Multi-factor-driven method for predicting concentration of multi-source harmful gas in underground cavern construction

By using a multi-factor driven machine learning approach, a model for predicting the concentration of harmful gases was constructed, which solved the problem of predicting the concentration of harmful gases during the construction of underground caverns and enabled real-time accurate prediction and safety management during the construction process.

CN118939936AInactive Publication Date: 2024-11-12CHINA THREE GORGES UNIV

Patent Information

Application Number
CN202410872121.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-01
Publication Date
2024-11-12
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

Existing technologies are insufficient to effectively predict the concentration of harmful gases during the construction of underground caverns, resulting in poor safety management and an inability to prevent problems such as excessive concentrations in advance.

Method used

A multi-factor driven machine learning approach is adopted. By collecting construction site data and preprocessing it, a machine learning model is built, including category boosting ensemble learning and deep learning modules. Combined with feature similarity calculation and principal component analysis, the concentration of harmful gases is predicted.

Benefits of technology

It enables real-time and accurate prediction of the concentration of harmful gases during the construction of underground caverns, improving the preventive nature of safety management, reducing training complexity, and enhancing prediction accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118939936A_ABST
    Figure CN118939936A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of underground cavern construction safety, in particular to a multi-factor-driven underground cavern construction multi-source harmful gas concentration prediction method, which comprises the following steps: collecting underground cavern construction site data, including harmful gas concentration monitoring values and factors related to harmful gas concentration values, comprise humidity, temperature and atmospheric pressure; the collected data are preprocessed; using the preprocessed data to construct a machine learning harmful gas concentration prediction model; performing harmful gas concentration prediction based on a machine learning harmful gas concentration prediction model; and important influence factors and important time nodes of harmful gas concentration are analyzed and predicted. The prediction method can realize harmful gas concentration prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of underground cavern construction safety, and in particular relates to a method for predicting the concentration of multi-source harmful gases in underground cavern construction driven by multiple factors. Background Art

[0002] The issue of underground cavern construction safety management has always attracted widespread attention from people inside and outside the industry. Harmful gases, as one of the major safety hazards in underground cavern construction, seriously affect the occupational health and safety of construction workers. At present, most construction sites still use "real-time monitoring before taking measures" as the main means of harmful gas safety management. This traditional safety management method often has poor prevention effects on harmful gases and is difficult to effectively prevent safety problems caused by excessive concentrations of harmful gases. Therefore, predicting the concentration of harmful gases in advance is of great significance for the safety management of underground cavern construction.

[0003] Most current technologies focus on the monitoring systems and treatment methods of harmful gases, and few use appropriate methods to achieve early prediction of harmful gases. The generation, dissipation and migration of harmful gases in underground caverns follow certain rules, which causes the concentration change of harmful gases to be related to time, and it has its own time series characteristics. The historical values ​​of harmful gas concentrations indicate the changes in harmful gas concentrations in the future to a certain extent, and different harmful gases have different patterns of change over time. At the same time, the changes in harmful gas concentrations are closely related to various factors such as temperature, humidity, atmospheric pressure, construction, and geology. Therefore, considering the multi-faceted influence of harmful gas concentrations, the time information of its concentration changes and other influencing factors are extracted to accurately predict its future values. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method for predicting the concentration of multi-source harmful gases in underground cavern construction driven by multiple factors, so as to realize the prediction of the concentration of harmful gases.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for predicting the concentration of multi-source harmful gases in underground cavern construction driven by multiple factors, comprising the following steps:

[0006] Step S1: Collecting underground cavern construction site data, including harmful gas concentration monitoring values ​​and factors related to the harmful gas concentration values, including humidity, temperature and atmospheric pressure;

[0007] Step S2: preprocessing the collected data;

[0008] Step S3: Use the preprocessed data to build a machine learning harmful gas concentration prediction model;

[0009] Step S4: Predict the concentration of harmful gases based on the machine learning harmful gas concentration prediction model.

