Impact penetration sounding data-driven ground mechanics parameter intelligent inversion method

Abstract

This invention provides an intelligent inversion method for ground mechanical parameters driven by impact penetration test data. It constructs a machine learning inversion model by learning and training the relationship between the time history curves of the penetrometer's dynamic response and ground mechanical property parameters in the interpretation model database. After acquiring measured cone resistance and acceleration time history data, these are input into the trained machine learning inversion model to directly obtain the ground mechanical property parameters. Furthermore, requiring only one input-output reverse transmission, the inversion speed can be reduced to the second level. The intelligent algorithm enables accurate interpretation of ground mechanical property parameters such as soil density, elastic modulus, Poisson's ratio, cohesion, and internal friction angle. This invention is rationally conceived and can efficiently, accurately, and intelligently interpret the ground mechanical property parameters of impact-penetrated soil.

Description

Technical Field

[0001] This invention relates to the field of ground mechanical surveying technology, specifically to an intelligent inversion method for ground mechanical parameters driven by impact penetration test data. Background Technology

[0002] The investigation and equipment accessibility assessment of ground bearing capacity in the field require obtaining ground mechanical property parameters such as soil density, elastic modulus, Poisson's ratio, cohesion, and internal friction angle. Conventional ground mechanical surveying techniques include non-contact and contact methods. Non-contact methods mainly include space-based remote sensing and airborne geophysical exploration, while contact methods mainly include conic index instruments, static cone penetration tests, and dynamic cone penetration tests. Non-contact measurements can achieve long-distance, wide-area geological information acquisition, but their non-contact nature makes it difficult to accurately obtain the mechanical information of the ground soil. Furthermore, in complex environments, they are easily affected by weather and human interference, resulting in lower data reliability and resolution. Contact measurements can achieve precise acquisition of soil type and mechanical parameters, but the relevant techniques require manual close-range detection, which cannot meet the needs of remote, real-time surveys.

[0003] To address this, in the invention patent "A Free-Fall Dynamic Penetration Tester," the inventors proposed a complete technical system for remote and rapid ground mechanics surveys in disaster areas and other regions difficult for personnel to access. The penetration tester can be carried by a drone to measure parameters such as penetration resistance, acceleration, and penetration depth of the soil in the field. As is known from basic soil mechanics theory, relevant parameters cannot be directly used to calculate mechanical property parameters such as cohesion and internal friction angle. An interpretation algorithm is needed to further invert the measured resistance, acceleration-time history curves, etc., into ground mechanical property parameters. In the invention patent "A Ground Mechanical Property Parameter Interpretation Method Based on Impact Penetration Testing," the inventors proposed a ground mechanical property parameter interpretation method based on a radial basis function neural network (RBF) model and a multi-objective genetic algorithm (NSGA-II). This method first generates resistance and acceleration-time history curves of the penetrometer under different mechanical parameters through finite element simulation, thereby constructing a database covering a wide range of ground soil mechanical properties. Then, the database is used to train the RBF model, constructing an alternative method for the finite element calculation process. The RBF model can directly convert ground mechanical property parameters into penetrometer resistance and acceleration-time history curves. Subsequently, the NSGA-II algorithm is used to provide the RBF model with different combinations of ground mechanical property parameters, allowing it to generate a large number of resistance and acceleration-time history curves. Then, the time history data curves of the experimental test cone tip resistance and acceleration to be interpreted are matched with the curves generated by the RBF model to find the generated curve with the smallest difference. Finally, the corresponding ground mechanical property parameters of the generated curves are assigned to the curves to be interpreted, completing the inversion of ground mechanical property parameters of the experimentally tested impact-penetrated soil.

[0004] The above interpretation method realizes the conversion of the penetrometer dynamic response time history curve into ground mechanical property parameters, but some problems still exist:

[0005] (1) The technical approach of this method is relatively complex. The interpretation path adopted requires the cooperation of multiple algorithms and models. When problems occur, the debugging is difficult, so it is difficult to apply to the actual interpretation process.

[0006] (2) Two types of errors will occur under this interpretation path: one is the error of the RBF substitution model. The RBF model is trained on the basis of the simulation database, and its accuracy is necessarily lower than that of the original simulation calculation; the other is the optimization error of the NSGA-II algorithm. The objective function and convergence criteria of the algorithm directly affect the accuracy of matching.

[0007] The accumulation of these two types of errors may lead to a decrease in the accuracy of the final ground mechanical property parameter inversion. Summary of the Invention

[0008] To address the technical problems existing in the background art mentioned above, this invention proposes an intelligent inversion method for ground mechanical parameters driven by impact penetration test data. Its concept is reasonable. It establishes machine learning inversion models through different neural network algorithms, and uses the interpretation concept of directly inverting ground mechanical property parameters using resistance and acceleration-time history curves. By utilizing the construction ideas of machine learning inversion models, the specific formats of input and output, and the method of calculating inversion errors, it can efficiently, accurately, and intelligently invert the ground mechanical property parameters of impact-penetrated soil.

[0009] To address the aforementioned technical problems, this invention provides an intelligent inversion method for ground mechanical parameters driven by impact penetration test data, which specifically includes the following steps:

[0010] (1) Numerical simulation calculation

[0011] Five basic ground mechanical property parameters are set: density, elastic modulus, cohesion, internal friction angle, and Poisson's ratio. Density gradient, elastic modulus gradient, cohesion gradient, internal friction angle gradient, and friction coefficient between the penetrometer and the soil are also set to simulate the densification and strengthening phenomenon of soil with depth.

[0012] A numerical simulation model was established using a non-uniform discretized finite element mesh. Then, the mechanical parameters of the penetrating soil were set to simulate the impact penetration test process. A fixed output time step was set to record the time history data of the nodal resistance at the cone tip and the overall acceleration of the penetrator from the moment of contact with the soil to the moment of final stopping. This yielded a set of corresponding time history curves of the penetrator's dynamic response and the ground mechanical property parameters. By changing the mechanical parameters of the penetrating soil, large-scale batch simulation calculations were performed to obtain data for different soil conditions, which effectively corresponded to the physical image of the actual impact penetration test soil in the field.

