A mine radon anomaly source positioning method and system based on a causal diagram
By constructing a directed acyclic graph using causal graphs and quantifying causal Shapley values, the problem of locating radon sources in mine monitoring caused by convergence of different sources was solved. This enabled efficient and accurate radon anomaly source location, reducing model complexity and cost.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANHUA UNIV
- Filing Date
- 2026-04-17
- Publication Date
- 2026-07-17
AI Technical Summary
In mine radon monitoring, the convergence of different sources makes it difficult to accurately locate the abnormal source using traditional methods, resulting in unstable early warning sensitivity. Increasing the number of monitoring points will increase investment and management costs.
A causal graph-based approach is adopted, which constructs a directed acyclic graph through a nonlinear causal learning algorithm, uses causal Shapley values to quantify feature importance, and builds a global distribution estimator and a final anomaly source model to achieve accurate localization of radon gas anomalies.
It improves the accuracy and reliability of anomaly source localization, reduces model complexity and cost, has strong generalization ability, and can accurately distinguish anomaly sources in different locations.
Smart Images
Figure CN122412876A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of mine radiation protection and safety monitoring technology, specifically relating to a method and system for locating abnormal radon sources in mines based on causal graphs. Background Technology
[0002] Radon (222Rn) is a colorless and odorless radioactive gas widely found in underground mines such as uranium mines, coal mines, and non-ferrous metal mines. Long-term inhalation of high concentrations of radon increases the risk of lung cancer in miners; therefore, radon monitoring is a crucial component of mine safety. Current monitoring methods typically involve deploying multiple monitoring points in mine roadways to monitor radon concentration, ventilation parameters (wind speed, wind pressure, temperature, and humidity) in real time. When abnormal radon concentrations are detected, it is necessary to quickly locate the source of the anomaly (such as rock wall fissures, goaf areas, abandoned roadways, etc.) to facilitate timely remedial measures.
[0003] However, radon monitoring in mines faces a core challenge: the "convergence of different sources" phenomenon. Radon anomalies (different sources) at different locations may exhibit highly similar temporal trends in their monitoring signals. For example, radon seepage from upstream rock wall fissures, radon accumulation in downstream goaf areas, and air leaks in the ventilation system can all lead to similar signal patterns at downstream monitoring points: "increasing radon concentration and decreasing wind pressure." This makes it difficult for traditional trend-based monitoring methods to distinguish the specific location of the anomaly, resulting in unstable early warning sensitivity and an inability to accurately pinpoint the anomaly source. While selecting certain points in the mine ventilation network for wind speed monitoring is beneficial for safety monitoring, increasing the number of monitoring points also increases investment and operational management costs, leading to economic losses. Summary of the Invention
[0004] The technical problem to be solved by the present invention is to provide a method and system for locating radon anomaly sources in mines based on causal graphs, which can accurately distinguish anomaly sources in different locations and achieve global convergence in radon anomaly source location.
[0005] A method for locating radon anomaly sources in mines based on causal graphs, comprising: Time series data of radon concentration at various monitoring points in the mine were collected to construct an original data matrix. The original data matrix was then standardized to obtain a feature matrix. The feature matrix is subjected to a nonlinear causal learning algorithm to construct an optimization objective with reconstruction loss, acyclic constraint, L1 regularization term, and L2 regularization term. The optimization objective is solved to obtain a weighted adjacency matrix between monitoring points. The weighted adjacency matrix is binarized to construct a directed acyclic graph. The global causal order, parent node set, and child node set are obtained based on the directed acyclic graph. The anomaly source prediction model is trained based on the feature matrix to obtain a pre-trained anomaly source prediction model. Based on the directed acyclic graph and the set of parent nodes, radon monitoring points are divided into root nodes without parent nodes and non-root nodes with parent nodes. A pre-trained anomaly source prediction model is used to construct a marginal distribution estimator based on radon concentration time series data for each root node, and a conditional distribution estimator with its parent node set as input is constructed for each non-root node. The marginal distribution estimators of all root nodes and the conditional distribution estimators of all non-root nodes are combined to form a global distribution estimator. Initialize the feature subset, select features from the feature matrix in causal order and add them to the feature subset, calculate the intervention distribution of the feature subset, and calculate the causal Shapley value based on the intervention distribution; All monitoring points are ranked by importance based on the causal Shapley values to obtain a list of feature importance. The top K features with the highest causal Shapley values are selected from the feature importance list as retraining features. The pretrained anomaly prediction model is retrained using the retraining features to obtain the final anomaly prediction model. The actual radon concentration time series data are input into the final anomaly source model to obtain the anomaly source location.
