Coal mine building foundation settlement prediction system and method based on multi-source intelligent sensor

Abstract

The application discloses a coal mine building foundation settlement prediction system and method based on a multi-source intelligent sensor, belongs to the technical field of coal mine building monitoring, and is characterized in that multi-source sensors such as settlement displacement, inclination, pore pressure, water potential and microseismicity are arranged in a building foundation area, non-settlement characteristic signals in a fixed time period are collected, and a multi-source original data set is constructed; non-displacement disturbance characteristics are extracted through wavelet transform and phase analysis to form a characteristic sequence matrix; the characteristic sequence matrix is input into a deep time graph neural network combined with time evolution weight and spatial coupling structure, and a foundation settlement prediction response sequence is output; error inversion is performed in combination with measured data, deep response factors are identified, and a cause graph is constructed; finally, a risk function is established based on risk factors and the prediction sequence, settlement risk grade division and regional early warning are realized, and the application has the advantages of response preposition, causal explainability and risk quantification and is suitable for settlement prediction and hierarchical control under complex geological environments.

Description

Technical Field

[0001] This invention relates to the field of coal mine building monitoring technology, specifically to a coal mine building foundation settlement prediction system and method based on multi-source intelligent sensors. Background Technology

[0002] Coal mine buildings are typically located above or around goaf areas, with complex geological structures and issues such as heterogeneous foundation soil, numerous weak interlayers, and frequent fluctuations in groundwater levels. As mining activities continue, stress release and surface subsidence often exhibit nonlinear and abrupt characteristics, seriously threatening the structural safety and service life of surface buildings. Especially under conditions of deep mining combined with rainwater infiltration during the rainy season, foundation settlement exhibits characteristics of rapid, small-scale abrupt changes, lack of warning, and difficulty in rebound, placing higher demands on traditional building monitoring systems.

[0003] Currently, common methods for monitoring building settlement mostly employ fixed-point displacement sensors, manual periodic measurements, or GNSS positioning. These methods suffer from drawbacks such as low monitoring frequency, limited spatial coverage, and inability to respond in real-time to multi-source coupled scenarios. They are particularly ineffective in predicting sudden, micro-scale, and multi-source induced complex settlement problems. Furthermore, existing technologies fail to fully integrate soil mechanics models with time-series data, making it difficult to form a dynamically evolving settlement prediction model, resulting in significant prediction errors.

[0004] Therefore, there is a need for a method to predict the foundation settlement of coal mine buildings by integrating multi-source intelligent sensors, possessing high-frequency sampling and multi-dimensional sensing capabilities, and combining machine learning and multi-physics field coupling analysis, so as to improve the prediction accuracy and response time under small settlement levels and ensure the safe operation of buildings in the mining area. Summary of the Invention

[0005] The purpose of this invention is to provide a coal mine building foundation settlement prediction system and method based on multi-source intelligent sensors to address the shortcomings in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting foundation settlement of coal mine buildings based on multi-source intelligent sensors, comprising:

[0007] S100. Deploy a multi-source sensor cluster in the foundation area of ​​the coal mine building to be tested to collect full-field non-settlement characteristic signals within a fixed time period and construct a multi-source raw dataset D0.

[0008] S200. Perform characteristic frequency decoupling and phase correlation analysis on the multi-source raw dataset D0, and extract the non-displacement feature sequence matrix F1={f1,f2,...,fi,...,fk} related to the settlement cause, where each fi corresponds to a higher-order perturbation mode of a physical field;

[0009] S300. Input the non-displacement feature sequence matrix F1 into the deep time-map neural network model based on the extended memory gating structure. Combine the time evolution weights and the spatial heterogeneous graph node relationships to output the target settlement prediction response value sequence P={p1,p2,...,pi,...,pn}, where pi represents the foundation settlement rate value at the i-th prediction time.

[0010] S400. The target settlement prediction response value sequence P is dynamically weighted and compared with the measured displacement sensor data. The induced response map is reconstructed through error inversion, and the deep foundation response factor group R={r1,r2,...,ri,...,rm} is extracted, where each ri represents a non-surface hidden factor affecting settlement.

[0011] S500. Based on the correlation between the deep foundation response factor group R and the target settlement prediction response value sequence P, a foundation settlement risk prediction function is constructed, and a risk level distribution map is generated in the entire monitoring area for graded alarm and pre-control point deployment.

[0012] Preferably, in S100, the multi-source sensor cluster includes a settlement displacement sensor, a tilt sensor, a soil water potential sensor, a pressure-compensated pore water pressure gauge, and a high-sensitivity micro-vibration sensor.

[0013] Preferably, in S200, continuous wavelet transform is performed on the multi-source original dataset D0, and Morlet mother wavelet function is used to perform multi-scale decomposition on the time series signals of various sensors to extract their dominant frequency components and corresponding instantaneous phase time series, wherein the dominant frequency component is the frequency center value in the frequency band where the cumulative energy value accounts for more than 95%.

[0014] Preferably, in S200, a frequency response matrix based on a time synchronization window is constructed, and the phase consistency between the dominant frequencies of different physical quantities is quantitatively analyzed using the phase synchronization index to identify highly correlated feature pairs with a synchronization index greater than 0.8, so as to construct a heterogeneous physical quantity coupling spectrum.

