Working face earth surface rock displacement digital intelligent observation and structure passing safety intelligent prediction method

Abstract

The invention discloses a working face earth surface rock movement digital intelligent observation and through-structure safety intelligent prediction method, and belongs to the technical field of mining industry and information, and the method comprises the following steps: determining a structure movement deformation safety early warning value; moving deformation data are collected and preprocessed; analyzing a dynamic change rule of surface movement deformation by applying finite element analysis; building a structure movement deformation prediction model; the structure movement deformation prediction model is trained and verified, and the trained model is used for structure movement deformation prediction; comprehensively evaluating safety and stability; and correcting parameters of the prediction model and continuously evaluating until the structure is safely pushed and mined. The invention provides a novel way with great potential for accurately predicting the rock migration trend and guaranteeing the safety of important facilities such as structures.

Description

Technical Field

[0001] The present invention belongs to the field of mining and information technology, and specifically relates to a method for digital intelligent observation of surface rock movement at a working face and intelligent prediction of safety of passing structures. Background Art

[0002] When there are important facilities such as structures near the coal mining area, the potential safety risks brought by surface rock movement in the process of "passing over structures" increase dramatically. The so-called "passing over structures" refers to the dynamic process in which surface rock movement gradually spreads from areas far away from structures to the location of structures and has a direct impact on them. For example, gas stations usually store a large amount of flammable and explosive oil products, and their facilities are extremely sensitive to surface deformation. When surface rock movement "passes over structures", even relatively small surface displacements may cause compression, stretching or distortion of the building structure of the gas station, causing cracks in the wall and uneven settlement of the foundation; collision and compression of oil storage equipment, causing deformation or even rupture; pulling and bending of oil pipelines, causing loose interfaces or pipeline ruptures. These damages can easily cause oil leaks. Once they encounter ignition sources such as open flames or static electricity, they may cause serious accidents such as fires or even explosions, which will not only cause huge economic losses, but also pose a serious threat to the surrounding environment and the safety of people's lives.

[0003] Traditional methods for observing and analyzing surface rock movement mainly include methods based on empirical formulas and some simple statistical analysis methods. Methods based on empirical formulas are often summarized under specific geological conditions, and their versatility is poor. When facing different geological environments or complex mining conditions, the prediction results may deviate greatly from the actual situation. Although simple statistical analysis methods can process historical data to a certain extent, they cannot effectively mine the deep information in the data, and it is difficult to capture the nonlinear characteristics and long-term dependencies in surface rock movement data. When dealing with complex time series data such as surface rock movement, traditional methods have many limitations. For example, for characteristics such as periodic fluctuations and trend changes in time series, it is difficult for traditional methods to accurately predict and model at the same time. Moreover, as time goes by, new data is constantly generated, and traditional methods cannot adaptively update the model to adapt to the new data pattern, resulting in the accuracy of the prediction gradually decreasing over time.

[0004] In recent years, with the development of machine learning and deep learning technologies, long short-term memory networks (LSTM) and their variants have shown excellent performance in processing time series data. In particular, stacked LSTM, by stacking multiple LSTM layers together, can gradually extract multi-level features from complex time series data. Each layer of LSTM can be regarded as a feature extractor. The lower-level LSTM can capture local features and short-term dependencies in the data. As the number of layers increases, the higher-level LSTM can further integrate these local features and mine longer-term and more complex dependencies and patterns. The multi-head attention mechanism of Transformer can process the feature sequence output by the LSTM layer in parallel, spanning different positions and time points, and comprehensively capturing the long-distance dependencies in the data, effectively compensating for the information transmission attenuation problem that may occur when LSTM processes long sequences. This multi-level feature extraction method gives it a significant advantage in processing complex time series data such as surface rock movement that is affected by multiple factors, and provides a new and promising way to accurately predict rock movement trends and ensure the safety of important facilities such as structures. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a method for digital observation of surface rock movement in working faces and intelligent prediction of the safety of passing structures. It mainly uses stacked long short-term memory networks and Transformer to establish a structure movement deformation prediction model based on movement deformation observation data mining, and uses this model to predict the movement deformation of structures. According to the prediction results of structure movement deformation, the safety and stability of the structure are comprehensively evaluated, and when the predicted value exceeds the warning value, corresponding technical measures are taken.

[0006] The technical solution of the present invention is as follows: A method for digital intelligent observation of rock movement on a working face and intelligent prediction of safety of passing structures comprises the following steps: Step 1: Determine the safety warning value of structure movement and deformation; Step 2: Collect movement deformation data and perform preprocessing; Step 3: Use finite element analysis to analyze the dynamic change law of surface movement and deformation; Step 4: Establish a prediction model for structure movement and deformation; Step 5: Train and verify the structure movement and deformation prediction model, and use the trained model to predict the structure movement and deformation; Step 6: Comprehensively evaluate safety and stability; Step 7: Modify the prediction model parameters and continue to evaluate until the structure is safely mined.

[0007] Furthermore, the specific process of step 1 is as follows: Step 1.1: Comprehensively consider the structural characteristics of the structure, the geological conditions and the surrounding environment to obtain the key factors affecting the stability and safety of the structure; the key factors include the foundation type of the structure, the characteristics of the building materials, the overall stiffness of the structure, the geological structure characteristics, and other surrounding engineering activities; Step 1.2: Identify the key indicators that affect the movement and deformation state of the structure, including displacement, inclination, curvature, and strain; determine the threshold of each key indicator in combination with the design specifications, usage requirements, and relevant industry standards of the structure. This threshold is the safety warning value for the movement and deformation of the structure.

[0008] Furthermore, the specific process of step 2 is as follows: Step 2.1, install various types of sensors at key structural parts of the structure, and collect daily surface movement and deformation data of the mined area in front of the structure through the sensors; the sensors include displacement sensors, inclination sensors, curvature sensors, and strain sensors; Step 2.2, perform outlier detection and processing; The process of outlier detection is as follows: define the displacement data sequence collected by the displacement sensor as , calculate the mean of the displacement data series and standard deviation , the formula is as follows: ; ; in, is the total number of displacement data points; For the displacement data points; For the displacement data points; for The displacement data points of the interval are marked as outliers; is the standard deviation The proceeds from the prescription; The process of outlier processing is as follows: For data with a stable trend, linear interpolation method is used to deal with outliers. Assume that at the first time point and the third time point The displacement data points are and , and the second time point If the displacement data points are abnormal or missing, Displacement data points It is calculated by the linear interpolation formula, which is: ; For data with unstable data change trends, the weighted linear interpolation method is used to process outliers. Middle Displacement data points The displacement data points that are identified as outliers and removed are called missing value points. Interpolate the position; suppose the time series corresponding to the displacement data series is , For the A point in time, is the total number of time points, one displacement data point corresponds to one time point, and the total number of time points corresponds to the total number of displacement data points; the weighted linear interpolation formula is: ; ; in, for The interpolation result of the location; for The weight of For the Time point To Time point time distance; For the Time point To Time point time distance; For the Time point The displacement data points; Step 2.3: Normalize the data. The formula is: ; in, is the normalized displacement data point; is the displacement data point; , are the minimum and maximum values ​​of the displacement data in the entire data set, respectively; For inclination, curvature, and strain, outlier detection and processing and data normalization are performed in the same way as displacement data.

[0009] Furthermore, the specific process of step 3 is as follows: Step 3.1, construct a geometric model, use the unit division strategy to discretize the units, and obtain a discretized model; the specific process is: first, build a two-dimensional or three-dimensional geometric model based on geological survey data and the design blueprint of the structure; then, use the unit division strategy to discretize the two-dimensional or three-dimensional geometric model to obtain a discretized model; the process of the unit division strategy is: let the total number of units be The total number of nodes is ; For each unit, the numbers of its three nodes are , , , No. The displacement vector of any node in a unit is , is the horizontal displacement, is the vertical displacement, , Respectively Horizontal displacement and vertical displacement of any node in a unit; ,node ,node The displacement vectors are , , ; , Node Horizontal displacement and vertical displacement; , Node Horizontal displacement and vertical displacement; , Node Horizontal displacement and vertical displacement; is the transposed sign; displacement vectors are connected through shape functions; for linear triangle elements, nodes The shape function is expressed as: ; ; ; ; in, For Node The shape function of is the area of ​​the triangular unit; For computing nodes The shape function is the intermediate parameter constructed; For computing nodes The correlation coefficient when the shape function is ; For computing nodes The correlation coefficient when the shape function is ; is the abscissa representing the horizontal position in the plane rectangular coordinate system; It is the ordinate representing the vertical position in the plane rectangular coordinate system; The nodes in the triangle element The horizontal axis of The nodes in the triangle element The horizontal axis of The nodes in the triangle element The vertical coordinate of The nodes in the triangle element The vertical coordinate of node Shape function ,node Shape function With Node The shape functions of are expressed in the same way; Final The displacement vector of any node in a unit is expressed as: ; At the same time, the strain-displacement relationship is: ; in, For the The strain vector of any node in a unit; is the strain-displacement matrix, which is given by: ; Step 3.2, establish the constitutive equation, equilibrium equation and overall stiffness matrix; Step 3.3: Solve the overall stiffness matrix and analyze the results to obtain the dynamic change law of surface movement and deformation.