[0006] Optionally, the step of obtaining the feature matrix after standardizing the original data matrix includes: Calculate the mean and standard deviation of the original data matrix; The original data matrix is standardized based on the mean and standard deviation to obtain the feature matrix.
[0007] Optionally, the optimization objective is solved to obtain a weighted adjacency matrix between monitoring points. The weighted adjacency matrix is then binarized to construct a directed acyclic graph, including: The optimization objective is expressed as:
[0008] in, This is the first layer weight matrix. This is a local weight matrix. It is a multilayer perceptron. For the i-th feature in the feature matrix, for Regularization coefficient, for Regularization coefficient, and To augment the Lagrange multipliers, For acyclic constraints, For reconstruction loss; The augmented Lagrange method is used to solve the optimization objective, and the weighted adjacency matrix between monitoring points is obtained. The weighted adjacency matrix is binarized and expressed as:
[0009] in, This is a binarized weighted adjacency matrix. For monitoring points For monitoring points The causal strength, The preset threshold; Construct a directed acyclic graph based on the binarized weighted adjacency matrix.
[0010] Optionally, obtaining the global causal order, the set of parent nodes, and the set of child nodes based on the directed acyclic graph includes: Based on the directed acyclic graph, obtain the node features; The node value is calculated based on the node characteristics; Perform topological sorting based on the node values to obtain a global causal order; Based on the causal strength corresponding to the nodes in the directed acyclic graph, obtain the set of parent nodes and the set of child nodes.
[0011] Optionally, radon monitoring points are divided into root nodes without parent nodes and non-root nodes with parent nodes; a marginal distribution estimator based on radon concentration time series data is constructed for each root node using a pre-trained anomaly source prediction model, and a conditional distribution estimator with its set of parent nodes as input is constructed for each non-root node, including: Nodes whose parent node set is empty are designated as root nodes, and nodes whose parent node set is not empty are designated as non-root nodes. For each root node, random samples are randomly sampled from the radon concentration time series data to obtain random samples; The random samples are input into the pre-trained anomaly source prediction model, and a marginal distribution estimator based on radon concentration time series data is constructed for each root node; For each non-root node, construct a conditional distribution estimator with its set of parent nodes as input.
[0012] Optionally, features are selected from the feature matrix in causal order and added to a feature subset. The intervention distribution of the feature subset is calculated, and the causal Shapley value is calculated based on the intervention distribution, including: Features are selected from the feature matrix in causal order and added to the feature subset. The intervention distribution of the feature subset is calculated and expressed as follows:
[0013] in, For the set of features without intervention, For nodes The set of parent nodes, Let P be a causal operator, and let P() be a probability distribution. For intervention operations, S is a subset of features. These are random variables representing uninterrupted characteristics; The causal Shapley value is calculated based on the intervention distribution.
[0014] Optionally, the causal Shapley value calculated based on the intervention distribution is expressed as:
[0015] in, To intervene in the distribution, For causal Shapley values, As expected.