[0015] Preferably, in the heterogeneous physical quantity coupling spectrum, the nodes represent the dominant frequency components of each physical quantity, and the edge weight is the ratio of the phase synchronization index to the spatial distance weighting function. The set of nodes with a degree value of not less than 3 and an average edge weight of greater than 0.85 is selected as the set of highly coupled feature nodes. For the physical quantities corresponding to the highly coupled nodes, the perturbation energy sequence, phase jump point density and frequency center drift rate in the dominant frequency band are extracted and normalized to form a non-displacement feature sequence matrix F1.

[0016] Preferably, in S300, the constructed graph neural network model is a deep temporal graph neural network model that combines graph convolution and extended memory gating structure. Each graph node contains physical type encoding, spatial coordinates and temporal perturbation features. The edge weights are jointly defined by the phase synchronization index and the node spatial distance. The temporal evolution weights use an exponential decay function to weight and encode the features of historical time steps.

[0017] Preferably, the extended memory gating structure includes an input gate, a forget gate, and a physical relevance attention gate, used to dynamically adjust the retention and update status of historical memory based on the perturbation change rate and the coupling strength of adjacent edges.

[0018] Preferably, in S400, a weighted feature analysis is performed on the settlement prediction error sequence using a sliding window method to extract three error indicators: local mean, variance, and maximum value. Ridge regression is then used to model the non-displacement feature sequence and the error evolution trend, identify perturbation factors that are highly correlated with the error evolution, and map them into an inducement response map in the heterogeneous graph structure.

[0019] Preferably, key disturbance nodes with activation frequencies exceeding a preset threshold are extracted from the induced response map to form a deep foundation response factor group R. Based on its activation frequency, disturbance amplitude, and evolution trend, a factor feature vector is constructed and input into a multivariate logistic regression model. Combined with the settlement prediction response value sequence P, a foundation settlement risk score sequence and level label are generated to complete the drawing of the full-field risk distribution map and the output of the early warning response.

[0020] This invention also provides a coal mine building foundation settlement prediction system based on multi-source intelligent sensors, comprising:

[0021] The data acquisition module deploys a cluster of multi-source sensors in the foundation area of ​​the coal mine building to be tested, which is used to collect non-settlement characteristic signals of the whole field within a fixed time period and construct a multi-source raw dataset D0.

[0022] The analysis module performs characteristic frequency decoupling and phase correlation analysis on the multi-source raw dataset D0, and extracts the non-displacement feature sequence matrix F1={f1,f2,...,fi,...,fk} related to the settlement cause, where each fi corresponds to a higher-order perturbation mode of a physical field;

[0023] The prediction module inputs the non-displacement feature sequence matrix F1 into a deep time-map neural network model based on an extended memory gating structure. Combining the temporal evolution weights and the spatial heterogeneous graph node relationships, it outputs the target settlement prediction response value sequence P={p1,p2,...,pi,...,pn}, where pi represents the foundation settlement rate value at the i-th prediction time.

[0024] The factor extraction module dynamically weights the target settlement prediction response value sequence P with the measured displacement sensor data, reconstructs the induced response map through error inversion, and extracts the deep foundation response factor group R={r1,r2,...,ri,...,rm}, where each ri represents a non-surface hidden factor affecting settlement;

[0025] The foundation settlement risk prediction module constructs a foundation settlement risk prediction function based on the correlation between the deep foundation response factor group R and the target settlement prediction response value sequence P, and generates a risk level distribution map in the entire monitoring area for graded alarm and pre-control point deployment.

[0026] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0027] 1. This invention constructs a data-driven settlement prediction system based on a multi-source intelligent sensor cluster, enabling the identification and trend modeling of micro-disturbances in the non-obvious settlement stage of coal mine building foundations. This overcomes the technical bottlenecks of traditional methods, such as dependence on obvious settlement, response lag, and untraceable causes. Compared to existing data sources that rely solely on displacement observations, this method introduces non-displacement physical quantities such as dip angle, pore pressure, water potential, and microseismic activity as the basis for cause detection. Combined with dominant frequency extraction and phase coupling analysis, it comprehensively uncovers the multi-physics field collaborative disturbance characteristics in the precursor stage of settlement, significantly improving the foresight of settlement identification.

[0028] 2. This invention effectively models the spatiotemporal correlation of multi-source disturbances in heterogeneous foundations by integrating a deep temporal neural network model with an extended memory gating structure. It also constructs an interpretable causal response map through an error inversion mechanism, extracting deep foundation response factors and realizing a closed-loop mechanism of prediction-interpretation-early warning. The risk assessment stage introduces a factor-driven risk function and dynamic threshold strategy, combining adaptability and real-time performance. This enables precise classification and optimized placement of settlement risks, improving the intelligence level and decision support capabilities of foundation safety monitoring in complex geological environments. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0030] Figure 1 This is a flowchart of the method of the present invention.

[0031] Figure 2 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Example 1, please refer to Figure 1 As shown in this embodiment, the method for predicting foundation settlement of coal mine buildings based on multi-source intelligent sensors includes:

[0034] S100. Deploy a multi-source sensor cluster in the foundation area of ​​the coal mine building to be tested to collect full-field non-settlement characteristic signals within a fixed time period and construct a multi-source raw dataset D0.

