A karst tunnel risk intelligent early warning method, device and equipment based on deep learning

Abstract

The present invention relates to the technical field of geological risk monitoring and early warning in tunnel construction, and specifically to a karst tunnel risk intelligent early warning method, device and equipment based on deep learning. The method comprises: obtaining multi-source coupling data; using a generative adversarial network to perform high-dimensional fusion on the processed multi-source coupling data, and reconstructing a high-precision three-dimensional model of the detection area; classifying the geological risk level of the detection area, and generating a corresponding risk distribution map; generating a dynamic risk evaluation index, and triggering an early warning signal when the risk level exceeds a preset threshold. The present invention comprises the steps of multi-source data acquisition, data preprocessing, high-precision three-dimensional model reconstruction, risk level classification, dynamic risk monitoring and early warning, and the like, and can solve the problems of single risk detection means, insufficient evaluation accuracy and poor real-time performance in existing karst tunnel construction.

Description

Technical Field

[0001] The present invention relates to the technical field of geological risk monitoring and early warning in tunnel construction, and in particular to a karst tunnel risk intelligent early warning method, device and equipment based on deep learning. Background Art

[0002] Karst tunnels have high-risk factors such as caves, underground rivers and water leakage due to the complexity of geological structure, which bring great safety hazards to tunnel construction. The morphology and distribution of karst development in karst tunnels are significantly random and uneven. Accurately detecting and evaluating the development degree of these risk areas is an important technical requirement to ensure construction safety.

[0003] Traditional geological exploration methods mainly rely on manual experience and detection data from a single device, such as ground penetrating radar (GPR), tunnel seismic probe (TSP) and tunnel resistivity probe (ERT). These methods usually lack the ability to comprehensively analyze multi-source data and are difficult to fully reveal the spatial distribution and dynamic changes of karst development under complex geological conditions. In addition, these methods are limited by equipment resolution and detection methods, making it difficult to accurately characterize the degree of karst development and conduct real-time dynamic monitoring. They often need to rely on subsequent data analysis and expert judgment, and are highly subjective and limited.

[0004] In the existing technology, some methods try to improve the detection accuracy by fusing multi-source data, but the spatial and temporal heterogeneity of multi-source data makes it difficult for traditional methods to handle it effectively. In addition, how to dynamically capture the risk change trend and trigger real-time warning signals through efficient algorithms is still a difficulty and research hotspot in the field of karst tunnel risk assessment. Therefore, there is an urgent need for an intelligent early warning method that combines real-time analysis of multi-source data, dynamic change monitoring and deep learning models to improve the accuracy and timeliness of geological risk assessment of karst tunnels and provide a scientific basis for construction safety. Summary of the invention

[0005] The technical problem to be solved by the present invention is that the existing geological exploration means in karst tunnel construction are single, data processing is not comprehensive, real-time performance is poor, and risk assessment is inaccurate. The purpose is to provide a karst tunnel risk intelligent early warning method, device and equipment based on deep learning, which realizes the efficient fusion of multi-source coupling data, accurate quantification of karst development degree, intelligent classification of geological risk levels and real-time monitoring and early warning of dynamic risk trends.

[0006] The present invention is achieved through the following technical solutions:

[0007] A karst tunnel risk intelligent early warning method based on deep learning, comprising:

[0008] Collect the geological electromagnetic field, elastic wave velocity field and resistivity of the tunnel floor and the front of the heading face, and integrate them to obtain the original multi-source coupling data; and pre-process the original multi-source coupling data to obtain multi-source coupling data;

[0009] Generative adversarial networks are used to perform high-dimensional fusion of the processed multi-source coupling data to reconstruct a high-precision three-dimensional model of the detection area;

[0010] Based on the deep residual network, combined with the reconstructed high-precision 3D model and its derived geometric and gradient features, the geological risk level of the detection area is classified and the corresponding risk distribution map is generated;

[0011] Based on time series analysis and spatial trend analysis, the dynamic change trend of risk level is monitored in real time, dynamic risk assessment indicators are generated, and early warning signals are triggered when the risk level exceeds the preset threshold.

[0012] Specifically, the original multi-source coupling data ,in, is the radar signal reflection intensity, is the elastic wave amplitude, is the resistivity value in space;

[0013] The methods for preprocessing the original multi-source coupling data include:

[0014] Perform discrete wavelet decomposition on the original multi-source coupling data, set a threshold and remove the wavelet components with wavelet coefficients less than the threshold, and reconstruct the signal ;

[0015] At each data point Selection Window , minimize the error function , fitting a local polynomial, where is the position in the window, are local polynomial coefficients, is the polynomial order; the original data points are replaced by fitted values ​​to generate smoothed data , is the order of the local polynomial;

[0016] Use exponential gain control to enhance the smooth data and obtain the final multi-source coupling data ,in, is the gain factor, is the nonlinear gain exponent.

