Rock-soil body parameter intelligent inversion system based on deep learning

Abstract

The invention discloses a rock-soil mass parameter intelligent inversion system based on deep learning, and relates to the technical field of intelligent exploration. A multi-source data acquisition module acquires multi-dimensional data and generates an original data set; the geological data preprocessing module cleans and standardizes the original data set; the deep learning model construction module adopts a hybrid convolution-graph neural network to extract the space-time correlation characteristics of the rock-soil body, combines the generalization ability of an adversarial training optimization model, and generates a high-precision inversion model; the parameter intelligent inversion module generates a parameter inversion report based on the model and the real-time data; and the dynamic feedback regulation and control module adaptively adjusts a data acquisition strategy by using reinforcement learning. According to the invention, through a modular cooperation and adaptive learning mechanism, the parameter inversion efficiency and precision are significantly improved, manual dependence is reduced, and intelligent support is provided for engineering design and safety evaluation under complex geological conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent exploration, and particularly to an intelligent inversion system for geotechnical parameters based on deep learning. Background Art

[0002] The inversion of geotechnical engineering parameters is a key technology in geological exploration and engineering design. Its goal is to accurately obtain the mechanical parameters of geotechnical materials by combining on-site monitoring data with theoretical models. These parameters have important guiding significance for engineering safety assessment, disaster warning, and construction design. However, with the increase in engineering complexity, 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 manual inversion techniques based on statistics. These methods have the following problems: low computational efficiency: When dealing with large-scale geotechnical data, the traditional finite element analysis method has a high computational complexity and is difficult to meet the real-time requirements; limited accuracy: Manual inversion methods based on statistics are highly dependent on data quality and prior knowledge, and it is difficult to handle the spatio-temporal correlation 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, and it is difficult to dynamically adjust monitoring strategies and model parameters.

[0004] In the prior art, inversion methods based on numerical simulation (such as genetic algorithms, particle swarm optimization) can partially improve the computational efficiency, but they rely on simplified models and static data input, and it is difficult to handle the spatio-temporal correlation of multi-source heterogeneous data. In addition, traditional systems lack modular collaboration and adaptive learning mechanisms, resulting in the disconnection between inversion results and engineering decisions and being unable to meet the high-precision and real-time engineering requirements under complex geological conditions.

[0005] The inversion of geotechnical engineering parameters is a key link in geological exploration and engineering design. Its core goal is to accurately obtain the mechanical parameters of geotechnical materials (such as elastic modulus, permeability coefficient, shear strength, etc.) by combining on-site monitoring data with theoretical models. With the increase in engineering complexity (including deep mining, tunnel excavation, and slope stability), traditional methods show obvious deficiencies in dealing with heterogeneous geotechnical materials, multi-field coupling effects, and dynamic geological changes, resulting in 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 regulation to break through the limitations of the prior art. Summary of the Invention

[0007] The object of the present invention is to provide an intelligent inversion system for geotechnical parameters based on deep learning. By means of multi-source data fusion, deep learning modeling and dynamic feedback regulation technology, it overcomes the defects of existing inversion systems in terms of low calculation efficiency, limited accuracy, insufficient adaptability, etc., thereby significantly improving the accuracy, real-time performance of parameter inversion and the robustness of the system, and 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 object, the present invention provides an intelligent inversion system for geotechnical parameters based on deep learning, which includes the following main modules: a multi-source data acquisition module, a geological data preprocessing module, a deep learning model construction module, a parameter intelligent inversion module, and a dynamic feedback regulation module;

[0009] The multi-source data acquisition module obtains sensor network data including the mechanical, hydrological and geological environment data of the geotechnical body through the sensor network sub-module, obtains remote sensing data including surface deformation and geological structure data through the remote sensing data sub-module, and spatiotemporally aligns the sensor network data and the remote sensing data to generate an original data set;

[0010] The geological data preprocessing module denoises the original data set using wavelet transform, performs feature fusion on the denoised original data using a feature fusion algorithm, and finally performs standardization to generate a high-quality data set;

