the differentiability of the output with respect to the input, which is convenient for calculating the gradient information required in physical constraints.

[0040] S4: The physical information neural network model is trained based on a composite loss function that incorporates physical constraints; the composite loss function includes a data fitting loss term and a physical constraint loss term; the physical constraint loss term includes monotonicity constraint loss, boundary constraint loss, and smoothness regularization loss; an adaptive weight scheduling strategy is used to dynamically adjust the weights of each physical constraint loss term.

[0041] The core innovation of this invention lies in embedding the physical laws governing the development of fissures in expansive soil into the loss function of a neural network. The composite loss function consists of data fitting loss and three physical constraint losses: ; Data fitting loss Mean squared error is used to measure the predicted value Compared with measured values Deviation between: ; in, Let be the measured value of the i-th sample. For the corresponding predicted value, This represents the number of training samples.

[0042] Because expansive soil exhibits irreversible and monotonically increasing crack development under repeated wet-dry cycles, i.e., crack ratio... The number of wet-dry cycles N increases monotonically: Therefore, a soft penalty approach is used to achieve monotonicity constraint loss. : ; When the predicted fracture rate decreases with the number of cycles (i.e.) This loss term incurs a penalty; the loss is zero when monotonicity is satisfied. In practice, the gradient... Obtained through PyTorch's automatic differentiation mechanism.

[0043] The predicted output must remain within a physically reasonable range. For the normalized predicted value y... i It should meet ,in Tolerance margin, thus setting boundary constraint loss. : ; Loss due to boundary constraints To prevent the model from producing negative fracture rates or predictions that exceed the physical upper limit.

[0044] The crack evolution process should be a smooth function of the input parameters to avoid drastic fluctuations in the prediction results; therefore, a smoothness regularization loss is set. : ; in, This is the gradient vector of the predicted crack rate with respect to the input parameters.

[0045] Specifically, to balance the relationship between data fitting and physical constraints, an adaptive weight scheduling strategy with linear preheating is adopted: when ; when ; in, For the current training round, For the number of rounds during the preheating period, As the initial weights, This is the final weight. , , All adopt the aforementioned adaptive weight scheduling strategy, and each independently selects from... linear growth to .

[0046] In the early stages of training (the first 500 rounds), the physical constraint weights are relatively small, and the model primarily focuses on fitting the data, quickly learning the basic patterns within it. As training progresses, the physical constraint weights increase linearly to their final value, gradually guiding the model to learn prediction patterns that conform to physical laws. This gradual approach avoids the optimization difficulties caused by excessively strong physical constraints in the early stages of training.

[0047] The specific configuration for model training is as follows: the optimizer uses Adam, and the initial learning rate is set to... The learning rate scheduling adopts the ReduceLROnPlateau strategy, the decay factor is set to 0.5, the patience value is set to 200 rounds, the maximum number of training rounds is set to 5000 rounds, and an early stopping strategy is adopted, that is, training stops when the verification loss does not decrease for 500 consecutive rounds. The training platform is implemented based on the PyTorch framework.

[0048] like Figure 4 As shown, the model training process terminated in round 523 due to an early stopping strategy. The training loss decreased rapidly in the first 100 rounds and then gradually stabilized. The physical constraint loss remained at a low level throughout the training process, indicating that the model successfully learned physically consistent prediction patterns. The training and validation losses are as follows: Figure 4 As shown in (a); data loss and physical constraint loss are as follows: Figure 4 (b) shows the learning rate scheduling as follows: Figure 4 As shown in (c); the training loss and validation results in the final training phase are as follows: Figure 4 As shown in (d).

[0049] S5: Input the input parameters of the working condition to be predicted into the trained physical information neural network model, and output the corresponding predicted values ​​of the fracture feature parameters.

[0050] Specifically, after the model training is completed, for the new working condition to be predicted, its input parameters are normalized in the same way as the training data and then input into the model. The model outputs the normalized prediction value, and then the prediction value of the actual physical quantity is obtained by inverse normalization.

[0051] To further illustrate the method provided in this embodiment, the prediction performance of the model is verified as follows: like Figure 5 As shown, the model's prediction performance was validated on the test set. The predicted fracture rate values ​​from the model and the measured values ​​are shown below. Figure 5 As shown in (a); the predicted value of the number of fractures predicted by the model and the measured value are as follows: Figure 5 As shown in (b); the predicted value of the average crack width predicted by the model is as follows: Figure 5 As shown in (c). The model's prediction performance on all datasets is shown in Table 2: Table 2 shows the model's prediction performance on all datasets.

[0052] The prediction performance on the training set is shown in Table 3: Table 3 shows the model's prediction performance on the training set.

[0053] The scatter plot of predicted and measured values ​​shows that the data points cluster around the 1:1 line, indicating a good consistency between the predicted and measured values.