[0016] A causal graph-based system for locating radon anomaly sources in mines, comprising: The data processing module is used to collect time series data of radon concentration at various monitoring points in the mine, construct an original data matrix, and obtain a feature matrix after standardization of the original data matrix. The directed acyclic graph (DAG) construction module is used to construct an optimization objective by using a nonlinear causal learning algorithm on the feature matrix, with reconstruction loss, acyclic constraint, L1 regularization term, and L2 regularization term. The optimization objective is solved to obtain a weighted adjacency matrix between monitoring points. The weighted adjacency matrix is binarized to construct a directed acyclic graph. The global causal order, parent node set, and child node set are obtained based on the directed acyclic graph. The pre-training module is used to train the anomaly source prediction model based on the feature matrix to obtain the pre-trained anomaly source prediction model. The global distribution estimator construction module is used to divide radon monitoring points into root nodes without parent nodes and non-root nodes with parent nodes based on the directed acyclic graph and the set of parent nodes; to construct a marginal distribution estimator based on radon concentration time series data for each root node using a pre-trained anomaly source prediction model, and to construct a conditional distribution estimator with the set of parent nodes as input for each non-root node; and to combine the marginal distribution estimators of all root nodes and the conditional distribution estimators of all non-root nodes to form a global distribution estimator. The causal calculation module is used to initialize the feature subset, select features from the feature matrix in causal order and add them to the feature subset, use a global distribution estimator to calculate the intervention distribution of the feature subset, and calculate the causal Shapley value based on the intervention distribution; The optimization module is used to sort all monitoring points by importance according to the causal Shapley value to obtain a feature importance list. The top K features with the highest causal Shapley value are selected from the feature importance list as retraining features. The retraining features are used to retrain the pre-trained anomaly source prediction model to obtain the final anomaly source model. The localization module is used to input the actual radon concentration time series data into the final anomaly source model to obtain the location of the anomaly source.
[0017] A terminal device includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor loads and executes the computer program, it employs a method for locating radon anomaly sources in mines based on causal graphs.
[0018] A computer-readable storage medium storing a computer program, which, when loaded and executed by a processor, employs a method for locating radon anomaly sources in mines based on causal graphs.
[0019] The beneficial effects of this invention are: The NOTEARS-MLP nonlinear algorithm is used to learn the true causal structure between monitoring points, and the causal Shapley value is used to quantify the causal impact of features on the prediction. This can accurately distinguish between causal effects and spurious correlations, avoiding misclassification of correlated but non-causal features as important monitoring points, thus improving the accuracy of anomaly source localization. By decomposing the Shapley value into direct and indirect effects, the transmission mechanism of anomalies affecting other monitoring points through causal paths can be traced, providing interpretable causal path analysis for anomaly source localization. The model is retrained based on the Top-K key features selected from the causal Shapley value, which significantly reduces the number of features while maintaining prediction accuracy, effectively reducing model complexity, training time, and storage costs. In addition, the NOTEARS algorithm automatically learns the causal structure from the data without the need for prior human knowledge, and has strong generalization ability and practicality. Finally, the NOTEARS algorithm with a multilayer perceptron structure can model nonlinear causal relationships, capturing complex causal patterns in real systems more accurately than traditional linear methods, thus improving the reliability and robustness of anomaly source localization. Attached Figure Description
[0020] Figure 1 This is a flowchart illustrating a method for locating anomaly sources of radon gas in mines based on a cause-effect graph, according to the present invention. Detailed Implementation
[0021] A method for locating radon anomaly sources in mines based on causal graphs, such as... Figure 1 As shown, it includes: S1. Collect time series data of radon concentration at each monitoring point in the mine, construct the original data matrix, and obtain the feature matrix after standardization of the original data matrix; The feature matrix is obtained after standardization of the original data matrix, including: Calculate the mean and standard deviation of the original data matrix; The original data matrix is standardized based on the mean and standard deviation to obtain the feature matrix.
[0022] Specifically, time-series data of radon concentration were collected from each monitoring point. Let the number of monitoring points be d and the number of sampling time points be T, and then construct the original data matrix. ,in, For the set of real numbers. Each element of the original data matrix is... express, For the first t The first time point j Radon concentration observations at each monitoring point.
[0023] Calculate the mean and standard deviation for each monitoring point (i.e., each column):
[0024]
[0025] in, The mean, The standard deviation is denoted as .
[0026] The standardized processing is represented as follows:
[0027] in, For standardized data, To prevent small constants from being divided by zero, the standardized data must have a mean of 0 and a standard deviation of 1.
[0028] Standardized data as feature matrix Each row represents a time sample, and each column represents a monitoring point, as shown below:
[0029] in, This represents the d-th monitoring data in row T.
