A method for predicting the spatiotemporal distribution of tunnel gas based on machine learning model

Abstract

A method for predicting the spatiotemporal distribution of gas in a tunnel based on a machine learning model belongs to the technical field of intelligent prevention and control of mine disasters, and solves the technical problems of accuracy and real-time prediction of gas concentration in the tunnel, including the following steps: discretizing the control equation in space and time, extracting the predicted value in the local area of ​​the flow field as a constraint block; decomposing the control equation into a constraint matrix and a prediction matrix; finally, constructing a hyperplane that conforms to the gas convection-diffusion control equation based on the constraint matrix, and projecting the prediction matrix onto the hyperplane to obtain a prediction that satisfies the constraints of the convection-diffusion equation. The present invention models the gas concentration in the tunnel from two perspectives: physical model prediction and big data prediction. It can fully extract and utilize the effective information in the potential physical process with the help of the physical model, and can combine the advantages of the recurrent neural network that is good at extracting time series data features, retaining the dependencies between data, and having relatively good generalization ability.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent prevention and control of mine disasters, relates to mine gas data prediction technology, and specifically relates to a method for predicting the spatiotemporal distribution of tunnel gas based on a machine learning model. Background Art

[0002] Gas disasters restrict the safe production of high-gas mines. Due to the complexity and uncertainty of the tunnel environment, how to accurately predict the gas concentration in the tunnel and promptly detect potential dangers is an urgent problem to be solved in the field of coal mine safety.

[0003] Although many patents on gas concentration prediction currently involve various methods and technologies such as machine learning, numerical simulation, data mining, and fuzzy mathematics, the accuracy of gas concentration prediction is still not high. The existing gas prediction methods have the following two defects:

[0004] 1. The prediction of gas concentration based on machine learning algorithms is a nonlinear system behavior that relies on a large amount of historical data. The characteristics and laws of gas concentration are extracted through data mining and machine learning technologies, which often need to consider the noise and uncertainty of the data. However, the number of existing gas concentration monitoring sensors in coal mines is still limited, making it difficult to achieve high-precision spatiotemporal distribution prediction.

[0005] 2. Currently, many models / research methods can only analyze data based on machine learning. In fact, they do not consider the underlying physical mechanism factors more, and have poor interpretability and generalizability. Summary of the invention

[0006] In order to overcome the shortcomings of the existing technology, improve the accuracy and real-time performance of tunnel gas concentration prediction, and provide reliable data support for safety decision-making, the present invention aims at the real-time monitoring and prediction problem of tunnel gas, and proposes a tunnel gas spatiotemporal distribution prediction method based on a machine learning model.

[0007] The design concept of the present invention is to model the tunnel gas concentration from the two perspectives of physical model prediction and big data prediction, and apply the recurrent neural network method based on physical processes, which can not only fully extract and utilize the effective information in the potential physical process with the help of the physical model, but also combine the advantages of recurrent neural networks in extracting time series data features, retaining the dependencies between data, and having relatively good generalization capabilities.

[0008] The present invention is achieved through the following technical solutions.

[0009] A method for predicting the spatiotemporal distribution of tunnel gas based on a machine learning model comprises the following steps:

[0010] S1. Collect historical data and establish a historical gas feature data set X (x, y, z, t, c) for big data prediction

[0011] Firstly, the spatial position, environmental parameters, CO concentration, dust concentration and gas concentration data of the tunnel are collected in real time by sensors and data acquisition equipment arranged in the tunnel as historical feature data c; secondly, the collected data are preprocessed by standardization method; thirdly, the principal component analysis method is used to extract features of the preprocessed data, and features related to gas concentration are extracted from the collected data, including historical gas concentration, temperature, humidity and wind speed; finally, the extracted features are transformed by data normalization method, and a historical gas feature data set X(x, y, z, t, c) for big data prediction is established, where x, y, z are the three-dimensional coordinate values ​​of the midpoint of the tunnel, t is time, and c is historical feature data;

[0012] S2. The control equation of gas distribution in coal mine tunnels is:

[0013]

[0014] Where ρ is the gas density; is the volume fraction of the gas component; D is the gas diffusion coefficient; x i 、x j (i=1, 2, 3; j=1, 2, 3) represent the x direction, y direction, and z direction respectively; u i (i=1, 2, 3) represents the projection of velocity u in the x-direction, y-direction, and z-direction respectively;

[0015] Boundary conditions of gas distribution in coal mine tunnels for:

[0016]

