Method and system for inverting settlement parameters of working face of well mine

Abstract

The invention relates to an inversion method and system for sedimentation parameters of a working face of a well mine. For a trend subsidence prediction model and a tendency subsidence prediction model, trend model parameter inversion and tendency model parameter inversion are respectively implemented by using a deep neural network, and the deep neural network used for trend model parameter inversion is mainly composed of a full connection layer and an activation function Leaky Relu. The deep neural network for tendency model parameter inversion is mainly composed of a pruned and optimized full connection layer and an activation function SILU, and a loss function is a mean square error between a predicted subsidence value and an actual subsidence monitoring result. According to the method, the parameters of the probability integral model can be learned and inverted from the monitoring data more accurately, so that a more efficient and robust technical means is provided for predicting the ground surface settlement of the mining area, the parameter operation speed is effectively increased, and the parameter error is reduced.

Description

Technical Field

[0001] The present invention relates to a method for inverting subsidence parameter of a fully mechanized coal face in an underground mine and a system for inverting parameters of a subsidence prediction model using the method, which can be mainly applied to the monitoring of surface subsidence in mining areas in the mining industry and belongs to the technical field of artificial intelligence. Background Art

[0002] Subsidence of a fully mechanized coal face in an underground mine refers to the phenomenon that the surface subsides due to underground mining activities during the mining of coal or other mineral resources. This kind of subsidence usually has an impact on the environment and buildings around the mining area, so it is necessary to monitor and predict. The existing surface subsidence monitoring technologies mainly include the Global Positioning System (GNSS), terrestrial laser scanning, leveling, tilt measurement, and Interferometric Synthetic Aperture Radar (InSAR) technology, etc. Due to the large magnitude of subsidence and severe deformation in the mining area, the traditional InSAR technology is prone to image decoherence and low monitoring accuracy during the monitoring process. Therefore, some scholars have proposed a method of combining the traditional InSAR technology with the probability integral method to construct a subsidence prediction model (probability integral model), but currently, such prediction models have problems such as slow parameter solving speed and relatively large errors. Summary of the Invention

[0003] The purpose of the present invention is to reduce the parameter error of the subsidence prediction model and improve the operation speed.

[0004] The technical solution of the present invention is: a method for inverting subsidence parameters of a fully mechanized coal face in an underground mine, which respectively performs inversion of the parameters of the strike model (the model parameters of the strike subsidence prediction model) and the parameters of the dip model (the model parameters of the dip subsidence prediction model) for the strike (main section) subsidence (surface subsidence) prediction model and the dip (any dip line) subsidence prediction model by using a deep neural network.

[0005] Preferably, both the strike subsidence prediction model and the dip subsidence prediction model are constructed based on the probability integral method.

[0006] Further, the strike subsidence model is:

[0007]

[0008] wherein, (x, z) is the (surface) coordinate of any point on the strike, y is the abscissa (horizontal coordinate, or strike coordinate), z is the ordinate (vertical coordinate), w is the maximum subsidence value, r is the main influence radius, and l is the mining calculation length when the strike is limited.

[0009] Further, the parameters of the strike model to be inverted are w, r, and l.

[0010] Preferably, the deep neural network for dip model parameter inversion mainly consists of fully connected layers and the activation function LeakyRelu.

[0011] Preferably, the parameter structure of the deep neural network for dip model parameter inversion is shown in the following table:

[0012]

[0013]

[0014] In the table, Linear represents the fully connected layer, and b is the batch size of the neural network data.

[0015] Preferably, the loss function for deep neural network training is:

[0016]

[0017] where (x i , z i ) is the measured data of the dip monitoring point i (the position dip coordinate x i and the measured subsidence amount z i ), i = 1, 2, 3,..., n, n is the total number of dip monitoring points, w i is the maximum subsidence amount fitted for the i-th sample, r i is the main influence radius fitted for the i-th sample, and l i is the dip mining length fitted for the i-th sample.

