Digital intelligent prediction method for surface subsidence of building

Abstract

The invention discloses a surface subsidence digital intelligent prediction method for a building, which belongs to the technical field of mining industry and information, and comprises the following steps: step 1, collecting multi-source data, and carrying out multi-source data fusion based on a Kalman filtering method; 2, processing the abnormal data based on the random forest; step 3, feature engineering and variable selection are carried out based on the graph neural network; and step 4, predicting surface subsidence based on the diffusion model and the quantum heuristic neural network. According to the method, the advantages of the diffusion model in the aspect of data generation and the powerful feature learning ability of the quantum heuristic neural network are fully utilized to predict the surface subsidence of the building, so that reliable basis and technical support are provided for safety assessment, engineering planning, disaster prevention and the like of the building; the precision and reliability of surface subsidence prediction are effectively improved, and intelligent development of related fields is promoted.

Description

Technical Field

[0001] The present invention belongs to the fields of mining and information technology, and particularly relates to a digital and intelligent prediction method for surface subsidence of buildings. Background Art

[0002] With the rapid development of urban construction and the continuous exploitation of underground resources, the problem of surface subsidence around buildings has become increasingly prominent. This surface subsidence not only poses a serious threat to the structural stability of buildings, but also may trigger a series of safety hazards. In order to reduce the adverse effects of surface subsidence, filling technology is usually used to control surface subsidence. However, the traditional prediction and evaluation methods have obvious deficiencies in accuracy, timeliness and intelligence, and it is difficult to meet the requirements of complex engineering environments. Therefore, it has become an urgent task to develop a high-precision and intelligent surface subsidence prediction method. Summary of the Invention

[0003] To solve the above problems, the present invention proposes a digital and intelligent prediction method for surface subsidence of buildings, which integrates a diffusion model and a quantum-inspired neural network, aiming to achieve accurate prediction of surface subsidence and comprehensive and objective evaluation of filling effects by integrating cutting-edge technologies and scientific algorithms.

[0004] The technical solution of the present invention is as follows: A digital and intelligent prediction method for surface subsidence of buildings, comprising the following steps: Step 1, collect multi-source data and perform multi-source data fusion based on the Kalman filtering method; Step 2, process abnormal data based on random forest; Step 3, perform feature engineering and variable selection based on a graph neural network; Step 4, predict surface subsidence based on a diffusion model and a quantum-inspired neural network.

[0005] Further, the specific process of Step 1 is as follows: Step 1.1, collect multi-source data, including surface settlement monitoring data, geological exploration data, underground mining activity information, and detailed information of the building itself; the surface settlement monitoring data includes the displacements and displacement change rates of a certain point on the surface in the x, y, and z directions; the geological exploration data includes the stratum structure and physical and mechanical parameters of rock and soil masses; the underground mining activity information includes the mining progress, mining method, and mining depth; the detailed information of the building itself includes the structure type, foundation form, and building materials; Step 1.2, use the Kalman filtering method to integrate the multi-source data and construct a unified data format and coordinate system; the specific process is as follows: Step 1.2.1, model the surface settlement system to determine the state equation and the observation equation; The state equation is as follows: ; Wherein, is the previous moment, is the current moment; , are the state vectors of the ground settlement system at the current moment and the previous moment respectively; is the state transition matrix; is the control input matrix; is the control input vector at the previous moment; is the process noise vector at the previous moment; The observation equation is as follows: ; Wherein, is the observation vector of all sensors at the current moment; is the observation matrix; is the observation noise vector at the current moment; Step 1.2.2. Initialize the parameters; specifically: set the state estimate at the initial moment to be , and set the initial estimated error covariance matrix to be ; Step 1.2.3. For each moment, perform the Kalman filter iteration process, and the iteration process includes a prediction step and an update step; Step 1.2.4. Continuously iterate to complete data integration.

[0006] Furthermore, the specific process of the said Step 1.2.3 is as follows: Step 1.2.3.1. The specific process of the prediction step is as follows: First, according to the state equation, calculate the prior state estimate at the current moment, and the formula is: ; Wherein, is the state estimate at the previous moment; Then, calculate the prior estimated error covariance matrix at the current moment, and the formula is: ; Wherein, is the error covariance matrix at the previous moment; is the transpose symbol; is the process noise covariance matrix; Step 1.2.3.2. The specific process of the update step is as follows: First, calculate the Kalman gain at the current moment , and the formula is: ; where, is the observation noise covariance matrix ; Then, combine the observation vector at the current moment to update the prior state estimate and obtain the optimal state estimate at the current moment , and the formula is: ; Finally, update the estimated error covariance matrix at the current moment , and the formula is: ; where, is the identity matrix; In the step 1.2.4, by continuously repeating the prediction step and the update step, iterative calculation is performed for each moment . As time progresses, the Kalman filter continuously fuses the observation information of different sensors until , the optimal state estimate at the final moment is obtained, which is the fused data result; is the total number of moments; Combine with the intermediate moment state estimate values, the original observation data of each sensor, the process and observation noise statistical information, and the relevant information of the state transition matrix and the observation matrix during the Kalman filtering process to obtain the fused data set .

[0007] Furthermore, the specific process of the step 2 is as follows: Step 2.1: Perform outlier detection on the fused data set based on the box plot method to construct a training data set; Suppose the fused data set contains samples, and each sample has features; traverse the fused data set , and mark the samples with outliers in all features; suppose there are samples with outliers, ; for the th sample containing outliers, , mark the outlier position of the th feature as empty; Construct the input feature vector : ; Among them, is the -th feature of the -th outlier-containing sample; At the same time, for these outlier-containing samples, collect the true values corresponding to their -th feature under the known non-missing condition as the target variable , and construct a training dataset : ; Step 2.2: Train a random forest model based on the training dataset; Step 2.3: Perform outlier prediction.

[0008] Furthermore, the specific process of the said Step 2.2 is as follows: Step 2.2.1: For the training dataset , use the bootstrap sampling method to generate the training sub-dataset for each decision tree; the process of the bootstrap sampling method is: for each decision tree to be constructed, randomly draw samples with replacement from to obtain the training sub-dataset for the -th decision tree; Step 2.2.2: For the -th decision tree, randomly select features from all input features for node splitting. Let the feature subset randomly selected by the -th decision tree be: ; Among them, is the index of the -th feature used for node splitting by the -th decision tree. This index corresponds to a certain feature in the dataset after removing the -th feature column with outliers; Step 2.2.3: Construct a decision tree based on the training sub-dataset and the randomly selected feature subset; If the predicted outliers are discrete categorical data, the Gini coefficient is used to determine the best splitting feature and splitting point during the node splitting process; at a certain node , let the class set be , is the number of classes, is the -th class; the sample set of node is , the total number of samples of node is , belonging to node the th category is , then the Gini coefficient calculation formula of node is: ; where is the Gini coefficient; When considering feature splitting, assume that based on feature at a certain splitting point the sample set of node is divided into two parts and , and the number of samples is and , respectively. Then the Gini index after splitting is: ; where is the Gini index after splitting; Traverse all randomly selected features and the corresponding splitting points, and select the feature and splitting point that make the smallest as the best splitting method for this node. Continuously repeat this process to split the decision tree nodes until the maximum depth is reached or the number of node samples is less than the pre-set sample number threshold; If the predicted outlier is a continuous numerical value, select the mean squared error to determine the best splitting feature and splitting point; at node , for a certain splitting point of feature , the samples are divided into two parts and , and the means of the corresponding predicted target variables are and , respectively. Then the mean squared error after splitting is: ; where is the mean squared error after splitting; is the input feature vector; is the target variable; Select the feature and splitting point that make the smallest for splitting, and continuously build the decision tree until the maximum depth is reached or the number of node samples is less than the pre-set sample number threshold; Through the above process, decision trees are constructed to form a random forest model ; is the th decision tree; The specific process of step 2.3 is as follows: First, for the th sample containing outliers, its corresponding input feature vector is sequentially input into each decision tree in the trained random forest model for prediction; after the input feature vector reaches the leaf node, each decision tree takes the mean of the target variables corresponding to all samples in the leaf node as the predicted value of the sample; Then, by synthesizing the predicted values of all decision trees for the sample, the estimated value of the outlier is determined, and the formula is: ; where, is the estimated value of the outlier of the th sample; is the predicted value of the th decision tree for the th sample.

