Multi-point water level prediction and sudden gushing water early warning method and device

Abstract

The application provides a multi-point water level prediction and sudden gushing water early warning method and device, which belongs to the technical field of the cross of hydrogeology, tunnel engineering and artificial intelligence, wherein the method comprises the following steps: obtaining underground water level time series data of multiple monitoring points in a research area; reconstructing missing data by adopting a multi-scale reconstruction method based on LOESS seasonal-trend decomposition and discrete wavelet transform; constructing a graph structure based on the spatial positions of the monitoring points, and extracting spatial dependence features by using a residual graph convolution network; combining a long short-term memory network and an attention mechanism to capture time sequence dependence relationship; and dividing sudden gushing water disaster warning grades according to water level prediction results and residual threshold values. The application does not need additional variable input, and can realize data reconstruction, rare monitoring network water level and tunnel sudden gushing water prediction, and sensor point optimization only by relying on multi-point observation data, and has good generalization ability and migration.

Description

Technical Field

[0001] This application relates to the field of hydrogeology, and in particular to a method and apparatus for multi-point water level prediction and early warning of sudden water inrush. This application also relates to a computing device and a computer-readable storage medium. Background Technology

[0002] During the construction or operation of underground engineering projects such as tunnels and mines, groundwater, driven by high pressure, breaches weak points in the geological structure and surges into the excavated space in a sudden and explosive manner, forming a sudden water inrush disaster. This disaster often occurs in karst development areas, fault fracture zones, and areas with confined aquifers. It is characterized by its sudden occurrence, strong destructive force, and wide range of impact, making it one of the most serious safety threats in underground engineering construction. As my country's transportation infrastructure and deep resource development continue to extend into complex geological areas, sudden water inrush disasters are becoming more frequent and severe, causing significant casualties and economic losses. Current technologies for monitoring and early warning of sudden water inrush disasters remain insufficient, especially under complex hydrogeological conditions, where the concealment and suddenness of the disaster pose a significant challenge to safety prevention and control.

[0003] Specifically, the occurrence of sudden water inrush disasters is closely related to the state of the groundwater system. Excavation and disturbance during underground engineering disrupt the original balance of the groundwater seepage field, leading to dynamic changes in groundwater level and pressure. When excavation exposes an aquifer or water-conducting channel, groundwater rapidly converges into the excavated space under the drive of pressure difference. Once the water pressure and inrush volume exceed the bearing capacity of the surrounding rock or supporting structure, a sudden water inrush disaster will be triggered. There is a significant dynamic response relationship between changes in inrush volume and the groundwater level of the aquifer. By analyzing the characteristics of water level changes, the inrush and filling patterns and their evolution can be effectively identified.

[0004] However, existing methods for predicting sudden water inrush disasters by groundwater level have two limitations: (1) they rely heavily on external variables such as precipitation, temperature, and soil moisture, which are difficult to obtain long-term consistent high-quality records in many regions, further limiting the performance of the model in data-scarce environments; (2) most methods only analyze water level data from a single monitoring point or discrete monitoring points independently, ignoring the hydraulic connections and spatial correlations between monitoring points, lacking effective fusion and collaborative analysis of multi-source water level data, and making it difficult to accurately capture the spatiotemporal evolution characteristics of the precursor information of sudden water inrush. Summary of the Invention

[0005] In view of this, embodiments of this application provide a method for multi-point water level prediction and sudden water inrush early warning to address the technical deficiencies in the prior art. Embodiments of this application also provide a multi-point water level prediction and sudden water inrush early warning device, a computing device, and a computer-readable storage medium.

[0006] According to a first aspect of the embodiments of this application, a method for multi-point water level prediction and sudden water inrush early warning is provided, including: Historical groundwater data from multiple monitoring points in the tunnel project were acquired, and a multi-point time series dataset was constructed based on the historical groundwater data. The multi-point time series dataset is processed using the LOESS-based seasonal trend decomposition algorithm and discrete wavelet transform to obtain the target time series data. A graph adjacency matrix is ​​constructed based on the coordinates of the monitoring points corresponding to the historical groundwater hydrological data. Based on the residual graph convolutional network, the spatial dependency features of the target time series data are extracted through the graph adjacency matrix, and the output of the residual graph convolutional network is processed based on the long short-term memory network to extract the hidden state sequence with time dependency. The hidden state sequence is weighted and aggregated based on the attention mechanism, and the processing result of the attention mechanism is mapped to the groundwater level prediction result through a fully connected layer. Based on the combined results of the groundwater level prediction and the actual measured water level, early warning information for sudden water inrush disasters is determined.

