Karst collapse risk analysis method and system based on geological big data mining

Abstract

The invention discloses a karst collapse risk analysis method and system based on geological big data mining, and the method comprises the steps: collecting multi-source heterogeneous geological data, and carrying out the data fusion through employing a Tucker decomposition-based tensor decomposition algorithm; cleaning, complementing and standardizing the fused data; feature selection and construction: selecting features related to the karst collapse risk; carrying out feature extraction on the preprocessed data; based on the features extracted by the feature engineering, constructing a risk assessment model by using a graph convolutional neural network, and optimizing model performance through an ensemble learning method; carrying out risk assessment and prediction by utilizing the trained model, and carrying out visualization through a geographic information system; and real-time risk assessment is realized based on an edge computing technology, and the model is continuously optimized. According to the method and the system, complex heterogeneous data can be processed, the time change trend of geological parameters can be accurately captured, the prediction accuracy under complex geological conditions is high, and the timeliness of early warning is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of engineering geological survey, and in particular to a karst collapse risk analysis method and system based on geological big data mining. Background Art

[0002] Karst collapse is a common geological disaster that poses a serious threat to the safety of water conservancy and hydropower projects. Traditional karst collapse risk assessment methods mainly rely on on-site investigations and expert experience, which have problems such as low efficiency, limited coverage, and insufficient prediction capabilities.

[0003] Several methods for karst collapse risk assessment have been disclosed in the prior art to solve the problem of relying mainly on field surveys and expert experience, such as Wang et al.'s "GIS-based multi-factor overlay analysis method" in Environmental Earth Sciences in 2019, Li et al.'s "artificial neural network-based risk prediction model" in Geomorphology in 2020, and Zhang et al.'s "risk assessment method based on remote sensing and ground subsidence monitoring" in Remote Sensing in 2021. These methods have improved the efficiency and accuracy of risk assessment to a certain extent, but still have the following shortcomings: (1) The fusion capability of multi-source heterogeneous data is limited, and the prediction accuracy under complex karst geological conditions is insufficient; (2) It is difficult to effectively capture the dynamic change characteristics of geological parameters, and lacks real-time risk assessment and early warning capabilities.

[0004] Therefore, there is an urgent need for a karst collapse risk analysis method that can make full use of geological big data and has high-precision prediction and real-time warning capabilities. Summary of the invention

[0005] In order to overcome the shortcomings of the above-mentioned technology, the purpose of the present invention is to provide a karst collapse risk analysis method and system based on geological big data mining to solve the problems of low prediction accuracy and inability to provide real-time warning in the prior art.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] A karst collapse risk analysis method based on geological big data mining includes the following steps:

[0008] Data collection and fusion steps: Collect multi-source heterogeneous geological data and use the tensor decomposition algorithm based on Tucker decomposition for data fusion;

[0009] Data preprocessing steps: cleaning, completion and standardization of the fused data;

[0010] Feature engineering steps: feature selection, selecting features related to karst collapse risk; and extracting features from the preprocessed data;

[0011] Model construction and training steps: Based on the features extracted by the feature engineering, a risk assessment model is constructed using a graph convolutional neural network, and the performance of the risk assessment model is optimized through an ensemble learning method;

[0012] Risk assessment and prediction steps: Use the trained risk assessment model to conduct risk assessment and prediction, and visualize it through geographic information system;

[0013] System optimization and iteration steps: Realize real-time risk assessment based on edge computing technology and continuously optimize the risk assessment model.

[0014] Preferably, in the data collection and fusion step, the multi-source heterogeneous geological data includes geological data, remote sensing data, hydrogeological data, engineering supervision data, and historical data.

[0015] Preferably, the geological data include geological maps, drilling data, geophysical data, lithology distribution, fault distribution; the remote sensing data include high-resolution satellite images and InSAR data; the hydrogeological data include groundwater level, flow rate, spring flow, water quality data; the engineering supervision data include tunnel internal deformation and pipeline stress; the historical data include historical collapse and subsidence event records.

[0016] Preferably, in the data preprocessing step, data cleaning includes using the IsolationForest algorithm improved by the LOF algorithm and / or the DTW algorithm to identify outliers and / or abnormal fluctuations; data completion includes using the LSTM autoencoder to fill missing values ​​of time series data, and using the thin plate spline interpolation method based on the variational method to complete spatial data (such as terrain) to fill missing parts in the spatial data; the data standardization processing method includes performing Z-score standardization on continuous data to make the data have zero mean and unit variance to reduce the impact of the dimension, performing one-hot encoding on discrete data to convert categorical variables into binary vectors to adapt to the processing requirements of the model, using the Robust Scaler method to standardize numerical data with outliers, and using the Target Encoding method to standardize categorical data.