[0009] Furthermore, the specific process of step 3 is as follows: Step 3.1, construct a graph neural network; the node set in the graph neural network contains various types of nodes, including monitoring point nodes , mining area nodes , and building nodes ; each monitoring point node contains its historical surface subsidence data sequence and geographical location information; the mining area node contains mining activity-related parameters and its geographical range information; the building node contains the structural characteristics, weight information, and location information of the building; the edges in the graph neural network include spatial relationship edges and time relationship edges; The construction process of the spatial relationship edges is as follows: For monitoring point nodes, if they are geographically adjacent, an undirected edge is established; the weight of the edge between monitoring point nodes is: ; where, is the weight of the edge between the th and the th monitoring point nodes; is the distance; , are respectively the th and the th monitoring point nodes; is a non-zero positive number; A directed edge is established between the mining area node and the monitoring point node within its mining influence range; the direction of the edge is from the mining area to the monitoring point, indicating the influence relationship of mining activities on the surface subsidence of the monitoring point; the weight of the edge between the mining area node and the monitoring point node is: ; Among them, is the weight of the edge between the th mining area node and the th monitoring point node; is the coefficient related to the mining characteristics of the th mining area node; is the mining intensity index of the th mining area node at the current time ; is the th mining area node; An edge is established between the building node and the adjacent monitoring point node, and the weight of the edge between the building node and the monitoring point node is: ; Among them, is the weight of the edge between the th building node and the th monitoring point node , and are the weight coefficients determined according to the actual situation, used to adjust the influence degree of the building weight and foundation stiffness on the edge weight; represents the weight of the building, represents the foundation stiffness of the building, represents the distance between the building and the monitoring point; The construction process of the time relationship edge is: establish a directed time edge between different time steps of the same monitoring point node; the weight of the edge between different time steps is set to 1; Step 3.2, construct a spatio-temporal graph convolutional network; Step 3.3, construct a loss function and train the spatio-temporal graph convolutional network; Step 3.4, perform feature selection.

[0010] Furthermore, in the step 3.2, in the spatio-temporal graph convolutional network, it includes several layers, and the connection between layers is realized by using the output of the previous layer as the input of the next layer; the spatio-temporal graph convolutional network includes a spatial convolution part and a time convolution part, and the specific working process is: Step 3.2.1, for the node at the th layer, first perform spatial convolution, and the formula is: ; Among them, is the spatial feature of node at the layer time step ; is the spatial feature of neighbor node of node at the layer time step ; is the spatial feature of node at the layer time step ; , are respectively the spatial neighbor node sets of node , neighbor node ; is the temporal neighbor node set of node ; is the temporal neighbor node set of node in a time step identifier; is the spatial weight matrix of the layer; is the temporal weight matrix of the layer; is the activation function; Let represent the spatial feature of node at time step , is the spatial feature of neighbor node of node at time step ; Step 3.2.2, then perform temporal convolution, and the formula is: ; Among them, represents the temporal feature of node at the layer time step ; is the weight of the temporal edge; is the temporal convolution weight matrix of the layer; is the temporal feature of node at the layer time step ; Step 3.2.3, finally fuse the results of spatial convolution and temporal convolution to obtain the fused feature of the node at the layer: ; Among them, is the node at the -th layer time step of the fused feature; is a fusion coefficient.

[0011] Furthermore, in step 3.3, a fully connected layer is connected to the output features of the spatio-temporal graph convolutional network to obtain the prediction result; for the training data containing graph samples, the mean square error is used as the loss function : ; Among them, is the true ground subsidence value of the -th graph sample; is the predicted ground subsidence value of the -th graph sample; The backpropagation algorithm is used to update the weights of each layer in the spatio-temporal graph convolutional network to minimize the loss function; the gradient descent method is used for updating: ; Among them, is the updated weight of the -th layer; is the weight of the -th layer; is the learning rate; In step 3.4, for the trained spatio-temporal graph convolutional network, the output by the last layer is the feature extracted related to the ground subsidence, is the node at the -th layer time step of the fused feature, is the total number of layers; feature selection is performed by calculating the correlation between these features and the predicted ground subsidence value; the Pearson correlation coefficient is used to measure the correlation between each feature and the predicted ground subsidence value. For the feature dimension of obtained after the spatio-temporal graph convolutional network of layers, for the -th feature and the predicted ground subsidence value , the calculation formula of the Pearson correlation coefficient is: ; Among them, is the -th feature with the feature dimension of and the predicted ground subsidence value and the predicted ground subsidence value Pearson correlation coefficient; After being calculated by the spatio-temporal graph convolutional network with layers, in the th graph sample, the eigenvalue of the th feature with a feature dimension of at time step ; Is the average value of the eigenvalues of the th feature with a feature dimension of after being calculated by the spatio-temporal graph convolutional network with layers in a batch of training data containing graph samples, at time step ; Is the average value of the predicted ground subsidence values in a batch of training data containing graph samples; Calculate the Pearson correlation coefficient between each feature and the predicted ground subsidence value, sort the features in descending order of the Pearson correlation coefficient, and select the first features as feature variables.

[0012] Furthermore, the specific process of step 4 is as follows: Step 4.1: Generate starting samples based on the forward diffusion process; the specific process is as follows: Step 4.1.1: Determine the number of forward diffusion steps and the noise parameter; the number of forward diffusion steps is: ; where represents rounding up; is the basic number of forward diffusion steps; and are different adjustment coefficients; is the prediction time range; is the data uncertainty measure; The noise scheduling parameter is: ; where is the time step in the forward diffusion process; is the set minimum value of the noise scheduling parameter; Step 4.1.2: The forward diffusion process is expressed as: ; where , are respectively the time steps , The noise-added sample vector generated at is the noise scheduling parameter at time step ; is the noise vector at time step ; Repeat the forward diffusion process times to obtain the starting sample ; Step 4.2. Generate a sequence of feature variables using the reverse diffusion process based on a quantum-inspired neural network; the specific process is as follows: Step 4.2.1. Initialize the parameters of the quantum-inspired neural network; First, use as the input feature vector for qubit representation; the initialization formula for the qubit state is as follows: ; ; ; where is the angular parameter used to initialize the state of the th qubit; is the loop index variable; , represents floor division, reflects a grouping or partitioning parameter when mapping the -dimensional to qubits; is the th element of ; , are the states of the th and th qubits at time step respectively; is the qubit state at time step ; represents the tensor product of quantum states; Then, construct a quantum-inspired gate operation layer based on quantum rotation gates and quantum controlled-NOT gates; The action formula of the quantum rotation gate on a single qubit is: ; where is the quantum rotation gate; is the rotation angle; The matrix representation of the quantum controlled-NOT gate is: ; where is the quantum controlled-NOT gate for the th qubit to control the state of the th qubit; The matrix representation of the first quantum heuristic gate operation layer is: ; where is the rotation angle of the first qubit to control the state of the th qubit; There are layers in the quantum heuristic gate operation layer, and other layers are constructed according to the process; the th quantum heuristic gate operation layer contains different combinations of quantum gate operations, and the parameters of these gates are determined during training and directly loaded during prediction; Step 4.2.2, reverse diffusion iteration; Starting from , first input into the quantum heuristic neural network for noise estimation; At the qubit representation layer, convert to the quantum state and then perform feature transformation through the quantum heuristic gate operation layer; For , , where , that is, the initial qubit state passes through the first quantum heuristic gate operation layer to obtain the first qubit state ; for , ; and so on until , ; , are the th and th qubit states respectively; is the th quantum heuristic gate operation layer; the sequence number of the qubit state corresponds to the sequence number of the quantum gate operation layer; After passing through quantum heuristic gate operation layers, use the quantum measurement method to perform quantum state conversion on the qubit state of the last layer to obtain the classical probability distribution ; Convert to the estimated noise vector with the same dimension as the noise vector; Let the weight matrix of the fully connected layer be , and the bias vector be , then: ; where is the -th component of the estimated noise vector ; is the weight of the -th element pair of with respect to the -th component of ; is the -th element of ; is the -th element in ; Then, update the sample according to the reverse diffusion formula to obtain the predicted sample vector: ; where , are the predicted sample vectors generated at time steps , respectively during the reverse diffusion process; is the noise scheduling parameter at time step ; is the estimated noise vector at time step ; Repeat the above steps, gradually decreasing from to to obtain the sequence of predicted surface subsidence related characteristic variables ; is the predicted sample vector generated at time step 0 during the reverse diffusion process; Step 4.3, Extract the predicted value of the surface subsidence amount; Step 4.4, Train the surface subsidence prediction model; Step 4.5, Input the real-time collected surface settlement monitoring data, geological exploration data, underground mining activity information, and detailed information of the building itself into the trained surface subsidence prediction model to predict the surface subsidence.