[0017] Specifically, the method for obtaining a high-precision three-dimensional model includes:

[0018] Based on the karst fractal characteristics and hydrochemical dynamics parameters, a karst development degree evaluation model is established; and the comprehensive index of karst development degree is calculated;

[0019] Combined with the comprehensive index of karst development degree, the deep deterministic policy gradient algorithm is used to obtain the detection path and measurement point planning;

[0020] The multi-source coupling data obtained based on the detection path and measurement point planning are used as the input condition of the generative adversarial network model to reconstruct a high-precision three-dimensional model of the detection area.

[0021] Specifically, the method for calculating the comprehensive index of karst development degree includes:

[0022] Space fractal dimension model ,in, is the grid length, To count the number of karst bodies filling the grid;

[0023] Water Chemistry Fractal Index Model ,in, is the water chemistry index value, is the probability density of the indicator;

[0024] Karst comprehensive fractal index model ,in, and is the weight parameter;

[0025] Calculation of comprehensive index of karst development degree ,in, It is an evaluation index for the possibility of karst development.

[0026] Specifically, the training method of the deep deterministic policy gradient algorithm includes:

[0027] Determine the reinforcement learning environment, state space , Action Space , reward function and the state transition function ,in, is the comprehensive index of karst development degree, is the current location of the device, The detection signal strength of the device at the current moment; is the three-dimensional moving step length of the device, is the moving cost of the equipment, and is the weight factor, is the state transfer function;

[0028] Determine the policy network for the deep deterministic policy gradient algorithm, ,in, is the policy function, are the parameters of the policy network;

[0029] Determine the value network of the deep deterministic policy gradient algorithm, which evaluates the value function for a given state-action pair ,in, is the state-action value function, is the parameter of the value network;

[0030] Determine the objective function of the deep deterministic policy gradient algorithm. The goal of the policy network is to maximize the cumulative reward and update the parameters of the policy network: ,in, is the expectation operation, is the objective function of the policy network;

[0031] The loss function of the value network is: ,in, , is the discount factor, and is the target network parameter, is the target value, is the number of samples of mini-batches randomly sampled from the experience replay pool, is the target value network, for the target strategy network;

[0032] Use soft update rules to gradually adjust the target network parameters, , ,in, is the soft update coefficient;

[0033] Initialize policy network parameters , Value Network Parameters , target network parameters and , Experience replay pool;

[0034] According to the current strategy network Determine the action space ;implement , update status , record reward ;

[0035] Will Store in the experience replay pool, and randomly sample small batches of data from the experience replay pool;

[0036] Update the value network and update the policy network parameters ,in, is the gradient operation;

[0037] Update the target network parameters according to the soft update rules , , and iteratively train until the policy network parameters and value network parameters Convergence, training completed.

[0038] Specifically, the training method of the generative adversarial network model includes:

[0039] Construct the generator and discriminator of the generative adversarial network; the generator includes an encoder and a decoder, and the discriminator includes a convolutional layer, a batch normalization layer, and a fully connected layer;

[0040] Extract high-dimensional features of multi-source data tensors through the encoder of the generator ,in, is an encoder network consisting of 3D convolutional layers and residual blocks;

[0041] The high-dimensional features are weighted and fused through the attention mechanism to obtain multi-source fusion features. ,in, and are the weights and biases of the attention mechanism, Calculate the attention weight;

[0042] The decoder of the generator decodes the multi-source fusion features to generate a 3D model ,in, is a decoder network consisting of a deconvolution layer and an upsampling module;

[0043] Determine the generator base loss function ,in, The three-dimensional model generated by the generator, is the expectation operation, For the generator according to the input data The generated 3D model, Generate 3D model for discriminator pair Authenticity rating;

[0044] Determine the reconstruction loss function ;

[0045] Determine the multi-scale perceptual loss function ,in, is the total number of discriminator layers in the perceptual loss, is the first Layer extraction module;

[0046] Determine the generator loss function ,in, and is the weight factor;

[0047] Determine the discriminator loss function ,in, is the distribution of the real 3D model, For the discriminator to the real 3D model Ratings;

[0048] Use the existing multi-source coupling data and real 3D models as training samples;

[0049] Coupling data from multiple sources Converted into a unified multi-source data tensor , and normalize it; input it into the generator, and the generator outputs the generated three-dimensional model;

[0050] Update the discriminator using the real 3D model and the generated 3D model, and update the generator and discriminator through the generator loss function and the discriminator loss function;

[0051] The training steps are iterated until the loss function converges to obtain the generative adversarial network model.