[0011] The deep learning model construction module uses a hybrid convolutional-graph neural network to extract the spatiotemporal features of the geotechnical body in the high-quality data set, and combines the adversarial training strategy of the generative adversarial network to optimize the generalization ability of the pre-trained model to generate a high-precision inversion model; the pre-trained model refers to the initial models of the hybrid convolutional-graph neural network and the generative adversarial network;

[0012] The parameter intelligent inversion module dynamically predicts the elastic modulus, permeability coefficient and shear strength parameters of the geotechnical body based on the inversion model and the high-quality data set, and generates a parameter inversion report;

[0013] The dynamic feedback regulation module adaptively adjusts the acquisition frequency of the sensor network data using a reinforcement learning algorithm to achieve closed-loop optimization of the system.

[0014] Furthermore, the sensor network sub-module is composed of multiple types of sensors and is deployed in the target geotechnical body area according to a three-dimensional grid topology;

[0015] The multiple types of sensors include strain sensors, pore water pressure sensors and ground-penetrating radars, which respectively collect sensor network data including the mechanical, hydrological and geological environment data of the geotechnical body.

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

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

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

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

[0020]

[0021] Among them, (x, y) is the coordinate point to be unified, z i is the known remote sensing and sensor network data, w i is the interpolation weight;

[0022] Finally, perform spatio-temporal alignment, that is, for each time point of the remote sensing data and the sensor network data and the spatial point (x, y), calculate the fused data, and the formula is as follows:

[0023]

[0024] Among them, is the fused data, is the remote sensing data, is the sensor network data, is the weight coefficient.

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

[0026] The standardization uses the Z-Score normalization algorithm to unify the data dimensions of the original dataset and generate a high-quality dataset.

[0027] Furthermore, the deep learning model construction module includes a spatio-temporal feature extraction sub-module and an adversarial training sub-module;

[0028] The spatio-temporal feature extraction sub-module, the spatio-temporal features are extracted by a hybrid convolutional-graph neural network;

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

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

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

[0032] The real-time data stream processing sub-module uses a sliding window technique and a time series segmentation algorithm to divide the input data at the current moment into time series segments. The input data is the 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 sub-module combines the trained high-precision inversion model to perform multi-task prediction of rock and soil mass parameters. The multi-tasks include prediction of the elastic modulus of the rock and soil mass, prediction of the permeability coefficient, and prediction of shear strength parameters.

[0034] The uncertainty quantification sub-module generates a confidence interval for parameter prediction based on the Monte Carlo Dropout technique to quantify the credibility.

[0035] The dynamic report generation sub-module automatically marks the parameters exceeding the preset rock and soil mass parameter threshold as abnormal parameters and generates a visual inversion report.

[0036] Furthermore, the inversion report includes:

[0037] Parameter prediction values: elastic modulus, permeability coefficient, shear strength.

[0038] Confidence interval: the confidence interval of parameter prediction.

[0039] Abnormal mark: parameters exceeding the preset rock and soil mass parameter threshold.

[0040] Visualization chart: line chart: the changing trend of parameters over time.

[0041] Furthermore, the dynamic feedback regulation module, based on the reinforcement learning algorithm, instructs the multi-source data acquisition module to adjust the sensor network data acquisition frequency according to the inversion error. The inversion error is obtained by calculating the difference between the predicted value and the true value of the rock and soil mass parameters by the multi-task prediction sub-module.