[0013] Furthermore, the specific process of the said Step 4.3 is: Redefine the sequence of predicted surface subsidence related characteristic variables as , and construct a linear regression model: ; where is the predicted surface subsidence amount; is the regression coefficient, reflecting the th characteristic variable 's influence on the ground subsidence volume ; is the error term; Construct the design matrix : ; Among them, is the number of samples; is the th characteristic variable of the th sample; the target vector , where is the th true ground subsidence volume corresponding to the th sample; According to the least squares method, calculate the estimated value of the regression coefficient ; ; ; Among them, is the th characteristic variable of the th sample; is the th true ground subsidence volume corresponding to the Finally, substitute into the linear regression model to obtain the final predicted ground subsidence volume: ; Among them, is the final predicted ground subsidence volume; In the step 4.4, construct the loss function of the ground subsidence prediction model as: ; Among them, represents the expectation, is to take the expectation of the joint distribution of different time steps , the sequence of characteristic variables related to ground subsidence and the noise vector ; , are the noise vector and the estimated noise vector at the time step respectively; Use the Adam optimization algorithm to update the parameters in the quantum-inspired neural network, and the update formula is as follows: Calculate the first-order moment estimate and the second-order moment estimate of the gradient: ; ; Among them, and are the first - moment estimate and the second - moment estimate of the time step respectively; and are the first - moment estimate and the second - moment estimate of the time step respectively; is the gradient of the time step ; and are the decay rate hyperparameters of the first - moment estimate and the second - moment estimate respectively; Correct the first - moment estimate and the second - moment estimate: ; ; Among them, and are the corrected first - moment estimate and the second - moment estimate of the time step respectively; and are the decay rate hyperparameters of the first - moment estimate and the second - moment estimate of the time step respectively; Update the parameters of the surface subsidence prediction model: ; Among them, and are the parameters of the surface subsidence prediction model at the time step and the time step respectively; is the learning rate; Repeat the processes of forward diffusion, reverse diffusion, loss calculation, and parameter update until the preset number of training rounds is reached or the loss function converges to a preset threshold, and output the trained surface subsidence prediction model.

[0014] Beneficial technical effects brought by the present invention: In terms of data processing, the Kalman filtering method is used to fuse multi-source data, greatly improving the accuracy and integrity of data, fully integrating multi-channel data resources such as various sensors and geographic information, and laying a foundation for accurate analysis; the abnormal data processing based on random forest effectively identifies and corrects outliers, ensuring the reliability of data quality and avoiding the misguidance of subsequent analysis by incorrect data. In terms of feature engineering, the graph neural network is used for feature engineering and variable selection, accurately extracting feature variables that have a key influence on surface subsidence, reducing the data dimension while improving the model training efficiency and prediction accuracy. In the surface subsidence prediction link based on the diffusion model and the quantum-inspired neural network, the data generation advantage of the diffusion model and the powerful feature learning and processing ability of the quantum-inspired neural network are fully utilized to achieve a more accurate prediction of the surface subsidence of buildings, and then provide valuable basis and technical support for building safety assessment, engineering planning, disaster prevention, etc., strongly promoting the development of related fields towards intelligence and high efficiency, and significantly improving the technical level and practical effect of the entire building surface subsidence prediction and related application fields. Brief Description of the Drawings

[0015] Figure 1 It is a flowchart of the method for digital prediction of surface subsidence of buildings according to the present invention. Detailed Embodiments

[0016] The present invention will be further described in detail below in conjunction with the drawings and specific embodiments: The present invention aims to accurately predict the surface subsidence condition of buildings and provide strong support for building safety assurance. A digital and intelligent prediction method for surface subsidence based on a diffusion model and a quantum-inspired neural network is proposed. The specific technical ideas are as follows: First, multi-source data collection and multi-source data fusion based on the Kalman filtering method are carried out. Abundant data are obtained by widely deploying various sensors, and relevant system and design parameter data are integrated. The Kalman filtering method is used to fuse the multi-source data to reduce data errors and uncertainties. Then, abnormal data processing based on random forest is performed. Part of the fused data is selected as the training set, and various feature vectors are extracted to train the random forest model, so as to determine abnormal data and perform corresponding processing to ensure data quality. Next, feature engineering and variable selection are carried out based on graph neural networks. The key elements of the building and its surrounding areas are abstracted as graph nodes and given features, the edges between nodes are constructed and weights are determined. The graph neural network is used to screen out the feature variables highly correlated with surface subsidence, improving the model training efficiency and prediction accuracy. Finally, surface subsidence prediction is realized based on the diffusion model and the quantum-inspired neural network. After determining the relevant parameters, forward and reverse diffusion are carried out, and the predicted sequence of feature variables related to surface subsidence is obtained through operations such as quantum measurement. Then, the predicted value of the surface subsidence amount is obtained through post-processing. The present invention comprehensively uses a variety of advanced technologies and methods, and can effectively and accurately predict the surface subsidence condition of buildings, providing strong technical support for building safety.

[0017] As Figure 1 shown, a digital and intelligent prediction method for surface subsidence of buildings specifically includes the following steps: Step 1: Multi-source data collection and multi-source data fusion based on the Kalman filtering method. Specifically, it includes the following steps: Step 1.1: Multi-source data collection. Collect surface settlement monitoring data: High-density monitoring points are arranged around the building and in areas that may be affected. High-precision measuring instruments such as intelligent total stations, Beidou satellite positioning systems (BDS), and synthetic aperture radar interferometry (InSAR) are used to conduct high-frequency surface settlement monitoring to obtain high-precision spatio-temporal sequence data, including information such as the displacement of a certain point on the surface in the x, y, and z directions and the displacement change rate. At the same time, collect geological exploration data (including formation structure, physical and mechanical parameters of rock and soil masses, etc.), underground mining activity information (such as mining progress, mining method, mining depth, etc.), and detailed information about the building itself (such as structural type, foundation form, building materials, etc.).

[0018] Step 1.2: Multi-source data fusion. The Kalman filtering method is used to integrate multi-source heterogeneous data, construct a unified data format and coordinate system, and lay a foundation for subsequent analysis and processing.

[0019] Step 1.2.1. System Modeling. First, it is necessary to model the dynamic system of the land subsidence under study and determine its state equation and observation equation.

[0020] Step 1.2.1.1. Construct the state equation. Let the state vector of the land subsidence system at the current moment be (for example it can include state information such as the displacements and displacement rates of a certain point on the ground in the x, y, and z directions), and its state transition satisfies a linear relationship. The state equation can be expressed as: ; where, is the state transition matrix, which describes the law of the state change of the system from the previous moment to the current moment. It depends on factors such as the physical characteristics of the land subsidence system itself. is the previous moment, is the current moment. is the state vector at the previous moment. For example, if the state vector at the current moment is ( is the dimension of the state vector, such as , then , , and respectively represent the displacements and displacement rates in the x, y, and z directions), then is a matrix. The element in the -th row and -th column of the state transition matrix reflects the influence degree of the state vector in the -th column at the previous moment on the state vector in the -th row at the current moment. is the transpose symbol. is the control input matrix. In some cases, if there are external control factors (such as some measures taken artificially to slow down the subsidence) affecting the land subsidence, it can be regarded as the control input. is the control input vector at the previous moment. Generally, if there is no such control input, can be set as the zero matrix, is the zero vector. is the process noise vector at the previous moment, which is used to describe the uncertainty in the system modeling process and the errors caused by factors such as unmodeled dynamics. It is usually assumed to be Gaussian white noise with a mean of zero, and its covariance matrix is denoted as , characterizes the variance characteristics of the process noise. For example, is a A positive definite symmetric matrix, where the elements on the diagonal represent the process noise variances of the corresponding state variables.

[0021] Step 1.2.1.2: Construct the observation equation. Different sensors are used to observe the ground settlement system. Let the observation vector of all sensors (specifically sensors are set in the present invention) at the current moment be (for example each element in it corresponds to relevant observation information such as the ground settlement value measured by a sensor), and the observation equation can be expressed as: ; where is the observation matrix, which maps the state vector of the system at the current moment to the observation space, and its dimension is ( is the dimension of the observation vector, i.e., the number of sensors, is the dimension of the state vector), and the element in the -th row and -th column of the observation matrix represents the contribution weight of the state variable in the -th column at the current moment to the observation vector in the -th row at the current moment. For example, if the first sensor mainly observes the ground settlement in the x direction, then the elements in the first row of the matrix corresponding to the state variables related to the x-direction displacement are non-zero, and the others are zero (specifically depending on the physical relationship between the sensor observations and the state variables). is the observation noise vector at the current moment, representing the errors existing in the sensor measurements themselves. It is also assumed to be Gaussian white noise with a mean of zero, and the observation noise covariance matrix is denoted as is a positive definite symmetric matrix, where the elements on the diagonal represent the variances of the observation noises of each sensor, and the off-diagonal elements reflect the correlations between the observation noises of different sensors (usually when it is assumed that the sensor observations are independent, the off-diagonal elements are zero).

[0022] Step 1.2.2: Before starting the Kalman filter iteration, some parameters need to be initialized. Specifically: Step 1.2.2.1: Initial state estimation; Set the state estimation at the initial moment , which is usually given according to prior knowledge or preliminary measurements of the system. For example, it can be determined based on a small amount of sensor data at the initial moment or the empirical values of historical similar situations. The element values of

[0023] Step 1.2.2.2, Initial Estimation Error Covariance Matrix; Set the initial estimation error covariance matrix , which reflects the degree of uncertainty in the initial state estimation and is also a positive definite symmetric matrix. Its initial value can be reasonably set according to the accuracy of the initial understanding of the system and experience, etc. For example, if the initial state estimation is relatively inaccurate, then the element values of

[0024] Step 1.2.3, For each moment, perform the Kalman filter iteration process; The specific process is as follows: Step 1.2.3.1, Prediction Step (Time Update).

[0025] Step 1.2.3.1.1, Prior State Prediction; According to the state equation, calculate the prior state estimation at the current moment , and the calculation formula is: ; Here, the state estimation at the previous moment and the state transition matrix and the control input (if any) are used to predict the state at the current moment.