[0017] Preferably, in the feature engineering step, feature selection includes using the Boruta algorithm combined with expert knowledge to evaluate feature importance and obtain key features;

[0018] The feature extraction includes dynamic feature extraction, which includes extracting time series features, calculating spatial features and constructing new features; the time series feature extraction adopts a dynamic feature extraction algorithm that combines wavelet transform and LSTM network; the dynamic feature extraction algorithm includes performing wavelet decomposition on time series data, and then inputting the decomposed subsequences into the LSTM network for feature learning and extraction.

[0019] Preferably, the key characteristics include one or more of the degree of karst development, thickness of the covering layer, annual variation of the groundwater level, rainfall infiltration coefficient, fault density, degree of cave development, surface subsidence rate, spring flow rate change rate, hydrochemical type, tunnel surrounding rock stability, groundwater flow velocity, groundwater quality changes, terrain undulation, soil moisture content, regional geostress field distribution, geological structure complexity, rock and soil strength, seasonal precipitation, rock layer inclination, and fault zone development status.

[0020] Preferably, the computing spatial features utilizes a PointNet++ network to extract terrain features from point cloud data; the point cloud data is acquired through remote sensing or on-site mapping in the data collection and fusion step.

[0021] Preferably, the new feature includes a comprehensive karst vulnerability index, which is calculated as follows:

[0022] Comprehensive karst vulnerability index = karst development degree × 0.3 + cover layer thickness × 0.2 + groundwater level fluctuation × 0.2 + fault density × 0.15 + cave development degree × 0.15.

[0023] Among them, the comprehensive karst vulnerability index aims to quantify the vulnerability of the karst geological environment in a region. Due to the complexity of the karst geological environment and the many influencing factors, there are significant differences and correlations between various parameters. Using a single parameter often cannot fully and accurately reflect the risk of karst collapse. The purpose of constructing a comprehensive karst vulnerability index is to:

[0024] 1. Integrate multiple key geological factors: Comprehensively consider the main factors affecting karst collapse, such as karst development degree, cover layer thickness, groundwater level fluctuation, fault density, and cave development degree, and combine them into a quantifiable vulnerability indicator through weighted method.

[0025] 2. Provide a unified risk assessment standard: Through the calculation of this index, a variety of complex geological conditions can be simplified into a numerical value, which is convenient for horizontal comparison and quantitative assessment of karst vulnerability in different regions and under different geological conditions.

[0026] 3. Improve the accuracy and efficiency of risk assessment: Through the calculation of the comprehensive index, the influence of subjective evaluation factors can be effectively reduced, and the scientificity and objectivity of karst collapse risk assessment can be further improved.

[0027] The construction of a comprehensive karst vulnerability index can achieve the following effects:

[0028] 1. Comprehensively reflect the vulnerability of karst geology: By combining key parameters such as karst development degree, cover layer thickness, groundwater level fluctuation, etc., the comprehensive karst vulnerability index can comprehensively reflect the regional geological vulnerability. Its weighted coefficient setting can highlight the influence of different parameters, making the risk assessment results more in line with the actual situation.

[0029] 2. Optimize the risk assessment model: As an important input feature in the model, the index can significantly improve the model's ability to capture risks in complex geological environments, especially under multi-source heterogeneous data conditions, and can improve the accuracy of risk prediction and the generalization ability of the model.

[0030] 3. Easy to visualize and support decision-making: Through the calculation of this index, a clear numerical standard can be provided for the visualization of geological risks, and it is convenient to provide a quantitative decision-making basis in actual projects, helping managers to quickly identify high-risk areas and take corresponding protective measures.

[0031] In summary, the construction of the comprehensive karst vulnerability index not only improves the systematic and scientific nature of karst collapse risk assessment, but also provides a more reliable guiding basis for engineering implementation.

[0032] Preferably, in the model construction and training steps, the construction of the risk assessment model adopts a graph convolutional neural network as the basic model, and combines ConvLSTM, LightGBM and CatBoost to build an integrated model; the integrated model adopts the Stacking integration method, with the outputs of the graph convolutional neural network, ConvLSTM, LightGBM and CatBoost as features, and trains an XGBoost model as the final predictor.

[0033] Preferably, the graph convolutional neural network includes an input layer, a hidden layer, a ReLU activation function, a regularization layer and an output layer; the regularization layer includes a BatchNormalization layer and a Dropout layer, and the parameter rate value of the Dropout layer is 0.3; the hidden layer includes 4 graph convolution layers, each layer has 64, 128, 256, and 128 neurons respectively, and a ReLU activation function, a BatchNormalization layer and a Dropout layer are sequentially set after each graph convolution layer; the output layer includes 1 neuron, and uses the Sigmoid activation function to output the risk probability.