[0018] Preferably, the activation function LeakyRelu is adjusted to the following form:

[0019]

[0020] Furthermore, the dip subsidence model is:

[0021]

[0022] where (y, z) is the coordinate of any point on the dip (surface coordinate), y is the abscissa (dip coordinate), z is the ordinate (vertical coordinate), w is the maximum subsidence value on this dip, α is the coal seam dip angle, D1 is the dip length, r1 is the influence radius of the downhill boundary, r2 is the influence radius of the uphill boundary, s1 is the inflection point offset in the downhill direction, s2 is the inflection point offset in the uphill direction, and θ is the mining influence propagation angle.

[0023] Furthermore, the dip model parameters to be inverted are w, r1, r2, s1, s2, and θ.

[0024] Preferably, the deep neural network for the inversion of the tendency model parameters mainly consists of a fully connected layer optimized by pruning and the activation function SILU.

[0025] Preferably, the parameter structure of the deep neural network for the inversion of the tendency model parameters is shown in the following table:

[0026]

[0027]

[0028] In the table, Linear represents the fully connected layer, and b is the batch size of the neural network data.

[0029] Preferably, the loss function for the training of the deep neural network is:

[0030]

[0031] where (y j , z) are the measured data of the tendency monitoring point j (the position tendency coordinate y j and the measured subsidence amount z i ), j = 1, 2, 3,..., m, m is the total number of tendency monitoring points, w i is the maximum subsidence amount fitted for the j-th sample, r 1j is the influence radius of the downhill boundary fitted for the j-th sample, r 2j is the influence radius of the uphill boundary fitted for the j-th sample, D 1j is the tendency length fitted for the j-th sample, s 1j is the inflection point offset distance in the downhill direction fitted for the j-th sample,, s 2j is the inflection point offset distance in the uphill direction fitted for the j-th sample, θ j is the mining influence propagation angle fitted for the j-th sample.

[0032] The subsidence prediction model parameter inversion system based on a deep neural network implements the subsidence prediction model parameter inversion by using any one of the subsidence prediction model parameter inversion methods disclosed in the present invention, including

[0033] a monitoring data acquisition unit, configured to obtain the measured data of each strike monitoring point and tendency monitoring point, where the strike monitoring points are distributed on the (surface) strike line of the (strike) main section, and the tendency monitoring points are one or more groups, and the tendency monitoring points in the same group are distributed on the same tendency line;

[0034] The strike model parameter inversion unit constructs a strike model parameter inversion model (a deep neural network model for strike model parameter inversion) based on the strike subsidence prediction model and strike measured data (the measured data from strike monitoring points), and implements strike model parameter inversion by using the strike model parameter inversion model and the strike measured data;

[0035] The dip model parameter inversion unit constructs a dip model parameter inversion model (a deep neural network model for dip model parameter inversion) based on the dip subsidence prediction model and dip measured data (the measured data from dip monitoring points), and implements dip model parameter inversion by using the dip model parameter inversion model and the dip measured data.

[0036] A human-computer interaction unit, a networking and / or communication unit, a storage unit, etc. can be set according to actual needs, and technical means such as cloud platforms, cloud computing, and cloud storage can be adopted.

[0037] The beneficial effects of the present invention are as follows: A neural network model for subsidence prediction model parameter inversion is constructed based on the subsidence prediction model, and corresponding neural network structures are respectively constructed for strike model parameter inversion and dip model parameter inversion, which not only greatly improves the speed of parameter solution but also enhances the stability of the model fitting degree. Through the adaptive adjustment ability of the deep learning algorithm, it can more accurately learn from the monitoring data and invert the parameters of the probability integral model, thereby providing a more efficient and robust technical means for predicting the surface subsidence in mining areas, effectively improving the parameter operation speed and reducing the parameter error. Description of the Drawings

[0038] Figure 1 is a schematic diagram of the parameter inversion process involved in the present invention (taking the strike as an example);

[0039] Figure 2 is an example of the layout of strike monitoring points and dip monitoring points involved in the present invention;

[0040] Figure 3 is a schematic diagram of the observed data and the subsidence curve fitting involved in the present invention;

[0041] Figure 4 is an example of the dip subsidence curve involved in the present invention;