[0013] (2) Construction of the interpretation model database

[0014] After acquiring a large amount of data from numerical simulation calculations, the simulation data needs to be processed uniformly and stored in an interpreter model database for easy modeling, retrieval, and access. The specific process for constructing the interpreter model database is as follows:

[0015] First, determine the data format and data type for importing the data into the interpretation model database; the time history curves of the dynamic response of each penetrometer and the ground mechanical property parameters are stored in a single CSV file and managed in a table format; the table columns from left to right are time, cone tip resistance, acceleration and ground mechanical property parameters, and labels are set for different data; the ground mechanical property parameter columns from top to bottom are density, density gradient, elastic modulus, elastic modulus gradient, Poisson's ratio, cohesion, cohesion gradient, internal friction angle, internal friction angle gradient and penetrometer surface friction coefficient; since a fixed output time step is set during the numerical simulation calculation in step (1), the time interval of all time history curves is consistent, which is 0.125ms; after determining the storage file, import it into the interpretation model database for management, and set the first-level directory for different soil types and the second-level directory for the initial velocity range;

[0016] (3) Data preprocessing

[0017] Before the data in the interpreter model database is applied to the construction and training of machine learning inversion models, it needs to undergo preprocessing in accordance with the input format of neural network models in order to incorporate all the data in the interpreter model database into an array matrix that conforms to the input format of neural network models.

[0018] (4) Construction of machine learning inversion model

[0019] By using different deep learning neural network algorithms, a machine learning inversion model of drag, acceleration-time history curves to ground mechanical property parameters is directly established. The data used to establish the machine learning inversion model is divided into training set and validation set in a 6:1 ratio, and 210 sets of data are reserved as test set for accuracy verification after training. Then, the machine learning inversion model based on artificial neural network, one-dimensional convolutional neural network and bidirectional recurrent neural network is trained.

[0020] (5) Inversion of ground mechanical property parameters

[0021] The reserved test set is retrieved and input into the machine learning inversion model of the artificial neural network, one-dimensional convolutional neural network and bidirectional recurrent neural network that have been trained. The 10 ground mechanical property parameters corresponding to each set of curves in the inversion test set are retrieved.

[0022] The inversion errors of the ground mechanical property parameters of each area in the test set under different machine learning inversion models are calculated, including the mean absolute percentage error and the normalized root mean square error. The calculation formula is as follows:

[0023] ;

[0024] ;

[0025] In the above formula, The normalized inversion value, The normalized true value is S=210, which is the number of parameters in the inversion.

[0026] (6) Validation of experimental test data

[0027] A machine learning-based intelligent ground mechanical property parameter inversion model was constructed using an interpretation model database. The effectiveness of the machine learning-based intelligent ground mechanical property parameter inversion model was then verified on indoor experimental data. The measured ground mechanical property parameters were then interpreted using the existing machine learning inversion model.

[0028] As a preferred embodiment of the present invention, after establishing the numerical simulation model in step (1), the mechanical parameters of the penetrated soil are set, and the specific method for simulating the impact penetration process is as follows: the coupled Euler-Lagrange finite element method is used to handle large deformation problems without causing mesh distortion, and a three-dimensional CEL model is established to reproduce the impact penetration process. Considering the axisymmetric nature of the penetration, the analysis domain is set to one-quarter of the complete soil mass; the soil is described by an Euler continuum, and its size is consistent with the actual indoor test soil tank; the penetrometer is idealized as a rigid Lagrange body, and an empty Euler mesh is set above the ground to allow the soil to bulge and move upward during the impact process; the numerical simulation... The mesh discretization of the true model is non-uniform to reconcile the local accuracy and overall efficiency of the numerical model; mesh refinement is performed in the penetration region where large strain localization and steep stress gradients are likely to occur, while outside the refined region, a gradually coarsening mesh is used to reduce the total number of meshes while maintaining sufficient resolution of the global response; the bottom of the soil region is fully constrained, while the vertical side boundaries are constrained horizontally to prevent lateral diffusion and avoid material loss through the Euler mesh; symmetry constraints are applied to the two cross-sections of the quarter-3D CEL model to ensure equivalence of the fully axisymmetric problem; at the start of each analysis, the penetrometer is positioned directly above the soil surface and an initial downward velocity is applied to penetrate the soil.

[0029] As a preferred embodiment of the present invention, the batch simulation calculation in step (1) is performed using the iSight simulation analysis workflow automation tool. A set of soil mechanical parameters are randomly generated and input into the same three-dimensional CEL finite element model to calculate the penetration cone tip resistance and acceleration-time history.

[0030] As a preferred embodiment of the present invention, the specific process of step (3) is as follows:

[0031] (3.1) Determine the length of the data with the longest ingress time in the interpretation model database. Based on this, unify the length of all data and supplement the missing parts with fixed values ​​in front of the original data.

[0032] (3.2) Normalize the input time history curve data to unify the scale of the numerical features to a specific range;

[0033] (3.3) Randomize the data extracted from the interpretation model database and shuffle the order of different groups to avoid the situation where the machine learning inversion model only learns a large number of single soils or specific speeds during training, resulting in poor model generalization ability and to avoid the model overfitting to the data order.

[0034] As a preferred embodiment of the present invention, the specific process of step (3.2) is as follows:

[0035] Min-Max normalization is used to linearly scale the data to the [0, 1] interval, as shown in the following formula:

[0036] ;

[0037] In the above formula, These are the original data points; and These are the minimum and maximum values ​​in the dataset, respectively. It is the normalized value.