[0052] Specifically, the method for generating the risk distribution map includes:

[0053] Output voxel data of high-precision 3D models and its gradient characteristics Combined into input feature tensor , ;

[0054] Construct a deep residual network consisting of several residual blocks; the residual block includes ,in, is a three-dimensional convolutional layer, is the normalization layer, is the activation function, is the output of the residual block;

[0055] Add a global average pooling layer to the network output layer and compress the spatial feature vector into a global feature vector , and input the global feature vector into the fully connected layer to generate the classification score of the risk level ;in, , and are the height, width, and depth of the tensor, respectively. , and is the index variable of the tensor, and are the weights and biases of the fully connected layer;

[0056] Use the Softmax function to normalize the classification scores and generate the probability distribution of risk levels. ,in, The detection area belongs to The probability of class risk, is the total number of risk level categories, Among all the classification scores generated by the network output layer, The classification score of the class;

[0057] According to the classification results, the detection area is divided into different risk levels and the corresponding three-dimensional risk distribution map is generated. ,in, The midpoint of the detection area The risk level, To select the maximum value.

[0058] Specifically, the method of triggering the early warning signal includes:

[0059] Convert risk distribution map into dynamic input tensor ,in, is the probability distribution of risk level at the current moment, The geometric features of the high-precision 3D model. is the gradient of the geometric feature;

[0060] Construct a time series model based on a recurrent neural network to extract the trend of risk level over time ,in, is the hidden state at the current moment, and is the weight matrix of the time series model, is the bias term, is the activation function;

[0061] Extracting spatial variation features through 3D convolutional networks , is a three-dimensional convolutional layer;

[0062] Integrate time series and spatial characteristics to generate dynamic risk assessment indicators ,in, is the weight of the time series, is the weight of the spatial feature;

[0063] According to the dynamic risk assessment indicators, it is judged in real time whether the risk level exceeds the preset threshold. ,like , then the warning signal is triggered; otherwise, it is not triggered.

[0064] A deep learning-based intelligent early warning device for karst tunnel risks comprises a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method described above is implemented.

[0065] A computer-readable storage device stores a computer program, and when the computer program is executed by a processor, the method described above is implemented.

[0066] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0067] The present invention includes the steps of multi-source data collection, data preprocessing, high-precision three-dimensional model reconstruction, risk level classification, dynamic risk monitoring and early warning, etc., which can solve the problems of single risk detection means, insufficient assessment accuracy and poor real-time performance in existing karst tunnel construction; by combining deep learning technology and multi-source data fusion, the quantification ability of karst development degree under complex geological conditions and the classification accuracy of risk levels are improved; the generative adversarial network is used to realize the efficient fusion of multi-source data and the high-precision reconstruction of the three-dimensional model, and through the dynamic risk monitoring mechanism combined with time series and spatial trend characteristics, the risk change trend can be captured in real time and early warning can be quickly triggered, which greatly improves the safety and management efficiency of the construction site. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The accompanying drawings illustrate exemplary embodiments of the present invention and, together with the description thereof, are used to explain the principles of the present invention. These drawings are included to provide a further understanding of the present invention, and the accompanying drawings are included in and constitute a part of this specification and do not constitute a limitation of the embodiments of the present invention.

[0069] Figure 1 It is a flow chart of a deep learning-based intelligent early warning method for karst tunnel risks according to the present invention.

[0070] Figure 2 It is a flowchart of the training method of the generative adversarial network model according to the present invention. DETAILED DESCRIPTION

[0071] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and implementation methods. It is understood that the specific implementation methods described herein are only used to explain the relevant content, rather than to limit the present invention.

[0072] It should also be noted that, for the convenience of description, only the parts related to the present invention are shown in the drawings.

[0073] In the absence of conflict, the embodiments and features of the embodiments of the present invention may be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0074] Embodiment 1, as Figure 1 As shown, a karst tunnel risk intelligent early warning method based on deep learning is provided, comprising:

[0075] The geological electromagnetic field, elastic wave velocity field and resistivity of the tunnel floor and in front of the heading face are collected and integrated to obtain the original multi-source coupling data; and the original multi-source coupling data are pre-processed to obtain multi-source coupling data.

[0076] During tunnel construction, geological electromagnetic fields, elastic wave velocity fields, and resistivity are collected by ground penetrating radar (GPR), tunnel seismic wave detector (TSP), and tunnel resistivity detector (ERT), respectively. These physical fields reflect the different characteristics of tunnel geology: electromagnetic fields are used to detect differences in rock electrical properties, wave velocity fields are used to capture density and elastic changes in geological structures, and resistivity reflects the conductivity of geological materials. The format and unit of the original data may be inconsistent, so they need to be unified and integrated to form original multi-source coupled data. Subsequently, the data is preprocessed by discrete wavelet decomposition denoising, local polynomial fitting smoothing, and exponential gain enhancement to improve data quality and generate unified multi-source coupled data.

[0077] A generative adversarial network is used to perform high-dimensional fusion on the processed multi-source coupling data to reconstruct a high-precision three-dimensional model of the detection area.