[0042] Figure 5 is an example of the subsidence curve fitting involved in the present invention, where each dot corresponds to the measured data;

[0043] Figure 6 is an example of the neural network framework for inverting the parameters of the strike subsidence prediction model;

[0044] Figure 7 is an example of the neural network framework for inverting the parameters of the dip subsidence prediction model;

[0045] Figure 8 It is a schematic diagram of the parameter inversion system architecture involved in the present invention. Detailed implementation manners

[0046] After the present invention conducts a fusion analysis on D-InSAR analysis and offset tracking technology, and uses the probability integral method to invert parameters. When solving the probability integral parameters, a deep neural network is applied to the parameter inversion of the subsidence prediction model of the probability integral method, so as to improve the solution speed, fitting stability and calculation accuracy of the probability integral model parameters.

[0047] Refer to Figure 1 , the present invention mainly includes the following steps:

[0048] S1. Obtain subsidence data through fusion analysis.

[0049] The differential interferometric synthetic aperture radar (D-InSAR) technology and the offset tracking technology are used for fusion analysis. This comprehensive method can comprehensively collect the subsidence (or settlement) data above the mining face, providing an accurate data basis for subsequent analysis.

[0050] S2. Mathematically model the subsidence area.

[0051] After obtaining the subsidence data, the subsidence area is mathematically modeled by the probability integral method. Specifically, mathematical models (subsidence prediction models) are respectively established for the subsidence deformation laws in the main section along the strike of the mining area and the inclination direction of the working face.

[0052] The actual subsidence value (which can be represented by the z coordinate) and the deformation position (which can be represented by the x coordinate or y coordinate) can be used to estimate and fit the parameters of the model formula, and invert each model parameter, so as to obtain a non-linear equation describing the surface subsidence (which can be called the surface deformation equation / function, or the subsidence curve / surface function, etc. The fitting curve can be referred to Figure 5 .

[0053] S3. Construct a neural network model (neural network framework).

[0054] For the parameter inversion of the subsidence prediction model along the strike, based on the characteristics of the surface deformation equation along the strike, the constructed neural network structure is mainly composed of a fully connected layer and the activation function LeakyRelu (refer to Figure 6 ).

[0055] For the inversion of the parameters of the inclination subsidence prediction model, based on the characteristics of the surface deformation equation in the inclination direction, compared with the neural network structure for the inversion of the parameters of the strike subsidence prediction model, there are mainly two changes: (1) To prevent the occurrence of overfitting due to the increase in parameters, pruning optimization is carried out to make the model have a lower number of parameters; (2) The activation function SILU is used instead of the LeakyRelu activation function, making the parameter adjustment smoother during the training process and also avoiding the occurrence of the gradient vanishing phenomenon (see Figure 7 ).

[0056] S4. Data fitting and parameter optimization.

[0057] Input the sample data into the neural network for fitting. Through calculation, implement the gradient descent algorithm to continuously update the prediction results. This process ensures that the parameters output by the neural network can achieve non - linear fitting of the surface deformation function, thereby improving the accuracy of surface subsidence prediction.

[0058] Sample data of the strike and inclination can be obtained through surveying means such as RTK surveying instrument. Among them, the (x i , z i )i = 1, 2,..n are the strike sample data, where x i is the position coordinate of the strike, and z i is the corresponding settlement value. And the (y j , z j )j = 1, 2,..m are the inclination sample data, y j is the position coordinate of the inclination, and zj is the corresponding settlement value. It is also possible to obtain the sample data required for model training by using other suitable existing technologies, such as using D - InSAR settlement monitoring. At the center line of the working face strike or at the 1 / 3 position of the inclination, the InSAR settlement values and the corresponding strike and inclination position coordinates are collected near the selected strike line or inclination line.

[0059] Set strike monitoring points and inclination monitoring points on the working face according to actual needs. For example, Figure 2 In the example, the working face is rectangular in shape, with the length direction being the working face strike and the width direction being the working face inclination. To monitor the subsidence of this working face, monitoring points are selected on the working face (surface), such as Figure 2 D1 to D17 in are the strike (main section) monitoring points, and coal mining is from D17 to D1. Two inclination monitoring points are set near D12 and D6 respectively, namely A1 to A7 and A8 to A12, and B1 to B3 and B4 to B6. D12 and D6 can be used as the detection points for the corresponding inclinations at the same time.