[0026] Step 1.2.3.1.2, Prior Estimation Error Covariance Prediction; Calculate the prior estimation error covariance matrix at the current moment , and the calculation formula is: ; This step reflects the change of the state estimation error covariance with the dynamic evolution of the system, considering the error covariance matrix at the previous moment through the propagation of the state transition matrix and the new uncertainty brought by the process noise covariance matrix .

[0027] Step 1.2.3.2, Update Step (Measurement Update).

[0028] Step 1.2.3.2.1, Kalman Gain Calculation; Calculate the Kalman gain at the current moment. It is a key parameter that weighs the proportion of the predicted state and the observed value in the final state estimation, and the calculation formula is: ; The Kalman gain at the current moment depends on the prior estimation error covariance matrix at the current moment, the observation matrix and the observation noise covariance matrix , which functions to reasonably adjust the utilization degree of the observation vector according to the reliability of the current observation information (reflected by ) and the uncertainty of the predicted state (reflected by ).

[0029] Step 1.2.3.2.2, State Estimation Update; Use the Kalman gain at the current moment to combine with the observation vector at the current moment to update the prior state estimate and obtain the optimal state estimate at the current moment. The calculation formula is as follows: ; Here represents the residual between the observation vector and the predicted observation value based on the prior state prediction. Multiplying the Kalman gain by this residual and adding it to the prior state prediction value gives the optimal state estimate that comprehensively considers prediction and observation.

[0030] Step 1.1.2.3.2.3, Estimation Error Covariance Update; Finally, update the estimation error covariance matrix at the current moment. The calculation formula is as follows: ; This step is used to update the uncertainty degree of the state estimation to prepare for the next iteration. is the identity matrix. In the ideal case, through such an update, the error covariance can be gradually reduced as new observation information is continuously incorporated, indicating that the estimation of the system state is becoming more and more accurate.

[0031] Step 1.2.4, Continuous Iteration and Data Integration. By continuously repeating the above prediction steps and update steps, iterative calculations are performed for each moment . As time progresses, the Kalman filter continuously fuses the observation information from different sensors (i.e., multi-source heterogeneous data sources) until , and the optimal state estimate at the final moment is obtained. The data from different data sources are integrated in a reasonable manner considering factors such as the reliability of each data and the dynamic characteristics of the system. is the total number of moments.

[0032] Take And the state estimation values at intermediate moments, the original observation data of each sensor, the process and observation noise statistical information, the state transition matrix, and the observation matrix related information, etc., are sorted through rules such as data screening, alignment, and feature combination, and are stored in a table or database storage format. After clarifying the format of each feature data type, a fused data set is obtained. . At the same time, in order to construct a unified data format and coordinate system, the state vector and other related parameters can be set according to the pre-set unified format and coordinate reference. For example, it is stipulated that the displacements in the x, y, and z directions are represented in a unified geographic coordinate system, and the data of all sensors are converted to this unified coordinate system when collected and participating in the Kalman filter calculation. Naturally, the finally fused data is in a unified data format and coordinate system, laying a foundation for subsequent analysis and processing.

[0033] Step 2. Anomaly data processing based on random forest; the specific process is as follows: Preprocess the collected original data, including noise removal, outlier detection and processing, missing value imputation, etc. Outlier detection uses the method based on the box plot (Box-Plot). If a data point exceeds the upper or lower limit of the box plot, it is identified as an outlier; the calculation formulas for the upper and lower limits are respectively: ; ; Among them, is the upper limit; is the lower quartile; is the interquartile range; is the lower limit; is the upper quartile.

[0034] For outliers (including missing values), random forest is used for imputation according to the distribution characteristics of the data. Random forest is an ensemble learning algorithm that makes predictions by constructing multiple decision trees and integrating their prediction results. When dealing with outliers, the feature containing the outlier is used as the target variable, and the remaining features are used as input variables to train the model, and then the trained model is used to predict the outlier.

[0035] Step 2.1. Data preparation. The data set fused in Step 1 contains samples, each sample has features, and the th feature of the th sample is denoted as , , . Mark the data columns corresponding to the features containing outliers.

[0036] Mark outlier samples and construct the relevant vectors of the training set. First, traverse the data set , and mark out all samples with outliers in all features. Suppose there are samples with outliers, . For the th sample with an outlier, , temporarily mark the position of the outlier in the th feature as empty (or represent the missing value with a specific symbol, such as NaN).

[0037] Construct the input feature vector , which is composed of the remaining features of the th sample with an outlier except the th feature, that is: ; wherein, is the th feature of the th sample with an outlier; At the same time, for these samples with outliers, collect the true values corresponding to their th feature under the known non-missing condition (partial complete data samples can be used) as the target variable , and construct the training data set : ; Step 2.2, Random forest model training.

[0038] Step 2.2.1, Bootstrap Sampling to construct a sub-data set. For the training data set , the bootstrap sampling method is used to generate the training sub-data set for each decision tree. The specific operation is: for each decision tree to be constructed (assuming decision trees are to be constructed, ), randomly draw samples with replacement from (here is the number of samples with outliers, and the number of sampling times is the same as the number of samples in the original data set, so that some samples may be drawn multiple times and some samples may not be drawn at all during the sampling process), and obtain the training sub-data set for the th decision tree.

[0039] Step 2.2.2, Random Feature Selection. When constructing each decision tree, random feature selection is also required to increase the diversity of the decision trees and reduce overfitting. For the th decision tree, randomly select features from all input features (excluding the th feature with outliers) (usually , here take the integer) for node splitting. Let the subset of features randomly selected for the th decision tree be : ; where is the index of the th feature used for node splitting in the th decision tree, and this index corresponds to a certain feature in the dataset after excluding the th feature column with outliers; Step 2.2.3, Decision Tree Construction. Each decision tree starts to be constructed based on the corresponding training sub-dataset and the randomly selected subset of features. For example, the th decision tree starts to be constructed based on the corresponding training sub-dataset and the randomly selected subset of features . Taking node splitting as an example, common metrics such as the Gini Index are used to determine the best splitting feature and splitting point.

[0040] The calculation process of the Gini Index is as follows: For a classification problem (assuming that the predicted outliers are discrete categorical data), at a certain node , let the set of classes be , is the number of classes, is the th class; the sample set of node is , the total number of samples in node is , and the number of samples belonging to the th class in node is . Then the Gini Index calculation formula for node is: ; where is the Gini Index; When considering feature splitting, assume that based on feature at a certain splitting point the node The sample set is divided into two parts and (corresponding to different branches), and the number of samples is and , respectively. Then the Gini index after splitting is: ; Among them, is the Gini index after splitting; Traverse all randomly selected features and the corresponding possible splitting points, and select the feature and splitting point that make the smallest as the best splitting method for this node. Continuously repeat this process to split the decision tree nodes until the stopping condition is met. The stopping condition is reaching the maximum depth or the number of samples in the node is less than the pre-set sample number threshold.

[0041] If the predicted outlier is a continuous numerical value, common measurement indicators can be selected such as the mean squared error (MSE), etc. For example, at the node , for a certain splitting point of the feature , the sample is divided into two parts and , and the means of the corresponding predicted target variables (i.e., the true values corresponding to the outliers) are and , respectively. Then the mean squared error after splitting is: ; Among them, is the mean squared error after splitting; is the input feature vector; is the target variable; Similarly, select the feature and splitting point that make the smallest for splitting, and continuously construct the decision tree until the stopping condition is met.

[0042] Through the above process, decision trees are constructed to form a random forest model . is the th decision tree.

[0043] Step 2.3. Outlier prediction.

[0044] Step 2.3.1. Single decision tree prediction. For the th sample containing outliers (including missing values), , input its corresponding input feature vector into each decision tree in the trained random forest model in turn, Perform prediction in []. The present invention relates to the problem of numerical prediction. After the input feature vector reaches the leaf node, each decision tree takes the mean value of the target variables (i.e., the known true values of outliers) corresponding to all samples in the leaf node as the predicted value of the sample.

[0045] Step 2.3.2: Determine the outlier estimate value based on the comprehensive predicted value. Determine the outlier estimate value by comprehensively considering the predicted values of all decision trees for the sample. For the prediction of numerical outliers, the simple average method is used, that is, add up the predicted values of all decision trees and take the average to obtain the final outlier estimate value. The calculation formula is: ; where, is the outlier estimate value of the th sample; is the predicted value of the th decision tree for the th sample.

[0046] Step 3: Feature engineering and variable selection based on graph neural network; the specific process is as follows: Extract feature variables closely related to surface subsidence from the preprocessed data, such as formation lithology characteristics (extract main components through principal component analysis), mining activity characteristics (mining speed, change rate of mining area, etc.), building characteristics (building weight, foundation stiffness, etc.), and time characteristics (seasons, interannual changes, etc.). Considering that surface subsidence data has complex spatial and correlation relationships, for example, the spatial distribution of formation structures, the geographical relationship between buildings and mining areas, etc., can be constructed as a graph structure. Graph neural networks can effectively process this type of graph-structured data and update node features by aggregating information from neighboring nodes. In this process, complex relationships in the data can be automatically learned and meaningful features can be extracted.