[0078] Generative adversarial network (GAN) is a deep learning model composed of generator and discriminator. The generator is responsible for converting multi-source data tensors into three-dimensional models, including using encoders to extract high-dimensional features of multi-source data, performing weighted fusion through attention mechanism, and decoding to generate high-precision three-dimensional models; the discriminator is used to evaluate the difference between the generated model and the real model, thereby guiding the continuous optimization of the generator. Through the high-dimensional data fusion capability of GAN, the karst development morphology of the detection area can be accurately reconstructed to generate a realistic and high-resolution three-dimensional model.

[0079] Based on the deep residual network, combined with the reconstructed high-precision 3D model and its derived geometric and gradient features, the geological risk level of the detection area is classified and the corresponding risk distribution map is generated.

[0080] The deep residual network (ResNet) is a model widely used in deep neural networks. It solves the gradient vanishing problem in deep networks by introducing residual connections. In this embodiment, the voxel data and its gradient features output by the reconstructed three-dimensional model are used as input tensors. After extracting features through several residual blocks, they are compressed into global feature vectors through the global average pooling layer. Then, the risk level classification score is generated using the fully connected layer, and the risk probability distribution is calculated through the Softmax function, and finally a three-dimensional risk distribution map reflecting the geological risk levels of different regions is generated.

[0081] Based on time series analysis and spatial trend analysis, the dynamic change trend of risk level is monitored in real time, dynamic risk assessment indicators are generated, and early warning signals are triggered when the risk level exceeds the preset threshold.

[0082] Time series analysis uses recurrent neural networks (such as LSTM) to model and extract its dynamic characteristics by analyzing the changing trends of risk distribution at different time points. Spatial trend analysis uses a three-dimensional convolutional network to capture the changing patterns of risk distribution in space. After fusing time series features with spatial features, dynamic risk assessment indicators are generated to reflect the comprehensive changes in regional risks in real time. When the dynamic risk assessment indicator exceeds the preset threshold, the system triggers an early warning signal to indicate that there is a potential danger in the construction area so that corresponding safety measures can be taken.

[0083] Example 2: Original multi-source coupling data ,in, It is the radar signal reflection intensity, collected by ground penetrating radar (GPR), reflecting the differences in electromagnetic characteristics of different media in the tunnel face and floor geology. It is the elastic wave amplitude, collected by the tunnel seismic probe (TSP), reflecting the characteristics of the elastic modulus and wave velocity field of the geological material. It is the resistivity value in space, collected by the tunnel resistivity detector (ERT), reflecting the conductivity of the geological body and the uniformity of its internal structure.

[0084] The methods for preprocessing the original multi-source coupling data include:

[0085] Perform discrete wavelet decomposition on the original multi-source coupling data to extract the wavelet coefficients of each scale. Set a threshold and remove the wavelet components whose wavelet coefficients are less than the threshold, filter out the noise signal, and reconstruct the signal. .

[0086] The local polynomial fit is done by Select a window around , fitting local characteristics to eliminate subtle random fluctuations. Minimize the error function , fitting a local polynomial, where is the position in the window, are local polynomial coefficients, is the polynomial order; the specific steps are: for each data point Fit the polynomial and minimize the error function, replace the original data points with the fitted polynomial values ​​to generate smooth data , is the order of the local polynomial;

[0087] The resolution and contrast of the data are improved by nonlinear enhancement methods, making the subtle geological features more prominent. The smoothed data is enhanced by exponential gain control to obtain the final multi-source coupled data. ,in, is the gain factor, which is used to control the overall amplification strength; is a nonlinear gain index used to enhance the contrast of local details.

[0088] This method combines wavelet decomposition, local polynomial fitting and smoothing enhancement techniques to extract the core information of geological characteristics from complex multi-source data.

[0089] Wavelet decomposition is used to eliminate noise and retain meaningful low-frequency and medium-frequency features; local polynomial fitting further smoothes the data and reduces the impact of random errors on the overall trend; exponential gain control enhances the data to make key geological features more prominent.

[0090] In the third embodiment, a method for obtaining a high-precision three-dimensional model includes:

[0091] Based on the fractal characteristics of karst and hydrochemical kinetic parameters, a karst development degree evaluation model was established, and the comprehensive index of karst development degree was calculated.

[0092] Combined with the comprehensive index of karst development degree, the deep deterministic policy gradient algorithm is used to obtain the detection path and measurement point planning.

[0093] The multi-source coupling data obtained based on the detection path and measurement point planning are used as the input condition of the generative adversarial network model to reconstruct a high-precision three-dimensional model of the detection area.

[0094] The specific method description is given below.