[0030] S2. The feature matrix is subjected to a nonlinear causal learning algorithm to construct an optimization objective with reconstruction loss, acyclic constraint, L1 regularization term, and L2 regularization term. The optimization objective is solved to obtain the weighted adjacency matrix between monitoring points. The weighted adjacency matrix is binarized to construct a directed acyclic graph. The global causal order, parent node set, and child node set are obtained from the directed acyclic graph. Solving for the optimization objective yields a weighted adjacency matrix between monitoring points. Binarizing this matrix, a directed acyclic graph is constructed, including: The optimization objective is expressed as:
[0031] in, For and This is the weight matrix. It is a multilayer perceptron. For the i-th feature in the feature matrix, for Regularization coefficient, for Regularization coefficient, and To augment the Lagrange multipliers, For acyclic constraints, For reconstruction loss;
[0032]
[0033] in, The elements are weighted adjacency matrix elements, where m is the number of neurons and d is the dimension. The square of the weight.
[0034] The augmented Lagrange method is used to solve the optimization objective, and the weighted adjacency matrix between monitoring points is obtained. Specifically, the augmented Lagrange method is used to alternately update the model parameters and Lagrange multipliers:
[0035]
[0036] Where t is the number of iterations. and For the augmented Lagrange multipliers after round t+1, Let be the weighted adjacency matrix after the t-th node.
[0037] The causal structure matrix is extracted from the trained neural network model and represented as follows:
[0038] in, For monitoring points For monitoring points The causal strength.
[0039] The L-BFGS optimizer is used to update the model parameters at each step until the convergence condition is met. Or it may reach the maximum number of iterations.
[0040] Binarize the weighted adjacency matrix as follows:
[0041] in, This is a binarized weighted adjacency matrix. For monitoring points For monitoring points The causal strength, The preset threshold; Construct a directed acyclic graph based on the binarized weighted adjacency matrix.
[0042] Obtain the global causal order, parent node set, and child node set from a directed acyclic graph, including: Obtain node features from the directed acyclic graph; The node value is calculated based on the node characteristics; Perform topological sorting based on node values to obtain the global causal order; Based on the causal strength of the nodes in the directed acyclic graph, obtain the set of parent nodes and the set of child nodes.
[0043] Specifically, the parent node set and child node set of each monitoring point are extracted based on the binary adjacency matrix, as follows: Parent node set:
[0044] Child node set:
[0045] in, Indicates direct impact on monitoring points The set of all monitoring points Indicates monitoring point The set of monitoring points that are directly affected.
[0046] A topological sorting algorithm is used to obtain the causal order. Node features are extracted from the directed acyclic graph, node values are calculated based on these features, and a topological sort is performed based on these node values to obtain the global causal order, represented as:
[0047] in, For nodes The node value.
[0048] Satisfy any directed edge In the causal order Ranked Previously, that is .
[0049] The global causal order is represented as:
[0050] S3. Train the anomaly source prediction model based on the feature matrix to obtain the pre-trained anomaly source prediction model; Specifically, the training and test data are first processed by standardizing parameters to obtain the standardized feature matrix. ) and test matrix ( The target variable is the anomaly category label (Y). An anomaly source prediction model was trained using an XGBoost multi-class classification model. After hyperparameter optimization, the model parameters were set as follows: number of trees (T=200), maximum depth (D=6), learning rate = 0.1, subsample ratio 0.8, column sampling ratio 0.8, and the objective function was multi-class log loss.
[0051] Specifically, the XGBoost gradient boosting tree algorithm is used to train the multi-class prediction model. ,in, For the number of monitoring points, To account for the number of anomaly types, the model consists of multiple decision trees:
[0052] Where T is the number of numbers. Let t be the prediction result for the t-th tree. It is the mode function. The radon concentration features are obtained from the feature matrix.
[0053] The model training uses log loss as the objective function:
[0054] in, Indicates the first Is the label of the sample the same as the first sample? kind, The model predicts the first The sample belongs to the first The probability of a class is the loss value, and n is the number of samples.
[0055] A well-trained model can output a value for a given input. Category probability distribution:
[0056] in, The condition is represented by Y, where Y is the label and X is the radon concentration feature. This represents the category probability distribution.