[0017] In the formula, x BC ,y BC , z BC Respectively represent the coordinates of the boundary area in the x, y, and z directions, Indicates the volume fraction of the gas component in the boundary area;

[0018] Initial Conditions of Gas Distribution in Coal Mine Tunnel for:

[0019]

[0020] S3. The control equation of gas distribution in coal mine tunnels is discretized in space and time respectively by finite difference method. The spatial dimension is processed by second-order central difference method, and the time is discretized by first-order backward difference method. The points in the vicinity of any point (determined by the results of time and space discretization) whose distance from the point is equal to the corresponding interval difference Δx, Δy, Δz, Δt are defined as constraint blocks, and the finite difference structural equation is constructed.

[0021]

[0022] In the formula, u x ,u y ,u z Respectively represent the partial derivatives of velocity u in the x, y, and z directions;

[0023] S4. Using matrix decomposition method, the control equation of gas distribution in coal mine tunnel is transformed into a hyperplane, so that the obtained mapping prediction value conforms to the physical mechanism. The prediction value on the constraint block in the neural network can be extracted as the prediction matrix H and the constraint matrix A:

[0024]

[0025] S5. Transform from the constraint block to the projection block, derive the mapping matrix P, and obtain the projection prediction value T:

[0026] P=(1-A T (AA T ) -1 A);

[0027] In the formula, P is the mapping matrix, A is the constraint matrix, and T is the projection prediction value;

[0028] S6. Randomly allocate the area in the coal mine tunnel and obtain the coordinates of the allocation points (x j ,y j , z j , t j ), boundary condition coordinates (x k ,y k , z k , t k ) and the initial value condition coordinates (x l ,y l , z l , t l ), complete data sampling;

[0029] S7. Determine the observation loss function MSE DATA :

[0030]

[0031] Where N is the number of observation points, NN(X) is the neural network algorithm, and Y obs represents the observed value of gas concentration;

[0032] S8. Determine the control equation loss function MSE PDE :

[0033]

[0034] Where N f is the number of points of action of the control equation;

[0035] S9. Determine the boundary condition loss and initial condition loss function:

[0036] Determine the boundary condition loss function MSE BC :

[0037]

[0038] Where N BC is the number of points in the boundary condition region, is the gas concentration of the boundary condition, is the i-th prediction value of the neural network model;

[0039] Determine the initial condition loss function MSE IC :

[0040]

[0041] Where N IC is the number of action points in the initial condition region, is the gas concentration at the initial condition, is the i-th prediction value of the neural network model;

[0042] S10, introduce the penalty parameter, and obtain the total loss function L of the prediction model:

[0043] L=λ DATA MSE DATA +λ PDE MSE PDE +λ BC MSE BC +λ IC MSE IC ;

[0044] In the formula, λ DATA , PDE , BC , IC are penalty parameters, and are all hyperparameters;

[0045] S11. According to the result of the total loss function L of the prediction model, the back propagation method and the Adam optimizer are used to update the weights and biases in the neural network until the total loss function L of the prediction model converges, and a trained model can be obtained;

[0046] S12. Use the validation set to validate the trained prediction model and adjust the hyperparameters;

[0047] S13. Input the real-time collected data into the trained model to obtain the prediction results of the spatiotemporal distribution of gas concentration.

[0048] The beneficial effects of the present invention are:

[0049] The method of the present invention takes advantage of the physical model and deep learning, and incorporates the gas convection diffusion control equation as a constraint condition into the neural network through projection. Specifically, the control equation is discretized in space and time, and the predicted value in the local area of ​​the flow field is extracted as a constraint block; the control equation is decomposed into a constraint matrix reflecting the physical mechanism and a prediction matrix composed of the output of the neural network on each constraint block. Finally, a hyperplane that conforms to the gas convection diffusion control equation is constructed based on the constraint matrix, and the prediction matrix is ​​projected onto the hyperplane to obtain a prediction that satisfies the convection diffusion equation constraint. The present invention can ensure higher prediction accuracy, strictly satisfy the control equation in a local area or the entire area, require very little sample data, have higher prediction accuracy, stronger robustness, and stronger generalizability, so as to better predict the roadway gas, provide reliable data support for risk decision-making in daily mine management and the entire process of disaster occurrence, achieve the purpose of moving the risk threshold forward and preventing disasters, and ensure the life safety of mine workers. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0051] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0052] like Figure 1 A method for predicting the spatiotemporal distribution of tunnel gas based on a machine learning model is shown, comprising the following steps:

[0053] S1. Collect historical data and establish a historical gas feature data set X (x, y, z, t, c) for big data prediction

[0054] Firstly, the spatial position, environmental parameters, CO concentration, dust concentration and gas concentration data of the tunnel are collected in real time by sensors and data acquisition equipment arranged in the tunnel as historical feature data c; secondly, the collected data are preprocessed by standardization method; thirdly, the principal component analysis method is used to extract features of the preprocessed data, and features related to gas concentration are extracted from the collected data, including historical gas concentration, temperature, humidity and wind speed; finally, the data normalization method is used to transform the extracted features, and a historical gas feature data set X(x, y, z, t, c) for big data prediction is established, where x, y, z are the three-dimensional coordinate values ​​of the midpoint of the tunnel respectively; t is time, and c is the historical feature data;

[0055] S2. The control equation of gas distribution in coal mine tunnels is:

[0056]

[0057] Where ρ is the gas density; is the volume fraction of the gas component; D is the gas diffusion coefficient; x i 、x j (i=1, 2, 3; j=1, 2, 3) represent the x direction, y direction, and z direction respectively; u i (i=1, 2, 3) represents the projection of velocity u in the x-direction, y-direction, and z-direction respectively; according to the wind speed distribution in the tunnel, the solution process is simplified, and the velocity u(x, y, z) is a function of the coordinate position and has nothing to do with time t;

[0058] Boundary conditions of gas distribution in coal mine tunnels for:

[0059]

[0060] In the formula, x BC ,y BC , z BC Respectively represent the coordinates of the boundary area in the x, y, and z directions, Indicates the volume fraction of the gas component in the boundary area;

[0061] Initial Conditions of Gas Distribution in Coal Mine Tunnel for:

[0062]

[0063] S3. The control equation of gas distribution in coal mine tunnels is discretized in space and time respectively by finite difference method. The spatial dimension is processed by second-order central difference method, and the time is discretized by first-order backward difference method. The points in the vicinity of any point (determined by the results of time and space discretization) whose distance from the point is equal to the corresponding interval difference Δx, Δy, Δz, Δt are defined as constraint blocks, and the finite difference structural equation is constructed.

[0064]

[0065] In the formula, u x ,u y ,u z Respectively represent the partial derivatives of velocity u in the x, y, and z directions;

[0066] S4. Using matrix decomposition method, the control equation of gas distribution in coal mine tunnel is transformed into a hyperplane, so that the obtained mapping prediction value conforms to the physical mechanism. The prediction value on the constraint block in the neural network can be extracted as the prediction matrix H and the constraint matrix A:

[0067]

[0068] S5. Transform from the constraint block to the projection block, derive the mapping matrix P, and obtain the projection prediction value T:

[0069] P=(1-A T (AA T ) -1 A);

[0070] In the formula, P is the mapping matrix, A is the constraint matrix, and T is the projection prediction value;

[0071] S6. Randomly allocate the area in the coal mine tunnel and obtain the coordinates of the allocation points (x j ,y j , z j , t j ), boundary condition coordinates (x k ,y k , z k , t k ) and the initial value condition coordinates (x l ,y l , z l , t l ), complete data sampling;

[0072] S7. Determine the observation loss function MSE DATA :

[0073]

[0074] Where N is the number of observation points, NN(X) is the neural network algorithm, and Y obs represents the observed value of gas concentration;

[0075] S8. Determine the control equation loss function MSE PDE :

[0076]

[0077] Where N f is the number of points of action of the control equation;

[0078] S9. Determine the boundary condition loss and initial condition loss function:

[0079] Determine the boundary condition loss function MSE BC :

[0080]

[0081] Where N BC is the number of points in the boundary condition region, is the gas concentration of the boundary condition, is the i-th prediction value of the neural network model;

[0082] Determine the initial condition loss function MSE IC :

[0083]

[0084] Where N IC is the number of action points in the initial condition region, is the gas concentration at the initial condition, is the i-th prediction value of the neural network model;

[0085] S10, introduce the penalty parameter, and obtain the total loss function L of the prediction model:

[0086] L=λ DATA MSE DATA +λ PDE MSE PDE +λ BC MSE BC +λ IC MSE IC ;

[0087] In the formula, λ DATA , PDE , BC , IC are penalty parameters, and are all hyperparameters;

[0088] S11, according to the result of the total loss function L of the prediction model, use the back propagation method and Adam optimizer to update the weights and biases in the neural network until the total loss function L of the prediction model converges;

[0089] S12. Use the validation set to validate the trained prediction model and adjust the hyperparameters;

[0090] S14. Input the real-time collected data into the trained model to obtain the prediction results of the spatiotemporal distribution of gas concentration.