[0010] In a preferred embodiment, in step S2, the operation steps are as follows:

[0011] S201, performing correlation analysis on all features of the field data to find a significant correlation feature group;

[0012] S202, using principal component analysis (PCA) method to perform dimensionality reduction processing on the correlation significant feature group;

[0013] S203, rescale all data after dimensionality reduction processing to the range of [0, 1].

[0014] In a preferred solution, in step S201, Pearson correlation analysis is performed on all features to calculate the Pearson correlation coefficient between any two sets of data, and the expression is as follows:

[0015]

[0016] In the formula, r represents the Pearson correlation coefficient between any two sets of data, and its value range is [-1, 1], 1 represents a perfect positive correlation, -1 represents a perfect negative correlation, and 0 represents no linear correlation; n is the number of data; x and y are two different variables.

[0017] In the preferred solution, in step S202, the covariance matrix in the PCA method is C ij , perform eigendecomposition on the covariance matrix to obtain eigenvectors. The eigenvectors represent the directions of the principal components of the data, and the eigenvalues ​​represent the variances in each direction, that is, the degree of data dispersion. Select the eigenvector W corresponding to the maximum eigenvalue as the new feature space. Finally, multiply the original data matrix X by W to obtain the reduced-dimensional data Y, the covariance matrix C ij The expression is as follows:

[0018]

[0019] Where m is the number of samples, X ki is the i-th eigenvalue of the k-th sample, is the mean value of the ith feature, is the mean value of the jth feature.

[0020] In a preferred solution, in step S203, the scaling process expression of rescaling all processed data to the range of [0, 1] is as follows:

[0021]

[0022] Among them, x min and x max Represent the maximum and minimum values ​​in the variable, x i is the value of the variable before scaling, x k is the new variable value after scaling.

[0023] In a preferred solution, step S3 includes the following steps:

[0024] S301. Construct a machine learning harmful gas concentration prediction model including a category boosting ensemble learning module and a deep learning module. The category boosting ensemble learning module is based on a category boosting ensemble learning algorithm. The category boosting ensemble learning module is used to extract all factor information related to the harmful gas concentration. The deep learning module is composed of a convolutional neural network, a long short-term memory network, and a self-attention mechanism, and is used to extract time information about the harmful gas concentration.

[0025] S302, dividing the underground cavern harmful gas prediction data set processed in step S2 into a training set, a validation set and a test set, using the training set and the validation set for model training, and using the test set to verify the model prediction accuracy.

[0026] In a preferred solution, in step S301, the process of extracting factor information related to the concentration of harmful gases by the category promotion integrated learning module is as follows:

[0027] Assume that the harmful gas prediction data set D = {x i ,y i}(i=1,2,…,m) contains m samples, where x i are the construction characteristics and geological characteristics corresponding to the i-th sample, y i is the label value of the i-th sample, i.e., the harmful gas monitoring value; each x i Contains n-dimensional features, x ij is the j-th dimension feature of the i-th sample; randomly sort the data set D to obtain σ, σ={σ(1), σ(2), …, σ(m)}, and for each sample x i Train a decision tree model M i , and the residual error for each training is as follows:

[0028] r i =y i -M σ(i-1) (x i );

[0029] Where: M σ(n) (x i ) is the sample x in the sequence σ(n) i Trained decision tree model;

[0030] Harmful gas prediction data set D = {x i ,y i}(i=1,2,…,m) n features contain both numerical features and categorical features, k is the dimension of this feature when processing a categorical feature, For the sample corresponding to the sequence σ(p), The harmful gas detection values ​​are averaged and the priori value and the weight coefficient of the priori value are added to make a certain category feature of the sample Convert to numeric type, The expression is as follows:

[0031]

[0032] Where: For sample The corresponding harmful gas monitoring value is only when hour, is equal to 1, otherwise it is equal to zero; a>0, which indicates the weight of the prior p.

[0033] In the preferred solution, in step S302, when the prediction accuracy of the model is tested using the test set, the model accuracy is evaluated using the evaluation index determination coefficient R 2 , root mean square error RMSE and mean absolute error MAE three evaluation indicators, the expressions are as follows:

[0034]

[0035] Where: y pred,i is the predicted value, y obs,i is the observed value, is the mean of the observations, and n is the number of observations.