[0035] In this embodiment, the foundation area of ​​an office building in a coal mining industrial park is used as the monitoring object. This area has a complex geological structure, including old mining subsidence areas, active fault zones, and intermittent groundwater upwelling. The foundation soil settlement exhibits significant heterogeneity and suddenness. To achieve early identification and prediction of the building's foundation settlement behavior, a multi-source intelligent sensor cluster is deployed in the foundation area according to the method proposed in this invention, and a data foundation is constructed for subsequent analysis. The specific implementation steps are as follows:

[0036] Sensor type and function selection:

[0037] Settlement displacement sensor (LVDT type): used to acquire minute displacement changes in the vertical direction of each control node, with a range of 0-10mm and an accuracy of 0.01mm, to provide reference data on actual ground settlement.

[0038] Tilt sensor (MEMS triaxial): used to monitor the slight tilt of the structure or the deformation trend of the soil. It has a high response frequency, strong adaptability, and covers the slope and the bottom of the structure.

[0039] Soil water potential sensor: used to measure the migration trend of water in the shallow and middle soil layers of the foundation, and to determine whether there is upper limit infiltration or changes in buoyancy by combining meteorological fluctuations.

[0040] Air pressure compensated pore water pressure gauge: It is installed at different depths (1.5m, 3m, 5m) under the foundation to monitor the dynamic changes of pore water pressure and automatically eliminate atmospheric pressure interference, reflecting the impact of groundwater activity on the stability of soil structure.

[0041] High-sensitivity microseismic sensors (short-period seismometers): used to capture microseismic activity below the shallow foundation, including low-frequency disturbances, small-scale rock mass slippage, and non-obvious settlement causes such as the release of hidden fracture stress.

[0042] Sensor deployment design: A cross-shaped grid structure is used as a reference point around the center of the target building, with 28 sensors deployed within a 25-meter radius. Redundant sensors are placed at corners, edges, and structurally sensitive areas to ensure data continuity and integrity. All sensors transmit data to the edge data processing module via a wireless LoRa gateway node, avoiding construction disturbances and complex wiring issues.

[0043] Data Acquisition and Integration: All sensors were uniformly sampled at a frequency of 1Hz and synchronized with time calibration. The system ran continuously for 72 hours, recording only the dynamic response of the foundation structure under natural conditions in the absence of significant settlement events, constructing a full-field non-settlement characteristic signal dataset D0. Data content includes, but is not limited to: dip angle change rate, daily pore pressure variation, microseismic spectrum changes, static displacement values, and soil water potential gradient fluctuations.

[0044] In this embodiment, the rate of change of tilt angle is the rate at which the tilt angle of the structure or foundation changes per unit time, reflecting the trend of minute posture deformation. The daily variation amplitude of pore pressure represents the fluctuation amplitude of pore water pressure over 24 hours, indicating the dynamic stability of the groundwater system. The microseismic spectrum variation reflects the frequency drift characteristics of microseismic signal energy within the monitoring period, revealing the stress disturbance behavior within the foundation. The static displacement value is the average foundation displacement when there is no significant activity, serving as a reference indicator for analyzing the settlement initiation point or baseline. The soil water potential gradient fluctuation measures the change in soil water potential difference at different depths in the vertical profile over time, used to identify groundwater seepage or softening risks.

[0045] Dataset standardization: The various raw physical quantities in dataset D0 are normalized and preprocessed to unify the units and timestamp formats, remove human noise and outliers, and generate a standardized raw multi-source monitoring dataset, which serves as the basic input for subsequent feature decoupling and settlement prediction modeling.

[0046] Through the deployment scheme and data acquisition method of this embodiment, the full-field non-settlement response characteristics covering foundation physical disturbance, boundary coupling, and external environmental factors can be obtained without relying on actual foundation settlement events. This solves the lag problem of "analysis can only be performed after settlement occurs" in traditional methods, and provides data support and causal sources for subsequent microscale settlement trend prediction models.

[0047] S200. Perform characteristic frequency decoupling and phase correlation analysis on the multi-source original dataset D0, and extract the non-displacement feature sequence matrix F1={f1,f2,...,fi,...,fk} related to the settlement cause, where each fi corresponds to a higher-order perturbation mode of a physical field.

[0048] For each type of sensor time-series data in D0, the Morlet mother wavelet function is selected as the basic analysis function to perform continuous wavelet transform on the signal sequence. This process obtains the energy distribution and temporal evolution characteristics of the signal in different frequency bands by projecting the time-series signal onto the wavelet space at different scales. For tilt sensor signals, pore water pressure signals, soil water potential signals, and microseismic signals, their time-frequency spectra from the 1st to the 72nd hour are obtained, and the dominant frequency components are identified by local energy peak extraction. The dominant frequency component is defined as the center value of the frequency band in which the cumulative energy value is above 95%. Simultaneously, the instantaneous phase value of the corresponding frequency is deduced from the wavelet coefficients to construct the corresponding instantaneous phase time series for subsequent phase consistency quantification analysis.

[0049] All sensor data were standardized into 30-minute time windows to construct frequency response matrices for each physical quantity within the synchronization period. Each row of this matrix corresponds to a sensor physical quantity (such as tilt angle, water potential, pore pressure, etc.), and each column corresponds to the power density of the dominant frequency in the corresponding time window.

[0050] The value of each matrix element is the mean power spectral density of the dominant frequency of the physical quantity within the time window, which reflects the dominant response intensity of each physical quantity in a specific time period.