[0095] Methods for calculating the comprehensive index of karst development degree include:

[0096] The spatial fractal dimension is used to quantify the geometric complexity and self-similarity of karst bodies, and to reflect the distribution law of karst bodies at different scales. ,in, is the grid length, To count the number of karst bodies filling the grid; by dividing the grid at different scales and counting the number of grids filled by karst bodies, the logarithmic relationship between the grid length and the filling number is calculated to obtain the spatial fractal dimension.

[0097] Water Chemistry Fractal Index Model ,in, is the water chemistry index value, The hydrochemical fractal index describes the distribution complexity and probability characteristics of hydrochemical indicators in karst groundwater, and reflects the influence of hydrochemical dynamics on karst development. By statistically analyzing the distribution of hydrochemical indicators and calculating the logarithmic relationship between the indicator value and its probability density, the hydrochemical fractal index is obtained. The larger the value, the more complex the distribution of hydrochemical characteristics.

[0098] Karst comprehensive fractal index model ,in, and is the weight parameter; the comprehensive fractal index combines the spatial fractal dimension and the hydrochemical fractal index, quantifies the development degree of the karst body, assigns different weights, and integrates the characteristics of the two dimensions to generate a comprehensive fractal index that reflects the overall characteristics of karst development.

[0099] Comprehensive index of karst development degree for evaluating karst development degree ,in, The karst development possibility evaluation index is used to generate a unified index to characterize the degree and risk level of karst development by combining the comprehensive fractal index and the possibility of karst development.

[0100] The training methods of the deep deterministic policy gradient algorithm include:

[0101] Determine the reinforcement learning environment, state space , Action Space , reward function and the state transition function ,in, is the comprehensive index of karst development degree, is the current location of the device, The detection signal strength of the device at the current moment; is the three-dimensional moving step length of the device, is the moving cost of the equipment, and is the weight factor, is the state transfer function;

[0102] Determine the policy network for the deep deterministic policy gradient algorithm, ,in, Is the strategy function, used to Generate next action ; The parameters of the policy network. The actions of the device are output according to the policy function. , that is, the direction and step size of the next move.

[0103] Determine the value network of the deep deterministic policy gradient algorithm, which evaluates the value function for a given state-action pair ,in, is the state-action value function, is the parameter of the value network;

[0104] Determine the objective function of the deep deterministic policy gradient algorithm. The goal of the policy network is to maximize the cumulative reward and update the parameters of the policy network: ,in, It is an expectation operation, which is used to comprehensively consider all possible paths and rewards. is the objective function of the policy network;

[0105] The value network is optimized by minimizing the temporal difference error. The loss function of the value network is: ,in, , is the discount factor, and is the target network parameter, is the target value, is the number of samples of mini-batches randomly sampled from the experience replay pool, is the target value network, for the target strategy network;

[0106] In order to improve the stability of training, the target network parameters are gradually adjusted using the soft update rule. , ,in, is the soft update coefficient;

[0107] Initialize policy network parameters , Value Network Parameters , target network parameters and , Experience replay pool;

[0108] According to the current strategy network Determine the action space ;implement , update status , record reward ;

[0109] Will Store in the experience replay pool, and randomly sample small batches of data from the experience replay pool;

[0110] Update the value network through the value network loss function, update the value network, and update the policy network parameters ,in, is the gradient operation;

[0111] Update the target network parameters according to the soft update rules , , and iteratively train until the policy network parameters and value network parameters Convergence, training completed.

[0112] like Figure 2 As shown in Figure 1, the training method of the generative adversarial network model includes:

[0113] Construct the generator and discriminator of the generative adversarial network; the core goal of the generator is to generate a 3D model that is close to the real data distribution from the input data, including the encoder and decoder. The generator continuously generates a 3D model that is as similar as possible to the real model, and continuously improves the generation effect through adversarial training with the discriminator. The goal of the discriminator is to distinguish whether the input data is a 3D model generated by the generator or a real 3D model, including convolutional layers, batch normalization layers, and fully connected layers. The discriminator is continuously optimized to improve its accuracy in distinguishing the real model, while making the model generated by the generator closer to the real model.

[0114] The generator and the discriminator are trained through game theory to continuously optimize the loss function. When the distribution of the three-dimensional model generated by the generator is almost consistent with that of the real model, the generator and the discriminator reach a balance in the game.

[0115] Extract high-dimensional features of multi-source data tensors through the encoder of the generator ,in, is an encoder network consisting of 3D convolutional layers and residual blocks;

[0116] The high-dimensional features are weighted and fused through the attention mechanism to obtain multi-source fusion features. ,in, and are the weights and biases of the attention mechanism, Calculate the attention weight;

[0117] The decoder of the generator decodes the multi-source fusion features to generate a 3D model ,in, is a decoder network consisting of a deconvolution layer and an upsampling module;

[0118] Determine the generator base loss function ,in, The three-dimensional model generated by the generator, is the expectation operation, For the generator according to the input data The generated 3D model, Generate 3D model for discriminator pair The basic loss function expresses the ability of the 3D model generated by the generator to be judged as real by the discriminator.