[0036] The preferred solution also includes step S5, using a machine learning interpretability enhancement algorithm to analyze important influencing factors and important time nodes of harmful gas concentrations, including the following steps:

[0037] S501, integrating SHapley addition interpretation algorithm to analyze the important influencing factors of harmful gas concentration;

[0038] S502. Integrate the local interpretable model-independent explanation algorithm to analyze the important time nodes for predicting the concentration of harmful gases.

[0039] The invention provides a method for predicting the concentration of harmful gases from multiple sources in underground cavern construction driven by multiple factors, which has the following beneficial effects:

[0040] 1. In step 1 of the present invention, real-time data of harmful gas concentration and related factors during the construction process are collected, so that the existing data can be used to predict the future harmful gas concentration during the construction process. The related factors of harmful gas concentration are considered very comprehensively, involving geological conditions, construction methods, humidity, temperature, atmospheric pressure and other influencing factors.

[0041] 2. In step 2 of the present invention, two suitable methods, feature similarity calculation and principal component analysis dimensionality reduction, are used to simply and effectively implement feature and data preprocessing, thereby improving result accuracy while reducing training complexity.

[0042] 3. In step three of the present invention, artificial intelligence machine learning is used to process the preprocessed data, and the nonlinear relationship and time series characteristics implicit in the data are learned through the powerful learning ability of the machine, thereby constructing a harmful gas prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0044] Figure 1 This is a flow chart of the intelligent prediction method for the concentration of harmful gases in underground caverns proposed by the present invention;

[0045] Figure 2 This is a framework diagram of the intelligent prediction method for the concentration of harmful gases in underground caverns proposed by the present invention;

[0046] Figure 3 A comparison chart of PM10 prediction values ​​and actual values ​​for each model provided in an embodiment of the present invention;

[0047] Figure 4 A comparison chart of PM2.5 prediction values ​​and actual values ​​for each model provided in an embodiment of the present invention;

[0048] Figure 5 A comparison chart of evaluation indicators of various models provided in the embodiments of the present invention;

[0049] Figure 6 An analysis diagram of factors affecting the harmful gas PM10 provided in an embodiment of the present invention;

[0050] Figure 7 An analysis diagram of factors affecting PM2.5, a harmful gas, provided in an embodiment of the present invention;

[0051] Figure 8 A time impact analysis diagram of harmful gas PM10 provided in an embodiment of the present invention;

[0052] Fig. 9 This is a time impact analysis diagram of the harmful gas PM2.5 provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0053] Embodiment 1:

[0054] like Figure 1 As shown, a method for predicting the concentration of multi-source harmful gases in underground cavern construction driven by multiple factors includes the following steps:

[0055] Step S1: Collecting underground cavern construction site data, including harmful gas concentration monitoring values ​​and factors related to the harmful gas concentration values, including humidity, temperature and atmospheric pressure.

[0056] During the construction process, the concentration of harmful gases is monitored in real time, and data information is extracted, collected and organized from environmental monitoring, blasting monitoring, microseismic monitoring and other equipment, and linked to engineering geological forecast text data. The concentration of harmful gases is accurately predicted by considering the time characteristics of the concentration of harmful gases and the influence of complex multiple factors.

[0057] Step S2: pre-process the collected data. Use the feature data to perform feature correlation analysis to check the significance of the feature correlation, and perform PCA dimension reduction on the features with significant correlation. Perform pre-processing such as noise reduction, outlier exclusion, normalization, and standardization on the processed feature data and harmful gas monitoring data.

[0058] The steps are as follows:

[0059] S201. Since some features come from the same category, feature redundancy is inevitable. Using redundant features for model training will not only fail to improve the model training effect, but will waste training resources and affect the model training speed and effect. Therefore, Pearson correlation analysis is performed on all features to calculate the Pearson correlation coefficient between any two sets of data and find the feature group with significant correlation. The principle of Pearson correlation analysis is shown in the following formula:

[0060]

[0061] In the formula, r represents the Pearson correlation coefficient between any two sets of data, and its value range is [-1, 1], 1 represents a perfect positive correlation, -1 represents a perfect negative correlation, and 0 represents no linear correlation; n is the number of data; x and y are two different variables.