[0057] Satisfy the probability normalization condition:
[0058] S4. Based on the directed acyclic graph and the set of parent nodes, the radon monitoring points are divided into root nodes without parent nodes and non-root nodes with parent nodes. A marginal distribution estimator based on radon concentration time series data is constructed for each root node using a pre-trained anomaly source prediction model. A conditional distribution estimator with the set of parent nodes as input is constructed for each non-root node. The marginal distribution estimators of all root nodes and the conditional distribution estimators of all non-root nodes are combined to form a global distribution estimator. Radon monitoring points are divided into root nodes without parent nodes and non-root nodes with parent nodes. A pre-trained anomaly source prediction model is used to construct a marginal distribution estimator based on radon concentration time series data for each root node, and a conditional distribution estimator with its set of parent nodes as input is constructed for each non-root node, including: Nodes whose parent set is empty are designated as root nodes, and nodes whose parent set is not empty are designated as non-root nodes. For each root node, random samples are randomly sampled from the radon concentration time series data to obtain random samples; Random samples are input into a pre-trained anomaly prediction model, and a marginal distribution estimator based on radon concentration time series data is constructed for each root node. For each non-root node, construct a conditional distribution estimator with its set of parent nodes as input.
[0059] Specifically, the monitoring points are divided into two categories: Root node: A node whose parent node set is empty. .
[0060] Non-root node: A node whose parent node set is not empty. .
[0061] Estimate the marginal distribution of the root node: For the root node (Without a parent node), estimate its marginal distribution using historical data. Extracting the first from the feature matrix All observations in the column:
[0062] in, The first in the characteristic matrix The observations of the column, This refers to the number of samples, i.e., the length of the time series. This is the set of all observations extracted from the radon concentration time series data.
[0063] When sampling from the marginal distribution, Uniformly randomized sample selection:
[0064] in, It is a random sampling.
[0065] Estimate the conditional distribution of non-root nodes: For non-root nodes (With a parent node), estimate its conditional distribution. Let the set of parent nodes be:
[0066] in, It is the m-th parent node.
[0067] The conditional expectation function is fitted using a neural network:
[0068] in, As expected.
[0069] The neural network structure is as follows:
[0070] in, , and This is the weight matrix. It is the ReLU activation function. Non-root node The conditional expectation, that is, given the value of the parent node, The predicted mean.
[0071] Training a conditional distribution neural network: Train the Conditional Expectation Network using the training data. For each non-root node... Construct training set Minimize loss :
[0072] in, j For non-root node indexes, i For sample index, For the first i Nodes in each sample j The parent node feature vector, For the first i Nodes in each sample j The observed values.
[0073] The Adam optimizer was used with a learning rate of 0.01 for 200 training epochs. After training, the standard deviation of the residuals was calculated.
[0074] in, The residual standard deviation is calculated to provide the standard deviation of the noise term for the conditional distribution of non-root nodes, thereby allowing subsequent virtual sample sampling to have a real fluctuation range and making the generated virtual samples closer to the statistical regularity of real radon monitoring data.
[0075] Conditional distribution sampling: Given a parent node value Based on conditions sampling:
[0076] That is, the conditional expectation plus Gaussian noise, and the noise variance is the variance of the training residual.
[0077] Save the trained XGBoost model, label encoder, normalized parameters, and distribution estimators for all nodes as model data:
[0078] in, The mean, Standard deviation, For pre-training anomaly source prediction models, Let be the standard deviation of the residual at the j-th node. For the conditional expectation network of the j-th node, Residual standard deviation.
[0079] S5. Initialize the feature subset, select features from the feature matrix in causal order and add them to the feature subset, calculate the intervention distribution of the feature subset, and calculate the causal Shapley value based on the intervention distribution; Features are selected from the feature matrix in causal order and added to a feature subset. The intervention distribution of the feature subset is calculated, and the causal Shapley value is calculated based on the intervention distribution, including: Based on the causal structure, samples are taken from the intervention distribution.
[0080] Features are selected from the feature matrix in causal order and added to the feature subset. The intervention distribution of the feature subset is calculated and expressed as follows:
[0081] in, For the set of features without intervention, For nodes The set of parent nodes, Let P be a causal operator, and let P() be a probability distribution. For intervention operations, S is a subset of features. These are random variables representing uninterrupted characteristics; The causal Shapley value was calculated based on the intervention distribution.