[0034] Preferably, the ConvLSTM network comprises: 3 layers of ConvLSTM, each layer having 64 filters; 2 layers of 3D convolutional layers; 1 layer of global average pooling layer; and 2 layers of fully connected layers.

[0035] Preferably, in the model training: the data set is divided into 60% training set, 20% validation set, and 20% test set; the Focal Loss loss function is used to solve the category imbalance problem; the optimizer adopts the AdamW algorithm, the initial learning rate is 0.001, and the cosine annealing schedule is used; the training strategy adopts five-fold cross validation, 300 rounds of training per fold, and the batch size is 256; regularization uses weight decay and Dropout, and the parameter rate value of Dropout is 0.3.

[0036] Preferably, in the risk assessment and prediction step, the karst collapse risk is divided into multiple levels, and the specific classification criteria are dynamically adjusted according to the actual situation; in actual use, the multi-source heterogeneous geological data of a certain area acquired in real time is input into the trained risk assessment model for processing to obtain a risk probability in the numerical range of 0 to 1, and the karst collapse risk level of the area is determined based on the obtained numerical value.

[0037] Preferably, the karst collapse risk level classification method is as follows: multiply the risk probability value obtained by real-time processing by 100 to obtain an integer value, and classify the risk according to the range of the integer value:

[0038] When the integer value is in the range of 0 to 20, the karst collapse risk level of the area is low risk; or, when the integer value is in the range of 21 to 40, the karst collapse risk level of the area is medium-low risk; or, when the integer value is in the range of 41 to 60, the karst collapse risk level of the area is medium risk; or, when the integer value is in the range of 61 to 80, the karst collapse risk level of the area is medium-high risk; or, when the integer value is in the range of 81 to 100, the karst collapse risk level of the area is high risk.

[0039] Preferably, the visualization is implemented using a QGIS plug-in, and different karst collapse risk levels are distinguished by different colors to form a risk distribution map, thereby achieving dynamic visualization of risk scores; the risk distribution map includes a 2D risk distribution map and a 3D risk stereogram.

[0040] Preferably, in the system optimization and iteration step, the edge computing technology includes deploying an edge server on site, pre-installing a lightweight model, receiving multi-source heterogeneous geological data obtained by real-time monitoring for preliminary risk assessment, triggering an early warning when the risk score exceeds a preset threshold, uploading it to the cloud, and analyzing it through a trained risk assessment model; the lightweight model is obtained by compressing the trained risk assessment model;

[0041] The continuous optimization of the risk assessment model includes continuously using new multi-source heterogeneous geological data to continuously train and optimize the risk assessment model.

[0042] The present invention also provides a karst collapse risk analysis system based on geological big data mining, comprising:

[0043] Data collection and fusion module, used to collect multi-source heterogeneous geological data and perform data fusion;

[0044] Data preprocessing module, used to clean, complete and standardize the fused data;

[0045] The feature engineering module is used to select features related to karst collapse risk and perform feature extraction on the data processed by the data preprocessing module;

[0046] Model building and training module, used to build and optimize risk assessment models;

[0047] The risk assessment and prediction module is used to perform risk assessment and prediction and visualize them;

[0048] System optimization and iteration module, used to achieve real-time risk assessment and continuously optimize risk assessment models;

[0049] The system optimization and iteration module includes an edge computing unit, which is used to perform real-time data processing and preliminary risk assessment on site.

[0050] Preferably, it also includes a user interaction interface module for displaying risk assessment results and management suggestions, and receiving user input to optimize risk assessment model performance and user experience.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] The present invention adopts a multi-source heterogeneous geological data fusion method based on tensor decomposition and uses the Tucker decomposition algorithm to process high-dimensional geological data, effectively solving the limitations of traditional methods in processing multi-source heterogeneous data. This method can retain the complex associations between data and significantly improve the accuracy and efficiency of data fusion.

[0053] The present invention adopts a dynamic feature extraction algorithm based on time series analysis, which innovatively combines wavelet transform and LSTM network. First, the time series data is decomposed into multiple scales using wavelet transform to capture the features of different frequencies. Then, the decomposed subsequences are input into the LSTM network to learn long-term and short-term dependencies. This method can consider both the local details and global trends of the data, significantly improving the effect of feature extraction.

[0054] This paper adopts karst geological spatial relationship modeling based on graph convolutional neural network (GCN), innovatively applies GCN to karst geological environment, and effectively handles spatial correlation under complex karst geological conditions by constructing spatial relationship diagrams between geological elements. This method significantly improves the model's ability to handle non-Euclidean data, making risk assessment more accurate.