[0062] S202, using principal component analysis (PCA) method to perform dimensionality reduction processing on the correlation significant feature group;

[0063] The role of PCA is to reduce the number of variables in the data set while retaining the most important information in the data set, thereby generating more representative new features to solve the problem of feature redundancy. PCA achieves dimensionality reduction by identifying the main components (i.e., directions) in the data and projecting the data onto these components. The covariance matrix calculation in this method is shown in the following formula:

[0064]

[0065] Where m is the number of samples, X ki is the i-th eigenvalue of the k-th sample, is the mean value of the ith feature, is the mean value of the jth feature.

[0066] The covariance matrix is ​​decomposed to obtain the eigenvector. The eigenvector represents the direction of the principal component of the data. The eigenvalue represents the variance in each direction, that is, the degree of data dispersion. The eigenvector W corresponding to the maximum eigenvalue is selected as the new feature space. Finally, the original data matrix X is multiplied by W to obtain the reduced-dimensional data Y.

[0067] S203: rescale all the data after the dimension reduction process to the range of [0, 1]. This process is used to prevent the model performance from being affected by the imbalance of the feature data range.

[0068] The data scaling process expression is as follows:

[0069]

[0070] Among them, x min and x max Respectively represent the maximum and minimum values ​​in the variable, x i is the value of the variable before scaling, x k is the new variable value after scaling.

[0071] Step S3: Use the preprocessed data to build a machine learning harmful gas concentration prediction model.

[0072] The steps include:

[0073] S301. A new CatBoost-CNN-LSTM-SA harmful gas concentration prediction model is used for intelligent prediction of hazardous gas concentrations in underground caverns.

[0074] CatBoost, Categorical Boosting, means category boosting; CNN, Convolutional Neural Network, means convolutional neural network; LSTM, Long Short-Term Memory, means long short-term memory network; SA, Self-Attention, means self-attention mechanism.

[0075] In this model, the CatBoost integrated learning module is used to extract all the information about the factors related to the concentration of hazardous gases, while the CNN-LSTM-SA deep learning module is used to extract the time information about the concentration of hazardous gases. The model aims to comprehensively consider the impact of construction activities, geological conditions and time factors on the concentration of hazardous gases, so as to accurately predict the concentration of hazardous gases in underground caverns in real time.

[0076] The construction of a machine learning harmful gas concentration prediction model includes a category boosting ensemble learning module and a deep learning module. The category boosting ensemble learning module is based on the category boosting ensemble learning algorithm. The category boosting ensemble learning module is used to extract all factor information related to the harmful gas concentration. The deep learning module is composed of a convolutional neural network, a long short-term memory network, and a self-attention mechanism, which is used to extract time information about the harmful gas concentration.

[0077] The training process of CatBoost integrated learning algorithm module is as follows:

[0078] Based on the special categorical feature processing method of CatBoost (Categorical Boosting, CatBoost, category boosting) algorithm, the characteristic stratum, weathering load degree, fault condition, fracture development condition, rock mass structure and groundwater status are marked as categorical features during training. Compared with other ensemble learning algorithms, CatBoost algorithm can directly process categorical features and convert them into numerical features.

[0079] Assume that the harmful gas prediction data set D = {x i ,y i}(i=1,2,…,m) contains m samples, where x i are the construction characteristics and geological characteristics corresponding to the i-th sample, y i is the label value of the i-th sample, i.e., the harmful gas monitoring value. i Contains n-dimensional features, x ij is the j-th dimension feature of the i-th sample. Randomly sort the data set D to obtain σ, σ = {σ(1), σ(2), …, σ(m)}, and for each sample x i Train a decision tree model M i, and the residual error for each training is as follows:

[0080] r i =y i -M σ(i-1) (x i );

[0081] Where: M σ(n) (x i ) is the sample x in the sequence σ(n) i Trained decision tree model.