[0060] See Figure 3 , the strike (main section) settlement curve is:

[0061]

[0062] Among them, (x, z) are the coordinates of any point on the strike (main strike section) (surface coordinates, such as Figure 3 shown), x is the abscissa (strike coordinate), z is the ordinate (vertical coordinate), is the subsidence amount at the surface x of the main strike section under the condition of reaching full mining subsidence for the dip, w is the maximum subsidence value, r is the main influence radius, l is the mining calculation length when the strike is limited, and erf is the Gaussian error function.

[0063] Equation 1 can be regarded as a strike subsidence prediction model. The corresponding subsidence amounts can be obtained through the monitoring points on the strike of the working face (for example, Figure 2 D1 to D17 in the example), and through the positions of these points D i (the subscript i represents the i-th strike monitoring point, and the x coordinate xi of this monitoring point can also be used to i define / determine this monitoring point) and the corresponding subsidence amount zi i , the corresponding model parameters (subsidence prediction model parameters) w, r, and l are fitted, and the obtained subsidence prediction model parameters are substituted into Equation 1 for subsidence prediction at any point x on the strike.

[0064] See Figure 4 , the subsidence curve for the dip (any dip) is:

[0065]

[0066] Among them, (y, z) are the coordinates of any point on the dip (such as Figure 4 shown), y is the dip coordinate, z is the ordinate (vertical coordinate), is the subsidence amount at the surface y of this dip, w is the maximum subsidence value of this dip, α is the coal seam dip angle, D1 is the dip length, r1 is the influence radius of the lower mountain boundary, r2 is the influence radius of the upper mountain boundary, s1 is the inflection point offset distance in the lower mountain direction, s2 is the inflection point offset distance in the upper mountain direction, θ is the mining influence propagation angle, and erf is the Gaussian error function.

[0067] It can be seen from Equation 2 that different from the horizontal state of the coal seam in the strike direction, due to the dip angle of the coal seam in the dip direction, the subsidence curve for the dip is more complex than the subsidence curve for the strike.

[0068] Equation 2 can be regarded as a dip subsidence prediction model. The corresponding subsidence amounts can be obtained through the monitoring points on the same dip of the working face (for example, A1 - A7, D12, and A8 - A12, or, B1 - B3, D6, and B4 - B6), and through the positions (abscissa) y i and the subsidence amounts (ordinate) z i(When \(i\) is the monitoring point number, it represents the coordinates or other quantities of the monitoring point \(i\) when used as a subscript), the corresponding model parameters (subsidence prediction model parameters) \(w\), \(r_1\), \(r_2\), \(s_1\), \(s_2\) and \(\theta\) are fitted, and the obtained subsidence prediction model parameters are substituted into Equation 2 for the subsidence prediction of any point \(x\) on the corresponding trend.

[0069] When using a deep neural network for the inversion of the parameters of the subsidence prediction model (strike), for the stability of iteration, the sample data / observation data \((x i , z i ) is standardized (normalized). After standardization, \((x i , z i ) can be expressed as \((u i , v i ) as follows:

[0070] 1) \(x i is translated and scaled so that the transformed \(u i is within the range of \((-30, 50)\), that is, \(u i =(x i - x k ) / p, where \(x k is the translation amount of \(x\), and \(p\) is the scaling factor (multiple) of \(x\);

[0071] 2) \(z i is scaled so that the transformed \(v i is within the range of \((-1, 2)\), that is, \(v i = z i / q, where \(q\) is the scaling factor (multiple) of \(z\).

[0072] When using a neural network for parameter inversion, the subsidence prediction model based on \(xz\) coordinates can be converted into a subsidence prediction model based on \(uv\) coordinates, and equivalent changes are allowed for the convenience of data processing. For example, for the strike subsidence prediction model, from Equation 1, we can get:

[0073]

[0074] Among them, \(l_1\) is the offset of the left horizontal movement distance after standardization, and \(l_2\) is the offset of the right horizontal movement distance after standardization.