[0051] For the instantaneous phase sequences of the dominant frequencies of the physical quantities obtained above, the coupling relationship is identified using the phase synchronization index calculation method. The phase synchronization index is defined as the absolute value of the cosine average of the instantaneous phase difference between two frequency signals within a certain time period, used to measure whether the two signals change synchronously in that frequency band. Specifically, the calculation method is as follows: Let φ1(t) and φ2(t) be the instantaneous phase sequences of the two signals, then their phase synchronization index is: Phase synchronization index = absolute value (the cosine function acts on the phase difference and the average value is calculated within the selected time period); if the phase synchronization index is greater than 0.8, then the two physical quantities are considered to have strong phase consistency in that frequency band and belong to a potential coupling pair.

[0052] The identified pairs of high-phase-synchronous physical quantities are used as edges between nodes in the graph to construct a heterogeneous physical quantity coupling spectrum. Each node in the graph represents a frequency component of a physical quantity, and the weight of an edge is equal to its corresponding phase synchronization index. This graph is used to reveal the possible coupling response structures between different physical fields.

[0053] Subsequently, a set of nodes in the graph with a degree value (number of connected edges) greater than or equal to 3 and an average edge weight greater than 0.85 is selected as the candidate input set for subsequent non-displacement feature extraction, ensuring that the selected features have multi-party consistency response capability.

[0054] In the aforementioned candidate input set, for each highly coupled node physical quantity, three feature indicators are extracted: the perturbation energy sequence, the phase jump point density, and the frequency center drift rate within its corresponding dominant frequency band. These indicators are then standardized to form a non-displacement feature sequence matrix F1. Each feature vector fi represents a higher-order perturbation mode of a physical quantity at the dominant frequency, used to characterize its potential driving behavior in the precursory settlement process.

[0055] After determining the dominant frequency bands of various non-displacement physical quantities (such as tilt angle, pore pressure, water potential, and microseismic activity), their dynamic response characteristics within these frequency bands are extracted to characterize their implicit relationship with settlement inducing factors. The following three characteristics are all calculated based on a sliding time window.

[0056] A perturbation energy sequence refers to the time-frequency local energy value variation of a physical quantity within a specific dominant frequency band, reflecting the evolution trend of the physical field perturbation intensity. Extraction method:

[0057] Perform a continuous wavelet transform on the original time series signal and select the identified dominant frequency band (e.g., 0.5–2 Hz). Within this frequency band, integrate the wavelet energy density coefficients and calculate the local energy value within each time window (e.g., 5 minutes). The results constitute a perturbation energy sequence E={e1,e2,...,en}, which reflects the dynamic activation degree of the physical quantity at the dominant frequency.

[0058] Phase jump density represents the number of jump points (i.e., abrupt or rapid phase changes) in the instantaneous phase curve per unit time, and is an indicator of nonlinear coupling interference or potential instability. Extraction method:

[0059] Extract the instantaneous phase sequence corresponding to the dominant frequency from the wavelet transform results; perform first-order difference on the phase time sequence and record all points with a difference greater than π / 4; count the number of such phase jump points per minute to obtain the phase jump point density sequence D={d1,d2,...,dn}.

[0060] The frequency center drift rate reflects the rate of change of the dominant frequency of a physical signal within a continuous time window, describing the stability or drift trend of the main response frequency band when a physical quantity is affected by external coupling. Extraction method:

[0061] The dominant frequency of the signal within each sliding time window is located, and the frequency center value (e.g., the centroid of the power spectrum) is calculated. The frequency center values ​​between two consecutive time windows are differencing and divided by the time window width to obtain the drift rate. A frequency drift rate sequence S={s1,s2,...,sn} is formed, indicating whether a systematic shift has occurred in the dominant response frequency band of the physical quantity. The calculation method is as follows: the frequency drift rate at each time point is equal to the frequency center of the current time window minus the frequency center of the previous time window, and then divided by the interval between the two time windows.

[0062] The above three higher-order disturbance features were extracted using sliding window time-frequency analysis, and respectively characterized the dynamic changes of non-displacement physical quantities from three dimensions: energy response intensity, nonlinear catastrophe behavior, and frequency stability. By comparing the evolution trends of these features among multiple physical quantities, disturbance sources highly correlated with the causes of foundation settlement can be effectively identified, serving as the core input of the settlement prediction model.

[0063] Finally, the non-displacement feature sequence matrix F1={f1,f2,...,fi,...,fk} is obtained and used as input for the subsequent settlement trend modeling stage.

[0064] S300. Input the non-displacement feature sequence matrix F1 into the deep time-map neural network model based on the extended memory gating structure. Combine the temporal evolution weights and the spatial heterogeneous graph node relationships to output the target settlement prediction response value sequence P={p1,p2,...,pi,...,pn}, where pi represents the foundation settlement rate value at the i-th prediction time.

[0065] First, based on the source physical quantity (such as tilt angle, pore water pressure, soil water potential, microseismic spectrum) of each eigenvector in the non-displacement feature sequence matrix F1={f1,f2,...,fi,...,fk} and its corresponding sensor deployment location, a heterogeneous graph structure G(V,E) is constructed, where: each node vi∈V represents an eigenvector fi; node attributes include:

[0066] Physical source type coding (e.g., inclination angle is 1, pore pressure is 2, microseismic is 3, etc.);

[0067] The spatial coordinates of the sensor (xi, yi);

[0068] Eigenvalue sequences (such as perturbation energy, phase jump point density, frequency center drift rate);

[0069] An edge eij∈E between nodes is established when the following conditions are met:

[0070] The spatial Euclidean distance Dij between the sensors corresponding to nodes vi and vj is less than or equal to 5 meters;

[0071] The phase synchronization index of the dominant frequency of the two nodes is higher than 0.85 (which can be obtained by averaging the cosine of the continuous phase difference).