[0082] Harmful gas prediction data set D = {x i ,y i In the n features (i=1, 2, ..., m), there are both numerical features and categorical features, and k is the dimension of this feature when processing a categorical feature. For the sample corresponding to the sequence σ(p), The harmful gas detection values ​​are averaged and the priori value and the weight coefficient of the priori value are added to make a certain category feature of the sample Convert to numeric type, The expression is as follows:

[0083]

[0084] Where: For sample The corresponding harmful gas monitoring value is only when hour, is equal to 1, otherwise it is equal to zero; a>0, which indicates the weight of the prior p.

[0085] The CNN-LSTM-SA deep learning module is as follows:

[0086] The convolution layer is the core of CNN, which consists of multiple convolution kernels. These convolution kernels are convolved with the input information to extract the hidden features of the information. When using CNN to process time series, the 1D convolution kernel only moves along the time axis, so its convolution operation formula is as follows:

[0087]

[0088] Among them, f is the input time series, g is the convolution kernel, and n and m are the discrete time points of the time series.

[0089] The design of LSTM is inspired by the logic gates of computers. It regulates the flow of information by introducing a gating mechanism, allowing the network to learn how long to keep information, when to update information, and forget old information. The three gating structures are called output gate, input gate, and forget gate. The LSTM unit structure is as follows Figure 2During the calculation, the input of the current time step and the hidden state of the previous time step are fed into the gate of the long short-term memory network as data. They are processed by three fully connected layers with sigmoid activation functions to calculate the input gate I t 、Forget Gate F t and output gate O t The output gate acts on the hidden state H t , use the tanh activation function to ensure that H t The interval is within (-1, 1). When the output gate is close to 1, all memory information is effectively passed to the prediction part; when the output gate is close to 0, all information in the memory element is retained without updating the hidden state. The definitions of the input gate, forget gate, output gate and hidden state are as follows:

[0090] I t =σ(X t W xi +H t-1 W hi +b i );

[0091] F t =σ(X t W xf +H t-1 W hf +b f );

[0092] O t =σ(X t W xo +H t-1 W ho +b o );

[0093] H t =O t ⊙tanh(C t );

[0094] Among them, W xi , W xf , W xo and W hi , W hf , W ho is the weight parameter, b i 、b f 、b o is the bias parameter. C t Represents the cell state used in the long short-term memory neural network. σ represents the activation function sigmoid.

[0095] In addition to the three gates, there is also a candidate memory element in the structure. Its calculation method is similar to the gate, but the tanh function is used as the activation function, and the function value range is (-1, 1). Its formula at time step t is as follows:

[0096]

[0097] Among them, W xc and W hc is the weight parameter, b c is the bias parameter. Represents a candidate cell state, an intermediate calculated value used to update the cell state.

[0098] Input Gate I t Control how much is used from The new data, and the forget gate F t Controls how much past memory is retained t-1 The content of . Thus we get:

[0099]

[0100] Through the above operations, the model can selectively decide which information to forget and update during training, so as to handle important events with long intervals and delays in the time series.

[0101] The three key vector sets of the self-attention mechanism are query q, key K, and value V, which are obtained by multiplying the input sequence with the corresponding weight matrix. Self-attention first determines the importance of each element to other elements by calculating the dot product between the query and all keys, that is, assigning weights; then uses the scaling factor To improve stability, d k is the dimension of the key vector.

[0102]

[0103] Where: d k represents the dimension of key K, Q represents the query matrix, K represents the key matrix, and V represents the value matrix.

[0104] These scores are normalized by the softmax function so that the sum of all scores is 1. Each score can be used as a probability, indicating the degree to which the target element should "pay attention" to each element in the sequence, that is, the weight that the element has for other elements. Finally, the normalized scores are multiplied and summed with the value vector to obtain the output of self-attention. The output vector is the weighted sum of all positions in the input sequence, and the weights are determined by the relevance of all elements in the input sequence.

[0105] S302, dividing the underground cavern harmful gas prediction data set processed in step S2 into a training set, a validation set and a test set, using the training set and the validation set for model training, and using the test set to verify the model prediction accuracy.