[0038] As a preferred embodiment of the present invention, the specific process of establishing the machine learning inversion model of drag, acceleration-time history curves to ground mechanical characteristic parameters in step (4) is as follows:

[0039] The machine learning inversion model takes cone tip drag and acceleration time history as inputs and outputs 10 ground mechanical property parameters. It employs three different deep learning algorithms: artificial neural network, one-dimensional convolutional neural network, and bidirectional recurrent neural network. The artificial neural network, by simulating the information processing mechanism of a biological nervous system, constructs multi-layer nonlinear transformation units to approximate complex mapping relationships. It is suitable for the nonlinear inversion problem of penetrometer acceleration, drag time series signals, and soil mechanical parameters. The mathematical model of its basic unit neuron is as follows:

[0040] ;

[0041] In the above formula, w iHere, b is the connection weight, b is the bias term, σ(·) is the activation function, and M is the number of features. When stacking fully connected layers to build a deep architecture, a multi-layer ANN model is written as:

[0042] ;

[0043] In the above formula, h (l) and h (l-1) For the input and output of the l-th layer of a multi-layer ANN model, W (l) and b (l) These are the weight matrix and the corresponding bias vector;

[0044] One-dimensional convolutional neural networks are used to process sequential data. They extract hierarchical temporal features through local receptive fields and weight sharing mechanisms. The convolution operation is defined as follows:

[0045] ;

[0046] In the above formula, τ is the time offset, f(·) is the input signal, and g(·) is the convolution kernel; after discretization, the output mapped by the k-th convolution kernel in the l-th layer is:

[0047] ;

[0048] In the above formula, T is the convolution kernel size. This is the weight vector at time offset τ. For bias terms;

[0049] Bidirectional recurrent neural networks process sequences through both forward and backward time dimensions, extracting more complete global contextual features from time-series data with clear physical causality. For impact penetration testing, the penetrator response at a specific moment depends not only on the past penetration history but also on subsequent responses. The mathematical model of the bidirectional recurrent neural network is as follows:

[0050] ;

[0051] ;

[0052] ;

[0053] In the above formula, W αβ and b β These are the weight matrix and the corresponding bias vector, α={ , , , },β={ , }); U represents the merging method of the two hidden layer vectors.

[0054] As a preferred embodiment of the present invention, the training process of the machine learning inversion model in step (4) is as follows: the hyperparameters of the neural network play a decisive role in the model training efficiency and the final inversion accuracy; after determining the architecture of the neural network model, the hyperparameters of different models are changed by an enumeration strategy, and the model is trained on the existing training set to iteratively obtain different learning models with optimal performance; the Adam optimization algorithm is used to minimize the loss function during the training process, and the loss function is defined as the mean absolute error between the true parameters and the inversion parameters on the validation set; different learning models are trained using Keras 2.9.0 in the Python environment, and the model accuracy and loss on the validation set are recorded as a trend of change with epoch; the model training results under different hyperparameters are combined to determine the optimal hyperparameters and the optimal model accordingly; after determining the optimal model, different mechanical parameters on the test set are inverted, and the inversion accuracy of different neural network algorithms is compared.

[0055] As a preferred embodiment of the present invention, there are two ways to interpret the measured ground mechanical property parameters by means of existing machine learning inversion models in step (6):

[0056] ① The penetration resistance and acceleration-time history curves obtained from the experimental tests are directly input into the machine learning inversion model trained in the interpretation model database, and the mechanical parameters are directly output.

[0057] ② Construct a dataset using experimental test data, transfer the machine learning inversion model trained on the interpretation model database to the dataset constructed using experimental test data, further train it based on the existing weights and bias model parameters, and finally obtain a transfer intelligent inversion model that is more suitable for experimental data.

[0058] As a preferred embodiment of the present invention: the experimental dataset includes 134 sets of corresponding penetration resistance, acceleration-time history curves and ground mechanical property parameters; the soil types include aeolian sandy soil and riparian sandy soil.

[0059] By adopting the above technical solution, the present invention has the following beneficial effects:

[0060] This invention presents a well-conceived intelligent inversion method for ground mechanical parameters driven by impact penetration test data. It is a machine learning method that directly drives the inversion of ground mechanical property parameters using impact penetration test data. By learning and training the relationship between the time history curves of the penetrometer's dynamic response and ground mechanical property parameters in the interpretation model database, a machine learning inversion model is constructed. After acquiring measured cone tip resistance and acceleration time history data, these are input into the trained machine learning inversion model to directly obtain the ground mechanical property parameters, resulting in higher efficiency. It does not require the coordination of multiple algorithms and models, and only requires a single input-output reverse transmission, reducing the inversion speed to the second level. Furthermore, the intelligent algorithm enables accurate interpretation of ground mechanical property parameters such as soil density, elastic modulus, Poisson's ratio, cohesion, and internal friction angle.

[0061] This invention establishes a machine learning inversion model through different neural network algorithms, and uses the concept of directly inverting the ground mechanical property parameters by means of resistance and acceleration-time history curves. It utilizes the construction ideas of machine learning inversion model, the specific format of input and output, and the method of calculating inversion error to efficiently, accurately, and intelligently invert the ground mechanical property parameters of impact-penetrated soil: (1) The inversion model is established by means of machine learning algorithm, and the time history curves of the penetrometer acceleration and cone tip resistance are directly converted into ground mechanical property parameters; (2) The deep learning algorithm ensures the accuracy of the inversion; (3) The intelligent machine learning method automatically interprets the data without the need for manual calculation.

[0062] This invention enables the inversion of dozens of ground mechanical property parameters, including density, density gradient, elastic modulus, elastic modulus gradient, Poisson's ratio, cohesion, cohesion gradient, internal friction angle, internal friction angle gradient, and friction system. It has more inversion parameters, an average inversion error of less than 15%, and a single inversion time of less than 2 seconds. Attached Figure Description

[0063] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0064] Figure 1 This is a flowchart illustrating the construction process of the inversion model involved in the intelligent inversion method for ground mechanical parameters driven by impact penetration test data of the present invention.