[0075] See Figure 6 , the neural network framework (which can be called trend-net) for the inversion of the parameters of the strike subsidence prediction model is mainly composed of a fully connected layer and the activation function LeakyRelu.

[0076] The input data X is mapped through two fully connected layers respectively. The number of neurons C in the fully connected layer is 10. Feature fusion is achieved through stacking processing, and then the feature points are mapped and gradually amplified to 100 layers to increase the non-linear expression ability of the model and improve the recognition accuracy. The features of these 100 layers are mapped to 4 parameters w, r, l1, and l2 to be solved. At this time, the dimension of the output features is the amount of data B fed into the neural network and the number of output features C. Data normalization is achieved through dimension conversion, and then non-linear mapping of the data is achieved through the LeakyRelu activation function.

[0077] Considering the actual physical meaning of the model parameters in the function, the activation function LeakyRelu is adjusted, and the adjusted LeakyRelu is denoted as F(x). Then:

[0078]

[0079] Thus, according to the input sample data, the value ranges of the output parameters are all made greater than 0.

[0080] Table 1 is an embodiment of the trend-net parameter structure. In the table, Linear represents the fully connected layer, and b is the batch size of the neural network data. It can be seen from the table that the number of model parameters is low, effectively avoiding the problem of overfitting in training, and avoiding the problem of gradient disappearance compared with the gradient descent method.

[0081] Table 1 trend-net parameter structure

[0082]

[0083]

[0084] The model parameter values fitted (inverted) through the neural network can be brought into the corresponding subsidence prediction model to predict the subsidence of each monitoring point and compared with the corresponding settlement monitoring results. Loss calculation is performed according to the mean square error loss function of formula 5 to achieve gradient descent and update the prediction results, so that the finally output parameters can achieve non-linear fitting.

[0085]

[0086] Among them, (x i , z i ) is the measured data (position coordinates x i and measured subsidence amount z i ) of the monitoring point i in the strike direction, i = 1, 2, 3,..., n, where n is the total number of monitoring points in the strike direction, w i is the maximum subsidence amount fitted for the i-th sample, r i is the main influence radius fitted for the i-th sample, l iThe extraction length along the strike fitted for the \(i\)-th sample.

[0087] According to the actual situation, constraint conditions are added to the output results to make the output physically meaningful. Specifically, it can be as shown in Equation (6):

[0088] \(w\in(0, 3), r\in(0, 30), l_1\in(0, 30), l_2\in(0, 30)\) (6)

[0089] See Figure 7 , the neural network framework (which can be called tendency-net) for the inversion of parameters of the tendency subsidence prediction model is mainly composed of a fully connected layer and the activation function SILU. Compared with trend-net for the inversion of parameters of the strike subsidence prediction model, tendency-net has mainly made two changes: to prevent the appearance of overfitting due to the increase in parameters, pruning optimization is carried out to make the model have a lower number of parameters. In terms of the activation function, SILU is used to replace the LeakyRelu activation function, making the parameters adjust more smoothly during the training process and avoiding the appearance of the gradient vanishing phenomenon.

[0090] The training of tendency-net and the inversion of tendency model parameters can be implemented in the same or similar way as trend-net.

[0091] Table 2 is an example of the parameter structure of tendency-net. In the table, Linear represents the fully connected layer, and \(b\) is the batch size of the neural network data.

[0092] Table 2 Parameter Structure of tendency-net

[0093]

[0094]

[0095] The loss is calculated according to the mean square error loss function of Equation (7) to implement gradient descent and update the prediction results, so that the six parameters finally output can achieve non-linear fitting.