[0055] The present invention adopts a real-time risk assessment system driven by edge computing, proposes a distributed risk assessment architecture based on edge computing, and realizes real-time processing of the entire process from data collection to risk warning. The system significantly reduces data transmission delays, improves the timeliness of risk assessment, and provides technical support for the timely prevention of karst collapse.

[0056] In general, the method and system of the present invention have strong multi-source data fusion capabilities and can process more complex heterogeneous data; they introduce dynamic feature extraction to better capture the temporal variation trends of geological parameters; they adopt advanced deep learning models to improve the prediction accuracy under complex geological conditions; they realize real-time risk assessment based on edge computing, greatly improving the timeliness of early warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 is a flow chart of the karst collapse risk analysis method of the present invention;

[0058] Figure 2 This is a schematic diagram of multi-source heterogeneous data fusion in a specific implementation manner of the present invention;

[0059] Figure 3 This is an example of the dynamic feature extraction result in a specific embodiment of the present invention;

[0060] Figure 4 This is a risk assessment model architecture diagram based on GCN in a specific implementation manner of the present invention;

[0061] Figure 5 It is a real-time risk assessment system architecture diagram in a specific implementation manner of the present invention. DETAILED DESCRIPTION

[0062] In order to better explain the present invention, the main contents of the present invention are further explained below in conjunction with specific embodiments, but the contents of the present invention are not limited to the following embodiments.

[0063] like Figure 1 As shown, the present invention provides a karst collapse risk analysis method based on geological big data mining, comprising the following steps:

[0064] 1) Data collection and fusion: Collect multi-source heterogeneous geological data and use the tensor decomposition algorithm based on Tucker decomposition for data fusion.

[0065] Multi-source heterogeneous geological data include geological maps, lithology distribution, fault distribution, remote sensing images, groundwater level, flow velocity, water quality data and historical collapse event records.

[0066] 2) Data preprocessing: Clean, complete and standardize the fused data, including Z-score standardization of continuous data to make the data have zero mean and unit variance to reduce the impact of dimension; perform one-hot encoding on discrete data and convert categorical variables into binary vectors to meet the processing requirements of the model;

[0067] Data cleaning uses the Isolation Forest algorithm improved by the LOF algorithm and / or the dynamic time warping algorithm to identify outliers and / or abnormal fluctuations; data completion uses the LSTM autoencoder to fill in the missing values ​​of time series data, and the thin plate spline interpolation method of the variational method to complete the spatial data (such as terrain) and fill in the missing parts in the spatial data; data standardization processing methods also include using the Robust Scaler method to standardize numerical data with outliers and using the Target Encoding method to standardize categorical data.

[0068] 3) Feature engineering: Select and construct features related to karst collapse risk, and extract dynamic features using a dynamic feature extraction algorithm that combines wavelet transform and long short-term memory network (LSTM).

[0069] Feature selection uses the Boruta algorithm combined with expert knowledge to evaluate feature importance. Dynamic feature extraction includes extracting time series features, calculating spatial features, and constructing new features; the dynamic feature extraction algorithm includes wavelet decomposition of time series data, and then inputting the decomposed subsequences into the LSTM network for feature learning; calculating spatial features is to use the PointNet++ network to extract terrain features from point cloud data, where point cloud data is obtained through remote sensing or on-site mapping in the data collection and fusion steps.

[0070] Among them, the new features include the comprehensive karst vulnerability index, which is calculated as follows:

[0071] Comprehensive karst vulnerability index = karst development degree × 0.3 + cover layer thickness × 0.2 + groundwater level fluctuation × 0.2 + fault density × 0.15 + cave development degree × 0.15

[0072] 4) Model construction and training: Based on the features extracted by the feature engineering, a risk assessment model is constructed using a graph convolutional neural network (GCN), the spatial correlation under complex karst geological conditions is processed by constructing a spatial relationship graph between geological elements, and the model performance is optimized through an integrated learning method;

[0073] The risk assessment model is constructed by using graph convolutional neural network (GCN) as the basic model, and combining ConvLSTM, LightGBM and CatBoost to build an integrated model. The integrated model uses the Stacking method, takes the output of GCN, ConvLSTM, LightGBM and CatBoost as features, and trains an XGBoost model as the final predictor.

[0074] 5) Risk assessment and prediction: Use the trained risk assessment model to conduct risk assessment and prediction, and visualize it through geographic information system;

[0075] Risk scores are divided into multiple levels, and the specific classification criteria are dynamically adjusted based on actual conditions.