[0065] Figure 2 This is a flowchart of the intelligent inversion method for ground mechanical parameters driven by impact penetration test data of the present invention.

[0066] Figure 3This is a schematic diagram of the numerical simulation calculations involved in the intelligent inversion method for ground mechanical parameters driven by impact penetration test data of the present invention.

[0067] Figure 4 This document presents the data format and interface diagram of the interpretation model database involved in the intelligent inversion method for ground mechanical parameters driven by impact penetration test data of this invention.

[0068] Figure 5 This is a diagram of a one-dimensional convolutional neural network machine learning inversion model involved in the intelligent inversion method for ground mechanical parameters driven by impact penetration test data of the present invention.

[0069] Figure 6 This is an error diagram of the validation set for the model training process under different parameters involved in the intelligent inversion method for ground mechanical parameters driven by impact penetration test data of the present invention.

[0070] Figure 7 This is the inversion result of 210 sets of ground mechanical property parameters of the 1d-CNN model involved in the intelligent inversion method of ground mechanical parameters driven by impact penetration test data of the present invention;

[0071] Figure 8 This is a graph showing the average absolute percentage error of ground mechanical property parameter inversion under different machine learning inversion models involved in the intelligent inversion method for ground mechanical parameters driven by impact penetration test data of the present invention.

[0072] Figure 9 This is a diagram showing the inversion results of ground mechanical property parameters under different approaches in the intelligent inversion method for ground mechanical parameters driven by impact penetration test data of this invention (original model: 134 groups; transfer model: 70 groups, and the other 64 groups were used for training). Detailed Implementation

[0073] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0074] The present invention will be further explained below with reference to specific embodiments.

[0075] like Figure 1As shown in the figure, this embodiment provides an intelligent inversion method for ground mechanical parameters driven by impact penetration test data. It utilizes three deep learning algorithms—Artificial Neural Network (ANN), One-Dimensional Convolutional Neural Network (1d-CNN), and Bidirectional Recurrent Neural Network (Bi-RNN)—to establish a machine learning method for directly inverting ground mechanical property parameters such as cone tip drag and acceleration from the time history curves of the penetrometer's dynamic response obtained from impact penetration tests. This achieves rapid and accurate interpretation of ground mechanical property parameters. The specific process is as follows: Figure 1 As shown.

[0076] S100, numerical simulation calculation

[0077] The data in the interpreter model database comes from the numerical calculation results of the CEL (Coupled Euler-Lagrange) finite element model, such as... Figure 2 As shown. Considering more complex actual soil environments, in addition to setting the five basic ground mechanical property parameters of density, elastic modulus, cohesion, internal friction angle, and Poisson's ratio, additional parameters such as density gradient, elastic modulus gradient, cohesion gradient, internal friction angle gradient, and friction coefficient between the penetrometer and the soil are set to simulate the densification and strengthening phenomenon of soil with depth.

[0078] A numerical simulation model was established using a non-uniform discretized finite element mesh. Then, the mechanical parameters of the penetrating soil were set, and the impact penetration test process was simulated. A fixed output time step was set, and time history data such as the nodal resistance at the cone tip and the overall acceleration were recorded from the moment the penetrator touched the soil to the final stopping moment. This yielded a set of corresponding time history curves of the penetrator's dynamic response and ground mechanical property parameters. By changing the mechanical parameters of the penetrating soil, large-scale batch calculations were performed (specifically, the iSight simulation analysis workflow automation tool was used; a set of randomly generated soil mechanical parameters were input into the same three-dimensional CEL finite element model to calculate the penetrator cone tip resistance and acceleration-time history). Ultimately, 7000 sets of data for different soil conditions were obtained, covering various characteristics such as different soil types, different penetration velocities, and different soil mechanical parameters, effectively corresponding to the physical image of actual impact penetration test soil in the field.

[0079] The aforementioned numerical simulation model, after setting the mechanical parameters of the penetrating soil, uses the following method to simulate the impact penetration process: the Coupled Eulerian-Lagrange (CEL) finite element method can handle large deformation problems without causing mesh distortion, and is particularly suitable for simulating soil-structure interaction phenomena involving penetration or significant soil deformation; a three-dimensional CEL model is established to reproduce the impact penetration process, and considering the axisymmetric nature of the penetration, the analysis domain is set to one-quarter of the complete soil mass; the soil is described by an Eulerian continuum, the size of which is consistent with the actual indoor test soil tank; the penetrator is idealized as a rigid Lagrange body; in addition, an empty Eulerian mesh is set above the ground to allow the soil to bulge and move upward during the impact; the mesh discretization of the numerical simulation model is non-uniform to coordinate the local accuracy and overall efficiency of the numerical model; the mesh is refined in the penetration region where large strain localization and steep stress gradients are prone to occur, and outside the refined region, a gradually coarser mesh is used to reduce the total number of meshes while maintaining sufficient resolution of the global response. The bottom of the soil region is fully constrained, while the vertical lateral boundaries are constrained horizontally to prevent lateral diffusion and avoid material loss through the Euler mesh. Symmetry constraints are applied to two sections of the quarter-3D CEL model to ensure equivalence for a fully axisymmetric problem. At the start of each analysis, the penetrometer is positioned directly above the soil surface, applying an initial downward velocity to penetrate the soil.

[0080] S200, Interpretation Model Database Construction

[0081] After acquiring a large amount of data from numerical simulation calculations, the simulation data needs to be processed uniformly and stored in an interpreter model database for easy modeling retrieval and access. The specific process of database construction involves first determining the data format and data type for importing into the interpreter model database, such as... Figure 3 As shown; the dynamic response time history curves of each penetrometer and the ground mechanical property parameters are stored in a single CSV file and managed in a table format; the table columns from left to right are time, cone tip resistance, acceleration and ground mechanical property parameters, and labels are set for different data; the ground mechanical property parameter columns from top to bottom are density, density gradient, elastic modulus, elastic modulus gradient, Poisson's ratio, cohesion, cohesion gradient, internal friction angle, internal friction angle gradient and penetrometer surface friction coefficient; since a fixed output time step is set during the numerical simulation calculation of step (1), the time interval of all time history curves is consistent, which is 0.125 ms; after determining the storage file, it is imported into the interpretation model database software for management, and the first-level directory is set for different soil types, and the second-level directory is the initial velocity range.

