Intelligent inversion system for geotechnical parameters based on deep learning

Abstract

The application discloses a kind of intelligent inversion system of rock-soil mass parameter based on deep learning, it is related to intelligent exploration technical field, multi-source data acquisition module collects multidimensional data, generates original data set;Geological data preprocessing module washes and standardizes original data set;Deep learning model construction module adopts hybrid convolution-graph neural network to extract rock-soil mass space-time correlation characteristics, combined with the generalization ability of model optimization model is optimized, and generates high-precision inversion model;Parameter intelligent inversion module is based on model and real-time data, generates parameter inversion report;Dynamic feedback control module uses reinforcement learning to adaptively adjust data acquisition strategy.The application significantly improves parameter inversion efficiency and precision by modularization cooperation and adaptive learning mechanism, reduces artificial dependence, and provides intelligent support for engineering design and safety evaluation under complex geological conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent exploration technology, specifically to an intelligent inversion system for soil and rock parameters based on deep learning. Background Technology

[0002] Geotechnical parameter inversion is a key technology in geological exploration and engineering design. Its goal is to accurately obtain the mechanical parameters of soil and rock masses by combining field monitoring data with theoretical models. These parameters are of great guiding significance for engineering safety assessment, disaster early warning, and construction design. However, with the increasing complexity of engineering projects, traditional parameter inversion methods face severe challenges in terms of accuracy, efficiency, and adaptability.

[0003] Traditional methods mainly rely on empirical formulas, finite element analysis, and statistically based manual inversion techniques. These methods have the following problems: low computational efficiency: traditional finite element analysis methods have high computational complexity when processing large-scale soil and rock data, making it difficult to meet real-time requirements; limited accuracy: statistically based manual inversion methods are highly dependent on data quality and prior knowledge, making it difficult to cope with the spatiotemporal correlations of multi-source heterogeneous data (including mechanical, hydrological, and remote sensing data) under complex geological conditions; insufficient adaptability: existing systems have shortcomings in data preprocessing, model generalization ability, and real-time feedback, making it difficult to dynamically adjust monitoring strategies and model parameters.

[0004] In existing technologies, numerical simulation-based inversion methods (such as genetic algorithms and particle swarm optimization) can partially improve computational efficiency, but they rely on simplified models and static data inputs, making it difficult to handle the spatiotemporal correlations of multi-source heterogeneous data. Furthermore, traditional systems lack modular collaboration and adaptive learning mechanisms, leading to a disconnect between inversion results and engineering decisions, failing to meet the high-precision, real-time engineering requirements under complex geological conditions.

[0005] Geotechnical parameter inversion is a crucial step in geological exploration and engineering design. Its core objective is to accurately obtain the mechanical parameters of soil and rock masses (such as elastic modulus, permeability coefficient, and shear strength) by combining field monitoring data with theoretical models. With increasing engineering complexity (including deep mining, tunnel excavation, and slope stability), traditional methods have proven inadequate in dealing with heterogeneous soil and rock masses, multi-field coupling effects, and dynamic geological changes, leading to significant deviations between parameter inversion results and actual engineering requirements.

[0006] Therefore, there is an urgent need for an intelligent inversion system that integrates multi-source data fusion, deep learning modeling, and dynamic feedback control to overcome the limitations of existing technologies. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent inversion system for geotechnical parameters based on deep learning. By using multi-source data fusion, deep learning modeling, and dynamic feedback control technology, it overcomes the shortcomings of existing inversion systems such as low computational efficiency, limited accuracy, and insufficient adaptability, thereby significantly improving the accuracy, real-time performance, and robustness of parameter inversion. It is particularly suitable for engineering design and safety assessment scenarios under complex geological conditions such as deep mining, tunnel excavation, and slope stability.

[0008] To achieve the above objectives, this invention provides a deep learning-based intelligent inversion system for soil and rock parameters, comprising the following main modules: a multi-source data acquisition module, a geological data preprocessing module, a deep learning model construction module, an intelligent parameter inversion module, and a dynamic feedback control module.

[0009] The multi-source data acquisition module acquires sensor network data, including mechanical, hydrological and geological environment data of soil and rock, through the sensor network sub-module, and remote sensing data, including surface deformation and geological structure data, through the remote sensing data sub-module. It then aligns the sensor network data and the remote sensing data in time and space to generate the original dataset.