[0119] Determine the reconstruction loss function ;

[0120] Determine the multi-scale perceptual loss function ,in, is the total number of discriminator layers in the perceptual loss, is the first The multi-scale perceptual loss function captures the feature differences between the generated model and the true model at different scales.

[0121] Determine the generator loss function ,in, and is the weight factor;

[0122] Determine the discriminator loss function ,in, is the distribution of the real 3D model, For the discriminator to the real 3D model 's rating.

[0123] Use the existing multi-source coupling data and real 3D models as training samples;

[0124] Coupling data from multiple sources Converted into a unified multi-source data tensor , and normalize it; input it into the generator, and the generator outputs the generated three-dimensional model;

[0125] Use the real 3D model and the generated 3D model to update the discriminator, calculate its authenticity score, and update the generator and discriminator through the generator loss function and the discriminator loss function;

[0126] The training steps are iterated until the loss function converges to obtain the generative adversarial network model.

[0127] In the fourth embodiment, the method for generating a risk distribution map includes:

[0128] Output voxel data of high-precision 3D models and its gradient characteristics Combined into input feature tensor The voxel data output by the high-precision 3D model describes the geological characteristics of each 3D point in the detection area. The gradient feature reflects the rate of change of the attributes of each point in 3D space by calculating the gradient of the 3D model in space, capturing the boundary and mutation information of the geological features. The input feature tensor provides complete geometric and gradient information as the basic input for the generation of risk distribution maps.

[0129] Construct a deep residual network consisting of several residual blocks; the residual block includes ,in, is a three-dimensional convolutional layer, is the normalization layer, is the activation function, is the output of the residual block; the introduction of skip connection alleviates the gradient vanishing problem, enabling the deep network to better capture the complex risk features in three-dimensional space.

[0130] Add a global average pooling layer to the network output layer and compress the spatial feature vector into a global feature vector , and input the global feature vector into the fully connected layer to generate the classification score of the risk level ;in, , and are the height, width, and depth of the tensor, respectively. , and is the index variable of the tensor, and are the weights and biases of the fully connected layer.

[0131] Use the Softmax function to normalize the classification scores and generate the probability distribution of risk levels. ,in, The detection area belongs to The probability of class risk, is the total number of risk level categories, Among all the classification scores generated by the network output layer, The classification score of the class.

[0132] According to the classification results, the detection area is divided into different risk levels and the corresponding three-dimensional risk distribution map is generated. ,in, The midpoint of the detection area level of risk.

[0133] The voxel data and gradient features output by the high-precision 3D model are used to provide geometric and spatial change information of the detection area. The deep residual network extracts multi-level spatial risk features through convolution and residual connection. The global average pooling layer and the fully connected layer convert the spatial features into a global risk level score. The Softmax function normalizes the classification results and generates a risk probability distribution for each location. According to the maximum probability principle, the detection area is divided into different risk levels and a 3D risk distribution map is generated.

[0134] In embodiment 5, the method for triggering the warning signal includes:

[0135] Convert risk distribution map into dynamic input tensor ,in, It is the probability distribution of risk level at the current moment, reflecting the possibility of different risk levels at the current moment. The geometric features of the high-precision three-dimensional model represent the morphology and geological structure of the area. It is the gradient of geometric features, capturing the rate of change of geological features and key boundaries.

[0136] Construct a time series model based on a recurrent neural network to extract the trend of risk level over time ,in, is the hidden state at the current moment, and is the weight matrix of the time series model, is the bias term, is the activation function; the hidden state can remember and predict the changing trend of the risk level in the time dimension.

[0137] Extracting spatial variation features through 3D convolutional network (3D-CNN) ;

[0138] Integrate time series and spatial characteristics to generate dynamic risk assessment indicators ,in, is the weight of the time series, is the weight of the spatial feature;

[0139] According to the dynamic risk assessment indicators, it is judged in real time whether the risk level exceeds the preset threshold. If the threshold is exceeded, it indicates that the risk level has reached a dangerous level and the system triggers an early warning signal; if the threshold is not exceeded, it indicates that the risk level is within a controllable range and no early warning needs to be triggered.

[0140] The dynamic input tensor integrates the risk distribution, geometric features and their gradients to form the basis of temporal and spatial changes. The recurrent neural network captures the temporal changes of the risk level, and the three-dimensional convolutional network extracts the spatial distribution characteristics to ensure the complete expression of temporal and spatial information. Combining temporal and spatial characteristics, the dynamic risk assessment index accurately evaluates the risk level changes at each point. Based on the threshold, real-time judgment and triggering of early warning signals provide timely and reliable safety guarantees for construction.