[0082] S300, Data Preprocessing

[0083] Before the data in the interpreter model database can be used to build and train machine learning inversion models, it needs to undergo preprocessing that conforms to the input format of neural network models. This mainly includes operations such as length unification, normalization, and randomization, so that all the data in the interpreter model database can be incorporated into an array matrix that conforms to the input format of neural network models.

[0084] ① Due to differences in the initial velocity of the penetrometer and the soil mechanical parameters, the time history curve lengths obtained from different simulation calculation examples vary. To ensure consistency in the dimensions of the input array matrix, the length of the longest penetration time in the interpretation model database was determined. Based on this, all data lengths were standardized, and any missing values ​​were supplemented with fixed values ​​(resistance: 0N, acceleration: -9.8 m / s²) before the original data. 2 ).

[0085] ② The input time-history curve data is normalized. The core objective is to unify the scale of numerical features to a specific range, eliminating problems such as slower model convergence, severely degraded algorithm performance, and unstable model weight updates caused by differences in the units and ranges of different features. Specifically, Min-Max normalization is used to linearly scale the data to the [0,1] interval, as shown in the following formula:

[0086] ;

[0087] in, These are the original data points. and These are the minimum and maximum values ​​in the dataset, respectively. It is the normalized value.

[0088] ③ Randomize the data extracted from the interpretation model database, shuffle the order of different groups, avoid the situation where the machine learning inversion model only learns a large number of single soils or specific speeds during training, resulting in poor model generalization ability, and can avoid the model overfitting to the data order to a certain extent.

[0089] In summary, data length consistency, normalization, and randomization are three indispensable steps before building an efficient machine learning inversion model. They respectively solve the problems of inconsistent data shape, inconsistent data scale, and biased data order, which helps to train a powerful and robust machine learning inversion model.

[0090] S400, Machine Learning Inversion Model Construction

[0091] The unique feature of this invention lies in its use of different deep learning neural network algorithms to directly establish a machine learning inversion model of drag, acceleration-time history curves, and ground mechanical property parameters. The data used to build the machine learning inversion model is divided into training and validation sets in a 6:1 ratio, with 210 sets of data retained as a test set for post-training accuracy verification. Taking a one-dimensional convolutional neural network as an example... Figure 4 As shown, the model includes an input layer, a one-dimensional convolutional layer, a max-pooling layer, a fully connected layer, and an output layer. By changing the parameters in different layers, the training of the model is optimized, improving the inversion accuracy for different ground mechanical property parameters.

[0092] The specific process of establishing a machine learning inversion model for drag and acceleration-time history curves to ground mechanical property parameters is as follows: the input of the machine learning inversion model is the cone drag and acceleration time history, and the output is 10 ground mechanical property parameters. Three different deep learning algorithms with different architectures are employed: Artificial Neural Network (ANN), One-Dimensional Convolutional Neural Network (1d-CNN), and Bidirectional Recurrent Neural Network (Bi-RNN). Among them, ANN, by simulating the information processing mechanism of a biological nervous system, constructs multi-layer nonlinear transformation units to approximate complex mapping relationships, and is suitable for the nonlinear inversion problem of penetrometer acceleration, drag time series signals, and soil mechanical parameters; its basic unit neuron mathematical model is as follows:

[0093] ;

[0094] In the formula, w i Here, b is the connection weight, b is the bias term, σ(·) is the activation function, and M is the number of features. When stacking fully connected layers to build a deep architecture, a multi-layer ANN model can be written as:

[0095] ;

[0096] In the formula, h (l) and h (l-1) For the input and output of the l-th layer of a multi-layer ANN model (i.e., the input and output of the l-th layer in a multi-layer ANN model constructed by stacking multiple fully connected layers), W (l) and b (l) These are the weight matrix and the corresponding bias vector.

[0097] 1d-CNN is a variant of convolutional neural networks specifically designed for processing sequential data. It extracts hierarchical temporal features through local receptive fields and weight sharing mechanisms. The convolution operation is defined as:

[0098] ;

[0099] In the formula, τ is the time offset, f(·) is the input signal, and g(·) is the convolution kernel. After discretization, the output mapped by the k-th convolution kernel in the l-th layer is:

[0100] ;

[0101] In the formula, T is the convolution kernel size. This is the weight vector at time offset τ. This is the bias term. The local connectivity of 1d-CNN makes it particularly suitable for capturing local waveform patterns (such as peaks, rising edges, and oscillation decay) in impact penetration signals. The convolutional kernel slides along the time axis (with a step size of 1), enabling it to automatically learn the dynamic response characteristics at different time scales, and then use these characteristics to invert relevant parameters such as soil strength.

[0102] Bi-RNN is a variant of recurrent neural networks that processes sequences through both forward and backward time dimensions, enabling the extraction of more complete global contextual features from time-series data with clear physical causality. For impact penetration testing, the penetrator response at a specific moment depends not only on the existing penetration history but also on subsequent responses. For example, the physical meaning of the initial peak drag needs to be inferred in reverse by considering the subsequent drag decay pattern. Therefore, the architecture of Bi-RNN helps extract richer physical information about the penetration process; its mathematical model is as follows:

[0103] ;

[0104] ;

[0105] ;

[0106] In the formula, W αβ and b β These are the weight matrix and the corresponding bias vector (here α = { , , , }, β = { , }). U represents the merging method of two hidden layer vectors, usually "concatenation". In summary, the input processing approaches of ANN, 1d-CNN, and Bi-RNN are respectively: ① Fully connected, ignoring the temporal nature of the data; ② Local perception of convolutional kernels, focusing on local waveform features; ③ Recurrent connections, establishing complete temporal dependencies. Among them, the Bi-RNN architecture is naturally suitable for building learning models with time histories as input features. Its capture of bidirectional dependency features of temporal data helps to understand the complex interaction mechanism between the soil and the penetrometer during the penetration process, thus demonstrating significant advantages in both physical meaning and data processing.