[0010] The geological data preprocessing module uses wavelet transform to denoise the original dataset, and then uses a feature fusion algorithm to fuse the features of the denoised original data. Finally, it is standardized to generate a high-quality dataset.

[0011] The deep learning model building module uses a hybrid convolutional-graph neural network to extract the spatiotemporal features of soil and rock in the high-quality dataset, and combines the adversarial training strategy of generative adversarial network to optimize the generalization ability of the pre-trained model and generate a high-precision inversion model; the pre-trained model refers to the initial model obtained by the hybrid convolutional-graph neural network and generative adversarial network.

[0012] The intelligent parameter inversion module dynamically predicts the elastic modulus, permeability coefficient, and shear strength parameters of soil and rock mass based on the inversion model and the high-quality dataset, and generates a parameter inversion report.

[0013] The dynamic feedback control module uses reinforcement learning algorithms to adaptively adjust the frequency of data acquisition from the sensor network, thereby achieving closed-loop optimization of the system.

[0014] Furthermore, the sensor network submodule consists of multiple types of sensors, deployed in the target soil and rock area according to a three-dimensional grid topology;

[0015] The various types of sensors include strain sensors, pore water pressure sensors, and ground-penetrating radar, which respectively collect sensor network data including mechanical, hydrological, and geological environmental data of the rock and soil.

[0016] Furthermore, the spatiotemporal alignment is achieved by combining timestamp matching with spatial interpolation, and the steps are as follows:

[0017] First, the timestamp matching is implemented as follows:

[0018] Select data that meets the requirements. The data ensures that the data is within the same time window. For remote sensing of time, For sensor network time, Maximum allowable time difference;

[0019] Secondly, the spatial interpolation technique unifies remote sensing data and sensor network data with different spatial resolutions into the same spatial grid. The spatial interpolation method is bilinear interpolation, and its formula z(·) is as follows:

[0020]

[0021] Where (x, y) are the coordinates that need to be unified, and z i Given remote sensing and sensor network data, w i For interpolation weights;

[0022] Finally, spatiotemporal alignment is performed, that is, aligning the time points of each remote sensing data and sensor network data. Given a spatial point (x, y), calculate the fused data using the following formula:

[0023]

[0024] in, For the merged data, For the remote sensing data, For the sensor network data, These are the weighting coefficients.

[0025] Furthermore, the feature fusion algorithm uses a long short-term memory network to extract temporal features from the original dataset;

[0026] The standardization process employs the Z-Score normalization algorithm to unify the data dimensions of the original dataset, thereby generating a high-quality dataset.

[0027] Furthermore, the deep learning model building module includes a spatiotemporal feature extraction submodule and an adversarial training submodule;

[0028] The spatiotemporal feature extraction submodule extracts spatiotemporal features using a hybrid convolutional-graph neural network.

[0029] The hybrid convolutional-graph neural network includes a convolutional neural network branch to extract local spatial pattern features and a graph neural network branch to extract unstructured topological relationship features of soil and rock. After extraction, the spatiotemporal features are dynamically weighted through an attention mechanism to generate a global feature vector.

[0030] The adversarial training submodule, based on the generative adversarial network (GAN) architecture, simulates the distribution of noisy data in the original dataset and improves the robustness of the hybrid convolutional-graph neural network to noisy data by minimizing the Wasserstein distance.

[0031] Furthermore, the parameter intelligent inversion module includes a real-time data stream processing submodule, a multi-task prediction submodule, an uncertainty quantification submodule, and a dynamic report generation submodule;

[0032] The real-time data stream processing submodule uses sliding window technology and time series segmentation algorithm to divide the input data at the current moment into time series segments. The input data is sensor network data and remote sensing data collected by the multi-source data acquisition module and processed by the geological data preprocessing module.

[0033] The multi-task prediction submodule combines the trained high-precision inversion model to perform multi-task prediction of soil and rock parameters; the multi-task includes prediction of soil and rock elastic modulus, permeability coefficient, and shear strength parameters.

[0034] The uncertainty quantification submodule generates confidence intervals for parameter predictions based on Monte Carlo Dropout technology, thus quantifying the credibility.