[0007] Optionally, the process of processing the multi-point time series dataset using a LOESS-based seasonal trend decomposition algorithm and discrete wavelet transform to obtain the target time series data includes: The LOESS-based seasonal trend decomposition algorithm is used to decompose the multi-point time series dataset into trend, seasonal, and residual terms. The residual term is processed based on the discrete wavelet transform to remove high-frequency noise and obtain a denoised residual term. The target time series data is obtained by superimposing the trend term, the seasonal term, and the denoised residual term.

[0008] Optionally, constructing a graph adjacency matrix based on the coordinates of the monitoring points corresponding to the historical groundwater hydrological data includes: Calculate the Euclidean distance dᵢ between the coordinates of the i-th monitoring point and the coordinates of the j-th monitoring point; Based on the Euclidean distance dᵢ and the scale parameter σ determined by the kernel density estimation, the graph adjacency matrix is ​​constructed as aᵢ=exp(−dᵢ² / σ²).

[0009] Optionally, the step of extracting the spatial dependency features of the target time series data through the graph adjacency matrix based on the residual graph convolutional network includes: The target time series data is graph convolved using the graph adjacency matrix through the residual graph convolutional network, which includes at least two graph convolutional layers and residual connections, wherein the residuals add the input features to the output features of the graph convolutional layers.

[0010] Optionally, the weighted aggregation of the hidden state sequence based on the attention mechanism includes: Attention weights are calculated for the hidden states of the Long Short-Term Memory network at each time step using an additive attention mechanism or a scaled dot product attention mechanism. The entire sequence of hidden states is weighted and summed according to the attention weights.

[0011] Optionally, the step of combining the predicted groundwater level with the measured groundwater level to determine the early warning information for sudden water inrush disasters includes: By comparing the predicted groundwater level with the measured groundwater level, the predicted residual time series is calculated and the early warning level of sudden water inrush disaster is classified. The groundwater level prediction results are processed using the Monte Carlo Dropout method to calculate the prediction confidence interval. The prediction confidence interval is then correlated with the sudden water inrush disaster warning level to obtain the sudden water inrush disaster warning information.

[0012] Optionally, it also includes: The SHAP method is used to quantify the contribution and interaction of the characteristics of each monitoring point to the groundwater level prediction results, screen the key observation points that affect the early warning of sudden water inrush, and optimize the sensor deployment based on the screening results.

[0013] According to a second aspect of the embodiments of this application, a multi-point water level prediction and sudden water inrush early warning device is provided, comprising: The acquisition module is configured to acquire historical groundwater data from multiple monitoring points in the tunnel project, and to construct a multi-point time series dataset based on the historical groundwater data. The processing module is configured to process the multi-point time series dataset using a LOESS-based seasonal trend decomposition algorithm and discrete wavelet transform to obtain the target time series data. The construction module is configured to construct a graph adjacency matrix based on the coordinates of the monitoring points corresponding to the historical groundwater data. The feature extraction module is configured to extract the spatial dependency features of the target time series data based on the residual graph convolutional network through the graph adjacency matrix, and process the output of the residual graph convolutional network based on the long short-term memory network to extract the hidden state sequence with time dependency. The mapping module is configured to perform weighted aggregation of the hidden state sequence based on the attention mechanism, and map the processing result of the attention mechanism to the groundwater level prediction result through a fully connected layer; The integrated module is configured to combine the predicted groundwater level with the measured groundwater level to determine early warning information for sudden water inrush disasters.

[0014] According to a third aspect of the embodiments of this application, a computing device is provided, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor executes the computer-executable instructions to implement the steps of the multi-point water level prediction and sudden water inrush early warning method.