[0107] The training process of the aforementioned machine learning inversion model is as follows: the hyperparameters of the neural network play a decisive role in the model training efficiency and the final inversion accuracy; after determining the architecture of the neural network model, an enumeration strategy is used to change the number of neural network layers, number of neurons, learning rate, batch size, and activation function of different models, and training is performed on the existing training set to iteratively obtain different learning models with optimal performance; the Adam optimization algorithm is used in the training process to minimize the loss function, which is defined as the mean absolute error between the true parameters and the inversion parameters on the validation set; Keras is used in the Python environment. 2.9.0 (a library directly callable in Python) was used to train different learning models, recording the trends of model accuracy and loss on the validation set with epochs. Based on the training results under different hyperparameters, the optimal hyperparameters were determined as follows: The ANN model includes an input layer, one fully connected layer, and an output layer, with 100 neurons in the fully connected layer and a batch size of 15; the 1d-CNN model includes an input layer, one one-dimensional convolutional layer, one one-dimensional max-pooling layer, and an output layer, with 75 convolutional kernels in the one-dimensional convolutional layer and a batch size of 15; the Bi-RNN model includes an input layer, one bidirectional LSTM layer, one unidirectional LSTM layer, and an output layer, with 25 units in each LSTM layer and a batch size of 35. The output layer activation function for all three learning models is 'sigmoid' to match the 0-1 output normalization range, and all models undergo flattening from a three-dimensional array to a two-dimensional array before the output layer. The learning rate is uniformly set to 0.001. After determining the optimal model, inversion was performed on different mechanical parameters of the test set, and the inversion accuracy of different neural network algorithms was compared.

[0108] The training error of the model on the validation set under different parameters is as follows: Figure 5 As shown in the figure, the left figure represents the validation set error of the artificial neural network model with different neurons. It can be seen that when the model has two fully connected layers with 75 neurons in each layer, the final mean absolute error (MAE) achieved by the model on the validation set is the lowest. The right figure represents the validation set error of the convolutional neural network model when adjusting the number of layers and the properties of each layer. It is found that the final MAE of the model with only one convolutional layer is the lowest. The above parameter tuning process ensures the improvement of the model's final prediction accuracy. During training, the accuracy of the validation set increases with the number of training steps, while the model error loss decreases with the number of training steps, proving that the model effectively captures information from the time-series data for prediction. The training loss history shows that the model does not exhibit problems such as "overfitting" or "underfitting".

[0109] S500, Ground Mechanical Property Parameter Inversion

[0110] After training the machine learning inversion models using artificial neural networks (ANN), one-dimensional convolutional neural networks (1d-CNN), and bidirectional recurrent neural networks (Bi-RNN), a reserved test set is retrieved and input into these models. The test set contains 10 ground mechanical property parameters corresponding to each curve. The prediction results of the 1d-CNN model are as follows: Figure 6 As shown in the figure, the inversion results of 210 sets of ground mechanical property parameters in the test set are compared with the true values. It can be seen that the predicted results are very close to the true values ​​for all parameters, and the predictions for the 210 sets of data can effectively match the changing trends of the true parameters.

[0111] The inversion errors of the ground mechanical property parameters of each area in the test set under different machine learning inversion models are calculated, including the mean absolute percentage error and the normalized root mean square error. The calculation formula is as follows:

[0112] ;

[0113] ;

[0114] in, and For the normalized inversion values ​​and the true values, S=210 is the number of parameters for the inversion.

[0115] Table 1 shows the RMSE of the ground mechanical property parameters obtained from the test set of 210 data points. Figure 7 The MAPE (Modular Optimization Equation) of five basic ground mechanical property parameters (density, elastic modulus, cohesion, internal friction angle, and Poisson's ratio) on a test set of 210 datasets under three machine learning inversion models is shown. It is evident that all three machine learning inversion models can effectively achieve accurate inversion of ground mechanical property parameters. Comparatively, the Bi-RNN model demonstrates better prediction accuracy than the ANN and 1d-CNN models. Calculations show that the average MAPE for the five ground mechanical property parameters under the three models is less than 15%. Furthermore, the inversion times for the ANN, 1d-CNN, and Bi-RNN machine learning inversion models are 0.19s, 0.17s, and 1.88s, respectively, indicating that a single inversion takes less than 2 seconds.

[0116] Table 1: Root mean square error of 10 ground mechanical property parameters retrieved under different machine learning inversion models

[0117]

[0118] S600, experimental test data verification

[0119] After constructing a machine learning-based intelligent ground mechanical property parameter inversion model using an interpretation model database, the model's effectiveness was verified on indoor experimental data. There are two approaches to interpreting measured ground mechanical property parameters using existing machine learning inversion models: one is to directly input the penetration drag and acceleration-time history curves from experimental tests into the inversion model trained in the interpretation model database, directly outputting the mechanical parameters; the other is to construct a dataset using experimental test data, transfer the machine learning inversion model trained in the interpretation model database to this dataset, and further train it based on existing weights and biases, ultimately obtaining a more suitable transferable intelligent inversion model for experimental data.