[0035] The dynamic report generation submodule automatically marks parameters that exceed the preset threshold for soil and rock parameters as abnormal parameters and generates a visual inversion report.

[0036] Furthermore, the inversion report includes:

[0037] Predicted values ​​of parameters: elastic modulus, permeability coefficient, and shear strength;

[0038] Confidence interval: The confidence interval for parameter prediction;

[0039] Anomaly marker: Parameters that exceed the preset threshold for soil and rock parameters;

[0040] Visual charts: Line chart: The trend of parameter changes over time.

[0041] Furthermore, the dynamic feedback control module is based on a reinforcement learning algorithm and adjusts the sensor network data acquisition frequency according to the inversion error. The inversion error is obtained by calculating the difference between the predicted and actual values ​​of the soil and rock parameters by the multi-task prediction submodule.

[0042] Furthermore, the reinforcement learning consists of a state space, an action space, a reward function, and a policy network;

[0043] The state space is defined as the current inversion error e of the system. t Resource utilization rate u t Sensor network health status h t , represented as s t =[e t ,u t ,h t ];

[0044] The resource utilization rate represents the resource utilization status of the system at the current moment, including the utilization rate of computing resources, storage resources and communication resources. The computing resource utilization rate includes the utilization rate of CPU and GPU, the storage resource utilization rate includes the utilization rate of memory and hard disk, and the communication resource utilization rate is the utilization rate of network bandwidth.

[0045] The sensor network health status includes sensor online rate, communication packet loss rate, and remaining battery power;

[0046] The action space: dynamically adjusts the data acquisition frequency f. t ∈{low frequency (1Hz), mid frequency (10Hz), high frequency (100Hz)};

[0047] The reward function balances inversion accuracy and resource consumption, and its formula is:

[0048]

[0049] in, f represents the square of the inversion error; t Indicates the data acquisition frequency; α, β, γ are weighting coefficients; For indicator functions, The value is determined by the preset inversion error threshold, the upper limit of resource utilization, and the lower limit of sensor network health status. When the inversion error is greater than the inversion error threshold, the resource utilization is greater than the upper limit of resource utilization, or the sensor network health status is lower than the lower limit of sensor network health status, the value is 0; otherwise, the value is 1.

[0050] The policy function employs a Deep Q-Network (DQN), with input state s. t Output the Q-value for each action, and select the action corresponding to the largest Q-value, as shown below:

[0051]

[0052] Among them, a t For the selected action, s tLet Q be the state, θ be the parameters of the deep Q-network, and θ be the parameters of the deep Q-network, updated through temporal difference (TD) error, as follows:

[0053]

[0054] Where, θ - The target network parameters for the deep Q-network are synchronized periodically. Indicates that in state s t+1 Choose the action that maximizes the Q value, r t For value function, It is the gradient, Q(s) t ,a t ;θ) represents the state s t And action a t The Q value under the given condition, where η is the learning rate.

[0055] Compared with the prior art, the advantages of the present invention are as follows:

[0056] (1) Multi-source data fusion and efficient preprocessing: This invention uses distributed sensor networks and remote sensing technology to collect multi-dimensional data in real time, and combines wavelet transform and feature fusion algorithms to achieve efficient data preprocessing. Compared with traditional methods that rely on a single data source and manual processing, this invention can integrate multi-source heterogeneous data such as mechanics, hydrology, and geological environment, significantly improving data quality and inversion accuracy.

[0057] (2) Deep Learning Modeling and High-Precision Inversion: This invention employs a Hybrid Convolutional-Graphical Neural Network (Hybrid CNN-GNN) to extract the spatiotemporal correlation features of soil and rock masses, and combines it with a Generative Adversarial Network (GAN) to enhance the model's generalization ability. Compared with traditional numerical simulation methods that rely on simplified models and static data input, this invention can capture the complex nonlinear characteristics of soil and rock masses, achieving high-precision parameter inversion, and is particularly suitable for heterogeneous soil and rock masses and multi-field coupled scenarios.

[0058] (3) Dynamic feedback control and closed-loop optimization: This invention uses reinforcement learning and Bayesian optimization algorithms to achieve data acquisition. Compared with traditional systems that lack a real-time feedback mechanism, this invention can dynamically adjust the monitoring strategy to ensure a high degree of consistency between the inversion results and engineering requirements, significantly improving the robustness and practicality of the system.