[0010] Further, in step 3.2, The constitutive equation of each unit is: ; in, For the The stress tensor of any node in an element; is the elasticity matrix, expressed as: ; in, is the elastic modulus; is Poisson's ratio; The specific expression of the equilibrium equation is: ; in, It is The volume of a unit; For virtual strain; is the strain vector; is the stress tensor; is the volume; It is The surface of each unit; is the surface force vector; For the surface; is the displacement vector; Is a node The force vector of The overall stiffness equation is: ; in, is the overall stiffness matrix; is the node displacement vector; is the force vector of the node; No. The stiffness matrix of the element The calculation formula is: ; Overall stiffness matrix The assembly process is to superimpose the stiffness matrix of each unit to the corresponding position of the overall stiffness matrix according to the corresponding relationship between the unit node and the overall node; In step 3.3, the overall stiffness equation is solved by direct method or iterative method. Get the node displacement vector ; After obtaining the node displacement vector, the unit strain vector and stress tensor are further calculated; the unit strain vector is calculated based on the strain-displacement relationship, and the unit stress tensor is calculated by the constitutive equation.

[0011] Furthermore, in step 4, the input variables of the structure movement deformation prediction model are the relevant parameters in the surface movement deformation observation data, and the output is the movement deformation value of the structure; the relevant parameters include displacement, inclination, curvature, and strain; The structure movement deformation prediction model includes multiple stacked LSTM layers, a Transformer layer, and a fully connected layer; LSTM is a long short-term memory network; Transformer is a sequence model based on the attention mechanism; The working process of the structure movement deformation prediction model is as follows: Step 4.1: Input the relevant parameters in the surface movement and deformation observation data into the first layer LSTM; the working process of the first layer LSTM is: Step 4.1.1: Preprocess the input related parameters. After preprocessing, the input related parameters are a vector at each time point. Define the first The input data for each time step is ,in is the dimension of the input data; For the time step, The Data of dimensions; Step 4.1.2: Calculate the forget gate. The formula is: ; in, is the forgotten gate in the first layer of LSTM Output of time steps; is the Sigmoid activation function; is the weight matrix of the forget gate in the first layer of LSTM; is the first layer of LSTM The hidden state of time steps; is the bias vector of the forget gate in the first layer of LSTM; Step 4.1.3: Perform input gate calculation to generate input gate control signals and candidate cell states. The formula is: ; ; in, is the first layer of LSTM The control signal of the input gate in time steps; is the weight matrix of the input gate in the first layer of LSTM; is the bias vector of the input gate in the first layer of LSTM; is the first layer of LSTM Candidate cell states for time steps; is the tanh activation function; is the weight matrix used to calculate the candidate cell state in the first layer of LSTM; is the bias vector used to calculate the candidate cell state in the first layer of LSTM; Step 4.1.4, update the cell state according to the output of the forget gate and the input gate: ; in, is the first layer of LSTM Update cell state at each time step; Represents an element-by-element multiplication operation; is the first layer of LSTM The cell state at each time step; Step 4.1.5: Perform output gate calculation to generate the output gate control signal and output hidden state. The formula is: ; ; in, is the first layer of LSTM The control signal of the output gate in time steps; is the weight matrix of the output gate in the first layer of LSTM; is the bias vector of the output gate in the first layer of LSTM; is the first layer of LSTM The hidden state of time steps; Step 4.2: Enter the second layer LSTM. The calculation formula of the second layer LSTM is as follows: ; ; ; ; ; ; in, For the forgotten gate in the second layer LSTM Output of time steps; is the weight matrix of the forget gate in the second layer of LSTM; is the bias vector of the forget gate in the second layer of LSTM; is the first The hidden state of time steps; is the first The control signal of the input gate in time steps; is the weight matrix of the input gate in the second layer of LSTM; is the bias vector of the input gate in the second layer of LSTM; is the first Candidate cell states for time steps; It is the weight matrix used to calculate the candidate cell state in the second layer LSTM; is the bias vector used to calculate the candidate cell state in the second layer LSTM; is the first Update cell state at each time step; is the first The cell state at each time step; is the first The control signal of the output gate in time steps; is the weight matrix of the output gate in the second layer of LSTM; is the bias vector of the output gate in the second layer of LSTM; In the second LSTM layer The hidden state of time steps Will be used as the input of the third layer LSTM; Step 4.3, The input of the LSTM layer is The output hidden state of the layer; The calculation formula of layer LSTM is as follows: ; ; ; ; ; ; in, For the Forgetting level in layer LSTM Output of time steps; It is The weight matrix of the forget gate in the layer LSTM; It is The bias vector of the forget gate in the layer LSTM; For the In the layer LSTM The control signal of the input gate in time steps; It is The weight matrix of the input gate in the layer LSTM; It is Bias vector of the input gate in layer LSTM; For the In the layer LSTM Candidate cell states for time steps; It is The weight matrix used to calculate the candidate cell state in the layer LSTM; It is The relevant bias vector used to calculate the candidate cell state in the layer LSTM; For the In the layer LSTM Update cell state at each time step; It is In the layer LSTM The cell state at each time step; For the In the layer LSTM The control signal of the output gate in time steps; It is The weight matrix of the output gate in the layer LSTM; It is The bias vector of the output gate in the layer LSTM; For the In the layer LSTM The hidden state of time steps; is the total number of time steps, and the final output of the entire stacked LSTM layer is In the layer LSTM The hidden state of time steps ; Step 4.4: Input Transformer layer and get the output matrix ; Step 4.5: Input the fully connected layer to predict the movement and deformation of the structure.

[0012] Furthermore, in step 4, the Transformer layer adopts a multi-head attention mechanism and a feedforward neural network; the specific process is: Step 4.4.1: First, the final output of the stacked LSTM layer Perform linear transformation to obtain the query matrix , key matrix Sum Matrix ;set up , , They are , , The weight matrix is: ; ; ; Then, suppose , , For the The query matrix, key matrix, and value matrix of each head, then The attention scores of the heads are: ; in, For the Attention score of the individual; is the softmax function; yes Dimensions; is the transpose symbol; is the number of heads; Finally, the output matrix of multi-head attention is for: ; in, For splicing operation; It is the weight matrix used to fuse multi-head information; For the Attention score of the individual; Step 4.4.2: Feedforward neural network processing is performed, and the formula is: ; in, It is a feed-forward neural network; , are two different weight matrices for a feedforward neural network; , are two different bias vectors for a feedforward neural network; In step 4.5, the fully connected layer will The features are converted into specific prediction values; The predicted value of the time step The calculation formula is: ; in, is the weight matrix of the fully connected layer; is the bias vector of the fully connected layer.

[0013] Furthermore, the specific process of step 5 is as follows: Step 5.1: Train the model by minimizing the mean square error loss function; let the predicted value sequence be , the true value sequence is , then the calculation formula of the mean square error loss function is: ; in, is the mean square error loss function value; For the The true value of the time step; For the The predicted value of time steps; At the same time, the mean absolute error loss function is used as a supplementary evaluation indicator, and the calculation formula is: ; in, is the mean absolute error loss function value; Step 5.2: Use the adaptive moment estimation optimization algorithm to update the parameters of the model; the parameters include the weight matrix and bias vector of each LSTM layer, Transformer layer, and output layer; the update process is as follows: First, calculate the update of the first-order moment vector: ; in, For the The first-order moment vector of time steps; For the The first-order moment vector of time steps; For the The gradient of time steps; Then, calculate the update of the second-order moment vector: ; in, For the The second-order moment vector of time steps; For the The second-order moment vector of time steps; and are two different hyperparameters used to control momentum; Next, the first-order moment vector and the second-order moment vector are corrected for deviation: ; ; in, , Respectively First-order moment vector deviation and second-order moment vector deviation for each time step; Finally, update the parameters: ; in, , Respectively time step, Parameters for time steps; is the learning rate; is a non-zero constant; Step 5.3: Hyperparameters including learning rate , the number of hidden units in the LSTM layer And the number of heads of the Transformer layer , set a series of value, Value and The value is evaluated on the validation set. The mean square error and mean absolute error are selected as the evaluation indicators. The best hyperparameter combination is selected. The number of training rounds when the loss of the validation set reaches the minimum value is selected as the final number of training rounds. The learning rate scheduling strategy is used to adjust the learning rate.

[0014] Furthermore, the specific process of step 6 is: comparing the model prediction value with the safety warning value determined in step 1. When the prediction value does not exceed the warning value, it indicates that the structure is currently in a safe state; when the prediction value exceeds the warning value, it indicates that there is a safety risk in the structure.

[0015] Furthermore, the specific process of step 7 is as follows: Step 7.1, during the mining process of the working face, continuously obtain the next stage of movement and deformation observation data of the surface and structures in the mined area; Step 7.2: Substitute the newly acquired observation data into the prediction model and modify the parameters of the prediction model through the Adam optimization algorithm so that the model can better adapt to the new data characteristics and actual deformation conditions; Step 7.3: Use the revised prediction model to predict the movement and deformation of the structure again, and re-evaluate its safety and stability; Step 7.4: Repeat step 7.3, continuously adjusting the model and evaluating the safety status until the working face is safely pushed through the structure, ensuring that the safety impact on the structure during the entire mining process is within a controllable range.