[0076] 6) System optimization and iteration: Based on edge computing technology, real-time processing of the entire process from data collection to risk warning is achieved, and the risk assessment model is continuously optimized.

[0077] Edge computing technology includes deploying edge servers on site, pre-installing lightweight models, receiving monitoring data in real time for preliminary risk assessment, triggering an early warning when the risk score exceeds a preset threshold, and uploading it to the cloud for in-depth analysis through the risk assessment model of the present invention.

[0078] A karst collapse risk analysis system based on geological big data mining, including:

[0079] Data collection and fusion module, used to collect and fuse multi-source heterogeneous geological data;

[0080] Data preprocessing module, used to clean, complete and standardize the fused data;

[0081] The feature engineering module is used to select and construct features related to karst collapse risk and to extract features from the data processed by the data preprocessing module;

[0082] Model building and training module, used to build and optimize risk assessment models;

[0083] The risk assessment and prediction module is used to perform risk assessment and prediction and visualize them;

[0084] System optimization and iteration module, used to achieve real-time risk assessment and continuously optimize the risk assessment model; the system optimization and iteration module includes an edge computing unit for real-time data processing and preliminary risk assessment on site;

[0085] The user interface module is used to display risk assessment results and management recommendations, while receiving user input to optimize risk assessment model performance and user experience.

[0086] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the karst collapse risk analysis method based on geological big data mining of the present invention.

[0087] The following is a further explanation of the karst collapse risk analysis method and system based on geological big data mining of the present invention by taking a large-scale inter-basin water diversion project as an example.

[0088] Background: This case involves a large-scale water diversion project across three provinces, with a total length of about 300 kilometers, of which about 50 kilometers pass through complex karst geological areas. The project involves multiple tunnels, aqueducts and culverts, and is particularly sensitive to karst collapse risks. In order to ensure the safety of the project, a long-term karst collapse risk assessment and real-time monitoring system needs to be established.

[0089] 1) Data collection and integration

[0090] 1.1) Data Collection

[0091] a. Geological data

[0092] 1:50000 geological map: covers the entire karst section, a total of 10 maps.

[0093] Drilling data: There is a borehole every 5 kilometers along the line, a total of 100 boreholes, with depths ranging from 50 to 200 meters.

[0094] Geophysical data: high-density electrical and seismic reflection detection profiles, with a total length of 100 kilometers.

[0095] b. Remote sensing data

[0096] High-resolution satellite images: 0.8-meter resolution panchromatic images and 3.2-meter resolution multispectral images acquired using the GF-2 satellite, updated every quarter.

[0097] InSAR data: C-band SAR data from the Sentinel-1 satellite is acquired every 12 days to monitor surface deformation.

[0098] c. Hydrogeological data

[0099] Groundwater level: 200 monitoring wells are set up along the route, and data is collected every hour.

[0100] Spring flow: monitor 50 important karst springs and measure flow once a day.

[0101] Water quality data: pH value, conductivity, main ion content, etc., sampled once a week.

[0102] d. Engineering monitoring data

[0103] Internal deformation of the tunnel: a monitoring section is set up every 100 meters, including parameters such as arch subsidence and peripheral convergence.

[0104] Pipeline stress: Set up a stress monitoring point every kilometer in the key section.

[0105] e. Historical data

[0106] A total of 305 collapse and subsidence incidents have been recorded in the region over the past 50 years.

[0107] 1.2) Data Fusion

[0108] like Figure 2 As shown in the figure, the improved Tucker decomposition algorithm is used to fuse multi-source heterogeneous data. A 6-order tensor is constructed, and the dimensions include space (x, y, z), time, data type, and data attribute. The core tensor and factor matrix are obtained through Tucker decomposition, and then reconstructed to obtain a unified data set after fusion.

[0109] 2) Data preprocessing

[0110] 2.1) Data cleaning

[0111] Combining Isolation Forest and LOF algorithms, an ensemble learning method is used to identify multidimensional outliers.

[0112] For groundwater level data, combined with rainfall data, the dynamic time warping (DTW) algorithm is used to detect abnormal fluctuations.

[0113] 2.2) Data completion

[0114] Time series data (such as groundwater levels): Use LSTM autoencoders to fill in missing values.

[0115] Spatial data (such as terrain): Thin-plate spline interpolation method using variational method.

[0116] 2.3) Data Standardization:

[0117] Numerical data: Use the Robust Scaler method to reduce the impact of outliers.

[0118] Categorical data: Use Target Encoding and consider the relationship between categories and target variables.