[0059] (4) Uncertainty Quantification and Risk Assessment: This invention generates confidence intervals for parameter predictions using Monte Carlo Dropout technology and combines them with engineering specification thresholds for risk assessment. Compared to traditional methods that lack uncertainty quantification mechanisms, this invention can provide a credibility assessment of the inversion results, providing a scientific basis for engineering safety decisions. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a system module diagram of the present invention;

[0062] Figure 2 This is a structural diagram of the deep learning model building module of the present invention;

[0063] Figure 3 This is a structural diagram of the dynamic feedback control module of the present invention. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present application.

[0065] Example 1: Please refer to Figure 1 As shown in this embodiment, a deep learning-based intelligent inversion system for soil and rock parameters is described. The system includes a multi-source data acquisition module, a geological data preprocessing module, a deep learning model construction module, an intelligent parameter inversion module, and a dynamic feedback control module.

[0066] The multi-source data acquisition module acquires sensor network data, including mechanical, hydrological and geological environment data of soil and rock, through the sensor network sub-module, and remote sensing data, including surface deformation and geological structure data, through the remote sensing data sub-module. It then aligns the sensor network data and the remote sensing data in time and space to generate an original dataset.

[0067] The sensor network submodule consists of multiple types of sensors, which are deployed in the target soil and rock area according to a three-dimensional grid topology.

[0068] The various types of sensors include strain sensors, pore water pressure sensors, and ground-penetrating radar, which respectively collect sensor network data including mechanical, hydrological, and geological environmental data of the rock and soil.

[0069] The spatiotemporal alignment is achieved by combining timestamp matching and spatial interpolation, and the steps are as follows:

[0070] First, the timestamp matching is implemented as follows:

[0071] Select data that meets the requirements. The data ensures that the data is within the same time window. For remote sensing of time, For sensor network time, Maximum allowable time difference;

[0072] Secondly, the spatial interpolation technique unifies remote sensing data and sensor network data with different spatial resolutions into the same spatial grid. The spatial interpolation method is bilinear interpolation, and its formula z(·) is as follows:

[0073]

[0074] Where (x, y) are the coordinates that need to be unified, and z i Given remote sensing and sensor network data, w i For interpolation weights;

[0075] Finally, spatiotemporal alignment is performed, that is, aligning the time points of each remote sensing data and sensor network data. Given a spatial point (x, y), calculate the fused data using the following formula:

[0076]

[0077] in, For the merged data, For the remote sensing data, For the sensor network data, These are the weighting coefficients.

[0078] The geological data preprocessing module uses wavelet transform to denoise the original dataset, and uses a feature fusion algorithm to fuse features in the denoised original data. Finally, it standardizes the data to generate a high-quality dataset.

[0079] The wavelet transform denoising algorithm eliminates noise in the original dataset; the core steps of the wavelet threshold denoising algorithm include wavelet decomposition: decomposing the sensor signal into detail coefficients and approximation coefficients at multiple scales through wavelet transform; thresholding: performing thresholding on the detail coefficients to retain significant features and remove noise components; wavelet reconstruction: reconstructing the processed coefficients into a denoised signal through inverse wavelet transform;

[0080] The feature fusion algorithm uses a long short-term memory network to extract temporal features from the original dataset; the standardization uses the Z-Score normalization algorithm to unify the data dimensions of the original dataset and generate a high-quality dataset.

[0081] The core steps of the Long Short-Term Memory Network include time series modeling:

[0082] By controlling the flow of information through input gates, forget gates, and output gates, long-term dependencies in time series are captured; Spatiotemporal alignment: the denoised original dataset is input into the LSTM network to generate aligned feature vectors.

[0083] The deep learning model building module, refer to Figure 2 As shown, a hybrid convolutional-graph neural network is used to extract the spatiotemporal features of soil and rock in the high-quality dataset, and the adversarial training strategy of generative adversarial network is combined to optimize the generalization ability of the pre-trained model and generate a high-precision inversion model; the pre-trained model refers to the initial model obtained by the hybrid convolutional-graph neural network and generative adversarial network.

[0084] The deep learning model building module includes a spatiotemporal feature extraction submodule and an adversarial training submodule;

[0085] The spatiotemporal feature extraction submodule extracts spatiotemporal features using a hybrid convolutional-graph neural network.