[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 t , resource utilization rate u t , and the health state h of the sensor network t , denoted as s t = [e t , u t , h t ;

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

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

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

[0047] The reward function: Balances the inversion accuracy and resource consumption. The formula is:

[0048]

[0049] Among them, represents the square of the inversion error; f t represents the data acquisition frequency; α, β, γ are weight coefficients; is an indicator function, The value is determined by a preset inversion error threshold, an upper limit of the resource utilization rate, and a lower limit of the health state of the sensor network. It takes 0 when the inversion error is greater than the inversion error threshold, the resource utilization rate is greater than the upper limit of the resource utilization rate, and the health state of the sensor network is lower than the lower limit of the health state of the sensor network, otherwise it takes 1;

[0050] The policy function: Adopts a deep Q network (DQN), takes the input state s t , outputs the Q values of each action, and selects the action corresponding to the maximum Q value, which is expressed as follows:

[0051]

[0052] Among them, a t is the selected action, s tis the state, Q is the Deep Q-Network, θ are the Deep Q-Network parameters, updated by the Temporal Difference (TD) error, expressed as follows:

[0053]

[0054] where θ - are the target network parameters of the Deep Q-Network, synchronized periodically denotes to select the action that maximizes the Q-value in state s t+1 , r t is the value function, is the gradient, Q(s t , a t ; θ) is the Q-value in state s t and action a t , and 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: The present invention collects multi-dimensional data in real time through a distributed sensor network and remote sensing technology, and combines wavelet transform and feature fusion algorithms to achieve efficient preprocessing of data. Compared with traditional methods that rely on a single data source and manual processing, the present 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: The present invention uses a Hybrid Convolutional- Graph Neural Network (Hybrid CNN-GNN) to extract spatio-temporal correlation features of rock and soil masses, and combines a Generative Adversarial Network (GAN) to improve the generalization ability of the model. Compared with traditional numerical simulation methods that rely on simplified models and static data input, the present invention can capture the complex non-linear features of rock and soil masses, achieve high-precision parameter inversion, and is especially suitable for heterogeneous rock and soil masses and multi-field coupling scenarios.

[0058] (3) Dynamic feedback regulation and closed-loop optimization: The present invention is based on reinforcement learning and Bayesian optimization algorithms to achieve data collection. Compared with traditional systems that lack a real-time feedback mechanism, the present 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: The present invention generates a confidence interval for parameter prediction through Monte Carlo Dropout technology and combines engineering specification thresholds for risk assessment. Compared with traditional methods that lack an uncertainty quantification mechanism, the present invention can provide a credibility assessment of the inversion results, providing a scientific basis for engineering safety decisions. Brief Description of the Drawings

[0060] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

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

[0062] Figure 2 It is the structure diagram of the deep learning model construction module of the present invention;

[0063] Figure 3 It is the structure diagram of the dynamic feedback regulation module of the present invention. Specific embodiments

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present application in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are some, but not all, of the embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts fall within the scope of protection of the present application.

[0065] Embodiment 1: Please refer to Figure 1 As shown, the intelligent inversion system for geotechnical parameters based on deep learning in this embodiment includes a multi-source data acquisition module, a geological data preprocessing module, a deep learning model construction module, a parameter intelligent inversion module, and a dynamic feedback regulation module;

[0066] The multi-source data acquisition module obtains sensor network data including the mechanical, hydrological, and geological environment data of the geotechnical body through the sensor network sub-module, and obtains remote sensing data including surface deformation and geological structure data through the remote sensing data sub-module, and spatiotemporally aligns the sensor network data and the remote sensing data to generate an original data set;

[0067] The sensor network sub-module is composed of multiple types of sensors and is deployed in the target geotechnical body area according to a three-dimensional grid topology;

[0068] The multiple types of sensors include strain sensors, pore water pressure sensors, and ground-penetrating radars, which respectively collect sensor network data including the mechanical, hydrological, and geological environment data of the geotechnical body;

[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 achieved as follows:

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

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

[0073]

[0074] where (x,y) is the coordinate point to be unified, and z i is the known remote sensing and sensor network data, and w i is the interpolation weight;

[0075] Finally, perform spatio-temporal alignment, that is, for each time point of the remote sensing data and the sensor network data and the spatial point (x,y), calculate the fused data, and the formula is as follows:

[0076]

[0077] where, is the fused data, is the remote sensing data, is the sensor network data, is the weight coefficient.