[0096]

[0097] where \((y\) j , \(z\) j ) are the measured data of the monitoring point \(j\) (the position coordinates \(y\) j and the measured subsidence amount \(z\) j ), \(j = 1, 2, 3, \cdots, m\), \(m\) is the total number of tendency monitoring points, \(w\) i is the maximum settlement amount fitted for the \(j\)-th sample, \(r\) 1j is the influence radius of the downhill boundary fitted for the \(j\)-th sample, \(r\)2j The influence radius of the uphill boundary fitted for the j-th sample, D 1j The dip length fitted for the j-th sample, s 1j The offset of the inflection point in the downhill direction fitted for the j-th sample, s 2j The offset of the inflection point in the uphill direction fitted for the j-th sample, θ j The propagation angle of mining influence fitted for the j-th sample.

[0098] According to the actual situation, add constraint conditions to the output results to make the output physically meaningful. Specifically, it can be as shown in Equation 8:

[0099] w ∈ (0, 3), r1 ∈ (0, 30), r2 ∈ (0, 30), s1 ∈ (0, 30), s2 ∈ (0, 30), θ ∈ (0, 628) (8)

[0100] The training of the neural network can be implemented in the following ways or any other suitable ways:

[0101] 1) Data input and batch processing

[0102] The input data is sent into the network in batch size, and each sample is the monitoring point of the strike (x i , z i ) or the monitoring point of the dip (y j , z j ). The initial feature dimension is adjusted so that the number of output features of each layer is a fixed value. For example, the dimension conversion is performed through a fully connected layer or a convolutional layer.

[0103] 2) Forward propagation layer by layer

[0104] The processing flow of each layer is as follows:

[0105] a) Dimension conversion: Map the input features to the corresponding dimensions through a linear transformation (such as a fully connected layer);

[0106] b) Normalization: Apply layer normalization (LayerNorm) or batch normalization (BatchNorm) to standardize the data, alleviate the internal covariate shift, and accelerate the training;

[0107] c) Nonlinear activation: Use the LeakyReLU function (for strike) or the SILU function (for dip) to enhance the nonlinear expression ability of the model.

[0108] 3) Optimization of the deep network

[0109] a) Gradient problem: Due to the multiple layers of the network depth, residual connections (Residual Connections) or gradient clipping (Gradient Clipping) may be introduced to prevent the gradient from vanishing / exploding;

[0110] b) Parameter initialization: Adopt methods such as initialization to adapt to the activation characteristics of LeakyReLU.

[0111] 4) Loss calculation and backpropagation

[0112] a) The output layer calculates the loss function according to the task design, such as Equation (5) or Equation (7);

[0113] b) Update the parameters of each layer through backpropagation, and the optimizer (such as Adam) adjusts the learning rate to iteratively optimize the model.

[0114] Application example. Apply the present invention to a fully-mechanized coal face in a mine. The strike length of the face in horizontal distance is 2,941.6 m, the inclined distance is 2,970 m, the average inclined distance in the dip direction is 283.7 m, and the horizontal distance is 271.9 m. The dip angle of the face is 15° - 33°, with an average of 20°. The average mining thickness is 4.1 m. The highest elevation of the face is 1,190.6 m, the lowest elevation is 877.1 m, and the maximum elevation difference of the face is 313.5 m.

[0115] A total of 17 observation points are arranged along the strike of the face, numbered from D01 to D17; a total of 16 observation points are arranged along two dip lines: D15-1 to D15-3, and B15-1 to B15-13. The overall layout method of the monitoring points is Figure 2 similar to the example.

[0116] To verify the accuracy of the surface deformation function fitted by the model parameter inversion method (deep learning) of the present invention, the common particle swarm optimization algorithm and least squares method fitting are implemented for the same example according to the prior art.

[0117] The comparison results in the strike direction are as follows:

[0118]

[0119] The comparison results in the dip direction are as follows:

[0120]

[0121] It can be seen from the above comparison results that the error of the predicted parameters of the probability integral model using the parameter inversion method based on deep learning of the present invention is smaller, and the fitting state is the best.

[0122] All the preferred and optional technical means disclosed in the present invention can be arbitrarily combined to form several different specific embodiments, except when specifically stated and when one preferred or optional technical means is a further limitation of another technical means.

Claims

1. An inversion method for subsidence parameters of a working face in an underground mine. For the strike subsidence prediction model and the dip subsidence prediction model, a deep neural network is used to perform the inversion of strike model parameters and dip model parameters respectively.