[0119] 3) Feature Engineering

[0120] 3.1) Feature Selection

[0121] Using the Boruta algorithm combined with expert knowledge, after 1,000 iterations, 20 key features were finally selected, including: karst development degree, cover layer thickness, annual groundwater level variation, rainfall infiltration coefficient, fault density, cave development degree, surface subsidence rate, spring flow rate change rate, hydrochemical type, tunnel surrounding rock stability, groundwater flow rate, groundwater quality changes, terrain undulation, soil moisture content, regional geostress field distribution, geological structure complexity, rock and soil strength, seasonal precipitation, rock layer inclination, and fault zone development status.

[0122] 3.2) Feature extraction

[0123] Time series features: Use the Tsfresh library to automatically extract time series features such as groundwater level and spring flow data. The extraction process is as follows Figure 3 As shown in the figure, the time series features are decomposed at multiple scales using wavelet transform to capture the features of different frequencies. The decomposed subsequences are then input into the LSTM network for feature learning to extract dynamic features, which include trend features, periodic features, mutation features, and correlation features.

[0124] Spatial features: Use the PointNet++ network to extract terrain features from point cloud data obtained from remote sensing or field mapping.

[0125] Image features: Use pre-trained ResNet50 to extract texture and structural features from satellite images.

[0126] New structural features: Definition of “comprehensive karst vulnerability index” = karst development degree × 0.3 + cover layer thickness × 0.2 + groundwater level fluctuation × 0.2 + fault density × 0.15 + cave development degree × 0.15.

[0127] 4) Model construction and training

[0128] 4.1) Graph Convolutional Neural Network (GCN) Construction

[0129] like Figure 4 As shown, the graph convolutional neural network (GCN) of the present invention includes:

[0130] Input layer: 20 features + adjacency matrix (based on the spatial relationship of geological units);

[0131] Hidden layer: 4 graph convolution layers, each with 64, 128, 256, and 128 neurons respectively;

[0132] Activation function: A ReLU activation function is set after each graph convolution layer;

[0133] Regularization: BatchNormalization and Dropout (rate = 0.3) are added after each ReLU activation function;

[0134] Output layer: 1 neuron, using Sigmoid activation function to output risk probability;

[0135] 4.2) Spatiotemporal sequence model

[0136] Construct a ConvLSTM network to process spatiotemporal sequence data. The ConvLSTM network is set up in sequence: 3 layers of ConvLSTM, each with 64 filters; 2 layers of 3D convolution layers; 1 layer of global average pooling layer; and 2 layers of fully connected layers.

[0137] The connection order of each layer of the ConvLSTM network is determined as follows:

[0138] 1. Step 1: 3 ConvLSTM layers

[0139] ConvLSTM (Convolutional Long Short-Term Memory) is used to process spatiotemporal sequence data. It has three consecutive layers, each with 64 filters. These three layers are connected layer by layer to extract spatiotemporal sequence features.

[0140] 2. Step 2: 2 3D convolutional layers

[0141] After the ConvLSTM layer, two layers of 3D convolution layers are connected to further capture the local spatial features in the spatiotemporal data. These two layers of 3D convolution can enhance the extraction of spatial features and process multi-dimensional spatiotemporal data.

[0142] 3. Step 3: Global Average Pooling Layer

[0143] After the 3D convolution layer, a global average pooling layer is connected to reduce the spatial information of each feature map to a global feature. The function of this layer is to reduce the data dimension by averaging and extract the global spatiotemporal features.

[0144] 4. Step 4: 2 fully connected layers

[0145] Finally, two fully connected layers are used to map the high-dimensional features extracted previously to the output space, refine the feature representation layer by layer, and finally obtain the desired output result.

[0146] The complete connection sequence is as follows:

[0147] 1. ConvLSTM layer (layer 1, 64 filters) → ConvLSTM layer (layer 2, 64 filters) → ConvLSTM layer (layer 3, 64 filters)

[0148] 2. 3D convolution layer (layer 1) → 3D convolution layer (layer 2)

[0149] 3. Global average pooling layer

[0150] 4. Fully connected layer (Layer 1) → Fully connected layer (Layer 2)

[0151] This order ensures that the spatiotemporal data undergoes sufficient feature extraction, which not only retains the dependencies of the spatiotemporal sequences, but also makes full use of the convolution operation to capture spatial features, and further refines the features through pooling layers and fully connected layers, ultimately forming an output that can be used for risk prediction.

[0152] 4.3) Integration strategy:

[0153] The Stacking integration method is adopted. The base models include: GCN model, ConvLSTM model, LightGBM (parameters: num_leaves=31, max_depth=-1, learning_rate=0.01, n_estimators=200), CatBoost (iterations 1000, learning rate 0.02); the meta-model uses XGBoost.