[0078] The geological data preprocessing module denoises the original data set using wavelet transform, and uses a feature fusion algorithm to fuse the features of the denoised original data, and finally performs standardization to generate a high-quality data set;

[0079] The wavelet transform denoising algorithm eliminates the noise in the original data set; the core steps of the wavelet threshold denoising algorithm include wavelet decomposition: decomposing the sensor signal into detail coefficients and approximation coefficients of multiple scales through wavelet transform; threshold processing: performing threshold processing on the detail coefficients, retaining significant features, and removing 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 the temporal features in the original data set; the standardization uses the Z-Score normalization algorithm to unify the data dimensions of the original data set to generate a high-quality data set;

[0081] The core steps of the long short-term memory network include time series modeling:

[0082] Controlling the information flow through the input gate, forget gate, and output gate to capture long-term dependencies in the time series; spatio-temporal alignment: inputting the denoised original dataset into the LSTM network to generate the aligned feature vectors.

[0083] The deep learning model construction module, referring to Figure 2 As shown, a hybrid convolutional-graph neural network is used to extract the spatio-temporal features of the rock and soil mass in the high-quality dataset, and the generalization ability of the pre-trained model is optimized by combining the adversarial training strategy of the generative adversarial network to generate a high-precision inversion model; the pre-trained model refers to the initial models of the hybrid convolutional-graph neural network and the generative adversarial network;

[0084] The deep learning model construction module includes a spatio-temporal feature extraction sub-module and an adversarial training sub-module;

[0085] The spatio-temporal feature extraction sub-module, where the spatio-temporal features are extracted by a hybrid convolutional-graph neural network;

[0086] The hybrid convolutional-graph neural network includes a convolutional neural network branch for extracting local spatial pattern features and a graph neural network branch for extracting the unstructured topological relationship features of the rock and soil mass; after extraction, the spatio-temporal features are dynamically weighted by an attention mechanism to generate global feature vectors;

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

[0088]

[0089] Among them, x is the input three-dimensional data, the weight parameters of the convolutional kernel, with dimensions (M, N, P), where M, N, and P are the sizes of the convolutional 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 relationship of the rock and soil mass, and the input of the graph neural network branch is the unstructured geological topological data of the rock and soil mass;

[0091] The construction process of the unstructured geological topological data is as follows:

[0092] Node construction and feature extraction: Each representative measurement point in the rock and soil mass is regarded as a node in the graph, and the geological structure area extracted by image segmentation for each sensor measurement point; construct a feature vector for each node, including spatial information, physical properties, and image features;

[0093] Node screening and aggregation: According to the requirements of the engineering, use the clustering algorithm to screen out the representative nodes;

[0094] Edge construction and definition of topological relationship: For each node, construct a preliminary connection according to spatial proximity, calculate the Euclidean distance between each node, and select several nodes with the closest distance to establish an edge; on the basis of the preliminary connection, if the areas of the two nodes are continuous in the geological structure (faults and crack extensions in this embodiment), strengthen the edge connection;

[0095] Edge weight calculation: Assign weights to each edge, and the weights are comprehensively calculated by the following factors: spatial distance, the closer the distance, the greater the weight; node feature similarity, geological connection indicators (including structural continuity, common fault identification), and finally form a weighted adjacency matrix;

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

[0097]

[0098] Among them, represents the new feature of node after aggregating the information of neighbor nodes, capturing the relationship between node and its neighbor nodes; σ is the activation function; represents the result of the linear transformation of the feature vector of node ; represents node and node the attention coefficient between; represents node 's neighbor nodes;

[0099] The graph convolutional layer is used to aggregate the information of neighbor nodes, which is expressed as follows:

[0100]

[0101] Among them, is the adjacency matrix A plus the identity matrix I, used to introduce self-connection; H (l) is the node feature matrix of the l-th layer; W (l) is the weight matrix of the l-th layer; σ' is the activation function; is the degree matrix, representing the degree of each node;

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

[0103] The generative adversarial network undergoes adversarial training between the generator network and the discriminator network;

[0104] The input of the generator network is the random noise z and the geological condition label c, and the output is the simulated anomaly data G(z, c);