[0141] Embodiment 6, a karst tunnel risk intelligent early warning device based on deep learning, includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the above method is implemented when the processor executes the computer program.

[0142] The memory can be used to store software programs and modules. The processor executes various functional applications and data processing of the terminal by running the software programs and modules stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, an execution program required for at least one function, etc.

[0143] The data storage area can store data created according to the use of the terminal, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0144] A computer-readable storage device stores a computer program, and the computer program implements the above method when executed by a processor.

[0145] Without loss of generality, computer readable media may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information such as computer readable instruction data structures, program modules or other data. Computer storage media include RAM, ROM, EPROM, EEPROM, flash memory or other solid-state storage technology, CD-ROM, DVD or other optical storage, cassettes, magnetic tapes, disk storage or other magnetic storage devices. Of course, those skilled in the art will appreciate that computer storage media are not limited to the above. The above-mentioned system memory and mass storage devices can be collectively referred to as memory.

[0146] In the description of this specification, the description with reference to the terms "one embodiment / method", "some embodiments / methods", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment / method or example are included in at least one embodiment / method or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment / method or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments / methods or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments / methods or examples described in this specification and the features of the different embodiments / methods or examples, unless they are contradictory.

[0147] It should be understood by those skilled in the art that the above embodiments are only for the purpose of clearly illustrating the present invention, and are not intended to limit the scope of the present invention. For those skilled in the art, other changes or modifications may be made based on the above invention, and these changes or modifications are still within the scope of the present invention.

Claims

1. A karst tunnel risk intelligent early warning method based on deep learning, characterized in that: include: Collect the geological electromagnetic field, elastic wave velocity field and resistivity of the tunnel floor and the front of the face, and integrate them to obtain the original multi-source coupling data; And preprocessing the original multi-source coupling data to obtain multi-source coupling data; Generative adversarial networks are used to perform high-dimensional fusion of the processed multi-source coupling data to reconstruct a high-precision three-dimensional model of the detection area; Based on the deep residual network, combined with the reconstructed high-precision 3D model and its derived geometric and gradient features, the geological risk level of the detection area is classified and the corresponding risk distribution map is generated; Based on time series analysis and spatial trend analysis, the dynamic change trend of risk level is monitored in real time, dynamic risk assessment indicators are generated, and early warning signals are triggered when the risk level exceeds the preset threshold; Among them, the methods of triggering the early warning signal include: Convert risk distribution map into dynamic input tensor ,in, is the probability distribution of risk level at the current moment, The geometric features of the high-precision 3D model. is the gradient of the geometric feature; Construct a time series model based on a recurrent neural network to extract the trend of risk level over time ,in, is the hidden state at the current moment, and is the weight matrix of the time series model, is the bias term, is the activation function; Extracting spatial variation features through 3D convolutional networks , It is a three-dimensional convolutional layer; Integrate time series and spatial characteristics to generate dynamic risk assessment indicators ,in, is the weight of the time series, is the weight of the spatial feature; According to the dynamic risk assessment indicators, it is judged in real time whether the risk level exceeds the preset threshold. ,like , then the warning signal is triggered; otherwise, it is not triggered.

2. According to claim 1, a karst tunnel risk intelligent early warning method based on deep learning is characterized in that: Original multi-source coupling data ,in, is the radar signal reflection intensity, is the elastic wave amplitude, is the resistivity value in space; The methods for preprocessing the original multi-source coupling data include: Perform discrete wavelet decomposition on the original multi-source coupling data, set a threshold and remove the wavelet components with wavelet coefficients less than the threshold, and reconstruct the signal ; At each data point Selection Window , minimize the error function , fitting a local polynomial, where is the position in the window, are local polynomial coefficients, is the polynomial order; the original data points are replaced by fitted values ​​to generate smoothed data , is the order of the local polynomial; Use exponential gain control to enhance the smooth data and obtain the final multi-source coupling data ,in, is the gain factor, is the nonlinear gain exponent.

3. According to claim 1, a karst tunnel risk intelligent early warning method based on deep learning is characterized in that: Methods for obtaining high-precision three-dimensional models include: Based on the karst fractal characteristics and hydrochemical dynamics parameters, a karst development degree evaluation model is established; and the comprehensive index of karst development degree is calculated; Combined with the comprehensive index of karst development degree, the deep deterministic policy gradient algorithm is used to obtain the detection path and measurement point planning; The multi-source coupling data obtained based on the detection path and measurement point planning are used as the input condition of the generative adversarial network model to reconstruct a high-precision three-dimensional model of the detection area.

4. According to claim 3, a karst tunnel risk intelligent early warning method based on deep learning is characterized in that: The methods for calculating the comprehensive index of karst development degree include: Space fractal dimension model ,in, is the grid length, To count the number of karst bodies filling the grid; Water Chemistry Fractal Index Model ,in, is the water chemistry index value, is the probability density of the indicator; Karst comprehensive fractal index model ,in, and is the weight parameter; Calculation of comprehensive index of karst development degree ,in, It is an evaluation index for the possibility of karst development.