[0047] Step 3.1: Construction of graph neural network.

[0048] Step 3.1.1: Node definition. The node set contains various types of nodes, such as monitoring point nodes , mining area nodes , building nodes , etc. Each monitoring point node contains features such as its historical surface subsidence data sequence, geographical location information, etc.; the mining area node has mining activity-related parameters (such as mining speed sequence, mining depth, mining method, etc.) and its geographical range information; the building node contains the structural characteristics of the building (such as height, number of floors, foundation type, etc.), weight information, and location information, etc.

[0049] Step 3.1.2, Edge construction. Spatial relationship edge: For monitoring point nodes, if they are geographically adjacent, an undirected edge is established. For example, , are the th and th monitoring point nodes respectively. If the distance between and is less than a pre-set distance threshold , then an undirected edge is established between and , that is, , represents the set of edges. The weight of the edge between monitoring point nodes can be determined according to the reciprocal of the distance or other spatial attenuation functions. The formula is: ; where is the weight of the edge between the th and th monitoring point nodes; is the distance; is a non-zero positive number, and its value is generally very small to prevent the denominator from being zero.

[0050] A directed edge is established between the mining area node and the monitoring point nodes within its mining influence range. The direction of the edge is from the mining area to the monitoring point, indicating the influence relationship of mining activities on the surface subsidence of the monitoring point. The weight of the edge between the mining area node and the monitoring point node can be determined according to factors such as the distance from the mining area to the monitoring point and the mining intensity. Specifically: ; where is the weight of the edge between the th mining area node and the th monitoring point node; is the coefficient related to the mining characteristics of the th mining area node; is the current time when mining the th mining area node, such as the mining volume or mining speed. is the th mining area node; Edges are established between building nodes and adjacent monitoring point nodes. The weight of the edge between building nodes and monitoring point nodes can be determined comprehensively according to the distance between the building and the monitoring point and the degree of influence of the building on the ground surface (such as factors like building weight and foundation stiffness). Assuming that represents the weight of the building, represents the building foundation stiffness, Indicates the distance between the building and the monitoring point. The weight calculation formula considering both the distance and the influence degree of the building is: ; Where is the weight of the edge between the th building node and the th monitoring point node . and are weight coefficients determined according to the actual situation, used to adjust the influence degree of the building weight and the foundation stiffness on the edge weight.

[0051] Time-relationship edge: A directed time edge is established between different time steps of the same monitoring point node. For the nodes at time step and time step and , an edge is established. The weight of the edge between different time steps can be set to 1, indicating the sequential relationship in time. For the mining area nodes and the building nodes, time edges can also be established similarly to reflect the relationship of the change of their own characteristics over time and the dynamic influence on the surface subsidence.

[0052] Step 3.2, Definition of the spatio-temporal graph convolutional network (STGCN) layer. In STGCN, the connection between layers is achieved by using the output of the previous layer as the input of the next layer.

[0053] Step 3.2.1, Spatial convolution part; For the feature update of the node at the th layer, spatial convolution is first considered. Let represent the spatial feature of the node at time step , is the spatial feature of the neighbor node of the node at time step . The spatial convolution formula is: ; Where is the spatial feature of the node at the th layer and time step ; is the spatial feature of the neighbor node of the node at the th layer and time step ; For the node at the layer time step spatial feature; and are the spatial neighbor node sets of the node and the neighbor node respectively; is the time neighbor node set of the node , that is, the set of the same node at different time steps; is the time neighbor node set of the node is a time step identifier in the time neighbor node set of the same node, representing the index of the same node at different time steps, and is used to aggregate and calculate the features of different time steps; is the spatial weight matrix of the layer; is the time weight matrix of the layer; is the activation function. The present invention can adopt the ReLU activation function. Here, the spatial convolution not only aggregates the features of spatial neighbor nodes, but also considers the features of time neighbor nodes, realizing the preliminary fusion of spatio-temporal features. The specific process is as follows: The spatial features of the upper layer node and its neighbor nodes after weighted aggregation (through the spatial weight matrix ) are fused and then passed through the activation function to obtain the spatial feature

[0054] of the nodes in this layer, which is used as one of the inputs for the spatial convolution of the next layer. Let represent the time feature of the node at the layer time step, and its update formula is: ; wherein, is the weight of the time edge; is the time convolution weight matrix of the layer; is the time feature of the node at the layer time step . The time convolution mainly focuses on the feature change relationship of the same node at different time steps, and further extracts the feature information in the time series. The specific process is as follows: The time features of the same node at different time steps in the upper layer pass through the time edge weight and the time convolution weight matrix ​After the function, through the activation function Obtain the time features of the nodes in this layer , and participate in the spatio-temporal fusion step.

[0055] Step 3.2.3, Spatio-temporal fusion; finally, fuse the results of spatial convolution and temporal convolution to obtain the fusion features of the nodes at the layer: ; Among them, is the fusion feature of node at the layer time step ; is a fusion coefficient used to balance the contributions of spatial and temporal features, which can be determined according to experimental or data characteristics.

[0056] Step 3.3, Training process and loss function. Suppose there is a supervised learning task related to surface subsidence, such as predicting the future surface subsidence value of a certain monitoring point. A fully connected layer can be connected to the output features of the STGCN to obtain the prediction result. For a batch of training data (including graph samples), the mean squared error (MSE) is used as the loss function : ; Among them, is the true surface subsidence value of the th graph sample; is the predicted surface subsidence value of the th graph sample.

[0057] Use the backpropagation algorithm to update the weights of each STGCN layer to minimize the loss function. For example, for the update of the weight of the layer, according to the gradient descent method: ; Among them, is the updated weight of the layer; is the learning rate; Step 3.4, Feature selection. For the trained STGCN, the output by the last layer is the extracted feature related to surface subsidence, is the fusion feature of node at the layer time step , is the total number of layers. Feature selection can be performed by calculating the correlation between these features and the predicted surface subsidence value. The Pearson correlation coefficient is used to measure the correlation between each feature and the predicted surface subsidence value. For the features obtained after the -layer spatio-temporal graph convolutional network calculation, the feature dimension is of the th feature and the predicted surface subsidence value , the calculation formula of the Pearson correlation coefficient is: ; where is the Pearson correlation coefficient between the th feature with a feature dimension of and the predicted surface subsidence value ; is the feature value of the th feature with a feature dimension of at time step in the rd graph sample after the calculation of the -layer spatio-temporal graph convolutional network; is the average value of the feature values of the th feature with a feature dimension of in a batch of training data containing graph samples after the calculation of the -layer spatio-temporal graph convolutional network at time step ; is the average value of the predicted surface subsidence values in a batch of training data containing graph samples; Calculate the Pearson correlation coefficient between each feature and the predicted surface subsidence value, and sort the features in descending order of the Pearson correlation coefficient. Select the top features with a high correlation with the predicted surface subsidence value as the feature variables that have a significant impact on the surface subsidence prediction for subsequent prediction model construction. The advantage of the spatio-temporal graph neural network is that it can automatically capture the spatio-temporal and correlation relationships in the data, and can extract more valuable features for the surface subsidence data containing complex geographical and physical relationships.

[0058] Step 4: Construct a surface subsidence prediction model based on the diffusion model and the quantum-inspired neural network. The specific process is as follows: Step 4.1: Generate the initial sample based on the forward diffusion process of the diffusion model.

[0059] Step 4.1.1: Determine the number of forward diffusion steps and the noise parameter.

[0060] Step 4.1.1.1, Determine the number of forward diffusion steps; Set the prediction time range as a variable representing the time span, and the data uncertainty measure is a quantitative indicator of the degree of data fluctuation or change. Determine the number of forward diffusion steps through and . If predicting the surface subsidence situation in the relatively short future and the data is relatively stable, a smaller value can be set; conversely, if predicting a longer time range and the data fluctuates greatly, a larger value is selected. Determine a basic number of forward diffusion steps based on experience or data characteristic analysis, and then adjust to obtain the actual number of forward diffusion steps through the following method : ; wherein, represents rounding up, and are different adjustment coefficients, determined according to experiments or domain knowledge, controls the influence weight of the prediction time range on the number of steps, reflects the influence weight of data uncertainty on the number of steps. For example, if predicting the surface subsidence situation in the next 10 months, the historical data standard deviation , , , , then .

[0061] Step 4.1.1.2, Determine the noise scheduling parameter; The noise scheduling parameter is set in a cosine annealing manner: ; wherein, is the time step in the forward diffusion process, and its value range is from 1 to ; is the minimum value of the set noise scheduling parameter, for example, taking the value 0.001.