[0120] The experimental dataset used in this invention includes 134 sets of corresponding penetration resistance, acceleration-time history curves, and ground mechanical property parameters. The soil types include aeolian sandy soil and riparian sandy soil. The ground mechanical property parameters of the experimental dataset are retrieved using the two methods mentioned above and different machine learning inversion models. Figure 8 The results of ground mechanical property parameter inversion obtained using two different approaches are presented. It is evident that the effectiveness of parameter inversion is limited when directly using the original model trained on the interpreter model database. However, after transfer learning, the accuracy of parameter prediction significantly improves. Table 2 shows the inversion errors of the five basic ground mechanical property parameters under different models and approaches. Table 2 reveals the substantial improvement in accuracy after using the transfer model, and this improvement is consistent across different machine learning inversion models. Ultimately, the mean MAPE of the five ground mechanical property parameters obtained using the transfer model can reach less than 10%.

[0121] Figure 9 The inversion results of ground mechanical property parameters tested under different approaches (original model: 134 groups; transfer model: 70 groups, and the other 64 groups were used for training).

[0122] Table 2: Inversion errors of experimentally tested ground mechanical property parameters under different machine learning inversion models and approaches

[0123]

[0124] This invention establishes a machine learning inversion model using different neural network algorithms. It utilizes the concept of directly inverting ground mechanical property parameters using resistance and acceleration-time history curves. By employing the construction ideas of the machine learning inversion model, the specific input and output formats, and the method for calculating inversion errors, it can efficiently, accurately, and intelligently invert the ground mechanical property parameters of impact-penetrated soil.

[0125] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent inversion of ground mechanical parameters driven by impact penetration test data, characterized in that, The inversion method specifically includes the following steps: (1) Numerical simulation calculation Five basic ground mechanical property parameters are set: density, elastic modulus, cohesion, internal friction angle, and Poisson's ratio. Density gradient, elastic modulus gradient, cohesion gradient, internal friction angle gradient, and friction coefficient between the penetrometer and the soil are also set to simulate the densification and strengthening phenomenon of soil with depth. A numerical simulation model was established using a non-uniform discretized finite element mesh. Then, the mechanical parameters of the penetrating soil were set to simulate the impact penetration test process. A fixed output time step was set to record the time history data of the nodal resistance at the cone tip and the overall acceleration of the penetrator from the moment of contact with the soil to the moment of final stopping. This yielded a set of corresponding time history curves of the penetrator's dynamic response and the ground mechanical property parameters. By changing the mechanical parameters of the penetrating soil, large-scale batch simulation calculations were performed to obtain data for different soil conditions, which effectively corresponded to the physical image of the actual impact penetration test soil in the field. (2) Construction of the interpretation model database After acquiring a large amount of data from numerical simulation calculations, the simulation data needs to be processed uniformly and stored in an interpreter model database for easy modeling, retrieval, and access. The specific process for constructing the interpreter model database is as follows: First, determine the data format and data type for importing the data into the interpretation model database; the time history curves of the dynamic response of each penetrometer and the ground mechanical property parameters are stored in a single CSV file and managed in a table format; the table columns from left to right are time, cone tip resistance, acceleration and ground mechanical property parameters, and labels are set for different data; the ground mechanical property parameter columns from top to bottom are density, density gradient, elastic modulus, elastic modulus gradient, Poisson's ratio, cohesion, cohesion gradient, internal friction angle, internal friction angle gradient and penetrometer surface friction coefficient; since a fixed output time step is set during the numerical simulation calculation in step (1), the time interval of all time history curves is consistent, which is 0.125ms; after determining the storage file, import it into the interpretation model database for management, and set the first-level directory for different soil types and the second-level directory for the initial velocity range; (3) Data preprocessing Before the data in the interpreter model database is applied to the construction and training of machine learning inversion models, it needs to undergo preprocessing in accordance with the input format of neural network models in order to incorporate all the data in the interpreter model database into an array matrix that conforms to the input format of neural network models. (4) Construction of machine learning inversion model By using different deep learning neural network algorithms, a machine learning inversion model of drag, acceleration-time history curves to ground mechanical property parameters is directly established. The data used to establish the machine learning inversion model is divided into training set and validation set in a 6:1 ratio, and 210 sets of data are reserved as test set for accuracy verification after training. Then, the machine learning inversion model based on artificial neural network, one-dimensional convolutional neural network and bidirectional recurrent neural network is trained. (5) Inversion of ground mechanical property parameters The reserved test set is retrieved and input into the machine learning inversion model of the artificial neural network, one-dimensional convolutional neural network and bidirectional recurrent neural network that have been trained. The 10 ground mechanical property parameters corresponding to each set of curves in the inversion test set are retrieved. The inversion errors of the ground mechanical property parameters of each area in the test set under different machine learning inversion models are calculated, including the mean absolute percentage error and the normalized root mean square error. The calculation formula is as follows: ； ； In the above formula, The normalized inversion value, The normalized true value is S=210, which is the number of parameters in the inversion. (6) Validation of experimental test data A machine learning-based intelligent ground mechanical property parameter inversion model was constructed using an interpretation model database. The effectiveness of the machine learning-based intelligent ground mechanical property parameter inversion model was then verified on indoor experimental data. The measured ground mechanical property parameters were then interpreted using the existing machine learning inversion model.

2. The intelligent inversion method for ground mechanical parameters driven by impact penetration test data as described in claim 1, characterized in that, After establishing the numerical simulation model in step (1), the mechanical parameters of the penetrating soil are set. The specific method for simulating the impact penetration process is as follows: the coupled Euler-Lagrange finite element method is used to handle large deformation problems without causing mesh distortion. A three-dimensional CEL model is established to reproduce the impact penetration process. Considering the axisymmetric nature of the penetration, the analysis domain is set to one-quarter of the complete soil mass. The soil is described by an Euler continuum, and its size is consistent with the actual indoor test soil tank. The penetrator is idealized as a rigid Lagrange body, and an empty Euler mesh is set above the ground to allow the soil to bulge and move upward during the impact. The mesh of the numerical simulation model is as follows: Discretization is non-uniform to reconcile the local accuracy and overall efficiency of the numerical model; mesh refinement is performed in penetration regions prone to large strain localization and steep stress gradients, while outside the refined regions, a gradually coarsening mesh is used to reduce the total number of meshes while maintaining sufficient resolution of the global response; the bottom of the soil region is fully constrained, while the vertical lateral boundaries are constrained horizontally to prevent lateral diffusion and avoid material loss through the Euler mesh; symmetry constraints are applied to the two cross-sections of the quarter-3D CEL model to ensure equivalence of the fully axisymmetric problem; at the start of each analysis, the penetrometer is positioned directly above the soil surface, applying an initial downward velocity to penetrate the soil.