[0016] The beneficial technical effects brought about by the present invention: The present invention proposes a method for digital observation of surface rock movement at the working face and intelligent prediction of the safety of passing structures. By utilizing the powerful processing capabilities of stacked LSTM and Transformer for complex time series data, it is able to deeply mine the multi-level features in the surface rock movement data. By stacking multiple layers of LSTM and Transformer, the model can gradually extract long-term change trends from short-term rock movement fluctuations. This multi-level information extraction method makes the prediction of rock movement more accurate. Compared with traditional prediction methods, it can more accurately capture subtle changes in rock movement, reduce prediction errors, and provide a reliable basis for subsequent safety assessments. By comparing the predicted rock movement data with the safety threshold of the structure in real time, potential safety threats can be discovered in a timely and accurate manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1It is a flow chart of the method for digital intelligent observation of surface rock movement at the working face and intelligent prediction of safety of passing structures according to the present invention. DETAILED DESCRIPTION

[0018] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments: The present invention aims to realize reasonable and accurate dynamic safety early warning of structures, and proposes a dynamic safety early warning and protection technology for structures based on movement deformation prediction-correction-re-prediction. The specific technical ideas are as follows: First, analyze the impact of surface movement deformation on the target structure, clarify the main indicators affecting the structure, and determine the threshold of each indicator, that is, the safety early warning value of structure movement deformation. Secondly, on the one hand, based on the daily surface movement deformation observation data of the mined area in front of the structure, analyze the dynamic change law of surface movement deformation, and predict the movement deformation of the structure accordingly; on the other hand, establish a structure movement deformation prediction model based on movement deformation observation data mining to predict the movement deformation of the structure; according to the prediction results of structure movement deformation, comprehensively evaluate the safety and stability of the structure, and take corresponding technical measures when the prediction value exceeds the warning value. Finally, according to the surface movement deformation observation data of the mined area and the structure in the next stage, correct the prediction model parameters, and then predict the movement deformation of the structure and its safety and stability evaluation until the working face safely pushes through the structure.

[0019] like Figure 1 As shown, a method for digital intelligent observation of rock movement on a working face and intelligent prediction of safety of passing structures comprises the following steps: Step 1: Determine the safety warning value of structure movement and deformation. Specifically, it includes the following steps: Step 1.1: Analyze the impact of ground movement and deformation on the target structure in detail. By comprehensively considering the structural characteristics, geological conditions and surrounding environment of the structure, the key factors affecting its stability and safety are obtained; the key factors finally determined are: the foundation type of the structure (such as shallow foundation and deep foundation, different foundation forms have different bearing and adaptability to ground deformation), building material properties (such as the strength grade of concrete, the yield strength of steel, etc., the material properties determine its performance under stress and deformation), the overall stiffness of the structure (including the cross-sectional dimensions and layout of components such as beams and columns, etc., the stiffness affects the ability of the structure to resist deformation), geological structure characteristics (such as the stability of the stratum, whether there are faults or weak interlayers, etc., the geological structure will aggravate or alleviate the transmission effect of ground movement and deformation on the structure), and other surrounding engineering activities (such as whether there are other underground engineering constructions and traffic loads in the adjacent area, these external factors may have a superimposed effect with the ground movement and deformation caused by coal mining).

[0020] Step 1.2, clarify the key indicators that affect the movement and deformation state of the structure, including displacement (horizontal displacement and vertical displacement), inclination, curvature, strain, etc. Combined with the design specifications, use requirements and relevant industry standards of the structure, determine the threshold of each key indicator, which is the safety warning value of the movement and deformation of the structure. For example, for ordinary brick-concrete residential buildings, the horizontal displacement threshold can be set to 10 mm, the vertical displacement threshold is 15 mm, the inclination threshold is 0.003 (the inclination ratio is calculated based on the height of the structure), the curvature threshold is 0.0002 (indicating the degree of bending per unit length), and the strain threshold is 0.001 (reflecting the degree of deformation of the structural material); for large steel structure frames in industrial plants, considering that their structural stiffness is relatively large but the deformation accuracy requirements are high, the horizontal displacement threshold can be set to 8 mm, the vertical displacement threshold is 12 mm, the inclination threshold is 0.002, the curvature threshold is 0.00015, and the strain threshold is 0.0008.

[0021] Step 2: Collect movement deformation data and perform preprocessing. The specific process is as follows: Step 2.1. Install various types of sensors at key structural parts of the structure, including but not limited to high-precision displacement sensors (such as laser displacement sensors, capacitive displacement sensors, etc.), inclination sensors (such as dual-axis inclinometers), curvature sensors, strain sensors (such as fiber Bragg grating strain sensors), etc. These sensors can collect daily surface movement and deformation data of the mined area in front of the structure in real time or regularly, and these data cover a sufficiently long time period and multiple working conditions. The arrangement of sensors needs to be carefully designed based on the structural mechanics model of the structure to ensure that the deformation state of the structure can be fully and accurately reflected. For large structures, such as high-rise buildings and large bridges, the arrangement of sensors needs to follow certain structural mechanics principles and monitoring objectives. For example, sensors are arranged at key parts such as bridge piers, main beam mid-span and ends, and cable towers to fully obtain deformation information of the bridge under different working conditions. For high-rise buildings, sensors are installed at corner columns, core tube peripheries, key beams and columns on different floors of the structure. The frequency of data collection depends on the type of structure and monitoring requirements. For large buildings with slow changes, the collection frequency can be once every hour or several hours; for structures in complex environments or those that are sensitive to deformation, the collection frequency may be increased to once per minute or even higher. At the same time, the collection timestamp of each data point needs to be recorded in order to subsequently construct accurate time series data.

[0022] Step 2.2: Perform outlier detection and processing.

[0023] Step 2.2.1. Outlier detection. Define the displacement data sequence collected by the displacement sensor as , assuming that under normal circumstances, the displacement data sequence follows a normal distribution . Calculate the mean of the displacement data series and standard deviation , the formula is as follows: ; ; in, is the total number of displacement data points; For the displacement data points; For the displacement data points; According to the properties of normal distribution, about 99.7% of the displacement data points should fall within For displacement data points outside this interval, they are initially marked as outliers. is the standard deviation The result of taking the square root.

[0024] Step 2.2.2, outlier processing. When the sensor has communication failure, power supply problem or other reasons that cause abnormal or missing data, the outlier needs to be processed. For short-term outliers (such as several consecutive time points), if the data change trend is relatively stable, a simple linear interpolation method is used. Assume that at the first time point and the third time point The displacement data points are and , and the second time point If the displacement data points are abnormal or missing, Displacement data points It can be calculated by the linear interpolation formula, which is: ; In order to improve the accuracy of interpolation, weighted linear interpolation is considered to assign different weights according to the time distance between the displacement data points and the outliers.

[0025] Assume the displacement data sequence Middle Displacement data points The displacement data points that are identified as outliers and removed are called missing value points. Interpolate the position. Suppose the time series corresponding to the displacement data series is , For the A point in time, is the total number of time points, which corresponds to the total number of displacement data points. Time point The displacement data points are , , No. The time point is , is the time point corresponding to the outlier point or missing value point, , No. Time point To Time point The time distance is .calculate Weight ,The weight is usually calculated based on a distance function, such as , For the Time point To Time point Here the denominator is a normalization factor that ensures that the sum of all weights is 1, i.e. The weighted linear interpolation formula is: ; in, for The interpolation result of the position; the meaning of this formula is that the interpolation result Value is another valid data point The weighted sum of It depends on the time distance between the data point and the missing value point. The closer the time distance is, the greater the weight of the data point is, and the greater the contribution to the interpolation result is.

[0026] For inclination, curvature, and strain, outlier detection and processing can be performed in the same way as displacement.

[0027] Step 2.3, perform data normalization. The specific process is: for each selected feature, that is, the data features related to the key indicators that can reflect the movement and deformation state of the structure collected by various sensors, including displacement (horizontal displacement and vertical displacement), inclination, curvature, strain, etc., normalization is performed separately. At the same time, the dimensional differences and data range differences between different sensor data need to be considered to ensure that the normalization process does not lose the original information of the data.

[0028] The minimum-maximum normalization method is used to map the data to the interval [0, 1]. Let the minimum value of the displacement data in the entire data set be , the maximum value is For each displacement data point in the data set , and the normalized result is expressed by the formula Calculated, is the normalized displacement data point. This normalization method can make feature data of different magnitudes comparable and help improve the efficiency and stability of model training. For example, the displacement data point may be at the millimeter level, while the strain data may be a very small proportional value. Through normalization, they can be unified to the same numerical range. When calculating the minimum and maximum values, it is necessary to ensure that there are no outliers in the data set, otherwise the normalization result may be inaccurate. If there are outliers in the data set, it is necessary to process the outliers before normalization.

[0029] For features such as inclination, curvature, strain, etc., data normalization can be performed in the same way as displacement features.