[0106] The model accuracy was evaluated using R 2 , RMSE and MAE three evaluation indicators, the expressions are as follows:

[0107]

[0108] Where: y pred,i is the predicted value, y obs,i is the observed value, is the mean of the observations, and n is the number of observations.

[0109] Step S4: Predict the concentration of harmful gases based on the machine learning harmful gas concentration prediction model.

[0110] By considering the time characteristics, an intelligent prediction model for underground harmful gas concentration is constructed to achieve real-time and accurate prediction of the concentration of harmful gases in underground caverns. Through the analysis of model results, the key influencing factors of the concentration of harmful gases in underground caverns and the time characteristics of the concentration of harmful gases are explored to provide a basis for the safe management of harmful gases in underground caverns.

[0111] The prediction results were compared between the existing six models and the prediction model of the present invention. Figures 3-4 As shown in the figure, the harmful gas monitoring values ​​fluctuate greatly, the predicted values ​​of harmful gases are still highly matched with the true values, and there are small deviations between a few predicted values ​​and the true values. The seven models used in the figure are all currently advanced time series prediction models, namely CatBoost-CNN-LSTM-SA, CNN-LSTM-SA, CatBoost-CNN-LSTM, GRU (Gated Recurrent Unit), CNN, LSTM, and CNN-GRU.

[0112] Among them, the CatBoost-CNN-LSTM-SA model adopted by the present invention is proved to be the optimal harmful gas concentration prediction model.

[0113] like Figure 5 As shown, the evaluation indicators of the calculated model are 91.36%, 307.43, 473.47 and 89.07%, 248.05, 444.87 respectively in predicting harmful gases PM10 and PM2.5. The prediction model adopted by the present invention is higher than all other advanced models.

[0114] Embodiment 2:

[0115] Different from Example 1, the method further includes step S5: using a machine learning interpretability enhancement algorithm to analyze important influencing factors and important time nodes of harmful gas concentrations, including the following steps:

[0116] S501, integrate SHapley addition interpretation algorithm to analyze the important influencing factors of harmful gas concentration.

[0117] The process of integrating SHapley addition interpretation algorithm to analyze the important influencing factors of harmful gas concentration is as follows:

[0118] SHAP Definition Where S is the set of non-zero indicators in z', assuming that the jth influencing feature of the i-th harmful gas sample is x i,j , φ i Assign factors affecting harmful gas concentration such as temperature, humidity, atmospheric pressure, and geological conditions:

[0119]

[0120] Where: x is the feature of the input sample, N is the feature dimension, M is the set of all features in the harmful gas prediction dataset D, S is a feature subset of M, and f x (S) represents the fitting result when using feature set S and f x (S∪{x i,j}) means adding feature x to the set S i,j The fitting results after .

[0121] S502. Integrate the local interpretable model-independent explanation algorithm to analyze the important time nodes for predicting the concentration of harmful gases.

[0122] The process of integrating the local interpretable model-independent interpretation algorithm and analyzing the important time nodes for predicting harmful gas concentrations is as follows:

[0123] The core idea of ​​LIME is to convert the behavior of a complex model near a specific data point into a simple linear model through local linear approximation, thereby providing an explanation. For a given harmful gas concentration data x, LIME will generate a set of data points Z similar to x, which can be obtained by perturbing the characteristics of the harmful gas concentration data point x, that is, the historical time nodes.

[0124] Apply the complex model f to the generated neighborhood data Z for prediction and obtain the predicted value f(Z).

[0125] In order to fit the local linear model, LIME weights the neighborhood data Z, and the weights are determined according to the distance between the neighborhood data and the original data point x. Usually, the Gaussian kernel function ω(z) is used to calculate the weights, and the weighted least squares method g(z) is used to fit the local linear model:

[0126]

[0127] g(z)=β 0 +β 1 z 1 +β 2 z 2 +…+β p z p ;

[0128] where D(z, x) is the distance between z and x (e.g., Euclidean distance), σ is a tuning parameter, g is a local linear model, and z 1 , z 2 , …, z p are the features of the data points and β are the model parameters.