[0091] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for predicting the spatiotemporal distribution of tunnel gas based on a machine learning model, characterized in that: The following steps are involved: S1. Collect historical data and establish a historical gas feature data set X (x, y, z, t, c) for big data prediction First, the spatial position, environmental parameters, CO concentration, dust concentration and gas concentration data of the tunnel are collected in real time through sensors and data acquisition equipment arranged in the tunnel as historical characteristic data c; Secondly, the collected data is preprocessed by using a standardization method. Thirdly, the principal component analysis method is used to extract features from the preprocessed data, and features related to gas concentration are extracted from the collected data, including historical gas concentration, temperature, humidity, and wind speed. Finally, the data normalization method is used to transform the extracted features and establish a historical gas feature data set X(x, y, z, t, c) for big data prediction, where x, y, and z are the three-dimensional coordinate values ​​of the midpoint of the tunnel, t is time, and c is historical feature data. S2. The control equation of gas distribution in coal mine tunnels is: Where ρ is the gas density; is the volume fraction of the gas component; D is the gas diffusion coefficient; x i 、x j (i=1, 2, 3; j=1, 2, 3) represent the x direction, y direction, and z direction respectively; u i (i=1, 2, 3) represents the projection of velocity u in the x-direction, y-direction, and z-direction respectively; Boundary conditions of gas distribution in coal mine tunnels for: In the formula, x BC ,y BC , z BC Respectively represent the coordinates of the boundary area in the x, y, and z directions, Indicates the volume fraction of the gas component in the boundary area; Initial Conditions of Gas Distribution in Coal Mine Tunnel for: S3. The control equation of gas distribution in coal mine tunnels is discretized in space and time by finite difference method. The spatial dimension is processed by second-order central difference method, and the time is discretized by first-order backward difference method. The points in the vicinity of any point whose distance from the point is equal to the corresponding interval difference Δx, Δy, Δz, Δt are defined as constraint blocks, and the finite difference structural equation is constructed. In the formula, u x ,u y ,u z Respectively represent the partial derivatives of velocity u in the x, y, and z directions; S4. Using matrix decomposition method, the control equation of gas distribution in coal mine tunnel is transformed into a hyperplane, so that the obtained mapping prediction value conforms to the physical mechanism. The prediction value on the constraint block in the neural network can be extracted as the prediction matrix H and the constraint matrix A: S5. Transform from the constraint block to the projection block, derive the mapping matrix P, and obtain the projection prediction value T: P=(1-A T (CHALLENGE ACCEPTED T ) -1 A); In the formula, P is the mapping matrix, A is the constraint matrix, and T is the projection prediction value; S6. Randomly allocate the area in the coal mine tunnel and obtain the coordinates of the allocation points (x j ,y j , z j , t j ), boundary condition coordinates (x k ,y k , z k , t k ) and the initial value condition coordinates (x l ,y l , z l , t l ), complete data sampling; S7. Determine the observation loss function MSE DATA : Where N is the number of observation points, NN(X) is the neural network algorithm, and Y obs represents the observed value of gas concentration; S8. Determine the control equation loss function MSE PDE : Where N f is the number of points of action of the control equation; S9. Determine the boundary condition loss and initial condition loss function: Determine the boundary condition loss function MSE BC : Where N BC is the number of points in the boundary condition region, is the gas concentration of the boundary condition, is the i-th prediction value of the neural network model; Determine the initial condition loss function MSE IC : Where N IC is the number of action points in the initial condition region, is the gas concentration at the initial condition, is the i-th prediction value of the neural network model; S10, introduce the penalty parameter, and obtain the total loss function L of the prediction model: L=λ DATA MSE DATA +λ PDE MSE PDE +λ BC MSE BC +λ IC MSE IC ; In the formula, λ DATA , PDE , BC , IC are penalty parameters, and are all hyperparameters; S11. According to the result of the total loss function L of the prediction model, the back propagation method and the Adam optimizer are used to update the weights and biases in the neural network until the total loss function L of the prediction model converges, and a trained model can be obtained; S12. Use the validation set to validate the trained prediction model and adjust the hyperparameters; S13. Input the real-time collected data into the trained model to obtain the prediction results of the spatiotemporal distribution of gas concentration.

Images