[0054] To verify the effectiveness of the physical information method, the PINN model of this invention is compared with several traditional machine learning methods, including Random Forest (RF), Gradient Boosting Regression (XGBoost), Support Vector Regression (SVR), and standard Multilayer Perceptron (MLP, without physical constraints). All models are trained and tested using the same data partitioning.

[0055] The comparison results are shown in Table 4: Table 4 Comparison results of the PINN model of this invention with various traditional machine learning methods

[0056] The following conclusions can be drawn from the comparison results:

[0057] (1) Overall performance: The PINN model of this invention has the following overall performance: The score reached 0.834, which is better than Random Forest (0.793) and XGBoost (0.758), demonstrating the effectiveness of incorporating physical constraints in improving prediction accuracy.

[0058] (2) Equilibrium prediction: Random forest and XGBoost achieved extremely high gap ratios. (0.988~0.990), but performed extremely poorly in terms of crack width ( (or negative values). In contrast, PINN provides a more balanced prediction across all output variables.

[0059] (3) Necessity of physical constraints: Standard MLPs without physical constraints perform the worst (overall). ), all individual outputs All values ​​are negative, indicating that physical constraints are indispensable in few-sample learning problems.

[0060] (4) Generalization ability: The physical information method effectively regularizes the model, prevents overfitting, and achieves better generalization performance on a limited test set.

[0061] like Figure 6 As shown, the trained PINN model is used to predict the evolution of the porosity with the number of wet-dry cycles under different initial moisture contents.

[0062] like Figure 6 As shown in (a), for moderately expansive soils ( Under the conditions of 90% compaction degree and 2cm sample thickness: With an initial moisture content of 30%, the porosity increased from approximately 0.5% in the first cycle to approximately 1.8% in the fifth cycle. Initial moisture content 35%: the fracture rate increased from approximately 0.8% in the first cycle to approximately 2.5% in the fifth cycle; Initial moisture content 40%: the fracture rate increased from approximately 1.2% in the first cycle to approximately 3.2% in the fifth cycle; like Figure 6 As shown in (b), for highly expansive soils ( Under the same conditions, the overall fissure rate of soil is higher than that of medium expansive soil, and the increase is greater.

[0063] The PINN model successfully captures the characteristic pattern of fracture development: rapid development of cyclic fractures in the early stage, followed by gradual stabilization in the later stage. The prediction curve satisfies the monotonicity constraint applied during training, proving that the physical information method effectively prevents unreasonable predictions.

[0064] Sensitivity analysis was conducted on the impact of each input variable on the fracture rate prediction. With other variables fixed as the mean, each individual variable was varied one by one, and the changes in the predicted fracture rate were observed. Specific analysis results show that:

[0065] (1) Soil type: The fissure rate of strong expansive soil is significantly higher than that of medium expansive soil, which is consistent with the higher free expansion rate and greater volume change potential of strong expansive soil.

[0066] (2) Initial moisture content: The porosity increases with the increase of initial moisture content. Higher moisture content leads to a larger moisture content gradient during drying, resulting in more severe shrinkage and crack development.

[0067] (3) Sample thickness: Thicker samples develop higher crack ratios, which is attributed to a larger moisture content gradient and differential shrinkage between the surface and the interior.

[0068] (4) Compaction degree: The compaction degree has a relatively small impact on crack development, which is consistent with the experimental observation results.

[0069] The sensitivity ranking is as follows: soil expansibility > initial moisture content > sample thickness > compaction degree.

[0070] Example 2

[0071] like Figure 7 As shown, this embodiment provides a slope monitoring system, the system for performing the steps of the method described above including: The data acquisition module is used to collect information on soil type, moisture content, compaction degree, soil layer thickness, and wet-dry cycle number of expansive soils on slopes, including: Soil type identification unit is used to determine the type (moderate or strong expansive) of expansive soil on slopes through free expansion rate test. Moisture content sensor is used to monitor changes in the moisture content of slope soil in real time; The compaction testing unit is used to test the compaction degree of soil using the sand cone method or a nuclear density meter. Soil thickness measurement unit, used to determine the thickness of expansive soil layers by borehole sampling or ground-penetrating radar;

[0072] The wet-dry cycle counting unit is used to count the number of wet-dry cycles based on rainfall-evaporation records.

[0073] The early warning module is used to compare the predicted values ​​of the fracture characteristic parameters output by the prediction module with preset thresholds, and to issue an early warning signal when the fracture characteristic parameters reach the preset early warning threshold.

[0074] The prediction module normalizes the information collected by the data acquisition module, inputs it into a pre-trained physical information neural network model, and outputs predicted values ​​for fracture rate, number of fractures, and average fracture width. For example... Figure 8 As shown, the prediction module can be deployed on edge computing devices or cloud servers to achieve real-time or periodic predictions.