[0129] The coefficients β of the fitted local linear model g can be used as an explanation for the behavior of the complex model f near the data point x.

[0130] Based on the above model, the SHapley Additive exPlanations (SHAP) algorithm and the Local Interpretable Model-agnosticExplanations (LIME) algorithm are integrated to deeply explore the relationship between features and results. The generation mechanism of harmful gas concentration in underground caverns is studied, and the influence mechanism of various factors on it and its specific time characteristics are analyzed.

[0131] Figures 6-7 The SHAP values ​​of all features affecting the concentration of each harmful gas are shown. The higher the feature ranking in the figure, the greater the impact of the feature on the result. The SHAP values ​​of each sample feature used to predict harmful gases are plotted as a scatter plot, and the size of the feature value and the relationship between the feature and the harmful gas prediction result are represented by color and value. This is used to analyze the key influencing factors of the concentration of harmful gases in underground caverns and provide more basis for the safety management of harmful gases in underground caverns.

[0132] Figures 8-9 The time-influence information of harmful gas concentration is displayed, and the time step importance distribution diagram shows the influence of each feature and each historical node of harmful gas concentration on the current prediction. From the time series characteristic analysis, the key influence moments of the historical concentration of harmful gases on the future concentration of harmful gases can be found, which is of great significance for the management and prediction of harmful gas concentration in underground cavern construction.

[0133] The above-mentioned embodiments are only preferred technical solutions of the present invention and should not be regarded as limitations of the present invention. The embodiments and features in the embodiments of the present application can be arbitrarily combined with each other without conflict. The protection scope of the present invention shall be the technical solutions recorded in the claims, including the equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A method for predicting the concentration of harmful gases from multiple sources in underground cavern construction driven by multiple factors, characterized in that: The following steps are involved: Step S1: Collecting underground cavern construction site data, including harmful gas concentration monitoring values ​​and factors related to the harmful gas concentration values, including humidity, temperature and atmospheric pressure; Step S2: preprocessing the collected data; Step S3: Use the preprocessed data to build a machine learning harmful gas concentration prediction model; Step S4: Predict the concentration of harmful gases based on the machine learning harmful gas concentration prediction model.

2. The method for predicting the concentration of harmful gases from multiple sources in underground cavern construction driven by multiple factors according to claim 1 is characterized in that: In step S2, the operation steps are as follows: S201, performing correlation analysis on all features of the field data to find a significant correlation feature group; S202, using principal component analysis to perform dimensionality reduction on the correlation-significant feature group; S203, rescale all data after dimensionality reduction processing to the range of [0, 1].

3. The method for predicting the concentration of harmful gases from multiple sources in underground cavern construction driven by multiple factors according to claim 2 is characterized in that: In step S201, Pearson correlation analysis is performed on all features to calculate the Pearson correlation coefficient between any two sets of data. The expression is as follows: In the formula, r represents the Pearson correlation coefficient between any two sets of data, and its value range is [-1, 1], 1 represents a perfect positive correlation, -1 represents a perfect negative correlation, and 0 represents no linear correlation; n is the number of data; x and y are two different variables.

4. The method for predicting the concentration of harmful gases from multiple sources in underground cavern construction driven by multiple factors according to claim 2 is characterized in that: In step S202, the covariance matrix in the principal component analysis method is C ij , perform eigendecomposition on the covariance matrix to obtain eigenvectors. The eigenvectors represent the directions of the principal components of the data, and the eigenvalues ​​represent the variances in each direction, that is, the degree of data dispersion. Select the eigenvector W corresponding to the maximum eigenvalue as the new feature space. Finally, multiply the original data matrix X by W to obtain the reduced-dimensional data Y, the covariance matrix C ij The expression is as follows: Where m is the number of samples, X ki is the i-th eigenvalue of the k-th sample, is the mean value of the ith feature, is the mean value of the jth feature.