[0105] The input of the discriminator network is the real geotechnical parameter data x and the generated data G(z, c), and the outputs are the discrimination results D(x) and D(G(z, c));

[0106] The Wasserstein distance is expressed as follows:

[0107]

[0108] where x is the real data, D(x) is the result of the discriminator when the input is x; G(z) is the data generated by inputting the generator random noise z; D(G(z)) is the result after inputting the discriminator, is to find the average output of the discriminator D on the real data and the generated data; by optimizing the Wasserstein distance, the generalization ability of the model to the real geotechnical parameter data is improved;

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

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

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

[0112] The input data is the sensor network data and remote sensing data collected by the multi-source data acquisition module and processed by the geological data preprocessing module;

[0113] The time series segments, for adapting to the model inference requirements, can be input into the model in real time after segmentation to meet the low-latency inference requirements;

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

[0115] The time series segmentation algorithm consists of a dynamic adjustment mechanism: identifying periodic features in the data through fast Fourier transform and dynamically adjusting the window length and segmentation strategy; constraint conditions: setting the minimum window length (10 minutes) and the maximum segmentation frequency (once per hour) in combination with engineering experience to avoid information loss caused by over-segmentation;

[0116] The multi-task prediction sub-module combines the trained high-precision inversion model to perform multi-task prediction of rock and soil mass parameters; the multi-tasks include prediction of the elastic modulus of the rock and soil mass, prediction of the permeability coefficient, and prediction of shear strength parameters;

[0117] The uncertainty quantification sub-module generates a confidence interval for parameter prediction based on the Monte Carlo Dropout technique to quantify the credibility of the inversion results; in the Monte Carlo, Dropout is activated during the inference stage, and the input data is run through the network multiple times, collecting the output results each time, and then performing statistical analysis (mean and variance) on these results to evaluate the prediction and uncertainty of the model;

[0118] The dynamic report generation sub-module automatically marks the parameters that exceed the preset rock and soil mass parameter threshold as abnormal parameters and generates a visual inversion report;

[0119] The abnormal parameters are automatically marked as abnormal when the predicted values of the rock and soil mass parameters exceed the threshold or the confidence interval covers the threshold;

[0120] The inversion report includes:

[0121] Parameter prediction values: elastic modulus, permeability coefficient, shear strength;

[0122] Confidence interval: the confidence interval of parameter prediction;

[0123] Abnormal mark: parameters that exceed the preset rock and soil mass parameter threshold;

[0124] Visualization chart: line chart: the changing trend of parameters over time.

[0125] The dynamic feedback regulation module adaptively adjusts the frequency of data acquisition of the sensor network using the reinforcement learning algorithm to achieve closed-loop optimization of the system.

[0126] The dynamic feedback control module, referring to Figure 3 As shown, based on the reinforcement learning algorithm, according to the inversion error, it instructs the multi-source data acquisition module to adjust the data acquisition frequency of the sensor network; the inversion error is obtained by calculating the difference between the predicted value and the true value of the geotechnical parameters by the multi-task prediction sub-module.

[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 , the resource utilization rate u t , the health status h of the sensor network t , denoted as s t =[e t , u t , h t ;

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

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

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

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

[0133]

[0134] Where represents the square of the inversion error; f t represents the data acquisition frequency; α, β, γ are weight coefficients; is an indicator function, The value is determined by the preset inversion error threshold, the upper limit of resource utilization rate, and the lower limit of the health status of the sensor network. It takes 0 when the inversion error is greater than the inversion error threshold, the resource utilization rate is greater than the upper limit of resource utilization rate, and the health status of the sensor network is lower than the lower limit of the health status of the sensor network, otherwise it takes 1;

[0135] The policy function: adopts a deep Q network (DQN), and the input state is s t, output the Q-values of each action, and select the action corresponding to the maximum Q-value, which is expressed as follows:

[0136]

[0137] where a t is the selected action, s t is the state, Q is the deep Q-network, θ is the network parameter, and it is updated through the temporal difference (TD) error, which is expressed as follows:

[0138]

[0139] where θ - is the target network parameter and is synchronized periodically denotes selecting the action that maximizes the Q-value in state s t+1 , r t is the value function, is the gradient, Q(s t , a t ; θ) is the Q-value in state s t and action a t cases, and η is the learning rate.