[0086] The hybrid convolutional-graph neural network includes a convolutional neural network branch to extract local spatial pattern features and a graph neural network branch to extract unstructured topological relationship features of soil and rock. After extraction, the spatiotemporal features are dynamically weighted through an attention mechanism to generate a global feature vector.

[0087] The convolutional neural network branch has the following structure: the input is local spatial data of the soil and rock mass; the structure includes a three-dimensional convolutional layer, a pooling layer, and an output layer, wherein the three-dimensional convolutional layer is represented as follows:

[0088]

[0089] Where x is the input 3D data, the weight parameters of the convolution kernel, and the dimension (M, N, P), where M, N, P are the dimensions of the convolution kernel in the height, width, and depth directions, respectively, and m, n, p are indices; b is the bias term, used to adjust the numerical range of the output feature map; y is the output feature map, representing the local spatial features after the convolution operation; v, j, k are the indices of the output feature map, representing the positions in the height, width, and depth directions, respectively;

[0090] The graph neural network branch is used to model the unstructured geological topological relationships of soil and rock masses, and the input of the graph neural network branch is the unstructured geological topological data of soil and rock masses;

[0091] The process of constructing the unstructured geological topology data is as follows:

[0092] Node construction and feature extraction: Representative measuring points in the rock and soil mass are regarded as nodes in the image. Each sensor measuring point is extracted from the geological structure region through image segmentation. A feature vector is constructed for each node, which includes spatial information, physical properties and image features.

[0093] Node selection and aggregation: Based on the requirements of the dense project, clustering algorithms are used to select representative nodes;

[0094] Edge construction and topological relationship definition: For each node, a preliminary connection is constructed based on spatial proximity, the Euclidean distance between each node is calculated, and several nodes with the closest distance are selected to establish an edge; on the basis of the preliminary connection, if the regions to which two nodes belong have continuity in geological structure (fault and fracture extension in this embodiment), then the edge connection is strengthened.

[0095] Edge weight calculation: Each edge is assigned a weight, which is calculated by combining the following factors: spatial distance (the closer the distance, the greater the weight); node feature similarity; and geological connection indicators (including structural continuity and common fault identification). Finally, a weighted adjacency matrix is ​​formed.

[0096] The graph neural network branch includes a graph attention mechanism and a graph convolutional layer; the graph attention mechanism is used to model the relationships between nodes, and the formula is:

[0097]

[0098] in, Represents a node The new features, obtained by aggregating neighbor node information, capture the nodes. The relationship between it and its neighboring nodes; σ is the activation function; Represents a node The result of linear transformation of the feature vectors; Represents a node and nodes Attention coefficient between them; Represents a node The neighboring nodes;

[0099] The graph convolutional layer is used to aggregate information about neighboring nodes, as shown below:

[0100]

[0101] in, Add the identity matrix I to the adjacency matrix A to introduce a self-join; H (l) W is the feature matrix of the nodes in the l-th layer; (l) Let σ' be the weight matrix of the l-th layer; σ' is the activation function. This is a degree matrix, representing the degree of each node;

[0102] The adversarial training submodule, based on a generative adversarial network architecture, simulates the distribution of noisy data in the original dataset and improves the robustness of the hybrid convolutional-graph neural network to noisy data by minimizing the Wasserstein distance.

[0103] The generative adversarial network is trained through adversarial training between the generator network and the discriminator network.

[0104] The generator network takes random noise z and geological condition labels c as input and outputs simulated anomaly data G(z,c).

[0105] The discriminator network takes real soil and rock parameter data x and generated data G(z,c) as input and outputs the discrimination results D(x) and D(G(z,c)).

[0106] The Wasserstein distance is expressed as follows:

[0107]

[0108] Where x represents the real data, D(x) is the discriminator's result when x is input; G(z) is the data generated by inputting random noise z into the generator; and D(G(z)) is the result after inputting into the discriminator. The goal is to calculate the average output of the discriminator D on both real and generated data; and to improve the model's generalization ability to real geotechnical parameter data by optimizing the Wasserstein distance.

[0109] The intelligent parameter inversion module dynamically predicts the elastic modulus, permeability coefficient, and shear strength parameters of soil and rock mass based on the inversion model and the high-quality dataset, and generates a parameter inversion report.