[0030] Step 3: Use finite element analysis to analyze the dynamic changes of surface movement and deformation. The core principle of finite element analysis is to discretize the continuous research object (such as an area containing surface and underground geological structures) into a finite number of small units, and finally solve the mechanical response of the entire system under given loads and boundary conditions by establishing mechanical equilibrium equations at the unit level and considering the connection and interaction between units, so as to obtain the dynamic changes of surface movement and deformation. It is based on the variational principle or the weighted residual method to convert the partial differential equations describing surface deformation into a set of algebraic equations for solution. This process can simulate the stress-strain state of complex geological bodies under the influence of various factors (such as mining activities, groundwater level changes, structure loads, etc.) and the resulting surface displacement changes. The specific process is as follows: Step 3.1: Construct a geometric model, use the unit division strategy to discretize the unit, and obtain a discretized model. The specific process is: Step 3.1.1. Construct a two-dimensional or three-dimensional geometric model based on detailed geological survey data, including information on stratum distribution, lithology parameters (such as elastic modulus, Poisson's ratio, density, etc.), geological structures (such as faults, folds, etc.), and precise design blueprints of structures (covering structural form, size, material properties, etc.). For example, when studying the impact of coal mining on surface structures, it is necessary to determine the depth, thickness, and inclination of the coal seam, the shape and range of the goaf, and the location, foundation form, and height of the surface structure. For three-dimensional models, the undulation of the terrain must also be considered, and terrain elevation information can be obtained through survey data or geographic information systems (GIS) for modeling.

[0031] Step 3.1.2, the process of unit division strategy is: select the appropriate unit type and divide it according to the characteristics and accuracy requirements of the research problem. After the division is completed, the unit discretization is realized, and then the discretization model is obtained. For complex shapes and stress concentration areas, triangular units are more advantageous. Assume that the total number of units is The total number of nodes is For each unit, the three nodes are numbered , , , No. The displacement vector of any node in a unit is , is the horizontal displacement, is the vertical displacement, , Respectively Horizontal displacement and vertical displacement of any node in a unit; ,node ,node The displacement vectors are , , ; , Node Horizontal displacement and vertical displacement; , Node Horizontal displacement and vertical displacement; , Node Horizontal displacement and vertical displacement; is the transposed symbol. The displacement vectors are connected through shape functions, which are functions based on the coordinates of the unit nodes and have the characteristics of taking values ​​between 0 and 1 within the unit and taking values ​​of 1 at the corresponding nodes. For linear triangular elements, the nodes The shape function can be expressed as: ; ; ; ; in, For Node The shape function of is the area of ​​the triangular unit; For computing nodes The shape function is the intermediate parameter constructed; For computing nodes The correlation coefficient when the shape function is ; For computing nodes The correlation coefficient when the shape function is ; is the abscissa representing the horizontal position in the plane rectangular coordinate system; It is the ordinate representing the vertical position in the plane rectangular coordinate system; The nodes in the triangle element The horizontal axis of The nodes in the triangle element The horizontal axis of The nodes in the triangle element The vertical coordinate of The nodes in the triangle element The vertical coordinate of .

[0032] node Shape function ,node Shape function With Node The shape functions of are expressed in the same way; Final The displacement vector of any node in a unit can be expressed as: ; At the same time, the strain-displacement relationship is: ; in, For the The strain vector of any node in a unit; is the strain-displacement matrix, which consists of the partial derivatives of the shape function with respect to the coordinates and is given by: ; Step 3.2: Establish the constitutive equation, equilibrium equation and overall stiffness matrix. The specific process is as follows: Step 3.2.1, determine the constitutive equation. Determine the constitutive equation according to the characteristics of the geological material. For elastic materials, the stress-strain relationship follows the generalized Hooke's law, and the constitutive equation can be expressed as: ; in, is the stress tensor; is the strain tensor; is the elastic matrix. For an isotropic elastic material, the elastic matrix The expression is: ; in, is the elastic modulus; is Poisson's ratio; these parameters are determined through geological surveys and material tests.

[0033] No. The constitutive equation of each unit is: ; in, For the The stress tensor of any node in an element; Step 3.2.2, establish the equilibrium equation and the overall stiffness matrix. Based on the principle of virtual work, the equilibrium equation is established. For the discretized model obtained in step 3.1, the virtual displacement Under this condition, the virtual work of external forces is equal to the virtual work of internal forces. The virtual work of external forces includes the virtual work done by surface forces and concentrated forces, while the virtual work of internal forces is the work done by the internal stress of the unit on the virtual strain. The specific expression of the equilibrium equation is: ; in, It is The volume of a unit; For virtual strain; is the strain vector; is the stress tensor; is the volume; It is The surface of each unit; is the surface force vector; For the surface; is the displacement vector; Is a node The force vector.

[0034] The stress-strain relationship Substituting into the equilibrium equation and using the strain-displacement relationship , the unit integral is converted into the relationship between the node force and displacement through the shape function, and the overall stiffness equation is obtained: ; in, is the overall stiffness matrix, which is assembled from the stiffness matrices of each unit. The elements in the overall stiffness matrix are Representation Node Unit displacement at the node The force generated on is the node displacement vector; is the force vector of the node. The stiffness matrix of the element The calculation formula is: ; Overall stiffness matrix The assembly process is to superimpose the stiffness matrix of each unit to the corresponding position of the overall stiffness matrix according to the corresponding relationship between the unit node and the overall node. The three nodes of a unit are numbered , , , each element in The stiffness matrix of the element The location is , these elements are added to the overall stiffness matrix The corresponding position Among them, For the Nodes of each unit Unit displacement at the node The force generated on Step 3.3: Solve the overall stiffness matrix and analyze the results to obtain the dynamic change law of surface movement and deformation. The specific process is: Step 3.3.1. Select the solution method. Solve the overall stiffness equation Get the node displacement vector , usually solved by direct method (such as Gaussian elimination method) or iterative method (such as conjugate gradient method, Jacobi iteration method, etc.). Direct method is more efficient when dealing with small and medium-scale problems, but for large-scale sparse matrices (common in finite element analysis), iterative method is often more advantageous because it can use the sparsity of the matrix to reduce the amount of calculation and storage requirements. In the solution process, it is necessary to consider the processing of boundary conditions. For example, for fixed boundaries of the ground surface, the corresponding node displacement components are set to zero; for symmetric boundaries, the amount of calculation can be reduced according to the symmetry conditions.

[0035] Step 3.3.2, result analysis. After obtaining the node displacement vector, the strain vector and stress tensor of the unit can be further calculated. The unit strain vector can be calculated based on the strain-displacement relationship, and the unit stress tensor is calculated by the constitutive equation. For surface movement deformation analysis, the focus is on the displacement of the surface nodes. For example, when calculating the vertical displacement of a point on the surface, it is obtained by interpolating the node displacement and shape function of the unit where the point is located. Suppose the coordinates of a point on the surface are , the unit is , then the vertical displacement of the point , For Node The vertical displacement of the structure can be determined by finite element analysis at different time points or under different working conditions. The dynamic change law of surface movement and deformation can be obtained, such as the development process of surface settlement basins, the change direction and magnitude of displacement vectors, etc. These results can be output in the form of data files for subsequent use as input for the structure movement and deformation prediction model based on stacked LSTM. At the same time, the distribution and changes of stress concentration areas can also be analyzed to determine whether they will cause geological body damage or instability of the structure foundation, providing a basis for engineering design and safety assessment. According to the rules obtained from the analysis, combined with factors such as the relative position of the structure and the mining area, the similarity of geological conditions, etc., the movement and deformation of the structure are predicted.

[0036] Step 4: Establish a prediction model for structure movement and deformation; Establish a prediction model for structure movement and deformation based on movement and deformation observation data mining. When constructing the model, the relevant parameters in the surface movement and deformation observation data are used as input variables, and the movement and deformation values ​​of the structure are used as output variables; the relevant parameters include displacement (horizontal displacement and vertical displacement), inclination, curvature, strain, etc.; When predicting the movement and deformation of a structure, its deformation process is affected by a variety of complex factors, such as the structural characteristics of the structure itself (including the materials and structural forms used in the structure), external loads (wind loads, earthquake effects, thermal stress caused by temperature changes, etc.), and changes in geological conditions (foundation settlement, soil creep, etc.). These factors are intertwined in the time dimension, making the movement and deformation data of the structure present complex time series characteristics. The stacked LSTM model can effectively capture these complex time series relationships, thus providing the possibility for accurate deformation prediction. Although LSTM can handle long-term dependencies in time series data, its performance may be limited when facing very long sequences or complex multi-factor interactions. The Transformer layer combined with the LSTM layer can use the advantages of the Transformer layer to supplement the shortcomings of the LSTM. For example, LSTM may gradually forget early information when processing long sequences, while the long-distance dependency capture ability of the Transformer layer can retrieve these important early information and integrate them into the subsequent feature representation. When the model predicts the movement and deformation of a structure, this combination can more comprehensively and accurately consider the influence of various factors at different times, thereby improving the accuracy of the prediction; The model as a whole consists of multiple stacked LSTM layers, a Transformer layer, and a fully connected layer. Transformer is a sequence model based on the attention mechanism. First, each layer of LSTM uses its unique gating mechanism to preliminarily process the input data, and with its expertise in processing time series data, it captures local dependencies and short-term change trends in the data. As the number of LSTM layers increases, the model can gradually extract more representative and abstract features, and preliminarily mine the time series laws in the data. Subsequently, the Transformer layer plays a key role. Its multi-head attention mechanism can process the feature sequence output by the LSTM layer in parallel, spanning different positions and time points, and comprehensively capturing the long-distance dependencies in the data, effectively compensating for the information transmission attenuation problem that may occur when LSTM processes long sequences. At the same time, the Transformer layer assigns different weights to different feature dimensions through the self-attention mechanism, which can focus on the more critical information parts in the data, further extract global features and complex interactions, and integrate the local features extracted by LSTM with more extensive contextual information, thereby building a more comprehensive and in-depth understanding of the data. This hierarchical structure, which combines stacked LSTM and Transformer, constitutes an efficient hierarchical learning system for data features. Starting from the original multi-source sensor data, it gradually digs out deep structural information and hidden patterns, effectively responding to the complex deformation patterns caused by the interweaving of multiple complex factors during the movement and deformation of structures, and provides a solid foundation for accurately predicting the trend of movement and deformation of structures; The working process of the structure movement deformation prediction model is as follows: Step 4.1: Input the relevant parameters in the surface movement and deformation observation data into the first layer LSTM; the working process of the first layer LSTM is: Step 4.1.1: Preprocess and initialize. Preprocess the input parameters. After preprocessing, the input parameters are initialized at each time point. is a vector that defines the first The input data for each time step is ,in is the dimension of the input data, i.e., the number of selected sensor data features; For the time step, The At the same time, initialize the initial hidden state of the first layer LSTM is a zero vector with dimension (For example, you can set , the specific value can be determined according to model adjustment and experiment). is the number of hidden units in the LSTM layer; the initial cell state Also initialized to a zero vector, the dimension is also The initialization value here is chosen to give the model a starting state. The zero vector is a simple and commonly used initialization method, but in some cases, random initialization or initialization based on a specific distribution can also be used to improve the performance of the model.