5. The method for predicting the concentration of harmful gases from multiple sources in underground cavern construction driven by multiple factors according to claim 2 is characterized in that: In step S203, the scaling process expression of rescaling all processed data to the range of [0, 1] is as follows: Among them, x min and x max Represent the maximum and minimum values ​​in the variable, x i is the value of the variable before scaling, x k is the new variable value after scaling.

6. The method for predicting the concentration of harmful gases from multiple sources in underground cavern construction driven by multiple factors according to claim 1 is characterized in that: The step S3 includes the following steps: S301. Construct a machine learning harmful gas concentration prediction model including a category boosting ensemble learning module and a deep learning module. The category boosting ensemble learning module is based on a category boosting ensemble learning algorithm. The category boosting ensemble learning module is used to extract all factor information related to the harmful gas concentration. The deep learning module is composed of a convolutional neural network, a long short-term memory network, and a self-attention mechanism, and is used to extract time information about the harmful gas concentration. S302, dividing the underground cavern harmful gas prediction data set processed in step S2 into a training set, a validation set and a test set, using the training set and the validation set for model training, and using the test set to verify the model prediction accuracy.

7. The method for predicting the concentration of harmful gases from multiple sources in underground cavern construction driven by multiple factors according to claim 6 is characterized in that: In step S301, the process of extracting factor information related to the concentration of harmful gases by the category promotion integrated learning module is as follows: Assume that the harmful gas prediction data set D = {x i ,y i }(i=1,2,…,m) contains m samples, where x i are the construction characteristics and geological characteristics corresponding to the i-th sample, y i is the label value of the i-th sample, i.e., the harmful gas monitoring value; each x i Contains n-dimensional features, x ij is the j-th dimension feature of the i-th sample; Randomly sort the data set D to obtain σ, σ = {σ(1), σ(2), …, σ(m)}, and for each sample x i Train a decision tree model M i , and the residual error for each training is as follows: r i =y i -M σ(i-1) (x i ); Where: M σ(n) (x i ) is the sample x in the sequence σ(n) i Trained decision tree model; Harmful gas prediction data set D = {x i ,y i }(i=1,2,…,m) n features contain both numerical features and categorical features, k is the dimension of this feature when processing a categorical feature, For the sample corresponding to the sequence σ(p), The harmful gas detection values ​​are averaged and the priori value and the weight coefficient of the priori value are added to make a certain category feature of the sample Convert to numeric type, The expression is as follows: Where: For sample The corresponding harmful gas monitoring value is only when hour, is equal to 1, otherwise it is equal to zero; a>0, which indicates the weight of the prior p.

8. The method for predicting the concentration of harmful gases from multiple sources in underground cavern construction driven by multiple factors according to claim 6 is characterized in that: In step S302, when the test set is used to test the prediction accuracy of the model, the model accuracy evaluation adopts the evaluation index determination coefficient R 2 , root mean square error RMSE and mean absolute error MAE three evaluation indicators, the expressions are as follows: Where: y pred,i is the predicted value, y obs, i is the observation value, is the mean of the observations, and n is the number of observations.

9. The method for predicting the concentration of harmful gases from multiple sources in underground cavern construction driven by multiple factors according to claim 1, characterized in that: The method further includes step S5, using a machine learning interpretability enhancement algorithm to analyze important influencing factors and important time nodes of harmful gas concentrations, including the following steps: S501, integrating SHapley addition interpretation algorithm to analyze the important influencing factors of harmful gas concentration; S502. Integrate the local interpretable model-independent explanation algorithm to analyze the important time nodes for predicting the concentration of harmful gases.

Citation Information

Patent Citations

  • Method for predicting concentration of TVOC in atmosphere based on deep learning

    CN118114047A

Cited By

  • Air quality prediction method suitable for underground building space

    CN119539202A

  • A method for predicting air quality in underground building spaces

    CN119539202B

  • Power transmission line deep foundation pit environment safety situation awareness method and device based on space coupling, storage medium and terminal

    CN119942739A

  • Prediction method, medium and system for chlorophyll concentration of fresh tobacco leaves

    CN120031822A

  • Testing method of high-precision greenhouse gas concentration observation system

    CN120102806A