[0110] The parameter intelligent inversion module includes a real-time data stream processing submodule, a multi-task prediction submodule, an uncertainty quantification submodule, and a dynamic report generation submodule.

[0111] The real-time data stream processing submodule uses sliding window technology and time series segmentation algorithm to divide the input data at the current moment into time series segments;

[0112] The input data consists of sensor network data and remote sensing data acquired by the multi-source data acquisition module and processed by the geological data preprocessing module.

[0113] The time-series segments are adapted to the model's inference requirements. The segmented time-series segments can be input into the model in real time to meet the requirements of low-latency inference.

[0114] The sliding window technology includes: window length: adaptively adjusted according to the dynamic characteristics of the soil and rock mass (stress relaxation time and seepage stabilization time in this embodiment), with a default range of 1 hour to 24 hours; window overlap ratio: dynamically set based on the data change rate (50% overlap) to ensure the continuity and integrity of time series segments; data segmented output: the data in each window is treated as an independent input unit and directly input into the inversion model for real-time prediction;

[0115] The time series segmentation algorithm includes a dynamic adjustment mechanism: identifying periodic features in the data through fast Fourier transform and dynamically adjusting the window length and segmentation strategy; constraints: setting a minimum window length (10 minutes) and a maximum segmentation frequency (once per hour) based on engineering experience to avoid over-segmentation leading to information loss;

[0116] The multi-task prediction submodule combines the trained high-precision inversion model to perform multi-task prediction of soil and rock parameters; the multi-task includes prediction of soil and rock elastic modulus, permeability coefficient, and shear strength parameters.

[0117] The uncertainty quantification submodule generates confidence intervals for parameter predictions based on Monte Carlo Dropout technology, quantifying the credibility of the inversion results. Monte Carlo activates Dropout during the inference phase and runs the input data through the network multiple times, collecting the output results each time. Then, these results are statistically analyzed (mean and variance) to evaluate the model's prediction and uncertainty.

[0118] The dynamic report generation submodule automatically marks parameters that exceed the preset threshold of soil and rock parameters as abnormal parameters and generates a visual inversion report.

[0119] The abnormal parameters are automatically marked as abnormal when the predicted value of the soil and rock parameters exceeds the threshold or the confidence interval coverage threshold.

[0120] The inversion report includes:

[0121] Predicted values ​​of parameters: elastic modulus, permeability coefficient, and shear strength;

[0122] Confidence interval: The confidence interval for parameter prediction;

[0123] Anomaly marker: Parameters that exceed the preset threshold for soil and rock parameters;

[0124] Visual charts: Line chart: The trend of parameter changes over time.

[0125] The dynamic feedback control module uses reinforcement learning algorithms to adaptively adjust the frequency of data acquisition from the sensor network, thereby achieving closed-loop optimization of the system.

[0126] The dynamic feedback control module, referencing Figure 3 As shown, based on the reinforcement learning algorithm, the multi-source data acquisition module adjusts the sensor network data acquisition frequency according to the inversion error indication; the inversion error is obtained by calculating the difference between the predicted values ​​and the true values ​​of the soil and rock parameters by the multi-task prediction submodule.

[0127] The reinforcement learning consists of a state space, an action space, a reward function, and a policy network;

[0128] The state space is defined as the current inversion error e of the system. t Resource utilization rate u t Sensor network health status h t , represented as s t =[e t ,u t ,h t ];

[0129] The resource utilization rate represents the resource utilization status of the system at the current moment, including the utilization rate of computing resources, storage resources and communication resources. The computing resource utilization rate includes the utilization rate of CPU and GPU, the storage resource utilization rate includes the utilization rate of memory and hard disk, and the communication resource utilization rate is the utilization rate of network bandwidth.

[0130] The sensor network health status includes sensor online rate, communication packet loss rate, and remaining battery power;

[0131] The action space: dynamically adjusts the data acquisition frequency f. t ∈{low frequency (1Hz), mid frequency (10Hz), high frequency (100Hz)};

[0132] The reward function balances inversion accuracy and resource consumption, and its formula is:

[0133]

[0134] in, f represents the square of the inversion error; t Indicates the data acquisition frequency; α, β, γ are weighting coefficients; For indicator functions, The value is determined by the preset inversion error threshold, the upper limit of resource utilization, and the lower limit of sensor network health status. When the inversion error is greater than the inversion error threshold, the resource utilization is greater than the upper limit of resource utilization, or the sensor network health status is lower than the lower limit of sensor network health status, the value is 0; otherwise, the value is 1.