[0082] For a given sample feature vector and feature subset Define the intervention value function:
[0083] in, To intervene in the distribution, For intervention value.
[0084] in Represents a set The features in the sample are intervened, and their values are fixed to the sample values. The corresponding feature values.
[0085] The feature subset is initially empty. Features are selected and added to the feature subset according to causal order before calculation. For example, if the i-th feature is selected to add to the feature subset, then the first i features constitute the feature subset, and the remaining features constitute the uninterrupted feature set. .
[0086] Calculate the causal Shapley value The Shapley value is calculated using a consistent causal order (symmetric version):
[0087] in, This represents the causal Shapley value.
[0088] in In the causal order, the feature is ranked The previous feature set. This value represents the feature. Marginal contribution to the model output.
[0089] In intervention set Below, fixed Other variables remain as follows Distributed sampling.
[0090] S6. Sort all monitoring points by importance according to the causal Shapley value to obtain a feature importance list. Select the top K features with the highest causal Shapley value from the feature importance list as retraining features. Use the retraining features to retrain the pretrained anomaly prediction model to obtain the final anomaly model. Specifically, after sorting, a list of feature importance is obtained, and the top [feature importance] is selected. Using these features as anomaly source localization results, and based on the importance ranking results, different numbers of Top-K features are selected to retrain the prediction model. Given a parameter for the number of features... Select the top from the feature importance list These features constitute the retraining feature set. Extract columns of selected features from the feature matrix:
[0091]
[0092] in, For the characteristic matrix, To test the feature matrix, This is a matrix slicing operation; the test feature matrix is proportionally divided from the feature matrix. This is the set of features for retraining.
[0093] S7. Input the actual radon concentration time series data into the final anomaly source model to obtain the anomaly source location.
[0094] A causal graph-based system for locating radon anomaly sources in mines, comprising: The data processing module is used to collect time series data of radon concentration at various monitoring points in the mine, construct the original data matrix, and obtain the feature matrix after standardization of the original data matrix; The directed acyclic graph (DAG) construction module is used to construct an optimization objective by using a nonlinear causal learning algorithm on the feature matrix, with reconstruction loss, acyclic constraint, L1 regularization term, and L2 regularization term. Solving the optimization objective yields a weighted adjacency matrix between monitoring points. After binarizing the weighted adjacency matrix, a DAG is constructed. Based on the DAG, the global causal order, parent node set, and child node set are obtained. The pre-training module is used to train the anomaly source prediction model based on the feature matrix to obtain the pre-trained anomaly source prediction model. The global distribution estimator construction module is used to divide radon monitoring points into root nodes without parent nodes and non-root nodes with parent nodes based on the directed acyclic graph and the set of parent nodes. It uses a pre-trained anomaly source prediction model to construct a marginal distribution estimator based on radon concentration time series data for each root node, and constructs a conditional distribution estimator with the set of parent nodes as input for each non-root node. The marginal distribution estimators of all root nodes and the conditional distribution estimators of all non-root nodes are combined to form the global distribution estimator. The causal calculation module is used to initialize the feature subset. It selects features from the feature matrix in causal order and adds them to the feature subset. It uses a global distribution estimator to calculate the intervention distribution of the feature subset and calculates the causal Shapley value based on the intervention distribution. The optimization module is used to sort all monitoring points by importance based on the causal Shapley value to obtain a list of feature importance. The top K features with the highest causal Shapley values are selected from the feature importance list as retraining features. The pre-trained anomaly prediction model is retrained using the retraining features to obtain the final anomaly prediction model. The localization module is used to input the actual radon concentration time series data into the final anomaly source model to obtain the location of the anomaly source.
[0095] This application also discloses a terminal device, including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor loads and executes the computer program, it employs a method for locating radon anomaly sources in mines based on causal graphs.
[0096] The terminal device can be a computer device such as a desktop computer, a laptop computer, or a cloud server. The terminal device includes, but is not limited to, a processor and a memory. For example, the terminal device may also include input / output devices, network access devices, and buses.
[0097] The processor can be a central processing unit (CPU). Of course, depending on the actual use, it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc., and this application does not limit it.