[0015] According to a fourth aspect of the embodiments of this application, a computer-readable storage medium is provided that stores computer-executable instructions, which, when executed by a processor, implement the steps of the multi-point water level prediction and sudden water inrush early warning method.

[0016] According to a fifth aspect of the present application, a chip is provided that stores a computer program, which, when executed by the chip, implements the steps of the multi-point water level prediction and sudden water inrush early warning method.

[0017] The multi-point water level prediction and sudden water inrush early warning method provided in this application acquires historical groundwater hydrological data from multiple monitoring points in a tunnel project and constructs a multi-point time series dataset based on the historical groundwater hydrological data. The multi-point time series dataset is processed using a LOESS-based seasonal trend decomposition algorithm and discrete wavelet transform to obtain target time series data. A graph adjacency matrix is ​​constructed based on the coordinates of the monitoring points corresponding to the historical groundwater hydrological data. Spatial dependency features of the target time series data are extracted using the graph adjacency matrix based on a residual graph convolutional network, and the output of the residual graph convolutional network is processed using a long short-term memory network to extract hidden state sequences with time dependencies. The hidden state sequences are weighted and aggregated based on an attention mechanism, and the processing result of the attention mechanism is mapped to groundwater level prediction results through a fully connected layer. The groundwater level prediction results are combined with measured water level values ​​to determine sudden water inrush disaster early warning information. No additional variable input is required; data reconstruction, prediction of water levels in scarce monitoring networks and tunnel sudden water inrushes, and optimization of sensor locations can be achieved solely based on multi-point observation data. It has good generalization ability and transferability. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a multi-point water level prediction and sudden water inrush early warning method provided in an embodiment of this application; Figure 2 This is a schematic diagram of time series missing data reconstruction for a multi-point water level prediction and sudden water inrush early warning method provided in an embodiment of this application; Figure 3 This is a structural diagram of a multi-point water level prediction model for a multi-point water level prediction and sudden water inrush early warning method provided in an embodiment of this application; Figure 4 This is a multi-point water level prediction model prediction result diagram of a multi-point water level prediction and sudden water inrush early warning method provided in an embodiment of this application; Figure 5 This is a schematic diagram of the feature importance analysis based on SHAP of a multi-point water level prediction and sudden water inrush early warning method provided in an embodiment of this application; Figure 6 This is a schematic diagram illustrating the importance of observation point features for point optimization in a multi-point water level prediction and sudden water inrush early warning method provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of a multi-point water level prediction and sudden water inrush early warning device provided in an embodiment of this application; Figure 8 This is a structural block diagram of a computing device provided in one embodiment of this application. Detailed Implementation

[0020] Many specific details are set forth in the following description to provide a full understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this application; therefore, this application is not limited to the specific embodiments disclosed below.

[0021] The terminology used in one or more embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the scope of one or more embodiments of this application. The singular forms “a,” “the,” and “the” used in one or more embodiments of this application and in the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” used in one or more embodiments of this application refers to and includes any or all possible combinations of one or more associated listed items.

[0022] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this application, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this application, and similarly, second may also be referred to as first.

[0023] This application provides a method for multi-point water level prediction and sudden water inrush early warning. This application also relates to a multi-point water level prediction and sudden water inrush early warning device, a computing device, and a computer-readable storage medium, which will be described in detail in the following embodiments.

[0024] Figure 1 The flowchart illustrates a multi-point water level prediction and sudden water inrush early warning method according to an embodiment of this application, which specifically includes the following steps: Step S102: Obtain historical groundwater data from multiple monitoring points in the tunnel project, and construct a multi-point time series dataset based on the historical groundwater data; Step S104: The multi-point time series dataset is processed by the LOESS-based seasonal trend decomposition algorithm and discrete wavelet transform to obtain the target time series data. Step S106: Construct a graph adjacency matrix based on the coordinates of the monitoring points corresponding to the historical groundwater data; Step S108: Based on the residual graph convolutional network, extract the spatial dependency features of the target time series data through the graph adjacency matrix, and process the output of the residual graph convolutional network based on the long short-term memory network to extract the hidden state sequence with time dependency. Step S110: The hidden state sequence is weighted and aggregated based on the attention mechanism, and the processing result of the attention mechanism is mapped to the groundwater level prediction result through a fully connected layer; Step S112: Based on the groundwater level prediction results and the measured water level values, determine the early warning information for sudden water inrush disasters.