[0135] The policy function employs a Deep Q-Network (DQN), with input state s. tOutput the Q-value for each action, and select the action corresponding to the largest Q-value, as shown below:

[0136]

[0137] Among them, a t For the selected action, s t Let Q be the state, Q be a deep Q-network, and θ be the network parameters, updated through temporal difference (TD) error, as follows:

[0138]

[0139] Where, θ - Synchronize the target network parameters periodically. Indicates that in state s t+1 Choose the action that maximizes the Q value, r t For value function, It is the gradient, Q(s) t ,a t ;θ) represents the state s t And action a t The Q value under the given condition, where η is the learning rate.

[0140] All the above formulas are dimensionless and use only numerical values ​​for calculation. These formulas are derived from a large amount of data and through software simulation, aiming to approximate reality as closely as possible. The preset parameters in the formulas can be adjusted by those skilled in the art according to specific needs.

[0141] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0142] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A deep learning-based intelligent inversion system for soil and rock parameters, characterized in that, The system includes: The multi-source data acquisition module acquires sensor network data, including mechanical, hydrological and geological environment data of soil and rock, through the sensor network sub-module, and remote sensing data, including surface deformation and geological structure data, through the remote sensing data sub-module. It then aligns the sensor network data and the remote sensing data in time and space to generate the original dataset. The geological data preprocessing module uses wavelet transform to denoise the original dataset, and then uses a feature fusion algorithm to fuse the features of the denoised original data. Finally, it is standardized to generate a high-quality dataset. The deep learning model building module uses a hybrid convolutional-graph neural network to extract the spatiotemporal features of soil and rock in the high-quality dataset, and combines the adversarial training strategy of generative adversarial network to optimize the generalization ability of the pre-trained model and generate a high-precision inversion model; the pre-trained model is the initial model of the hybrid convolutional-graph neural network and generative adversarial network. The intelligent parameter inversion module dynamically predicts the elastic modulus, permeability coefficient, and shear strength parameters of soil and rock mass based on the inversion model and the high-quality dataset, and generates a parameter inversion report. The dynamic feedback control module uses a reinforcement learning algorithm to adaptively adjust the frequency of data acquisition from the sensor network, thereby achieving closed-loop optimization of the system. The deep learning model building module includes a spatiotemporal feature extraction submodule and an adversarial training submodule; The spatiotemporal feature extraction submodule extracts spatiotemporal features through a hybrid convolutional-graph neural network; The hybrid convolutional-graph neural network includes a convolutional neural network branch to extract local spatial pattern features and a graph neural network branch to extract unstructured topological relationship features of soil and rock masses. After extraction, spatiotemporal features are dynamically weighted through an attention mechanism to generate a global feature vector. The graph neural network branch is used to construct edge connections between nodes based on the continuity of geological structures, and is weighted by spatial distance, node feature similarity, and geological connection indicators to reflect the unstructured topological relationships of soil and rock masses. The adversarial training submodule, based on the generative adversarial network (GAN) architecture, simulates the distribution of noisy data in the original dataset and improves the robustness of the hybrid convolutional-graph neural network to noisy data by minimizing the Wasserstein distance; The parameter intelligent inversion module includes a real-time data stream processing submodule, a multi-task prediction submodule, an uncertainty quantification submodule, and a dynamic report generation submodule. The real-time data stream processing submodule uses sliding window technology and time series segmentation algorithm to divide the input data at the current moment into time series segments. The input data is sensor network data and remote sensing data collected by the multi-source data acquisition module and processed by the geological data preprocessing module. The multi-task prediction submodule performs multi-task prediction of soil and rock parameters in conjunction with the trained high-precision inversion model. The multi-task includes prediction of the elastic modulus of soil and rock mass, prediction of permeability coefficient, and prediction of shear strength parameters. The uncertainty quantification submodule generates confidence intervals for parameter predictions based on Monte Carlo Dropout technology, thus quantifying the credibility. The dynamic report generation submodule automatically marks parameters that exceed the preset threshold for soil and rock parameters and parameters whose confidence intervals cover the preset threshold as abnormal parameters, and generates a visual inversion report. The dynamic feedback control module, based on a reinforcement learning algorithm, adaptively adjusts the frequency of data acquisition from the sensor network according to the abnormal parameter labels and inversion errors, thereby achieving closed-loop optimization of the system.