[0072] The edge weight Wij is calculated as follows: Let the phase synchronization index be Sij and the spatial distance be Dij, then the edge weight is Wij = Sij ÷ (1 + Dij).

[0073] The graph structure is represented in the form of an adjacency matrix for use as input to the graph neural network. The node attributes are encoded as feature matrices of a uniform dimension. The graph structure input to the network is G=(A,X), where A is the edge weight matrix and X is the node feature matrix.

[0074] To express the dynamic evolution of features over time, a temporal evolution weighting mechanism is introduced to extend the static graph structure into a time graph sequence:

[0075] For each node fi, extract its historical value sequence over T consecutive time steps, where T can be set to 12, representing the past 6 hours, and each step is 30 minutes.

[0076] Constructing a time window matrix ,in This represents the eigenvalue at time step t; the time evolution weight wt is set to exponential decay. Where α is the initial weight (which can be set to 1) and β is the decay coefficient (which can be set to 0.9); finally, the weighted features fused under the historical time window are expressed as: The weighted historical features of all nodes are merged to form the time-series graph node features, which are then input into the subsequent model. This step encodes the trend information of non-displacement disturbance features in the time dimension into each graph node.

[0077] This embodiment employs a neural network model that integrates a graph convolution structure and an extended gated recurrent unit (E-GRU).

[0078] The network structure includes the following layers:

[0079] Graph Convolutional Layer (GCN Layer):

[0080] The system receives a graph structure G=(A,X') as input and performs feature aggregation calculation based on the adjacency edge weights for each node. This calculation is performed as follows: the embedding of each node vi is represented as a weighted average of its own features and the features of its neighboring nodes, with the weights derived from the edge weight matrix A.

[0081] Extended Memory Gated Unit Layer (E-GRU Layer): Based on the traditional gated recurrent unit, a physical correlation attention gate is added to adjust the memory retention level based on the coupling strength of adjacent nodes.

[0082] Within each time step, perform the following steps:

[0083] Input gate: Determines the amount of information input based on the rate of change of the current node's features;

[0084] Forgetting Gate: The degree of historical memory erasure is determined by fluctuations in the time series.

[0085] Note the gate: Combine the average weight of all adjacent edges of this node in the edge weight matrix to enhance the influence of nodes with strong correlation with other physical quantities on the retention of memory state;

[0086] Update state: Integrate the above three factors and update the node's memory state vector.

[0087] Output layer: Perform linear transformation and activation on the node state vector of the last time step to obtain the predicted settlement rate.

[0088] The final model takes the full map state within each time window as input and outputs a sequence of predicted settlement rates for the next n time steps, P={p1,p2,...,pi,...,pn}, where pi represents the foundation settlement rate value at the i-th prediction time.

[0089] The output frequency and time window remain consistent with the input, such as updating the prediction results every 30 minutes;

[0090] The network training uses measured ground settlement data as a supervision signal, and the loss function is weighted mean square error, prioritizing the fitting of settlement abrupt change points.

[0091] In this embodiment, the training method includes: the network training uses the Adam optimizer, and the learning rate is initially set to 0.001; all network parameters are initialized using a Xavier uniform distribution; the dataset is divided into 70% training, 15% validation, and 15% testing; the model performance evaluation metrics include mean squared error, mean absolute error, and early warning rate.

[0092] S400. The target settlement prediction response value sequence P is dynamically weighted and compared with the measured displacement sensor data. The induced response map is reconstructed through error inversion, and the deep foundation response factor group R={r1,r2,...,ri,...,rm} is extracted, where each ri represents a non-surface hidden factor affecting settlement.

[0093] First, obtain the settlement prediction response value sequence P, and then obtain the measured settlement rate data D={d1,d2,...,dn} collected by LVDT settlement sensors deployed at the center of the building foundation during the corresponding time period. Calculate the error between the predicted and measured values, and construct the prediction error sequence Eabs={e1,e2,...,en}, where: each error term... , representing the magnitude of the deviation between the model prediction and the actual measurement at the i-th time step. This error sequence is used for subsequent sliding window analysis to characterize the relationship between the error evolution trend and the perturbation causes.

[0094] To extract the structural trend of prediction error over time, a sliding window weighting process is applied to the error series Eabs to extract local statistical features. The details are as follows:

[0095] Set the window length W=6 (each step is 30 minutes, corresponding to a 3-hour sliding window), and the step size is 1; for each window, the error sequence segment... Calculate the following three characteristic indicators: Local mean μe: reflects the overall level of prediction bias over that period; Local variance : Represents the fluctuation range of prediction error, reflecting the stability of disturbance; Local maximum emax: Represents whether there is sudden prediction distortion; The three indicators of each window are combined into a three-dimensional error vector. The error feature vectors of all windows constitute the time-varying error feature sequence E={E1,E2,...,Ek}, where k=n-W+1. This sequence is used to find which physical perturbation features frequently exhibit correlation peaks during the error evolution process.