3. The intelligent inversion method for ground mechanical parameters driven by impact penetration test data as described in claim 1, characterized in that, The batch simulation calculation in step (1) uses the iSight simulation analysis workflow automation tool to randomly generate a set of soil mechanical parameters and input them into the same three-dimensional CEL finite element model to calculate the penetration tip resistance and acceleration-time history.

4. The intelligent inversion method for ground mechanical parameters driven by impact penetration test data as described in claim 1, characterized in that, The specific process of step (3) is as follows: (3.1) Determine the length of the data with the longest ingress time in the interpretation model database. Based on this, unify the length of all data and supplement the missing parts with fixed values ​​in front of the original data. (3.2) Normalize the input time history curve data to unify the scale of the numerical features to a specific range; (3.3) Randomize the data extracted from the interpretation model database and shuffle the order of different groups to avoid the situation where the machine learning inversion model only learns a large number of single soils or specific speeds during training, resulting in poor model generalization ability and to avoid the model overfitting to the data order.

5. The intelligent inversion method for ground mechanical parameters driven by impact penetration test data as described in claim 4, characterized in that, The specific process of step (3.2) is as follows: Min-Max normalization is used to linearly scale the data to the [0, 1] interval, as shown in the following formula: ； In the above formula, These are the original data points; and These are the minimum and maximum values ​​in the dataset, respectively. It is the normalized value.

6. The intelligent inversion method for ground mechanical parameters driven by impact penetration test data as described in claim 1, characterized in that, The specific process of establishing the machine learning inversion model of drag, acceleration-time history curves to ground mechanical property parameters in step (4) is as follows: The machine learning inversion model takes cone tip drag and acceleration time history as inputs and outputs 10 ground mechanical property parameters. It employs three different deep learning algorithms: artificial neural network, one-dimensional convolutional neural network, and bidirectional recurrent neural network. The artificial neural network, by simulating the information processing mechanism of a biological nervous system, constructs multi-layer nonlinear transformation units to approximate complex mapping relationships. It is suitable for the nonlinear inversion problem of penetrometer acceleration, drag time series signals, and soil mechanical parameters. The mathematical model of its basic unit neuron is as follows: ； In the above formula, w i Here, b is the connection weight, b is the bias term, σ(·) is the activation function, and M is the number of features. When stacking fully connected layers to build a deep architecture, a multi-layer ANN model is written as: ； In the above formula, h (l) and h (l-1) W represents the input and output of the l-th layer of a multi-layer ANN model. (l) and b (l) These are the weight matrix and the corresponding bias vector; One-dimensional convolutional neural networks are used to process sequential data. They extract hierarchical temporal features through local receptive fields and weight sharing mechanisms. The convolution operation is defined as follows: ； In the above formula, τ is the time offset, f(·) is the input signal, and g(·) is the convolution kernel; after discretization, the output mapped by the k-th convolution kernel in the l-th layer is: ； In the above formula, T is the convolution kernel size. This is the weight vector at time offset τ. For bias terms; Bidirectional recurrent neural networks process sequences through both forward and backward time dimensions, extracting more complete global contextual features from time-series data with clear physical causality. For impact penetration testing, the penetrator response at a specific moment depends not only on the past penetration history but also on subsequent responses. The mathematical model of the bidirectional recurrent neural network is as follows: ； ； ； In the above formula, W αβ and b β These are the weight matrix and the corresponding bias vector, α={ , , , },β={ , }); U represents the merging method of the two hidden layer vectors.

7. The intelligent inversion method for ground mechanical parameters driven by impact penetration test data as described in claim 1, characterized in that, The training process of the machine learning inversion model in step (4) is as follows: the hyperparameters of the neural network play a decisive role in the model training efficiency and the final inversion accuracy; after determining the architecture of the neural network model, the hyperparameters of different models are changed by an enumeration strategy, and the model is trained on the existing training set to iteratively obtain different learning models with the best performance. The training process uses the Adam optimization algorithm to minimize the loss function, which is defined as the mean absolute error between the true parameters and the inverted parameters on the validation set. Different learning models are trained using Keras 2.9.0 in the Python environment, and the model accuracy and loss on the validation set are recorded as a function of epochs. The training results of the models under different hyperparameters are combined to determine the optimal hyperparameters and the optimal model accordingly. After determining the optimal model, inversion is performed on different mechanical parameters on the test set, and the inversion accuracy of different neural network algorithms is compared.

8. The intelligent inversion method for ground mechanical parameters driven by impact penetration test data as described in claim 1, characterized in that, There are two approaches to interpreting measured ground mechanical property parameters using existing machine learning inversion models in step (6): ① The penetration resistance and acceleration-time history curves obtained from the experimental tests are directly input into the machine learning inversion model trained in the interpretation model database, and the mechanical parameters are directly output. ② Construct a dataset using experimental test data, transfer the machine learning inversion model trained on the interpretation model database to the dataset constructed using experimental test data, further train it based on the existing weights and bias model parameters, and finally obtain a transfer intelligent inversion model that is more suitable for experimental data.

9. The intelligent inversion method for ground mechanical parameters driven by impact penetration test data as described in claim 8, characterized in that: The experimental dataset includes 134 sets of corresponding penetration resistance, acceleration-time history curves, and ground mechanical property parameters; the soil types include aeolian sandy soil and riparian sandy soil.

Images