2. The system according to claim 1, characterized in that, The sensor network submodule consists of multiple types of sensors, which are deployed in the target soil and rock area according to a three-dimensional grid topology. The various types of sensors include strain sensors, pore water pressure sensors, and ground-penetrating radar, which respectively collect sensor network data including mechanical, hydrological, and geological environmental data of the rock and soil.

3. The system according to claim 2, characterized in that, The spatiotemporal alignment is achieved by combining timestamp matching and spatial interpolation, and the steps are as follows: First, the timestamp matching is implemented as follows: Select data that meets the requirements. The data ensures that the data is within the same time window. For remote sensing of time, For sensor network time, The maximum allowable time difference; Secondly, the spatial interpolation unifies remote sensing data and sensor network data with different spatial resolutions into the same spatial grid. The spatial interpolation method is bilinear interpolation, and its formula is... as follows: in, For the purpose of unifying the coordinates of spatial points, Given remote sensing and sensor network data, For interpolation weights, C represents the amount of remote sensing and sensor network data; Finally, spatiotemporal alignment is performed, that is, aligning the time points of each remote sensing data and sensor network data. and spatial points The formula for calculating the fused data is as follows: in, For the merged data, For the remote sensing data, For the sensor network data, These are the weighting coefficients.

4. The system according to claim 1, characterized in that, The feature fusion algorithm uses a long short-term memory network to extract temporal features from the original dataset; The standardization process employs the Z-Score normalization algorithm to unify the data dimensions of the original dataset, thereby generating a high-quality dataset.

5. The system according to claim 1, characterized in that, The inversion report includes: Predicted values ​​of parameters: elastic modulus, permeability coefficient, and shear strength; Confidence interval: The confidence interval for parameter prediction; Anomaly marker: Parameters that exceed the preset threshold for soil and rock parameters; Visualization chart: Line chart showing how parameters change over time.

6. The system according to claim 5, characterized in that, The dynamic feedback control module is based on a reinforcement learning algorithm and adjusts the sensor network data acquisition frequency according to the inversion error indication of the multi-source data acquisition module. The inversion error is obtained by calculating the difference between the predicted and actual values ​​of the soil and rock parameters by the multi-task prediction submodule.

7. The system according to claim 6, characterized in that, The reinforcement learning consists of a state space, an action space, a reward function, and a policy network; The state space is defined as the current inversion error of the system. Resource utilization rate Sensor network health status , represented as ; The resource utilization rate represents the resource utilization status of the system at the current moment, including the utilization rate of computing resources, storage resources and communication resources. The computing resource utilization rate includes the utilization rate of CPU and GPU, the storage resource utilization rate includes the utilization rate of memory and hard disk, and the communication resource utilization rate is the network bandwidth utilization rate. The sensor network health status includes sensor online rate, communication packet loss rate, and remaining battery power; The action space: dynamically adjusts the data acquisition frequency. ; The reward function balances inversion accuracy and resource consumption, and its formula is: in, Represents the square of the inversion error; Indicates the data acquisition frequency; These are the weighting coefficients; For indicator functions, The value is determined by the preset inversion error threshold, the upper limit of resource utilization, and the lower limit of sensor network health status. When the inversion error is greater than the inversion error threshold, the resource utilization is greater than the upper limit of resource utilization, or the sensor network health status is lower than the lower limit of sensor network health status, the value is 0; otherwise, the value is 1. The policy network employs a Deep Q-Network (DQN), with input state... Output the Q-value for each action, and select the action corresponding to the largest Q-value, as shown below: in, For the selected action, For a set of actions, For state, It is a deep Q-network. The parameters of the deep Q-network are updated using temporal difference (TD) error, as shown below: in, The target network parameters for the deep Q-network are synchronized periodically. , Indicates the state Choose the action that maximizes the Q value. For the reward function, It is the gradient. In the state and actions Q value under the condition, This is the learning rate.

Images