[0096] The obtained error feature vector sequence E is used as the dependent variable output vector group, and the values ​​of each feature in the non-displacement disturbance feature sequence matrix F1 within the corresponding time window are used as the explanatory variable input matrix. Each fi represents a non-displacement disturbance mode feature, such as: the disturbance energy sequence in the anomaly region of the dip angle change rate; the frequency sequence of pore pressure fluctuation points; the drift rate of the microseismic spectrum center, etc.

[0097] The modeling steps are as follows:

[0098] Unified Dimension: Locally aggregate all features in F1 within each error time window (e.g., average or maximum value) to align with the time dimension of E;

[0099] Regression method selection: Ridge regression is used for modeling, which is superior to linear regression and has the ability to resist multicollinearity. The regression target is each component of E.

[0100] Standardization: All feature values ​​are normalized to the [0,1] interval to avoid the influence of feature scale on weights; Feature weight extraction: Record the relationship between each non-displacement feature and each error index. The regression coefficient βi; if the absolute value of the average regression coefficient of a certain feature is greater than 0.6, or the regression coefficient exceeds 0.5 in more than 3 consecutive windows, then the perturbation feature is considered to have a significant driving force on the error and is marked as a "key perturbation factor".

[0101] The identified key perturbation features are located at their node positions vi in ​​the original heterogeneous graph G=(V,E), and all their directly connected neighboring nodes N(vi) are obtained. The causal response subgraph GR=(VR,ER) is constructed as follows:

[0102] The node set VR includes all critical perturbation nodes and their first-order physical adjacency nodes; the edge set ER inherits the edge structure and edge weights from the original graph G, representing the physical field coupling relationship; for each node vj in GR, the frequency Fj at which it is identified as a "critical perturbation node" within multiple error windows is counted. Let the total number of sliding windows be K, if... If the value is ≥0.3 (i.e., active within 30% or more of the window), then the disturbance characteristic corresponding to vj is identified as the foundation response factor rj and included in the foundation response factor group R={r1,r2,...,rm}. An example of the physical meaning of the response factor rj is given below:

[0103] r1: The microseismic dominant frequency drift rate increases significantly, suggesting intensified fracture activity or stress release;

[0104] r2: The pore water pressure exhibits a nonlinear abrupt change in the middle layer of soil, indicating that potential seepage causes structural instability;

[0105] r3: The amplitude of the rate of change of tilt angle increases but the direction frequently reverses, indicating that local deformation and springback have occurred in the structure.

[0106] r4: The soil water potential gradient remained high for a long time, suggesting the existence of a continuous upwelling path for groundwater.

[0107] These factors constitute the "physical cause inversion interpretation layer" in settlement prediction, which can be used for subsequent steps such as early warning identification, anomaly location, and monitoring strategy optimization.

[0108] In this embodiment, by using the prediction error as a reverse signal source, the action path of the disturbance factor in settlement modeling is dynamically tracked, realizing the cross-domain modeling strategy of "inverting causality from error". It has physical interpretability, real-time adaptability and graph structure visualization capability, and is suitable for complex situations of multi-field coupling and weak signal disturbance in coal mine foundation settlement.

[0109] S500. Based on the correlation between the deep foundation response factor group R and the target settlement prediction response value sequence P, a foundation settlement risk prediction function is constructed, and a risk level distribution map is generated in the entire monitoring area for graded alarm and pre-control point deployment.

[0110] This embodiment aims to construct a settlement risk prediction function Ψ(R,P) with dynamic threshold and adaptive scoring capability based on the aforementioned foundation response factor group R and target settlement prediction response value sequence P. This function is used to output accurate risk level labels and scores to assist in the deployment of foundation structure safety early warning and response strategies.

[0111] For each factor ri (i=1 to m) in the foundation response factor group R, the following three core physical characteristic indicators are extracted from the analysis period:

[0112] The activation frequency Ai is defined as: the number of times Ni, in which factor ri is identified as a high-activation perturbation factor in the past K sliding time windows, divided by the total number of time windows K;

[0113] The perturbation amplitude Mi is defined as the time mean of the perturbation energy sequence of the physical quantity corresponding to ri within the dominant frequency band. The method for obtaining it is to integrate the wavelet power spectral density within each time window within the dominant frequency band and then average it over the entire period.

[0114] The evolution trend Ti is defined as the linear regression slope of the perturbation amplitude change curve of ri within the analysis period. The calculation method is as follows: take the perturbation amplitude time series as the dependent variable and the time index as the independent variable, perform least squares fitting, and extract the first-order slope coefficient.

[0115] The above three indicators are combined to form a factor feature vector xi=[Ai,Mi,Ti]. Then, min-max standardization is performed on all features, that is:

[0116] Let the maximum value of a certain indicator be max, and the minimum value be min;

[0117] Perform normalization calculations on any numerical value v:

[0118] Standardized value = .

[0119] Finally, the normalized factor feature vector set XR={x1,x2,...,xm} is obtained, which is used for the construction of the risk function.

[0120] A multivariate logistic regression model is constructed using the factor feature vector set XR as input and the target settlement prediction response value sequence P as the reference output. Model input construction method:

[0121] Concatenate the feature vectors of all response factors within each time step ti into a set of input vectors Xi;

[0122] If there are m response factors, each with 3 indicators, the dimension of the input vector at each time step is 3×m.