5. The method for intelligent early warning of karst tunnel risks based on deep learning according to claim 3 is characterized in that: The training methods of the deep deterministic policy gradient algorithm include: Determine the reinforcement learning environment, state space , Action Space , reward function and the state transition function ,in, is the comprehensive index of karst development degree, is the current location of the device, The detection signal strength of the device at the current moment; is the three-dimensional moving step length of the device, is the moving cost of the equipment, and is the weight factor, is the state transfer function; Determine the policy network for the deep deterministic policy gradient algorithm, ,in, is the policy function, are the parameters of the policy network; Determine the value network of the deep deterministic policy gradient algorithm, which evaluates the value function for a given state-action pair ,in, is the state-action value function, is the parameter of the value network; Determine the objective function of the deep deterministic policy gradient algorithm. The goal of the policy network is to maximize the cumulative reward and update the parameters of the policy network: ,in, is the expectation operation, is the objective function of the policy network; The loss function of the value network is: ,in, , is the discount factor, and is the target network parameter, is the target value, is the number of samples of mini-batches randomly sampled from the experience replay pool, is the target value network, for the target strategy network; Use soft update rules to gradually adjust the target network parameters, , ,in, is the soft update coefficient; Initialize policy network parameters , Value Network Parameters , target network parameters and , Experience replay pool; According to the current strategy network Determine the action space ;implement , update status , record reward ; Will Store in the experience replay pool, and randomly sample small batches of data from the experience replay pool; Update the value network and update the policy network parameters ,in, is the gradient operation; Update the target network parameters according to the soft update rules , , and iteratively train until the policy network parameters and value network parameters Convergence, training completed.

6. The method for intelligent early warning of karst tunnel risks based on deep learning according to claim 4 is characterized in that: The training methods for the generative adversarial network model include: Construct the generator and discriminator of the generative adversarial network; the generator includes an encoder and a decoder, and the discriminator includes a convolutional layer, a batch normalization layer, and a fully connected layer; Extract high-dimensional features of multi-source data tensors through the encoder of the generator ,in, is an encoder network consisting of 3D convolutional layers and residual blocks; The high-dimensional features are weighted and fused through the attention mechanism to obtain multi-source fusion features. ,in, and are the weights and biases of the attention mechanism, Calculate the attention weight; The decoder of the generator decodes the multi-source fusion features to generate a 3D model ,in, is a decoder network consisting of a deconvolution layer and an upsampling module; Determine the generator base loss function ,in, The three-dimensional model generated by the generator, is the expectation operation, For the generator according to the input data The generated 3D model, Generate 3D model for discriminator pair Authenticity rating; Determine the reconstruction loss function ; Determine the multi-scale perceptual loss function ,in, is the total number of discriminator layers in the perceptual loss, is the first Layer extraction module; Determine the generator loss function ,in, and is the weight factor; Determine the discriminator loss function ,in, is the distribution of the real 3D model, For the discriminator to the real 3D model Ratings; Use the existing multi-source coupling data and real three-dimensional models as training samples; Coupling data from multiple sources Converted into a unified multi-source data tensor , and normalize it; input it into the generator, and the generator outputs the generated three-dimensional model; Update the discriminator using the real 3D model and the generated 3D model, and update the generator and discriminator through the generator loss function and the discriminator loss function; The training steps are iterated until the loss function converges to obtain the generative adversarial network model.

7. The method for intelligent early warning of karst tunnel risks based on deep learning according to claim 1 is characterized in that: The method for generating the risk distribution map includes: Output voxel data of high-precision 3D models and its gradient characteristics Combined into input feature tensor , ; Construct a deep residual network consisting of several residual blocks; the residual block includes ,in, is a three-dimensional convolutional layer, is the normalization layer, is the activation function, is the output of the residual block; Add a global average pooling layer to the network output layer and compress the spatial feature vector into a global feature vector , and input the global feature vector into the fully connected layer to generate Classification score of risk class ;in, , and are the height, width, and depth of the tensor, respectively. , and is the index variable of the tensor, and are the weights and biases of the fully connected layer; Use the Softmax function to normalize the classification scores and generate the probability distribution of risk levels. ,in, The detection area belongs to The probability of class risk, is the total number of risk level categories, Among all the classification scores generated by the network output layer, The classification score of the class, Generated for the fully connected layer Classification score of risk class; According to the classification results, the detection area is divided into different risk levels and the corresponding three-dimensional risk distribution map is generated. ,in, The midpoint of the detection area The risk level, To select the maximum value.

8. A karst tunnel risk intelligent early warning device based on deep learning, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

9. A computer-readable storage device storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images