[0098] The memory can be an internal storage unit of the terminal device, such as a hard disk or RAM of the terminal device, or an external storage device of the terminal device, such as a plug-in hard disk, smart memory card (SMC), secure digital card (SD), or flash memory card (FC) equipped on the terminal device. Furthermore, the memory can be a combination of internal storage units and external storage devices of the terminal device. The memory is used to store computer programs and other programs and data required by the terminal device. The memory can also be used to temporarily store data that has been output or will be output. This application does not limit this.
[0099] In this terminal device, a method for locating anomaly sources of radon gas in mines based on causal graphs, as described in the above embodiments, is stored in the memory of the terminal device and loaded and executed on the processor of the terminal device for convenient use.
[0100] This application also discloses a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it employs a mine radon anomaly source localization method based on causal graphs as described in the above embodiments.
[0101] The computer program can be stored in a computer-readable medium. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or certain middleware. The computer-readable medium includes any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the computer-readable medium includes, but is not limited to, the above-mentioned components.
[0102] The above-described method for locating radon anomaly sources in mines based on causal graphs is stored in the computer-readable storage medium and loaded and executed on the processor to facilitate the storage and application of the method.
[0103] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; under the concept of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.
[0104] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.
Claims
1. A method for locating anomaly sources of radon gas in mines based on causal graphs, characterized in that, include: Time series data of radon concentration at various monitoring points in the mine were collected to construct an original data matrix. The original data matrix was then standardized to obtain a feature matrix. The feature matrix is subjected to a nonlinear causal learning algorithm to construct an optimization objective with reconstruction loss, acyclic constraint, L1 regularization term, and L2 regularization term. The optimization objective is solved to obtain a weighted adjacency matrix between monitoring points. The weighted adjacency matrix is binarized to construct a directed acyclic graph. The global causal order, parent node set, and child node set are obtained based on the directed acyclic graph. The anomaly source prediction model is trained based on the feature matrix to obtain a pre-trained anomaly source prediction model. Based on the directed acyclic graph and the set of parent nodes, radon monitoring points are divided into root nodes without parent nodes and non-root nodes with parent nodes. A pre-trained anomaly source prediction model is used to construct a marginal distribution estimator based on radon concentration time series data for each root node, and a conditional distribution estimator with its parent node set as input is constructed for each non-root node. The marginal distribution estimators of all root nodes and the conditional distribution estimators of all non-root nodes are combined to form a global distribution estimator. Initialize the feature subset, select features from the feature matrix in causal order and add them to the feature subset, use the global distribution estimator to calculate the intervention distribution of the feature subset, and calculate the causal Shapley value based on the intervention distribution; All monitoring points are ranked by importance based on the causal Shapley values to obtain a list of feature importance. The top K features with the highest causal Shapley values are selected from the feature importance list as retraining features. The pretrained anomaly prediction model is retrained using the retraining features to obtain the final anomaly prediction model. The actual radon concentration time series data are input into the final anomaly source model to obtain the anomaly source location.
2. The method for locating radon anomaly sources in mines based on causal graphs according to claim 1, characterized in that, The step of obtaining the feature matrix after standardizing the original data matrix includes: Calculate the mean and standard deviation of the original data matrix; The original data matrix is standardized based on the mean and standard deviation to obtain the feature matrix.
3. The method for locating mine radon anomaly sources based on causal graphs according to claim 1, characterized in that, Solving for the optimization objective yields a weighted adjacency matrix between monitoring points. Binarizing this weighted adjacency matrix then constructs a directed acyclic graph, including: The optimization objective is expressed as: in, This is the first layer weight matrix. This is a local weight matrix. It is a multilayer perceptron. For the i-th feature in the feature matrix, for Regularization coefficient, for Regularization coefficient, and To augment the Lagrange multipliers, For acyclic constraints, For reconstruction loss; The augmented Lagrange method is used to solve the optimization objective, and the weighted adjacency matrix between monitoring points is obtained. The weighted adjacency matrix is binarized and expressed as: in, This is a binarized weighted adjacency matrix. For monitoring points For monitoring points The causal strength, The preset threshold; Construct a directed acyclic graph based on the binarized weighted adjacency matrix.