[0123] Model output definition: The target output of the model is a binary classification label Yi, indicating whether there is a settlement risk event at this time step;

[0124] The risk event is defined as whether the predicted settlement rate pi exceeds the historically set "warning threshold" (e.g., pi > 2.0 mm / hour);

[0125] The corresponding output risk score Si ∈ (0, 1), which is calculated by the logistic function.

[0126] Model function form: The input vector Xi at each time step takes the inner product with the weight vector W obtained through training, and the result is converted into a risk score Si through the sigmoid function;

[0127] The sigmoid function is: ; where b is the bias term and e is the base of the natural logarithm. The model parameters W and b are trained using historical data, and the goal is to minimize the cross-entropy loss function.

[0128] To achieve dynamic adjustment of the risk level, a quantile adaptive threshold division strategy based on the historical prediction error distribution is introduced:

[0129] Collect the prediction error sequence Ehist of the most recent M days (i.e., the absolute value of the difference between the predicted value and the measured value); construct a probability distribution histogram for this sequence.

[0130] Set the risk score boundaries corresponding to low risk, medium risk, and high risk as: P1 = the 70th percentile; P2 = the 90th percentile; obtain the dynamic thresholds: S < P1 → low risk level; P1 ≤ S < P2 → medium risk level; S ≥ P2 → high risk level. The threshold is recalculated every 7 days to make the model adapt to changes in prediction performance.

[0131] Package the aforementioned model logic into a risk prediction function Ψ(R, P), and its definition is as follows:

[0132] Input parameters: the ground response factor group R (including factor state characteristics), the predicted settlement response value sequence P;

[0133] Internal structure: factor index extraction and standardization; multivariate logistic regression model inference; quantile threshold division of risk levels;

[0134] Output: Risk score sequence S={s1,s2,...,sn}; risk level label L={l1,l2,...,ln}, where each li∈{mild, moderate, severe}. This function can be executed once in real time within each prediction period to generate a time-evolving settlement risk sequence, which can be used for intelligent early warning, site adjustment, and engineering response decision support.

[0135] This method, for the first time, uses the activity, disturbance intensity, and evolution direction of the foundation response factors as quantifiable modeling features. Through a logistic regression model and an adaptive threshold method driven by historical errors, a learnable, transferable, and interpretable settlement risk function Ψ(R,P) is formed, which significantly improves the early warning capability and risk classification sensitivity of the prediction system under sudden settlement conditions.

[0136] Example 2, please refer to Figure 2 As shown in this embodiment, the coal mine building foundation settlement prediction system based on multi-source intelligent sensors includes:

[0137] The data acquisition module deploys a cluster of multi-source sensors in the foundation area of ​​the coal mine building to be tested, which is used to collect non-settlement characteristic signals of the whole field within a fixed time period and construct a multi-source raw dataset D0.

[0138] The analysis module performs characteristic frequency decoupling and phase correlation analysis on the multi-source raw dataset D0, and extracts the non-displacement feature sequence matrix F1={f1,f2,...,fi,...,fk} related to the settlement cause, where each fi corresponds to a higher-order perturbation mode of a physical field;

[0139] The prediction module inputs the non-displacement feature sequence matrix F1 into a deep time-map neural network model based on an extended memory gating structure. Combining the temporal evolution weights and the spatial heterogeneous graph node relationships, it outputs the target settlement prediction response value sequence P={p1,p2,...,pi,...,pn}, where pi represents the foundation settlement rate value at the i-th prediction time.

[0140] The factor extraction module dynamically weights the target settlement prediction response value sequence P with the measured displacement sensor data, reconstructs the induced response map through error inversion, and extracts the deep foundation response factor group R={r1,r2,...,ri,...,rm}, where each ri represents a non-surface hidden factor affecting settlement;

[0141] The foundation settlement risk prediction module constructs a foundation settlement risk prediction function based on the correlation between the deep foundation response factor group R and the target settlement prediction response value sequence P, and generates a risk level distribution map in the entire monitoring area for graded alarm and pre-control point deployment.

[0142] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for predicting foundation settlement of coal mine buildings based on multi-source intelligent sensors, characterized in that: include: S100. Deploy a multi-source sensor cluster in the foundation area of ​​the coal mine building to be tested to collect full-field non-settlement characteristic signals within a fixed time period and construct a multi-source raw dataset D0. S200. Perform characteristic frequency decoupling and phase correlation analysis on the multi-source raw dataset D0, and extract the non-displacement feature sequence matrix F1={f1,f2,...,fi,...,fk} related to the settlement cause, where each fi corresponds to a higher-order perturbation mode of a physical field; S300. Input the non-displacement feature sequence matrix F1 into the deep time-map neural network model based on the extended memory gating structure. Combine the time evolution weights and the spatial heterogeneous graph node relationships to output the target settlement prediction response value sequence P={p1,p2,...,pi,...,pn}, where pi represents the foundation settlement rate value at the i-th prediction time. S400. The target settlement prediction response value sequence P is dynamically weighted and compared with the measured displacement sensor data. The induced response map is reconstructed through error inversion, and the deep foundation response factor group R={r1,r2,...,ri,...,rm} is extracted, where each ri represents a non-surface hidden factor affecting settlement. S500. Based on the correlation between the deep foundation response factor group R and the target settlement prediction response value sequence P, a foundation settlement risk prediction function is constructed, and a risk level distribution map is generated in the entire monitoring area for graded alarm and pre-control point deployment.