[0037] Step 4.1.2, perform forget gate calculation. The calculation of forget gate involves the concatenation of the input data at the current time point and the hidden state at the previous time point. Input data for time steps and the first layer of LSTM The hidden state of time steps Concatenate into a vector , whose dimensions are Then, we perform matrix multiplication with the weight matrix of the forget gate, add the bias vector, and finally get the output of the forget gate through the Sigmoid activation function: ; in, is the forgotten gate in the first layer of LSTM Output of time steps; is the Sigmoid activation function; is the weight matrix of the forget gate in the first layer of LSTM, with a dimension of , The initial values ​​of the elements can be generated by random initialization methods (such as Xavier initialization or He initialization). These initialization methods can reasonably set the initial range of weights according to the input and output dimensions to help the model converge faster; is the bias vector of the forget gate in the first layer of LSTM, with a dimension of .

[0038] During the training process, and The value of will be continuously adjusted according to the gradient descent algorithm. Each element of The cell state at each time step A dimension of is used to determine the degree of retention of the dimension information, 0 means completely discarded, 1 means completely retained. For example, if If an element of is close to 0, then the corresponding The information in will be largely forgotten at the current time point. This forgetting mechanism is one of the keys to LSTM's ability to handle long-term dependencies. It can prevent old and unimportant information from accumulating in long time series, thus affecting the model's predictive ability.

[0039] Step 4.1.3: Perform input gate calculation to generate input gate control signal and candidate cell state. The calculation method of input gate control signal is similar to that of forget gate. Multiply it with the weight matrix of the input gate, add the bias vector, and finally pass it through the Sigmoid activation function to get: ; in, is the first layer of LSTM The control signal of the input gate in time steps; is the weight matrix of the input gate in the first layer of LSTM, with a dimension of ; is the bias vector of the input gate in the first layer of LSTM; and The initialization method is similar to the parameters of the forget gate. The control signal takes values ​​between 0 and 1 and controls the extent to which new information flows into the cell state. When an element of is close to 1, it means that more new information is allowed to enter the cell state; when it is close to 0, it means that the inflow of new information is restricted.

[0040] Generate the candidate cell state at the current time point through a tanh activation function layer. Multiply it by the weight, add the bias vector, and then pass it through the tanh activation function to get: ; in, is the first layer of LSTM Candidate cell states for time steps, Contains the potential information contained in the current time step input data and the hidden state of the previous time step, and its value is between -1 and 1. For example, if a displacement feature value in the input data changes significantly, this change will be reflected in the multiplication of the weight matrix and the action of the tanh function. In this way, the model can dynamically adjust the candidate cell states according to the changes in the input data and prepare for the update of the cell state. is the tanh activation function; is the weight matrix used to calculate the candidate cell state in the first layer of LSTM, with a dimension of ; is the bias vector used to calculate the candidate cell state in the first layer of LSTM, with a dimension of .

[0041] Step 4.1.4, update the cell state according to the output of the forget gate and the input gate: ; in, is the first layer of LSTM Update cell state at each time step; Represents an element-by-element multiplication operation; is the first layer of LSTM The updating process is to first selectively forget part of the information in the cell state at the previous time point according to the output of the forget gate, and then integrate the new candidate cell state information according to the output of the input gate. For example, if The value of one dimension is 0.3. If the value of this dimension is 5, then when updating, this part of information will become 0.3×5=1.5; at the same time, if The value in the corresponding dimension is 0.8, When the value of this dimension is 0.6, the newly incorporated information is 0.8×0.6=0.48, and the final updated cell state value of this dimension is 1.5+0.48=1.98. This way of updating the cell state enables the model to flexibly adjust the degree of utilization of historical information and new information at each time point, thereby better handling changes in time series data.

[0042] Step 4.1.5: Perform output gate calculation to generate the output gate control signal and output hidden state. The calculation method of the output gate control signal is similar to that of the forget gate and input gate. Multiply it with the weight matrix of the output gate, add the bias vector, and finally pass it through the Sigmoid activation function to get: ; in, is the first layer of LSTM The control signal of the output gate in time steps; is the weight matrix of the output gate in the first layer of LSTM, with a dimension of ; is the bias vector of the output gate in the first layer of LSTM; Will Through the tanh function, a value between -1 and 1 is obtained, and then Multiply element by element to get the hidden state at the current time point: ; in, is the first layer of LSTM The hidden state of the time step contains not only the input information of the current time point, but also the information of the previous time point. It will be used as one of the inputs of the next time step and also as the output of the first layer LSTM to the next layer LSTM. For example, if The value of a dimension is 0.7. If the value of this dimension is 0.8, then The value of this dimension is 0.7×0.8=0.56. Through the output gate, the model can control the information contained in the hidden state so that only the information related to the current prediction task is passed to the next layer or as output, further improving the model's ability to filter and utilize information.

[0043] Step 4.2: Enter the second layer of LSTM, whose calculation process is similar to the first layer of LSTM, but the weight matrix and bias vector are different to adapt to the new input data dimension and further extract features. The calculation formula of the second layer of LSTM is as follows: ; ; ; ; ; ; in, For the forgotten gate in the second layer LSTM Output of time steps; is the weight matrix of the forget gate in the second layer LSTM, with dimension (Because the input is the hidden state of the previous layer, the dimension is ); is the bias vector of the forget gate in the second layer LSTM, with dimension . is the first The hidden state of time steps. Here and It is also generated by a suitable initialization method and adjusted during the training process. Similar to the forget gate in the first layer of LSTM, Each element determines the state of the second layer cell The corresponding dimension information is forgotten, but it is based on the output of the previous layer and the hidden state of the previous time step of the current layer. is the first The control signal of the input gate in time steps; is the weight matrix of the input gate in the second layer of LSTM; is the bias vector of the input gate in the second layer of LSTM. These two parameters determine the degree of control over the flow of new information into the cell state of the second layer. Its initialization and update method is similar to that of the first layer, but it is targeted at the input data dimension and structural characteristics of the second layer. is the first Candidate cell states for time steps; It is the weight matrix used to calculate the candidate cell state in the second layer LSTM; is the bias vector used to calculate the candidate cell state in the second layer LSTM. The hidden state of time steps and the second LSTM layer The hidden state of time steps The candidate cell state is generated by concatenating the first layer and the second layer (the concatenation of the first layer and the second layer) with a value range between -1 and 1, which contains the potential update content after further processing of the information passed from the first layer. is the first Update cell state at each time step; is the first This updating process is similar to that of the first layer. Through the forget gate and input gate, the cell state of the second layer LSTM is updated, integrating part of the information of the previous time step and the newly generated candidate information of the current time step, so that the cell state of the second layer LSTM can capture more complex changes in time series features. These features are further abstracted based on the first layer LSTM after processing the original input data. is the first The control signal of the output gate in time steps; is the weight matrix of the output gate in the second layer of LSTM; is the bias vector of the output gate in the second layer of LSTM. It generates a control signal based on the current input and determines the output degree of the current cell state.

[0044] In the second LSTM layer The hidden state of time steps It further captures more complex time series features, which will serve as the input of the third layer of LSTM. Compared with the first layer of LSTM, the second layer of LSTM can mine deeper information from the features extracted in the first layer, such as capturing more complex interactions between different sensor data features and long-term dependency patterns in time series. This hierarchical feature extraction capability enables the model to better understand the complex laws in the movement and deformation data of the structure. As the number of layers increases, the ability to abstractly represent the data is also continuously enhanced.