[0062] Step 4.1.2, Generate the starting sample. Combine the features selected in Step 3.4 to form a sequence of surface subsidence-related feature variables , which can be regarded as a high-dimensional vector, containing information on various features such as formation lithology and mining activities in the time series. In the forward diffusion process, a series of noise-added samples are generated by gradually adding Gaussian noise. The noise-added sample vector For passing The initial sample is obtained by adding Gaussian noise to the time step. The forward diffusion process is expressed as: ; in, , The time steps in the forward diffusion process are , The noised sample vector generated when ; The time step is The noise scheduling parameters when , which determines the degree of noise added at each time step; The time step is The noise vector is drawn from a standard Gaussian distribution Repeat the forward diffusion process times, get the starting sample .

[0063] Step 4.2: Generate feature variable sequence using back diffusion process based on quantum-inspired neural network.

[0064] Step 4.2.1. Initialization of quantum-inspired neural network parameters.

[0065] Step 4.2.1.1. Quantum bit representation layer. In order to utilize some characteristics of quantum computing (such as superposition and entanglement) to enhance the representation ability of the model, the noise sample vector in the diffusion model is transformed into Assume that the input feature vector dimension is , mapping it to qubits, Determined according to the characteristic dimension and the principles of quantum computing, e.g. The state of a qubit can be represented by the complex vector The initial quantum bit state is It can be obtained by encoding the input eigenvector, for example, using a simple linear encoding method to map the real eigenvalues ​​to the amplitude of the quantum state. Map to The quantum state of qubits. , for No. elements; the initialization formula of the quantum bit state is as follows: ; ; ; Among them, is the angular parameter for initializing the state of the -th qubit, which is calculated according to the ratio of the sum of partial elements of to the total modulus length, ; ; is the loop index variable. By taking values of from 1 to , the sum of elements at specific positions in is calculated; , represents floor function, reflects a grouping or partitioning parameter when mapping -dimensional to qubits; is the -th element of ; , are the states of the -th and -th qubits at time step respectively; is the state of the qubits at time step ; represents the tensor product of quantum states.

[0066] Step 4.2.1.2, Quantum-inspired Gate Operation Layer. Similar to the layers in traditional neural networks, the quantum-inspired gate operation layer is defined. For example, the quantum-inspired gate operation layer is constructed using the Rotation Gate and the CNOT Gate. For the Rotation Gate, let be the rotation angle (learnable parameter). The action formula of the Rotation Gate operation on a single qubit is: ; Let the first quantum-inspired gate operation layer apply qubits respectively with , , …, . Then the matrix representation of is: ; Among them, is the rotation angle for the first qubit to control the state of the -th qubit; The quantum controlled-NOT gate acts on two qubits and is used to implement entanglement operations. By combining multiple layers of quantum heuristic gate operations, complex transformations can be performed on the qubit states to extract and process features. The matrix representation of the quantum controlled-NOT gate (CNOT Gate) is: ; where is the quantum controlled-NOT gate for the -th qubit to control the state of the -th qubit; assuming there are layers in the quantum heuristic gate operation layer, other layers are constructed according to the process; the -th ( ) quantum heuristic gate operation layer contains different combinations of quantum gate operations, and the parameters of these gates (such as the rotation angle of the -th qubit to control the state of the -th qubit) have been determined during the training process and are directly loaded during prediction.

[0067] Step 4.2.2, Reverse diffusion iteration. After passing through multiple quantum gate layers, the qubit states need to be converted back to a classical probability distribution, which is achieved through quantum measurement. After quantum measurement, a classical probability distribution vector is obtained, which is used as an intermediate representation for estimating the noise vector at time step . For example, for a system with qubits, a probability distribution vector of length is obtained after measurement, and it can be converted into a vector with the same dimension as the noise vector through a fully connected layer for updating in the reverse diffusion process.

[0068] Starting from , first input into the quantum heuristic neural network for noise estimation. After converting to the quantum state in the qubit representation layer, feature transformation is performed through the quantum heuristic gate operation layer. For , , where , that is, the initial qubit state passes through the first quantum heuristic gate operation layer to obtain the first qubit state . For , . And so on, until , . and are the -th and -th qubit states, respectively; is the -th quantum heuristic gate operation layer; the sequence number of the qubit state corresponds to the sequence number of the quantum heuristic gate operation layer. After passing through quantum gate operation layers, the quantum measurement method is used to perform a quantum state transformation on the qubit state of the last layer, obtaining a classical probability distribution with a dimension of because the measurement results of qubits have possible states. The specific process of the quantum measurement method is as follows: the qubit state formed by qubits is expressed as a linear combination of the ground states using the computational basis, and the measurement probabilities of each ground state are calculated. During measurement, it collapses to a certain ground state according to the probability. After multiple measurements, the frequencies of each ground state appearance are statistically counted, obtaining the random process of the classical probability distribution . The result of each measurement is uncertain, and accurate classical probability distribution can only be obtained through a large number of repeated measurements, which is actually realized by relying on quantum measurement devices.

[0069] Then, through the fully connected layer of the measurement and output layer, is converted into an estimated noise vector with the same dimension as the noise vector ( -dimensional). Let the weight matrix of the fully connected layer be (with a dimension of ), and the bias vector be (with a dimension of ), then: ; where is the -th component of the estimated noise vector , which is obtained by performing a linear transformation on the probability distribution obtained by quantum measurement through the fully connected layer, and adding the bias vector of the fully connected layer to obtain the final estimated noise vector; is used to sequentially access each element in the probability distribution vector in order to calculate each component of the estimated noise vector through a linear transformation using the weight matrix of the fully connected layer; is the -th element of for The weight of the th component; is the th element of ; is the th element in The dimension of is ; Then, update the sample according to the reverse diffusion formula to obtain the predicted sample vector: ; Among them, , are the predicted sample vectors generated at time steps , respectively during the reverse diffusion process; is the noise scheduling parameter at time step , which is obtained by subtracting the estimated noise vector from and adjusting it according to a certain ratio. is the estimated noise vector at time step .

[0070] Repeat the above steps, gradually decreasing from to to obtain the predicted sequence of surface subsidence-related characteristic variables . is the predicted sample vector generated at time step 0 during the reverse diffusion process; Step 4.3, Extract the predicted value of surface subsidence. Extract the information for predicting the surface subsidence amount from the predicted sequence of surface subsidence-related characteristic variables, and obtain the predicted value of the surface subsidence amount for a future period through the linear regression layer. Redefine the predicted sequence of surface subsidence-related characteristic variables as (where is the sequence length), and construct a linear regression model: ; Among them, is the predicted surface subsidence amount; is the regression coefficient, reflecting the influence degree of the th characteristic variable on the surface subsidence amount ; is the error term.

[0071] Construct the design matrix : ; Among them, is the number of samples; is the th feature variable of the th sample; each row of the matrix represents the feature data of a sample plus the constant term 1. The target vector where is the true ground subsidence amount corresponding to the

[0072] According to the least squares method, calculate the estimated value of the regression coefficient: ; ; ; where is the th feature variable of the th sample; is the true ground subsidence amount corresponding to the th sample; first calculate the product of the transpose of the design matrix and itself and the product of the transpose of the design matrix and the target vector , and then solve the linear equation system to obtain the estimated value of the regression coefficient.

[0073] Finally, substitute into the linear regression model to obtain the final predicted ground subsidence amount: ; where is the final predicted ground subsidence amount.

[0074] Step 4.4, Model training of the ground subsidence prediction model.

[0075] Step 4.4.1, Loss function. Adopt a reconstruction loss similar to that in the variational autoencoder (VAE). In the diffusion model, the goal is to make the final sample generated by the reverse diffusion process as close as possible to the original sample. Therefore, the loss function of the ground subsidence prediction model can be defined as: ; where represents the expectation, is to take the expectation of the joint distribution of different time steps , the sequence of ground subsidence-related feature variables and the noise vector ; , are the noise vector and the estimated noise vector at time step , respectively.

[0076] Step 4.4.2, Optimization algorithm. Use the Adaptive Moment Estimation (Adam) optimization algorithm to update the parameters in the quantum-inspired neural network (including the parameters of quantum gates and the parameters of fully connected layers, etc.). The parameter update formula of the Adam optimization algorithm is as follows: Calculate the first-order moment estimate and the second-order moment estimate of the gradient: ; ; where , are the first-order moment estimate and the second-order moment estimate at time step , respectively; , are the first-order moment estimate and the second-order moment estimate at time step , respectively; is the gradient at time step ; and are the decay rate hyperparameters of the first-order moment estimate and the second-order moment estimate, respectively (usually , ).

[0077] Correct the first-order moment estimate and the second-order moment estimate: ; ; where , are the corrected first-order moment estimate and the second-order moment estimate at time step , respectively; , are the decay rate hyperparameters of the first-order moment estimate and the second-order moment estimate at time step , respectively; Update the parameters of the surface subsidence prediction model: ; where , are the parameters of the surface subsidence prediction model at time step and time step , respectively; is the learning rate.

[0078] Repeat the above processes of forward diffusion, reverse diffusion, loss calculation, and parameter update until the preset number of training epochs is reached or the loss function converges to a preset threshold, and output the trained surface subsidence prediction model.