4. The method for locating radon anomaly sources in mines based on causal graphs according to claim 1, characterized in that, Based on the directed acyclic graph, the global causal order, the set of parent nodes, and the set of child nodes are obtained, including: Based on the directed acyclic graph, obtain the node features; The node value is calculated based on the node characteristics; Perform topological sorting based on the node values to obtain a global causal order; Based on the causal strength corresponding to the nodes in the directed acyclic graph, obtain the set of parent nodes and the set of child nodes.
5. The method for locating mine radon anomaly sources based on causal graphs according to claim 1, characterized in that, Radon monitoring points are divided into root nodes without parent nodes and non-root nodes with parent nodes. A pre-trained anomaly source prediction model is used to construct a marginal distribution estimator based on radon concentration time series data for each root node, and a conditional distribution estimator with its set of parent nodes as input is constructed for each non-root node, including: Nodes whose parent node set is empty are designated as root nodes, and nodes whose parent node set is not empty are designated as non-root nodes. For each root node, random samples are randomly sampled from the radon concentration time series data to obtain random samples; The random samples are input into the pre-trained anomaly source prediction model, and a marginal distribution estimator based on radon concentration time series data is constructed for each root node; For each non-root node, construct a conditional distribution estimator with its set of parent nodes as input.
6. The method for locating mine radon anomaly sources based on causal graphs according to claim 1, characterized in that, Features are selected from the feature matrix in causal order and added to a feature subset. The intervention distribution of the feature subset is calculated, and the causal Shapley value is calculated based on the intervention distribution, including: Features are selected from the feature matrix in causal order and added to the feature subset. The intervention distribution of the feature subset is calculated and expressed as follows: in, For the set of features without intervention, For nodes The set of parent nodes, Let P be a causal operator, and let P() be a probability distribution. For intervention operations, S is a subset of features. These are random variables representing uninterrupted characteristics; The causal Shapley value is calculated based on the intervention distribution.
7. The method for locating mine radon anomaly sources based on causal graphs according to claim 6, characterized in that, The causal Shapley value calculated based on the intervention distribution is expressed as follows: in, To intervene in the distribution, For causal Shapley values, As expected.
8. A mine radon anomaly source localization system based on causal graphs, characterized in that, include: The data processing module is used to collect time series data of radon concentration at various monitoring points in the mine, construct an original data matrix, and obtain a feature matrix after standardization of the original data matrix. The directed acyclic graph (DAG) construction module is used to construct an optimization objective by using a nonlinear causal learning algorithm on the feature matrix, with reconstruction loss, acyclic constraint, L1 regularization term, and L2 regularization term. The optimization objective is solved to obtain a weighted adjacency matrix between monitoring points. The weighted adjacency matrix is binarized to construct a directed acyclic graph. The global causal order, parent node set, and child node set are obtained based on the directed acyclic graph. The pre-training module is used to train the anomaly source prediction model based on the feature matrix to obtain the pre-trained anomaly source prediction model. The global distribution estimator construction module is used to divide radon monitoring points into root nodes without parent nodes and non-root nodes with parent nodes based on the directed acyclic graph and the set of parent nodes; to construct a marginal distribution estimator based on radon concentration time series data for each root node using a pre-trained anomaly source prediction model, and to construct a conditional distribution estimator with the set of parent nodes as input for each non-root node; and to combine the marginal distribution estimators of all root nodes and the conditional distribution estimators of all non-root nodes to form a global distribution estimator. The causal calculation module is used to initialize the feature subset, select features from the feature matrix in causal order and add them to the feature subset, use a global distribution estimator to calculate the intervention distribution of the feature subset, and calculate the causal Shapley value based on the intervention distribution; The optimization module is used to sort all monitoring points by importance according to the causal Shapley value to obtain a feature importance list. The top K features with the highest causal Shapley values are selected from the feature importance list as retraining features. The retraining features are used to retrain the pre-trained anomaly source prediction model to obtain the final anomaly source model. The localization module is used to input the actual radon concentration time series data into the final anomaly source model to obtain the location of the anomaly source.
9. A terminal device, comprising a memory and a processor, characterized in that, The memory stores a computer program that can run on a processor, and when the processor loads and executes the computer program, it employs the method described in any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is loaded and executed by the processor, it employs the method described in any one of claims 1 to 7.