[0045] Step 4.3, The calculation process of the first LSTM layer is similar to that of the second layer, except that the input becomes The output hidden state of the layer. The calculation formula of layer LSTM is as follows: ; ; ; ; ; ; in, For the Forgetting level in layer LSTM Output of time steps; It is The weight matrix of the forget gate in the layer LSTM has dimensions according to the In the LSTM layer The hidden state of time steps The dimensions and The hidden state dimension in the LSTM layer is determined, and the dimension is ; It is The bias vector of the forget gate in the LSTM layer, with dimension . Decided from the In the layer LSTM Time step cell state The calculation of which information is discarded depends on the output of the L-1th layer and the hidden state of the Lth layer in the previous time step. The Sigmoid activation function is used to map the result to between 0 and 1 to control the forgetting of information. For the In the layer LSTM The control signal of the input gate in time steps; It is The weight matrix of the input gate in the layer LSTM; It is The bias vector of the input gate in the LSTM layer. This control signal is used to control the flow of new information into the The degree of cell state in the layer is based on the current input (including the The output of the layer LSTM and the The hidden state of the previous time step in the layer LSTM is calculated through specific weights and biases and activated by the Sigmoid function. For the In the layer LSTM Candidate cell states for time steps; It is The weight matrix used to calculate the candidate cell state in the layer LSTM; It is The relevant bias vector used to calculate the candidate cell state in the LSTM layer. The candidate cell state is generated by the tanh activation function, and its value reflects the potential update information contained in the current input, providing new content for the update of the cell state. For the In the layer LSTM Update cell state at each time step; It is In the layer LSTM According to the same update logic as the first and second layers, the cell state is updated according to the output of the forget gate and the input gate, so that the first The cell state of the LSTM layer combines historical information and new candidate information to further mine deep features in the data. For the In the layer LSTM The control signal of the output gate in time steps; It is The weight matrix of the output gate in the layer LSTM; It is The bias vector of the output gate in the LSTM layer. The control signal is calculated based on the current input to determine the output part of the cell state. In the layer LSTM The hidden state of the time step The state, as a higher-level feature representation of the entire stacked LSTM layer, contains the complex time series features extracted from the original input data after multi-layer processing. is the total number of time steps, and the final output of the entire stacked LSTM layer is In the layer LSTM The hidden state of time steps .

[0046] Step 4.4: Enter the Transformer layer. The Transformer layer uses a multi-head attention mechanism and a feedforward neural network. The specific process is: Step 4.4.1: First, the final output of the stacked LSTM layer Perform linear transformation to obtain the query matrix , key matrix Sum Matrix .set up , , They are , , The weight matrix is: ; ; ; For example, if is a Matrix, assuming , , , , are all 64×64 matrices, then , , Both are 100×64 matrices.

[0047] Then, calculate the attention score: ; in, Score for attention; is the softmax function; yes Dimensions; is the transposed symbol. In order to obtain the feature representation of different subspaces, a multi-head attention mechanism is usually used. , , Split into Head (e.g. ), each head calculates the attention score and then concatenates them. , , For the The query matrix, key matrix, and value matrix of each head, then The attention scores of the heads are: ; in, For the Attention score of the individual; Finally, the output matrix of multi-head attention is for: ; in, For splicing operation; It is the weight matrix used to fuse multi-head information; For the For example, , , Split into 8 heads by column, each head has a dimension of 100×8, and calculate the attention score of each head to get 8 100×8 matrices. Assume is a 64×64 matrix, and after splicing, a 100×64 output matrix is ​​obtained .

[0048] Step 4.4.2: Feedforward neural network processing is performed, and the formula is: ; in, It is a feed-forward neural network; , are two different weight matrices for a feedforward neural network; , are two different bias vectors for a feedforward neural network. For example, if is a 100×64 matrix, is a 64×256 matrix, is a 256-dimensional vector, is a 256×64 matrix, is a 64-dimensional vector, then first calculate Get a 100×256 matrix, after the ReLU activation function After that, Multiply and add The final output is still of dimension 100×64.

[0049] The Transformer layer can further explore the global features and complex relationships in the data through the multi-head attention mechanism and feedforward neural network, making up for the shortcomings of LSTM in processing long sequences.

[0050] Step 4.5: Input to the fully connected layer to predict the movement and deformation of the structure. Output matrix of the multi-head attention of the Transformer layer As a high-level feature representation of the entire model. According to the prediction target (such as predicting the displacement and tilt angle of a certain point of the structure), add a fully connected layer (Dense layer) to The features are converted into specific prediction values. Suppose the weight matrix of the fully connected layer is , the dimension is (in is the dimension of the predicted target, for example, if only displacement is predicted, can be 1), the bias vector of the fully connected layer is , the dimension is . No. The predicted value of the time step The calculation formula is: ; When constructing a fully connected layer, the weight matrix The initialization of is also important. You can use a random initialization method and adjust it appropriately according to the specific activation function (if any) and the input and output dimensions. Initialize to a small value, such as a constant close to 0. The fully connected layer maps the complex features contained in the output of the Transformer layer to the space of the prediction target, realizing the conversion from feature representation to specific prediction value.

[0051] Step 5: Train and verify the structure movement and deformation prediction model to ensure that the model has high accuracy and reliability. Use the trained model to predict the structure movement and deformation. The specific process is as follows: Step 5.1, loss function selection. For the problem of structure movement and deformation prediction, the commonly used loss function is the mean square error (MSE) loss function, which measures the average value of the square error between the predicted value and the true value. Suppose the predicted value sequence is , the true value sequence is , then the calculation formula of the mean square error loss function is: ; in, is the mean square error loss function value; For the The true value of the time step; For the The predicted value of time steps; The model is trained by minimizing the MSE loss function so that the model's prediction results are as close as possible to the actual movement and deformation values ​​of the structure. MSE squares the error, which makes larger errors account for a larger proportion in the loss calculation, thus prompting the model to pay more attention to reducing larger errors.

[0052] In addition, the mean absolute error (MAE) loss function is used as a supplementary evaluation indicator, and its calculation formula is: ; in, is the mean absolute error loss function value; MAE is more robust to outliers because it does not perform square operations and will not over-amplify the impact of large errors. Combining with MSE can more comprehensively evaluate the performance of the model. During the model training process, observe the changing trends of MSE and MAE at the same time. When MSE decreases and MAE increases, it may mean that the model's ability to handle outliers decreases, and further data analysis or adjustment of the model structure is required.

[0053] Step 5.2: Use the Adaptive Moment Estimation (Adam) optimization algorithm to update the model parameters (including each LSTM layer, Transformer layer, and the weight matrix and bias vector of the output layer). The Adam optimization algorithm combines the idea of ​​momentum method and adaptive learning rate. During the training process, the Adam algorithm automatically adjusts the learning rate of each parameter based on the first-order moment estimation and second-order moment estimation of the gradient, so that the model can converge faster and more stably. Specifically, the Adam algorithm maintains two momentum vectors for each parameter: The first-order moment vector of the time step (similar to speed) and The second-order moment vector of time steps (Similar to acceleration). For the parameter (This includes all weight matrices and bias vectors), in the The gradient of the time step is , the learning rate is , and are two different hyperparameters used to control momentum (usually close to 1, such as , ), is a non-zero constant (to prevent division by zero, such as ), the update process is as follows: First, calculate the update of the first-order moment vector: ; in, For the The first-order moment vector of the time step; this step is the exponentially weighted moving average of the gradient, and the new first-order moment vector is calculated based on the current gradient and the first-order moment vector of the previous time point, where Determines the degree of retention of past gradient information.

[0054] Then, calculate the update of the second-order moment vector: ; in, For the The second-order moment vector of time steps; this is the exponentially weighted moving average of the square of the gradient, which is used to measure the change in the gradient. It also determines the degree of retention of past information.

[0055] Next, the first-order moment vector and the second-order moment vector are corrected for deviation: ; ; in, , Respectively The first-order moment vector deviation and the second-order moment vector deviation at each time step are calculated; since the setting of the initial value when calculating the moving average of the first-order moment vector and the second-order moment vector will cause deviations in the early estimates, the deviation correction step can reduce the impact of this deviation, especially in the early stages of training.

[0056] Finally, update the parameters: ; in, , Respectively time step, The parameters of the time step are updated according to the modified first-order moment vector and second-order moment vector and the learning rate, so that the parameters move in the direction of decreasing the loss function.

[0057] Step 5.3: Hyperparameter adjustment. The specific process is: Step 5.3.1. For hyperparameter tuning, use grid search to form different hyperparameter combinations. Adjust the learning rate , the number of hidden units in the LSTM layer And the number of heads of the Transformer layer , set a series of Value (such as 0.001, 0.005, 0.01, etc.), Values ​​(such as 32, 64, 128, etc.) and values ​​(such as 4, 8, 16, etc.), and then evaluate the performance of the model under different combinations on the validation set. The evaluation indicators are MSE and MAE, and the hyperparameter combination with the best performance is selected.

[0058] Step 5.3.2. To alleviate the overfitting phenomenon, add a Dropout layer after each LSTM layer. Set a suitable probability threshold (such as 0.2-0.5). During the training process, these Dropout layers will randomly set the output of some neurons to 0, reduce the mutual dependence between neurons, and improve the generalization ability of the model. At the same time, consider adding a Dropout layer to some parts of the Transformer layer (such as after the multi-head attention mechanism or after the feedforward neural network). Its probability threshold can be the same as that of the Dropout layer after the LSTM layer or set separately. Determine whether to add a Dropout layer to some parts of the Transformer layer by evaluating on the validation set. When the evaluation result exceeds the probability threshold, add it, otherwise, do not add it.