[0079] Step 4.5: Input the newly collected real-time data into the trained surface subsidence prediction model for surface subsidence prediction. The new data includes surface settlement monitoring data, geological exploration data, underground mining activity information, detailed information of the building itself, etc. The model of the present invention based on the diffusion model and the quantum-inspired neural network combines the advantages of the diffusion model in generating data and the potentially powerful representation ability of quantum-inspired computing, providing a novel and exploratory method for surface subsidence prediction.

[0080] 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 those skilled in the art within the scope of the essence of the present invention should also fall within the protection scope of the present invention.

Claims

1. A digital intelligence prediction method for surface subsidence of buildings, characterized in that, It includes the following steps: Step 1: Collect multi-source data and perform multi-source data fusion based on the Kalman filtering method; Step 2: Process abnormal data based on random forests; Step 3: Conduct feature engineering and variable selection based on graph neural networks; Step 4: Predict surface subsidence based on diffusion models and quantum-inspired neural networks.

2. The digital prediction method for ground subsidence of a building according to claim 1, wherein The specific process of Step 1 is as follows: Step 1.1: Collect multi-source data, including surface subsidence monitoring data, geological exploration data, underground mining activity information, and detailed information about the building itself; the surface subsidence monitoring data includes the displacements of a certain point on the surface in the x, y, and z directions and the displacement change rate; the geological exploration data includes the stratum structure and the physical and mechanical parameters of rock and soil masses; the underground mining activity information includes the mining progress, mining method, and mining depth; the detailed information about the building itself includes the structure type, foundation form, and building materials; Step 1.2: Use the Kalman filtering method to integrate the multi-source data and construct a unified data format and coordinate system; the specific process is as follows: Step 1.2.1: Model the surface subsidence system to determine the state equation and the observation equation; The state equation is: ； Among them, is the previous moment, is the current moment; , are the state vectors of the ground settlement system at the current moment and the previous moment respectively; is the state transition matrix; is the control input matrix; is the control input vector at the previous moment; is the process noise vector at the previous moment; The observation equation is: ； wherein, is the observation vector of all sensors at the current moment; is the observation matrix; is the observation noise vector at the current moment; Step 1.2.

2. Initialize the parameters; specifically: Set the state estimate at the initial time to be , and set the initial estimated error covariance matrix to be ; Step 1.2.3: For each moment, perform the Kalman filtering iteration process, which includes a prediction step and an update step; Step 1.2.4: Continuously iterate to complete data integration.

3. The ground subsidence digital prediction method for buildings according to claim 2, characterized in that, The specific process of Step 1.2.3 is as follows: The specific process of the prediction step is: First, according to the state equation, calculate the prior state estimate at the current moment , and the formula is: ； Among them, is the state estimate at the previous moment ; Then, calculate the prior estimation error covariance matrix at the current moment , and the formula is: ； wherein, is the error covariance matrix at the previous moment; is the transpose symbol; is the process noise covariance matrix; The specific process of the update step is: First, calculate the Kalman gain at the current moment , and the formula is: ； Among them, is the observation noise covariance matrix ; Then, combine the observation vector at the current moment to update the prior state estimate and obtain the optimal state estimate at the current moment , and the formula is: ； Finally, update the estimation error covariance matrix at the current moment , and the formula is: ； Among them, is the identity matrix; In step 1.2.4, by continuously repeating the prediction step and the update step, for each moment Iterative calculations are performed. As time progresses, the Kalman filter continuously fuses the observation information of different sensors until At this time, the optimal state estimate at the final moment is obtained , That is the data result after fusion; is the total number of moments; Combine with the state estimation values at intermediate times during the Kalman filtering process, the original observation data of each sensor, the process and observation noise statistical information, and the relevant information of the state transition matrix and the observation matrix to obtain a fused data set .

4. The method for digital prediction of ground subsidence for buildings according to claim 3, characterized in that, The specific process of Step 2 is as follows: Step 2.1: Detect outliers in the fused dataset based on the box plot method and construct a training dataset; Let the merged dataset contain samples, and each sample has features; Traverse the fused dataset , and mark all samples with outliers in features; Suppose there are samples with outliers, ; For the th sample containing outliers, , mark the outlier position of the th feature as empty; Construct input feature vector : ； Among them, is the th feature of the th outlier-containing sample; Meanwhile, for this sample containing outliers, collect the true value corresponding to its th feature under the known non-missing condition as the target variable , and construct a training data set : ； Step 2.2: Train a random forest model based on the training dataset; Step 2.3: Perform outlier prediction.

5. The ground settlement digital prediction method for buildings according to claim 4, characterized in that The specific process of Step 2.2 is as follows: Step 2.2.1: For the training data set , the bootstrap sampling method is used to generate the training sub - data set for each decision tree; the process of the bootstrap sampling method is as follows: for each decision tree to be constructed, randomly draw samples with replacement from to obtain the training sub - data set for the th decision tree; Step 2.2.

2. For the th decision tree, randomly select input features from all features for node splitting. Let the feature subset randomly selected for the th decision tree be: ​ ； Among them, is the index of the th feature used by the th decision tree for node splitting, and this index corresponds to a certain feature in the dataset after removing the th feature column with outliers; Step 2.2.3: Construct a decision tree based on the training sub-dataset and a randomly selected feature subset; If the predicted outliers are discrete categorical data, the Gini coefficient is used in the node splitting process to determine the best splitting feature and splitting point; at a certain node , let the set of categories be , be the number of categories, be the -th category; the sample set of node is , the total number of samples of node is , and the number of samples belonging to the -th category of node is , then the calculation formula for the Gini coefficient of node is: ； Among them, is the Gini coefficient; When considering feature splitting, assume that based on feature at a certain splitting point the sample set of node is divided into two parts and , with the number of samples being and respectively. Then the Gini index after splitting is as follows: ； Among them, is the Gini index after splitting; Traverse all randomly selected features and the corresponding split points, and select the feature and split point that make the smallest as the best splitting method for this node. Continuously repeat this process to split the decision tree nodes until the maximum depth is reached or the number of node samples is less than the pre-set sample number threshold; If the predicted outlier is a continuous numerical value, the mean squared error is selected to determine the best splitting feature and splitting point; at node , for a certain splitting point of feature , the samples are divided into two parts and , and the means of the corresponding predicted target variables are and respectively. Then the mean squared error after splitting is: ； Among them, is the mean squared error after splitting; is the input feature vector; is the target variable; Select to split using the feature and split point that result in the minimum value, and continuously construct the decision tree until the maximum depth is reached or the number of samples in a node is less than the preset sample number threshold; Through the above process, a decision tree is constructed, and a random forest model is formed by ; is the th decision tree; The specific process of Step 2.3 is as follows: First, for the th sample containing outliers, input its corresponding input feature vector sequentially into each decision tree in the trained random forest model for prediction; after the input feature vector reaches the leaf node, each decision tree takes the mean of the target variables corresponding to all samples in the leaf node as the predicted value of the sample; Then, determine the estimated value of the outlier by synthesizing the predicted values of all decision trees for the sample. The formula is: ； Among them, is the outlier estimate of the th sample; is the predicted value of the th decision tree for the th sample.

6. The method for intelligent prediction of ground subsidence for buildings according to claim 5, characterized in that, The specific process of Step 3 is as follows: Step 3.1: Construct a graph neural network; the node set in the graph neural network contains multiple types of nodes, including monitoring point nodes , mining area nodes , and building nodes ; each monitoring point node contains its historical surface subsidence data sequence and geographical location information; the mining area node contains mining activity-related parameters and its geographical range information; The building nodes contain the structural characteristics, weight information, and location information of the building; the edges in the graph neural network include spatial relationship edges and temporal relationship edges; The construction process of the spatial relationship edges is: For the monitoring point nodes, if they are geographically adjacent, establish an undirected edge; the weight of the edge between the monitoring point nodes is: ； Among them, is the th, weight of the edge between the th and the th monitoring point nodes; is the distance; and are non-zero positive numbers; Establish a directed edge between the mining area node and the monitoring point nodes within its mining influence range; the direction of the edge is from the mining area to the monitoring point, indicating the influence relationship of the mining activity on the surface subsidence of the monitoring point; the weight of the edge between the mining area node and the monitoring point node is: ； Among them, is the weight of the edge between the th mining area node and the th monitoring point node; is a coefficient related to the mining characteristics of the th mining area node; is the mining intensity index of the th mining area node at the current time ; is the th mining area node; Establish an edge between the building node and the adjacent monitoring point nodes. The weight of the edge between the building node and the monitoring point node is: ； Among them, is the th building node and the th monitoring point node the weight of the edge between them, and are weight coefficients determined according to the actual situation, used to adjust the influence degree of building weight and foundation stiffness on the edge weight; represents the weight of the building, represents the building foundation stiffness, represents the distance between the building and the monitoring point; The construction process of the temporal relationship edges is: Establish directed temporal edges between different time steps of the same monitoring point node; the weight between different time steps is set to 1; Step 3.2: Construct a spatio-temporal graph convolutional network; Step 3.3: Construct a loss function and train the spatio-temporal graph convolutional network; Step 3.4: Conduct feature selection.