[0154] 4.4) Model training:

[0155] Dataset division: 60% training set, 20% validation set, 20% test set;

[0156] Loss function: Focal Loss, to solve the problem of category imbalance;

[0157] Optimizer: AdamW, initial learning rate 0.001, using cosine annealing schedule;

[0158] Training strategy: 5-fold cross validation, 300 rounds of training per fold, batch size 256;

[0159] Regularization: Use weight decay (L2 regularization) and Dropout (rate of 0.3).

[0160] 5) Risk assessment and prediction

[0161] 5.1) Regional division

[0162] The entire karst section was divided into 5000 grid cells of 200 m × 200 m.

[0163] 5.2) Risk score calculation

[0164] For each grid cell, its 20 eigenvalues ​​are input into the trained risk assessment model to obtain a risk probability of 0-1, which is then multiplied by 100 and rounded to an integer to obtain a risk score.

[0165] 5.3) Risk level classification

[0166] 0-20: Low risk, green

[0167] 21-40: medium-low risk, light green

[0168] 41-60: Medium risk, yellow

[0169] 61-80: Medium to high risk, orange

[0170] 81-100: High risk, red

[0171] 5.4) Dynamic risk map generation:

[0172] A plug-in was developed using QGIS to achieve dynamic visualization of risk scores, including 2D risk distribution maps and 3D risk stereo maps.

[0173] 6) System optimization and iteration

[0174] like Figure 5 As shown, system optimization and iteration include the following steps:

[0175] 6.1) Edge computing network deployment

[0176] 50 edge computing nodes are deployed in high-risk areas, each node is responsible for monitoring a 10km range.

[0177] Hardware configuration: NVIDIA Jetson Xavier NX, 6-core ARM processor, 8GB RAM, 32GB storage.

[0178] Software environment: Ubuntu 20.04LTS, Docker containerized deployment.

[0179] 6.2) Real-time monitoring and early warning

[0180] Data collection frequency: once every 10 minutes in normal state, once every 1 minute in warning state.

[0181] Lightweight model: Compress the trained risk assessment model to 10MB and deploy it on edge nodes.

[0182] The method to obtain the lightweight model is as follows:

[0183] 1. Model compression technology:

[0184] Model pruning: Reduce the size and complexity of the model by removing unimportant weights or neurons. After pruning, the model parameters are significantly reduced, thereby reducing the model size.

[0185] Quantization: Convert floating-point calculations to low-precision integer calculations (such as 8-bit quantization), which reduces the model size and improves the inference speed.

[0186] Knowledge distillation: By training a smaller model (called the student model) to imitate the behavior of a larger complex model (the teacher model). The student model is smaller but has performance close to that of the teacher model.

[0187] Through these techniques, the original deep learning model can be transformed into a lightweight model to reduce memory usage and computational requirements.

[0188] Early warning trigger mechanism:

[0189] Yellow warning: a single risk score exceeds 70 or exceeds 65 for three consecutive times;

[0190] Orange warning: single value exceeding 80 or two consecutive values ​​exceeding 75;

[0191] Red alert: exceeds 90 for a single time or exceeds 85 for two consecutive times.

[0192] 6.3) Model update and optimization

[0193] Incremental learning: Incremental training is performed on the central server using new data every week.

[0194] Model evaluation: Perform a comprehensive evaluation of the model once a month and calculate indicators such as AUC and F1 score.

[0195] Automatic deployment: If the AUC of the new model on the validation set improves by more than 2%, the production environment model is automatically updated.

[0196] 6.4) Human-computer interaction

[0197] Develop web applications and mobile apps to provide the following functions: real-time risk map display, historical data query and analysis, early warning information push, and expert remote consultation interface.

[0198] By implementing the above system, the large-scale water diversion project has greatly improved its ability to predict and prevent karst collapse risks during the construction and operation of the karst section. During the one-year operation, the system successfully warned of five medium-sized karst collapse events, with a prediction accuracy of 92.5%, 15 percentage points higher than the traditional method. The warning was issued 72 hours in advance on average, providing a strong guarantee for project safety. At the same time, the accumulated large amount of data and experience provides a valuable reference for risk assessment of similar projects.

Claims

1. A karst collapse risk analysis method based on geological big data mining, characterized by: The following steps are involved: Data collection and fusion steps: Collect multi-source heterogeneous geological data and use the tensor decomposition algorithm based on Tucker decomposition for data fusion; Data preprocessing steps: cleaning, completion and standardization of the fused data; Feature Engineering Steps : Feature selection, selecting features related to karst collapse risk; And perform feature extraction on the preprocessed data; Model building and training steps: Based on the features extracted by the feature engineering, a risk assessment model is built using a graph convolutional neural network, and the performance of the risk assessment model is optimized through an ensemble learning method; Risk assessment and prediction steps: Use the trained risk assessment model to conduct risk assessment and prediction, and visualize it through geographic information system; System optimization and iteration steps: Realize real-time risk assessment based on edge computing technology and continuously optimize the risk assessment model.