[0059] Step 5.3.3, determine the number of training rounds. The number of training rounds is an important hyperparameter. If the number of training rounds is too small, the model may not fully learn the patterns in the data, resulting in underfitting; if the number of training rounds is too large, it may lead to overfitting. Determine the appropriate number of training rounds by observing the loss curve on the validation set. Generally, the number of training rounds when the loss of the validation set reaches the minimum value is selected as the final number of training rounds. For example, during the training process, record the loss value of the validation set after each epoch, draw the loss curve, and stop training when the loss curve begins to rise or no longer decreases within multiple epochs, and select the number of training rounds corresponding to the epoch with the smallest loss before.

[0060] Step 5.3.4, setting up the learning rate scheduler. The learning rate scheduler uses a learning rate scheduling strategy to adjust the changes in the learning rate during the training process. Gradually reducing the learning rate as the number of training rounds increases can make the model converge more stably in the later stages of training. There are two types of learning rate scheduling strategies, namely the step decay strategy and the exponential decay strategy. For the step decay strategy, it is necessary to determine the decay step size (such as decaying once every 10 epochs) and the decay factor (such as 0.9). At each decay step, multiply the learning rate by the decay factor. The initial learning rate is ,go through After decay steps, the learning rate becomes For the exponential decay strategy, it is necessary to determine the decay rate (such as 0.95). The learning rate decreases exponentially with the number of training rounds, that is, , Indicates the current number of training rounds. By evaluating the performance of the model under different learning rate scheduling strategies and parameter settings on the validation set, the most appropriate learning rate scheduler parameters are selected to further optimize the model training process and improve the performance and generalization ability of the model.

[0061] Step 6: Comprehensively evaluate safety and stability; the specific process is: Based on the predicted value of structure movement and deformation obtained by the above method, comprehensively evaluate the safety and stability of the structure. Compare the model prediction value with the safety warning value determined in step 1. When the predicted value does not exceed the warning value, it indicates that the structure is currently in a relatively safe state; when the predicted value exceeds the warning value, it means that there is a safety risk to the structure and corresponding technical measures need to be taken, such as temporary reinforcement of the structure, adjustment of the mining plan, etc.

[0062] Step 7: Modify the prediction model parameters and continue to evaluate until the structure is safely mined. This includes the following steps: Step 7.1: During the mining process, continuously obtain the next stage of movement and deformation observation data of the surface and structures in the mined area. These new data include the surface deformation when closer to the structure during the mining process and the actual response of the structure itself; Step 7.2: Substitute the newly acquired observation data into the prediction model and modify the parameters of the prediction model through the Adam optimization algorithm so that the model can better adapt to the new data characteristics and actual deformation conditions; Step 7.3: Use the revised prediction model to predict the movement and deformation of the structure again, and re-evaluate its safety and stability; Step 7.4: Repeat step 7.3, continuously adjusting the model and evaluating the safety status until the working face is safely pushed through the structure, ensuring that the safety impact on the structure during the entire mining process is within a controllable range.

[0063] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for digital intelligent observation of rock movement on a working face and intelligent prediction of safety of passing structures, characterized in that: The steps include: Step 1: Determine the safety warning value of structure movement and deformation; Step 2: Collect movement deformation data and perform preprocessing; Step 3: Use finite element analysis to analyze the dynamic change law of surface movement and deformation; Step 4: Establish a prediction model for structure movement and deformation; Step 5: Train and verify the structure movement and deformation prediction model, and use the trained model to predict the structure movement and deformation; Step 6: Comprehensively evaluate safety and stability; Step 7: Modify the prediction model parameters and continue to evaluate until the structure is safely mined.

2. According to claim 1, the method for digital intelligent observation of rock movement on the working face and intelligent prediction of safety of passing structures is characterized by: The specific process of step 1 is as follows: Step 1.1: Comprehensively consider the structural characteristics of the structure, the geological conditions and the surrounding environment to obtain the key factors affecting the stability and safety of the structure; Key factors include the type of foundation of the structure, the characteristics of the building materials, the overall stiffness of the structure, the geological structure characteristics, and other surrounding engineering activities; Step 1.2: Identify the key indicators that affect the movement and deformation state of the structure, including displacement, inclination, curvature, and strain; determine the threshold of each key indicator in combination with the design specifications, usage requirements, and relevant industry standards of the structure. This threshold is the safety warning value for the movement and deformation of the structure.

3. According to claim 1, the method for digital intelligent observation of rock movement on the working face and intelligent prediction of safety of passing structures is characterized by: The specific process of step 2 is: Step 2.1, install various types of sensors at key structural parts of the structure, and collect daily surface movement and deformation data of the mined area in front of the structure through the sensors; the sensors include displacement sensors, inclination sensors, curvature sensors, and strain sensors; Step 2.2, perform outlier detection and processing; The process of outlier detection is as follows: define the displacement data sequence collected by the displacement sensor as , calculate the mean of the displacement data series and standard deviation , the formula is as follows: ； ； in, is the total number of displacement data points; For the displacement data points; For the displacement data points; for The displacement data points of the interval are marked as outliers; is the standard deviation The proceeds from the prescription; The process of outlier processing is as follows: For data with a stable trend, linear interpolation method is used to deal with outliers. Assume that at the first time point and the third time point The displacement data points are and , and the second time point If the displacement data points are abnormal or missing, Displacement data points It is calculated by the linear interpolation formula, which is: ； For data with unstable data change trends, the weighted linear interpolation method is used to process outliers. Middle Displacement data points The displacement data points that are identified as outliers and removed are called missing value points. Interpolate the position; suppose the time series corresponding to the displacement data series is , For the A point in time, is the total number of time points, one displacement data point corresponds to one time point, and the total number of time points corresponds to the total number of displacement data points; the weighted linear interpolation formula is: ； ； in, for The interpolation result of the location; for The weight of For the Time point To Time point time distance; For the Time point To Time point time distance; For the Time point The displacement data points; Step 2.3: Normalize the data. The formula is: ； in, is the normalized displacement data point; is the displacement data point; , are the minimum and maximum values ​​of the displacement data in the entire data set, respectively; For inclination, curvature, and strain, outlier detection and processing and data normalization are performed in the same way as displacement data.

4. According to claim 1, the method for digital intelligent observation of rock movement on the working face and intelligent prediction of safety of passing structures is characterized by: The specific process of step 3 is as follows: Step 3.1, construct a geometric model, use the unit division strategy to discretize the units, and obtain a discretized model; the specific process is: first, build a two-dimensional or three-dimensional geometric model based on geological survey data and the design blueprint of the structure; then, use the unit division strategy to discretize the two-dimensional or three-dimensional geometric model to obtain a discretized model; the process of the unit division strategy is: let the total number of units be The total number of nodes is ; For each unit, the numbers of its three nodes are , , , No. The displacement vector of any node in a unit is , is the horizontal displacement, is the vertical displacement, , Respectively Horizontal displacement and vertical displacement of any node in a unit; ,node ,node The displacement vectors are , , ; , Node Horizontal displacement and vertical displacement; , Node Horizontal displacement and vertical displacement; , Node Horizontal displacement and vertical displacement; is the transposed sign; displacement vectors are connected through shape functions; For linear triangular elements, the nodes The shape function is expressed as: ； ； ； ； in, For Node The shape function of is the area of ​​the triangular unit; For computing nodes The shape function is the intermediate parameter constructed; For computing nodes The correlation coefficient when the shape function is ; For computing nodes The correlation coefficient when the shape function is ; is the abscissa representing the horizontal position in the plane rectangular coordinate system; It is the ordinate representing the vertical position in the plane rectangular coordinate system; The nodes in the triangle element The horizontal axis of The nodes in the triangle element The horizontal axis of The nodes in the triangle element The vertical coordinate of The nodes in the triangle element The vertical coordinate of node Shape function ,node Shape function With Node The shape functions of are expressed in the same way; Final The displacement vector of any node in a unit is expressed as: ； At the same time, the strain-displacement relationship is: ； in, For the The strain vector of any node in a unit; is the strain-displacement matrix, which is given by: ； Step 3.2, establish the constitutive equation, equilibrium equation and overall stiffness matrix; Step 3.3: Solve the overall stiffness matrix and analyze the results to obtain the dynamic change law of surface movement and deformation.

5. According to claim 4, the method for digital intelligent observation of rock movement on the working face and intelligent prediction of safety of passing structures is characterized by: In step 3.2, The constitutive equation of each unit is: ； in, For the The stress tensor of any node in an element; is the elasticity matrix, expressed as: ； in, is the elastic modulus; is Poisson's ratio; The specific expression of the equilibrium equation is: ； in, It is The volume of a unit; For virtual strain; is the strain vector; is the stress tensor; is the volume; It is The surface of each unit; is the surface force vector; For the surface; is the displacement vector; Is a node The force vector of The overall stiffness equation is: ； in, is the overall stiffness matrix; is the node displacement vector; is the force vector of the node; No. The stiffness matrix of the element The calculation formula is: ； Overall stiffness matrix The assembly process is to superimpose the stiffness matrix of each unit to the corresponding position of the overall stiffness matrix according to the corresponding relationship between the unit node and the overall node; In step 3.3, the overall stiffness equation is solved by direct method or iterative method. Get the node displacement vector ; After obtaining the node displacement vector, the unit strain vector and stress tensor are further calculated; the unit strain vector is calculated based on the strain-displacement relationship, and the unit stress tensor is calculated by the constitutive equation.