[0140] The above formulas are all in dimensionless form and only use numerical values for calculation. These formulas are based on a large amount of data and obtained through software simulation, aiming to be as close to the actual situation 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, the descriptions referring to terms such as "one embodiment", "example", "specific example", etc. mean that the specific features, structures, materials, or characteristics described in connection with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0142] The above-described preferred embodiments of the present invention disclosed are only used to help explain the present invention. The preferred embodiments do not elaborate on all the details and do not limit the invention to only the specific implementation manners. Obviously, many modifications and variations can be made according to the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the present invention, so that those skilled in the art can well understand and utilize the present invention. The present invention is only limited by the claims and their full scope and equivalents.

Claims

1. An intelligent inversion system for geotechnical parameters based on deep learning, characterized in that, The system includes: A multi-source data acquisition module that obtains sensor network data including mechanical, hydrological, and geological environment data of the rock and soil mass through a sensor network sub-module, obtains remote sensing data including surface deformation and geological structure data through a remote sensing data sub-module, and aligns the sensor network data and the remote sensing data in space and time to generate an original data set; A geological data preprocessing module that denoises the original data set using wavelet transform, performs feature fusion on the denoised original data using a feature fusion algorithm, and finally standardizes it to generate a high-quality data set; A deep learning model construction module that extracts spatio-temporal features of the rock and soil mass in the high-quality data set using a hybrid convolutional-graph neural network, and optimizes the generalization ability of the pre-trained model by combining the adversarial training strategy of the generative adversarial network to generate a high-precision inversion model; the pre-trained model refers to the initial models of the hybrid convolutional-graph neural network and the generative adversarial network; A parameter intelligent inversion module that dynamically predicts the elastic modulus, permeability coefficient, and shear strength parameters of the rock and soil mass based on the inversion model and the high-quality data set, and generates a parameter inversion report; A dynamic feedback regulation module that adaptively adjusts the frequency of sensor network data acquisition using a reinforcement learning algorithm to achieve closed-loop optimization of the system.

2. The system according to claim 1, wherein The sensor network sub-module consists of multiple types of sensors and is deployed in a three-dimensional grid topology in the target rock and soil mass area; The multiple 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 environment data of the rock and soil mass.

3. The system according to claim 2, wherein The spatio-temporal alignment is achieved by combining timestamp matching and spatial interpolation, and the steps are as follows: First, the timestamp matching is achieved as follows: Select the data that satisfies |u remote -u sensor |≤Δu to ensure that the data is within the same time window, where u remote is the remote sensing time, and u sensor is the sensor network time, and Δu is the maximum allowable time difference; Second, the spatial interpolation technology unifies remote sensing data with different spatial resolutions and sensor network data into the same spatial grid. The spatial interpolation method is the bilinear interpolation method, and its formula z(·) is as follows: Among them, (x, y) are the spatial point coordinates to be unified, and z i is the known remote sensing and sensor network data, and w i is the interpolation weight, and C represents the quantity of remote sensing and sensor network data; Finally, spatio-temporal alignment is performed, that is, for each time point u and spatial point (x, y) of the remote sensing data and the sensor network data, the fusion data is calculated, and the formula is as follows: Among them, D fusion (x, y, u) is the fused data, and D remote (x, y, u) is the remote sensing data, and D sebsor (x, y, u) is the sensor network data, is the weight coefficient.

4. The system according to claim 1, wherein The feature fusion algorithm uses a long short-term memory network to extract the temporal features in the original data set; The standardization uses the Z-Score normalization algorithm to unify the data dimensions of the original data set to generate a high-quality data set.