2. The method for predicting foundation settlement of coal mine buildings based on multi-source intelligent sensors according to claim 1, characterized in that: In S100, the multi-source sensor cluster includes a settlement displacement sensor, a tilt sensor, a soil water potential sensor, a pressure-compensated pore water pressure gauge, and a high-sensitivity micro-vibration sensor.

3. The method for predicting foundation settlement of coal mine buildings based on multi-source intelligent sensors according to claim 1, characterized in that: In S200, continuous wavelet transform is performed on the multi-source original dataset D0. The Morlet mother wavelet function is used to perform multi-scale decomposition on the time series signals of various sensors to extract their dominant frequency components and corresponding instantaneous phase time series. The dominant frequency component is the frequency center value in the frequency band where the cumulative energy value accounts for more than 95%.

4. The method for predicting foundation settlement of coal mine buildings based on multi-source intelligent sensors according to claim 3, characterized in that: In S200, a frequency response matrix based on a time synchronization window is constructed, and the phase consistency between the dominant frequencies of different physical quantities is quantitatively analyzed using the phase synchronization index. Highly correlated feature pairs with a synchronization index greater than 0.8 are identified to construct a heterogeneous physical quantity coupling spectrum.

5. The method for predicting foundation settlement of coal mine buildings based on multi-source intelligent sensors according to claim 4, characterized in that: In the heterogeneous physical quantity coupling spectrum, nodes represent the dominant frequency components of each physical quantity, and the edge weight is the ratio of the phase synchronization index to the spatial distance weighting function. The set of nodes with a degree value of not less than 3 and an average edge weight of greater than 0.85 is selected as the set of highly coupled feature nodes. For the physical quantities corresponding to the highly coupled nodes, the perturbation energy sequence, phase jump point density and frequency center drift rate in the dominant frequency band are extracted and normalized to form a non-displacement feature sequence matrix F1.

6. The method for predicting foundation settlement of coal mine buildings based on multi-source intelligent sensors according to claim 1, characterized in that: In S300, the constructed graph neural network model is a deep temporal graph neural network model that combines graph convolution and extended memory gating structure. Each graph node contains physical type encoding, spatial coordinates and temporal perturbation features. The edge weights are jointly defined by the phase synchronization index and the node spatial distance. The temporal evolution weights use an exponential decay function to weight the features of historical time steps.

7. The method for predicting foundation settlement of coal mine buildings based on multi-source intelligent sensors according to claim 6, characterized in that: The extended memory gating structure includes an input gate, a forget gate, and a physical relevance attention gate, which are used to dynamically adjust the retention and update status of historical memory based on the perturbation change rate and the coupling strength of adjacent edges.

8. The method for predicting foundation settlement of coal mine buildings based on multi-source intelligent sensors according to claim 1, characterized in that: In S400, a weighted feature analysis is performed on the settlement prediction error sequence using a sliding window method to extract three error indicators: local mean, variance, and maximum value. Ridge regression is then used to model the non-displacement feature sequence and error evolution trend, identify perturbation factors that are highly correlated with error evolution, and map them into a causal response map in a heterogeneous graph structure.

9. The method for predicting foundation settlement of coal mine buildings based on multi-source intelligent sensors according to claim 1, characterized in that: Key disturbance nodes with activation frequencies exceeding a preset threshold are extracted from the induced response map to form a deep foundation response factor group R. Based on its activation frequency, disturbance amplitude, and evolution trend, a factor feature vector is constructed and input into a multivariate logistic regression model. Combined with the settlement prediction response value sequence P, a foundation settlement risk score sequence and level label are generated, and the entire field risk distribution map is drawn and the early warning response is output.

10. A coal mine building foundation settlement prediction system based on multi-source intelligent sensors, used to implement the coal mine building foundation settlement prediction method based on multi-source intelligent sensors as described in any one of claims 1-9, characterized in that: include: The data acquisition module deploys a cluster of multi-source sensors in the foundation area of ​​the coal mine building to be tested, which is used to collect non-settlement characteristic signals of the whole field within a fixed time period and construct a multi-source raw dataset D0. The analysis module performs characteristic frequency decoupling and phase correlation analysis on the multi-source raw dataset D0, and extracts the non-displacement feature sequence matrix F1={f1,f2,...,fi,...,fk} related to the settlement cause, where each fi corresponds to a higher-order perturbation mode of a physical field; The prediction module inputs the non-displacement feature sequence matrix F1 into a deep time-map neural network model based on an extended memory gating structure. Combining the temporal evolution weights and the spatial heterogeneous graph node relationships, it outputs the target settlement prediction response value sequence P={p1,p2,...,pi,...,pn}, where pi represents the foundation settlement rate value at the i-th prediction time. The factor extraction module dynamically weights the target settlement prediction response value sequence P with the measured displacement sensor data, reconstructs the induced response map through error inversion, and extracts the deep foundation response factor group R={r1,r2,...,ri,...,rm}, where each ri represents a non-surface hidden factor affecting settlement; The foundation settlement risk prediction module constructs a foundation settlement risk prediction function based on the correlation between the deep foundation response factor group R and the target settlement prediction response value sequence P, and generates a risk level distribution map in the entire monitoring area for graded alarm and pre-control point deployment.

Images