6. The method for digital intelligent observation of rock movement on the working face and intelligent prediction of safety of passing structures according to claim 5 is characterized in that: In step 4, the input variables of the structure movement deformation prediction model are the relevant parameters in the surface movement deformation observation data, and the output is the movement deformation value of the structure; the relevant parameters include displacement, inclination, curvature, and strain; The structure movement deformation prediction model includes multiple stacked LSTM layers, a Transformer layer, and a fully connected layer; LSTM is a long short-term memory network; Transformer is a sequence model based on the attention mechanism; The working process of the structure movement deformation prediction model is as follows: Step 4.1: Input the relevant parameters in the surface movement and deformation observation data into the first layer LSTM; the working process of the first layer LSTM is: Step 4.1.1: Preprocess the input related parameters. After preprocessing, the input related parameters are a vector at each time point. Define the first The input data for each time step is ,in is the dimension of the input data; For the time step, The Data of dimensions; Step 4.1.2: Calculate the forget gate. The formula is: ； in, is the forgotten gate in the first layer of LSTM Output of time steps; is the Sigmoid activation function; is the weight matrix of the forget gate in the first layer of LSTM; is the first layer of LSTM The hidden state of time steps; is the bias vector of the forget gate in the first layer of LSTM; Step 4.1.3: Perform input gate calculation to generate input gate control signals and candidate cell states. The formula is: ； ； in, is the first layer of LSTM The control signal of the input gate in time steps; is the weight matrix of the input gate in the first layer of LSTM; is the bias vector of the input gate in the first layer of LSTM; is the first layer of LSTM Candidate cell states for time steps; is the tanh activation function; is the weight matrix used to calculate the candidate cell state in the first layer of LSTM; is the bias vector used to calculate the candidate cell state in the first layer of LSTM; Step 4.1.4, update the cell state according to the output of the forget gate and the input gate: ； in, is the first layer of LSTM Update cell state at each time step; Represents an element-by-element multiplication operation; is the first layer of LSTM The cell state at each time step; Step 4.1.5: Perform output gate calculation to generate the output gate control signal and output hidden state. The formula is: ； ； in, is the first layer of LSTM The control signal of the output gate in time steps; is the weight matrix of the output gate in the first layer of LSTM; is the bias vector of the output gate in the first layer of LSTM; is the first layer of LSTM The hidden state of time steps; Step 4.2: Enter the second layer LSTM. The calculation formula of the second layer LSTM is as follows: ； ； ； ； ； ； in, For the forgotten gate in the second layer LSTM Output of time steps; is the weight matrix of the forget gate in the second layer of LSTM; is the bias vector of the forget gate in the second layer of LSTM; is the first The hidden state of time steps; is the first The control signal of the input gate in time steps; is the weight matrix of the input gate in the second layer of LSTM; is the bias vector of the input gate in the second layer of LSTM; is the first Candidate cell states for time steps; It is the weight matrix used to calculate the candidate cell state in the second layer LSTM; is the bias vector used to calculate the candidate cell state in the second layer of LSTM; is the first Update cell state at each time step; is the first The cell state at each time step; is the first The control signal of the output gate in time steps; is the weight matrix of the output gate in the second layer of LSTM; is the bias vector of the output gate in the second layer of LSTM; In the second LSTM layer The hidden state of time steps Will be used as the input of the third layer LSTM; Step 4.3, The input of the LSTM layer is The output hidden state of the layer; The calculation formula of layer LSTM is as follows: ； ； ； ； ； ； in, For the Forgetting level in layer LSTM Output of time steps; It is The weight matrix of the forget gate in the layer LSTM; It is The bias vector of the forget gate in the layer LSTM; For the In the layer LSTM The control signal of the input gate in time steps; It is The weight matrix of the input gate in the layer LSTM; It is Bias vector of the input gate in layer LSTM; For the In the LSTM layer Candidate cell states for time steps; It is The weight matrix used to calculate the candidate cell state in the layer LSTM; It is The relevant bias vector used to calculate the candidate cell state in the layer LSTM; For the In the LSTM layer Update cell state at each time step; It is In the LSTM layer The cell state at each time step; For the In the LSTM layer The control signal of the output gate in time steps; It is The weight matrix of the output gate in the layer LSTM; It is The bias vector of the output gate in the layer LSTM; For the In the LSTM layer The hidden state of time steps; is the total number of time steps, and the final output of the entire stacked LSTM layer is In the LSTM layer The hidden state of time steps ; Step 4.4: Input Transformer layer and get the output matrix ; Step 4.5: Input the fully connected layer to predict the movement and deformation of the structure.

7. The method for digital intelligent observation of rock movement on the working face and intelligent prediction of safety of passing structures according to claim 6 is characterized in that: In step 4, the Transformer layer uses a multi-head attention mechanism and a feedforward neural network; the specific process is: Step 4.4.1: First, the final output of the stacked LSTM layer Perform linear transformation to obtain the query matrix , key matrix Sum Matrix ;set up , , They are , , The weight matrix is: ； ； ； Then, suppose , , For the The query matrix, key matrix, and value matrix of each head, then The attention scores of the heads are: ； in, For the Attention score of the individual; is the softmax function; yes Dimensions; is the transpose symbol; is the number of heads; Finally, the output matrix of multi-head attention is for: ； in, For splicing operation; It is the weight matrix used to fuse multi-head information; For the Attention score of the individual; Step 4.4.2: Feedforward neural network processing is performed, and the formula is: ； in, It is a feed-forward neural network; , are two different weight matrices for a feedforward neural network; , are two different bias vectors for a feedforward neural network; In step 4.5, the fully connected layer will The features are converted into specific prediction values; The predicted value of the time step The calculation formula is: ； in, is the weight matrix of the fully connected layer; is the bias vector of the fully connected layer.

8. The method for digital intelligent observation of rock movement on the working face and intelligent prediction of safety of passing structures according to claim 7 is characterized in that: The specific process of step 5 is as follows: Step 5.1: Train the model by minimizing the mean square error loss function; let the predicted value sequence be , the true value sequence is , then the calculation formula of the mean square error loss function is: ； in, is the mean square error loss function value; For the The true value of the time step; For the The predicted value of time steps; At the same time, the mean absolute error loss function is used as a supplementary evaluation indicator, and the calculation formula is: ； in, is the mean absolute error loss function value; Step 5.2: Use the adaptive moment estimation optimization algorithm to update the parameters of the model; the parameters include the weight matrix and bias vector of each LSTM layer, Transformer layer, and output layer; the update process is as follows: First, calculate the update of the first-order moment vector: ； in, For the The first-order moment vector of time steps; For the The first-order moment vector of time steps; For the The gradient of time steps; Then, calculate the update of the second-order moment vector: ； in, For the The second-order moment vector of time steps; For the The second-order moment vector of time steps; and are two different hyperparameters used to control momentum; Next, the first-order moment vector and the second-order moment vector are corrected for deviation: ； ； in, , Respectively First-order moment vector deviation and second-order moment vector deviation for each time step; Finally, update the parameters: ； in, , Respectively time step, Parameters for time steps; is the learning rate; is a non-zero constant; Step 5.3: Hyperparameters including learning rate , the number of hidden units in the LSTM layer And the number of heads of the Transformer layer , set a series of value, Value and The value is evaluated on the validation set. The mean square error and mean absolute error are selected as the evaluation indicators. The best hyperparameter combination is selected. The number of training rounds when the loss of the validation set reaches the minimum value is selected as the final number of training rounds. The learning rate scheduling strategy is used to adjust the learning rate.

9. The method for digital intelligent observation of rock movement on working face and intelligent prediction of safety of passing structures according to claim 1 is characterized in that: The specific process of step 6 is: compare the model prediction value with the safety warning value determined in step 1. When the prediction value does not exceed the warning value, it indicates that the structure is currently in a safe state; when the prediction value exceeds the warning value, it indicates that there is a safety risk in the structure.

10. The method for digital intelligent observation of rock movement on working face and intelligent prediction of safety of passing structures according to claim 9 is characterized in that: The specific process of step 7 is as follows: Step 7.1, during the mining process of the working face, continuously obtain the next stage of movement and deformation observation data of the surface and structures in the mined area; Step 7.2: Substitute the newly acquired observation data into the prediction model and modify the parameters of the prediction model through the Adam optimization algorithm so that the model can better adapt to the new data characteristics and actual deformation conditions; Step 7.3: Use the revised prediction model to predict the movement and deformation of the structure again, and re-evaluate its safety and stability; Step 7.4: Repeat step 7.3, continuously adjusting the model and evaluating the safety status until the working face is safely pushed through the structure, ensuring that the safety impact on the structure during the entire mining process is within a controllable range.

Images