5. The system according to claim 1, characterized in that, The deep learning model construction module includes a spatio-temporal feature extraction sub-module and an adversarial training sub-module; The spatio-temporal feature extraction sub-module extracts the spatio-temporal features through a hybrid convolutional-graph neural network; The hybrid convolutional-graph neural network includes a convolutional neural network branch that extracts local spatial pattern features and a graph neural network branch that extracts unstructured topological relationship features of the rock and soil mass; after extraction, the spatio-temporal features are dynamically weighted through an attention mechanism to generate a global feature vector; The adversarial training sub-module, based on the generative adversarial network (GAN) architecture, simulates the noise data distribution in the original data set, and improves the robustness of the hybrid convolutional-graph neural network to noise data by minimizing the Wasserstein distance.

6. The system according to claim 1, characterized in that The parameter intelligent inversion module includes a real-time data stream processing sub-module, a multi-task prediction sub-module, an uncertainty quantification sub-module, and a dynamic report generation sub-module; The real-time data stream processing sub-module uses a sliding window technique and a 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 sub-module combines the trained high-precision inversion model to perform multi-task prediction of geotechnical parameters; The multi-tasks include prediction of geotechnical elastic modulus, permeability coefficient, and shear strength parameters; The uncertainty quantification sub-module generates a confidence interval for parameter prediction based on the Monte Carlo Dropout technique and quantifies the credibility; The dynamic report generation sub-module automatically marks the parameters that exceed the preset geotechnical parameter threshold as abnormal parameters and generates a visual inversion report.

7. The system according to claim 6, wherein The inversion report includes: Parameter prediction values: elastic modulus, permeability coefficient, shear strength; Confidence interval: the confidence interval of parameter prediction; Abnormal mark: the parameters that exceed the preset geotechnical parameter threshold; Visualization chart: a line chart showing the change of parameters over time.

8. The system according to claim 7, wherein The dynamic feedback regulation module, based on the reinforcement learning algorithm, instructs the multi-source data acquisition module to adjust the sensor network data acquisition frequency according to the inversion error; The inversion error is obtained by calculating the difference between the predicted value and the true value of the geotechnical parameters by the multi-task prediction sub-module.

9. The system according to claim 8, wherein 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 e of the system t , the resource utilization rate u t , the health state h of the sensor network t , denoted as s t = [e t , u t , h t ; The resource utilization rate represents the resource utilization situation of the system at the current moment, including the occupancy rates of computing resources, storage resources, and communication resources. The computing resource utilization rate includes the usage rates of CPU and GPU. The storage resource utilization rate includes the usage rates of memory and hard disk. The communication resource utilization rate is the occupancy rate of network bandwidth; The health status of the sensor network includes the sensor online rate, the packet loss rate of communication, and the remaining battery power; The action space: Dynamically adjust the data acquisition frequency f t ∈{low frequency (1Hz), medium frequency (10Hz), high frequency (100Hz)}; The reward function: balance the inversion accuracy and resource consumption, and the formula is: Among them, represents the square of the inversion error; f t represents the data acquisition frequency; α, β, γ are weight coefficients; is an indicator function, and its value is determined by a preset inversion error threshold, an upper limit of resource utilization rate, and a lower limit of the health state of the sensor network. It takes 0 when the inversion error is greater than the inversion error threshold, the resource utilization rate is greater than the upper limit of the resource utilization rate, and the health state of the sensor network is lower than the lower limit of the health state of the sensor network, otherwise it takes 1; The policy network: adopts a Deep Q-Network (DQN), takes the input state s t , outputs the Q-values of each action, and selects the action corresponding to the maximum Q-value, which is expressed as follows: where a t is the selected action, s t is the state, Q is the deep Q-network, θ is the deep Q-network parameter, and is updated by the temporal difference (TD) error, expressed as follows: Among them, θ - is the target network parameter of the deep Q-network, and θ - is regularly synchronized with θ (s t+1 , a; θ - ) represents selecting the action that maximizes the Q value in state s t+1 , r t is the value function, is the gradient, Q(s t , a t ; θ) is the Q value in the case of state s t and action a t , and η is the learning rate.

Images