2. The inversion method for the subsidence parameters of the working face in an underground mine according to claim 1, characterized in that The strike subsidence model is as follows: Among them, (x, z) is the coordinate of any point on the strike, z is the vertical coordinate, w is the maximum subsidence value, r is the main influence radius, and l is the mining calculation length when the strike is limited.

3. The subsidence prediction model parameter inversion method according to claim 2, characterized in that The deep neural network for strike model parameter inversion is mainly composed of a fully connected layer and the activation function LeakyRelu.

4. The inversion method for the subsidence parameters of the working face in an underground mine as claimed in claim 3, wherein The parameter structure of the deep neural network for strike model parameter inversion is shown in the following table: In the table, Linear represents the fully connected layer, and b is the batch size of the neural network data.

5. The inversion method for the subsidence parameters of the working face in an underground mine as described in claim 3, characterized in that The loss function for deep neural network training is: Among them, (x i , z i ) are the measured data of the monitoring point i along the strike (the position strike coordinates x i and the measured subsidence amount z i ), where i = 1, 2, 3, …, n, and n is the total number of monitoring points along the strike. w i is the maximum subsidence amount fitted for the i-th sample, r i is the main influence radius fitted for the i-th sample, and l i is the strike mining length fitted for the i-th sample.

6. The inversion method for the subsidence parameters of the working face in an underground mine as described in claim 1, characterized in that The dip subsidence model is as follows: Among them, (y, z) is the coordinate of any point on the dip, y is the dip coordinate, z is the vertical coordinate, w is the maximum subsidence value on this dip, α is the coal seam dip angle, D1 is the dip length, r1 is the influence radius of the downhill boundary, r2 is the influence radius of the uphill boundary, s1 is the inflection point offset in the downhill direction, s2 is the inflection point offset in the uphill direction, and θ is the mining influence propagation angle.

7. The inversion method for the subsidence parameters of the working face in an underground mine according to claim 6, characterized in that The deep neural network for dip model parameter inversion is mainly composed of a fully connected layer optimized by pruning and the activation function SILU.

8. The inversion method of the subsidence parameters of the working face in an underground mine according to claim 7, characterized in that The parameter structure of the deep neural network for dip model parameter inversion is shown in the following table: In the table, Linear represents the fully connected layer, and b is the batch size of the neural network data.

9. The inversion method for the subsidence parameters of the working face in an underground mine according to claim 7, characterized in that The loss function for deep neural network training is: where (y j , z) are the measured data of the inclination monitoring point j (the position inclination coordinate y j and the measured subsidence amount z i ), j = 1, 2, 3, …, m, m is the total number of inclination monitoring points, w i is the maximum subsidence amount fitted for the jth sample, r 1j is the influence radius of the downhill boundary fitted for the jth sample, r 2j is the influence radius of the uphill boundary fitted for the jth sample, D 1j is the inclination length fitted for the jth sample, s 1j is the inflection point offset distance in the downhill direction fitted for the jth sample, s 2j is the inflection point offset distance in the uphill direction fitted for the jth sample, θ j is the mining influence propagation angle fitted for the jth sample.

10. An inversion system for subsidence parameters of a working face in an underground mine. The inversion method for subsidence parameters of a working face in an underground mine described in any one of claims 1-9 is used to perform the inversion of subsidence prediction model parameters, including: A monitoring data acquisition unit for obtaining the measured data of each strike monitoring point and dip monitoring point. Each strike monitoring point is distributed on the strike line of the main section. The dip monitoring points are one group or multiple groups, and each dip monitoring point in the same group is distributed on the same dip line; A strike model parameter inversion unit for constructing a strike model parameter inversion model based on the strike subsidence prediction model and the strike measured data, and performing strike model parameter inversion using the strike model parameter inversion model and the strike measured data; A dip model parameter inversion unit for constructing a dip model parameter inversion model based on the dip subsidence prediction model and the dip measured data, and performing dip model parameter inversion using the dip model parameter inversion model and the dip measured data.

Images