2. The karst collapse risk analysis method according to claim 1, characterized in that: In the data collection and fusion step, the multi-source heterogeneous geological data includes geological data, remote sensing data, hydrogeological data, engineering supervision data, and historical data.

3. The karst collapse risk analysis method according to claim 1, characterized in that: In the data preprocessing step, data cleaning includes using the Isolation Forest algorithm improved by the LOF algorithm and / or the DTW algorithm to identify outliers and / or abnormal fluctuations; data completion includes using the LSTM autoencoder to fill the missing values ​​of time series data, and the thin plate spline interpolation method based on the variational method to complete the spatial data (such as terrain) to fill the missing parts in the spatial data; the data standardization processing method includes performing Z-score standardization on continuous data to make the data have zero mean and unit variance to reduce the impact of the dimension, performing one-hot encoding on discrete data to convert categorical variables into binary vectors to adapt to the processing requirements of the model, using the Robust Scaler method to standardize numerical data with outliers, and using the Target Encoding method to standardize categorical data.

4. The karst collapse risk analysis method according to claim 1, characterized in that: In the feature engineering step, feature selection includes using the Boruta algorithm combined with expert knowledge to evaluate feature importance and obtain key features; The feature extraction includes dynamic feature extraction, which includes extracting time series features, calculating spatial features and constructing new features; the time series feature extraction adopts a dynamic feature extraction algorithm that combines wavelet transform and LSTM network; the dynamic feature extraction algorithm includes performing wavelet decomposition on time series data, and then inputting the decomposed subsequences into the LSTM network for feature learning and extraction.

5. The karst collapse risk analysis method according to claim 4, characterized in that: The new features include a comprehensive karst vulnerability index, which is calculated as follows: Comprehensive karst vulnerability index = karst development degree × 0.3 + cover layer thickness × 0.2 + groundwater level fluctuation × 0.2 + fault density × 0.15 + cave development degree × 0.

15.

6. The karst collapse risk analysis method according to claim 1, characterized in that: In the model construction and training steps, the construction of the risk assessment model adopts a graph convolutional neural network as the basic model, and combines ConvLSTM, LightGBM and CatBoost to build an integrated model; the integrated model adopts the Stacking integration method, takes the output of the graph convolutional neural network, ConvLSTM, LightGBM and CatBoost as features, and trains an XGBoost model as the final predictor.

7. The karst collapse risk analysis method according to claim 1, characterized in that: In the risk assessment and prediction step, the karst collapse risk is divided into multiple levels, and the specific classification standards are dynamically adjusted according to the actual situation; in actual use, the multi-source heterogeneous geological data of a certain area acquired in real time is input into the trained risk assessment model for processing to obtain a risk probability in the numerical range of 0 to 1, and the karst collapse risk level of the area is determined based on the obtained value.

8. The karst collapse risk analysis method according to claim 1, characterized in that: In the system optimization and iteration step, the edge computing technology includes deploying edge servers on site, pre-installing lightweight models, receiving multi-source heterogeneous geological data obtained by real-time monitoring for preliminary risk assessment, triggering an early warning when the risk score exceeds a preset threshold, uploading it to the cloud, and analyzing it through a trained risk assessment model; the lightweight model is obtained by compressing the trained risk assessment model; The continuous optimization of the risk assessment model includes continuously using new multi-source heterogeneous geological data to continuously train and optimize the risk assessment model.

9. A karst collapse risk analysis system based on geological big data mining, characterized by: include: Data collection and fusion module, used to collect multi-source heterogeneous geological data and perform data fusion; Data preprocessing module, used to clean, complete and standardize the fused data; The feature engineering module is used to select features related to karst collapse risk and perform feature extraction on the data processed by the data preprocessing module; Model building and training module, used to build and optimize risk assessment models; The risk assessment and prediction module is used to perform risk assessment and prediction and visualize them; System optimization and iteration module, used to achieve real-time risk assessment and continuously optimize risk assessment models; The system optimization and iteration module includes an edge computing unit, which is used to perform real-time data processing and preliminary risk assessment on site.

10. The karst collapse risk analysis system according to claim 9, characterized in that: It also includes a user interface module for displaying risk assessment results and management recommendations, while receiving user input to optimize risk assessment model performance and user experience.

Images