7. The ground settlement digital prediction method for buildings according to claim 6, characterized in that In step 3.2, in the spatio-temporal graph convolutional network, there are several layers, and the connection between layers is realized by using the output of the previous layer as the input of the next layer; the spatio-temporal graph convolutional network includes a spatial convolution part and a temporal convolution part. The specific working process is as follows: Step 3.2.

1. For the node in the layer, for feature update, first perform spatial convolution. The formula is: ； Among them, is the node at the layer time step spatial feature; is the node neighbor node of at the layer time step spatial feature; is the node at the layer time step spatial feature; , are respectively the spatial neighbor node sets of the node , neighbor node ; is the temporal neighbor node set of the node ; is the temporal neighbor node set of the node a time step identifier in; is the spatial weight matrix of the layer; is the temporal weight matrix of the layer; is the activation function; Let represent the spatial feature of the node at the time step , is the neighbor node of the node at the time step spatial feature; Step 3.2.2: Then conduct temporal convolution. The formula is: ； Among them, represents the node at the layer time step time feature; is the weight of the time edge; is the layer time convolution weight matrix; is the time feature of node at the layer time step time feature; Step 3.2.

3. Finally, fuse the results of spatial convolution and temporal convolution to obtain the fused feature of the node at the th layer: ； Among them, is a node at the layer time step of the fused feature; is a fusion coefficient.

8. The intelligent prediction method for surface subsidence of a building according to claim 7, wherein In step 3.3, a fully connected layer is connected to the output features of the spatio-temporal graph convolutional network to obtain the prediction result; for the training data containing graph samples, the mean squared error is used as the loss function : ； Among them, is the true surface subsidence value of the th map sample; is the predicted surface subsidence value of the th map sample; Use the backpropagation algorithm to update the weights of each layer in the spatio-temporal graph convolutional network to minimize the loss function; use the gradient descent method for updating: ； Among them, is the updated weight of the layer; is the weight of the layer; is the learning rate; In step 3.4, for the trained spatio-temporal graph convolutional network, the output of the last layer is the feature extracted related to surface subsidence, which is the fused feature of node at the -th time step of layer ; is the total number of layers; feature selection is performed by calculating the correlation between these features and the predicted surface subsidence value; the Pearson correlation coefficient is used to measure the correlation between each feature and the predicted surface subsidence value. For the feature dimension of obtained after calculation by the spatio-temporal graph convolutional network with layers, for the -th feature and the predicted surface subsidence value , the calculation formula of the Pearson correlation coefficient is:​ ； Among them, is the Pearson correlation coefficient between the th feature with a feature dimension of and the predicted ground settlement value ; is the feature value of the th graph sample, the th feature with a feature dimension of at time step after being calculated by the spatio-temporal graph convolutional network through layers; is the average value of the feature values of the th feature with a feature dimension of at time step in a batch of training data containing graph samples, after being calculated by the spatio-temporal graph convolutional network through layers; is the average value of the predicted ground settlement values in a batch of training data containing graph samples; Calculate the Pearson correlation coefficient between each feature and the predicted surface subsidence value, sort the features in descending order of the Pearson correlation coefficient, and select the top features as feature variables.

9. The ground subsidence digital prediction method for buildings according to claim 8, characterized in that The specific process of step 4 is as follows: Step 4.1: Generate the initial sample based on the forward diffusion process. The specific process is as follows: Step 4.1.1, determine the number of forward diffusion steps and the noise parameter; the number of forward diffusion steps is: ； Among them, represents rounding up; is the number of forward diffusion steps based on; and are different adjustment coefficients; is the prediction time range; is the data uncertainty measure; Noise scheduling parameter It is as follows: ； Among them, is the time step in the forward diffusion process; is the minimum value of the set noise scheduling parameter; Step 4.1.2: The forward diffusion process is expressed as: ； Among them, , are the noise-added sample vectors generated at time steps , respectively during the forward diffusion process; is the noise schedule parameter at time step ; is the noise vector at time step . Repeat the forward diffusion process times to obtain the starting sample ; Step 4.2: Generate the feature variable sequence based on the quantum-inspired neural network using the reverse diffusion process. The specific process is as follows: Step 4.2.1: Initialize the parameters of the quantum-inspired neural network; First, take as the input feature vector for qubit representation; the initialization formula for the qubit state is as follows: ； ； ； Among them, is the angular parameter used to initialize the th qubit state; is the loop index variable; , denotes floor function, reflects a grouping or partitioning parameter when mapping a -dimensional to qubits; is the th element of ; , are the th and th qubit states at time step respectively; is the qubit state at time step ; represents the tensor product of quantum states; Then, construct the quantum-inspired gate operation layer based on the quantum rotation gate and the quantum controlled-NOT gate; The formula for the action of the quantum rotation gate on a single qubit is: ； Among them, is a quantum rotation gate; is the rotation angle; The matrix representation of the quantum controlled-NOT gate is: ； Among them, is the th qubit controlling the state of the th qubit, which is a quantum controlled-NOT gate; The first quantum-inspired gate operation layer The matrix representation of which is: ； Among them, is the rotation angle for the first qubit to control the state of the th qubit; The quantum-inspired gate operation layer has a total of layers, and other layers are constructed according to the process; the th quantum-inspired gate operation layer contains different combinations of quantum gate operations, and the parameters of these gates are determined during training and directly loaded during prediction; Step 4.2.2: Reverse diffusion iteration; From the beginning, first input into the quantum-inspired neural network for noise estimation; At the qubit representation layer, is converted into a quantum state After that, feature transformation is performed through the quantum-inspired gate operation layer; For , , where , i.e., the initial qubit state after the first layer of quantum-inspired gate operations results in the first qubit state ; For , ; and so on until , ; , are the -th and -th qubit states respectively; is the -th layer of quantum-inspired gate operations; the sequence number of the qubit state corresponds to the sequence number of the quantum gate operation layer; After passing through the quantum-inspired gate operation layer, a quantum measurement method is used to perform a quantum state transformation on the qubit states of the last layer to obtain a classical probability distribution ; Convert into an estimated noise vector with the same dimension as the noise vector ; Let the weight matrix of the fully connected layer be , and the bias vector be , then: ； Among them, is the -th component of the estimated noise vector ; is the weight of the -th element pair of with respect to the -th component of ; is the -th element of ; is the -th element in ; Then update the sample according to the reverse diffusion formula to obtain the predicted sample vector: ； Among them, , are the predicted sample vectors generated at time steps , respectively during the reverse diffusion process; is the noise scheduling parameter at time step ; is the estimated noise vector at time step . Repeat the above steps, starting from and gradually decreasing to to obtain the predicted sequence of characteristic variables related to surface subsidence ; is the predicted sample vector generated at time step 0 during the reverse diffusion process; Step 4.3: Extract the predicted value of the ground subsidence amount; Step 4.4: Train the ground subsidence prediction model; Step 4.5: Input the real-time collected ground settlement monitoring data, geological exploration data, underground mining activity information, and detailed information of the building itself into the trained ground subsidence prediction model to predict the ground subsidence.

10. The method for numerically intelligent prediction of ground settlement for buildings according to claim 9, characterized in that, The specific process of step 4.3 is as follows: Redefine the predicted sequence of characteristic variables related to surface subsidence as , and construct a linear regression model: ； Among them, is the predicted surface subsidence amount; is the regression coefficient, reflecting the th characteristic variable on the surface subsidence amount influence degree; is the error term; Construct a design matrix : ； Among them, is the number of samples; is the -th feature variable of the -th sample; the target vector where is the true ground subsidence amount corresponding to the -th sample; Calculate the estimated value of the regression coefficient according to the least squares method : ； ； ； Among them, is the -th feature variable of the -th sample; is the true ground settlement corresponding to the -th sample; Finally, substitute into the linear regression model to obtain the final predicted surface subsidence amount: ； Among them, is the final predicted ground settlement volume; In step 4.4, the loss function for constructing the surface subsidence prediction model is as follows: ； Among them, denotes expectation, which is to take the expectation of the joint distribution of sequences of characteristic variables related to surface subsidence and noise vectors at different time steps; , are the noise vector and the estimated noise vector at time step respectively. Use the Adam optimization algorithm to update the parameters in the quantum-inspired neural network. The update formula is as follows: Calculate the first-order moment estimate and the second-order moment estimate of the gradient: ； ； Among them, and are the first moment estimate and the second moment estimate of the time step respectively; and are the first moment estimate and the second moment estimate of the time step respectively; is the gradient of the time step ; and are the decay rate hyperparameters of the first moment estimate and the second moment estimate respectively; Correct the first-order moment estimate and the second-order moment estimate: ； ； Among them, and are the corrected first moment estimate and second moment estimate at time step respectively; and are the decay rate hyperparameters of the first moment estimate and second moment estimate at time step respectively. Update the model parameters of the ground subsidence prediction model: ； Among them, , are the predicted model parameters of ground subsidence at time steps and respectively; is the learning rate. Repeat the processes of forward diffusion, reverse diffusion and loss calculation, and parameter update until the preset number of training rounds is reached or the loss function converges to the preset threshold, and output the trained ground subsidence prediction model.

Images