[0075] Optionally, the early warning module can issue tiered early warnings based on the prediction results: Yellow alert: Triggered when the predicted value of the crack characteristic parameter reaches 80% of the historical maximum value of the monitoring point, indicating that the development of slope cracks is approaching the historical peak and the monitoring frequency needs to be increased; Red Alert: Triggered when the predicted value of the crack characteristic parameter reaches 100% of the historical maximum value, indicating that the development of slope cracks has reached or exceeded the historical peak value, and engineering measures must be taken immediately.

[0076] In practice, the warning signals are pushed to relevant management personnel via SMS, email, or monitoring platform.

[0077] Example 3

[0078] This embodiment also provides an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the above-described method for predicting the evolution of rock and soil fractures based on a physical information neural network.

[0079] The electronic device may be a server, industrial control computer, embedded computing device, or personal computer. The processor may be a central processing unit (CPU), graphics processing unit (GPU), or neural network processor (NPU).

[0080] Example 4

[0081] This embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method for predicting the evolution of rock and soil fractures based on a physical information neural network.

[0082] The computer-readable storage media include, but are not limited to, USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks or optical disks, and other media capable of storing program code.

[0083] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting the evolution of fractures in soil and rock masses based on physical information neural networks, characterized in that, Includes the following steps: S1. Obtain multi-factor test data for soil and rock, including input parameters and fracture characteristic parameters; among which, the input parameters include soil type, initial moisture content, compaction degree, sample thickness and number of wet-dry cycles; fracture characteristic parameters include fracture ratio, number of fractures and average fracture width. S2. Normalize the experimental data and divide the preprocessed experimental data into training set and test set according to a preset ratio. S3. Construct a physical information neural network model; the model includes an input layer, multiple hidden layers, and an output layer; the input layer receives normalized input parameters, and the output layer outputs predicted values ​​of fracture rate, number of fractures, and average fracture width. S4. The physical information neural network model is trained based on a composite loss function that incorporates physical constraints. The composite loss function includes a data fitting loss term and a physical constraint loss term. The physical constraint loss term includes monotonicity constraint loss, boundary constraint loss, and smoothness regularization loss. An adaptive weight scheduling strategy is used to dynamically adjust the weights of each physical constraint loss term. S5. Input the input parameters of the soil and rock under the working condition to be predicted into the trained physical information neural network model, and output the corresponding predicted values ​​of the fracture characteristic parameters.

2. The method according to claim 1, characterized in that, In S4, the monotonicity constraint loss is calculated as follows: ； in, For monotonicity-constrained loss, Let i be the predicted crack rate of the i-th sample. This refers to the number of wet and dry cycles. This represents the number of samples.

3. The method according to claim 1, characterized in that, The adaptive weight scheduling strategy is as follows: ； in, Here, is a general weight scheduling function, and t is the current training epoch. For the number of rounds during the preheating period, As the initial weights, This is the final weight.

4. The method according to claim 1, characterized in that, The expression for the composite loss function is: ； ； ； ； in, For composite loss function, For the sample size, For data fitting loss, For the normalized first Measured values ​​of a sample These are the normalized predicted values; The loss is due to monotonicity constraints. For boundary constraint loss, This is the tolerance margin; For smoothness regularization loss, Let x be the gradient vector of the input parameter vector. Let be the predicted crack rate of the i-th sample; , , These correspond to the weighting coefficients of each physical constraint term.

5. The method according to claim 1, characterized in that, Among the input parameters, the initial moisture content ranges from 30% to 40%, the compaction degree ranges from 80% to 90%, the sample thickness ranges from 1 to 4 cm, and the number of wet-dry cycles ranges from 1 to 5.

6. The method according to claim 1, characterized in that, The physical information neural network model consists of 2-5 hidden layers, each containing 32-256 neurons. The activation function is the hyperbolic tangent function (Tanh). The model is trained using the Adam optimizer with an initial learning rate of [missing value]. It employs the ReduceLROnPlateau learning rate scheduling strategy.

7. A slope monitoring system, said system being used to perform the steps of the method according to any one of claims 1-6, characterized in that, include: The data acquisition module is used to collect information on soil type, moisture content, compaction degree, soil layer thickness, and number of wet-dry cycles of expansive soil on slopes. The prediction module is used to input the information collected by the data acquisition module into the trained physical information neural network model, and output the predicted values ​​of fracture rate, number of fractures and average fracture width. The early warning module is used to compare the predicted values ​​of the fracture characteristic parameters output by the prediction module with the corresponding preset thresholds, and to issue an early warning signal when the fracture characteristic parameters reach the preset early warning thresholds.

8. The system according to claim 7, characterized in that, In the early warning module, preset early warning thresholds include a yellow warning threshold and a red warning threshold; a yellow warning is triggered when the predicted value of the fracture characteristic parameter reaches 80% of the historical maximum value, and a red warning is triggered when it reaches 100% of the historical maximum value.

Images