[0025] The monitoring points are the locations where groundwater level sensors are installed along the tunnel excavation route, fault fracture zones, and water-rich sections. Each point collects independent time-series groundwater hydrological data, and there is a hydraulic connectivity and spatial correlation between the points. The historical groundwater hydrological data consists of long-term continuous time-series monitoring values ​​of groundwater level, water pressure, flow rate, and permeability coefficient collected from each monitoring point. It is composed of historical groundwater level observation data and data on water inrush disasters, including seasonal fluctuations, long-term trends, equipment noise, and abnormal signals preceding sudden water inrush. The multi-point time-series dataset is a two-dimensional structured dataset constructed with the monitoring points as the spatial dimension and the collection time as the temporal dimension. It synchronously stores the hydrological observation values ​​of all points at the same time period, carrying spatiotemporal coupling information of multiple measuring points.

[0026] The Seasonal Trend Decomposition Algorithm (STL) of LOESS is an algorithm for time series decomposition based on Local Weighted Regression LOESS. It can separate non-stationary hydrological time series into long-term trend terms, periodic seasonal terms, and random residual noise terms, adapting to the decomposition of periodic fluctuations in groundwater. Discrete Wavelet Transform (DWT) is a multi-scale time-frequency analysis tool that uses high-pass and low-pass filtering to decompose time series residuals in a hierarchical manner, accurately separating high-frequency random noise from low-frequency precursory abrupt changes in water inrush signals, while preserving the characteristics of local abrupt changes in the time series. The target time series data is clean time series data reconstructed after STL decomposition and wavelet residual denoising, eliminating invalid noise such as sensor electromagnetic interference and acquisition jitter, and fully preserving the characteristics of water level trends, seasonal fluctuations, and precursory changes in water inrush. The coordinates of the monitoring points are the planar / spatial coordinates of each water level sensor in the three-dimensional geological coordinate system of the tunnel, used to quantify the spatial distance between measuring points and the strength of hydraulic connectivity. These coordinates are collected synchronously during the historical groundwater data collection.

[0027] The graph adjacency matrix is ​​a square matrix with monitoring points as graph nodes and the hydraulic connectivity strength of the points as edge weights. The matrix elements represent the degree of mutual influence between two monitoring points on groundwater and serve as the input basis for graph convolutional networks. The residual graph convolutional network, or ResGCN, is a multi-layer graph convolutional network with superimposed residual connections. It relies on the adjacency matrix to mine the nonlinear spatial dependencies between nodes, and the residual connections solve the problems of gradient vanishing and feature degradation in multi-layer graph convolution. The spatial dependency features are the coupling and correlation features between water levels of different monitoring points. Water levels of monitoring points that are close to each other and have strong hydraulic connectivity change synchronously, and faults and fissures will strengthen the spatial correlation of monitoring points. The long short-term memory network, or LSTM, is a recurrent neural network with built-in forget gates, input gates, and output gates. It solves the problem of long-term gradient vanishing and captures the long-term temporal evolution law of groundwater level and the temporal dependency relationship of the gradual precursor of sudden water inrush.

[0028] The hidden state sequence is a high-dimensional feature vector sequence output by the LSTM at each time step, storing the time-dependent features of historical water level time series, including both short-term fluctuations and long-term evolution information. The attention mechanism is an adaptive weighting algorithm that automatically assigns weights to the time series hidden states that contribute more to water level prediction, weakening ineffective stationary time series and amplifying the weights of key time series segments that are precursors to sudden water inrush. The weighted aggregation is a linear weighted summation of the hidden states at all time steps based on the attention weights, fusing effective time series features across all time periods to obtain a global time series representation vector. The fully connected layer is a multi-layer linear mapping network that maps the spatiotemporal fusion feature vector to a standardized groundwater level prediction numerical output. The groundwater level prediction result is the sequence of estimated water level values ​​for future time periods output by the model for each monitoring point. The measured water level value is the real-time groundwater level observation data collected by the sensors at the current moment. The sudden water inrush disaster early warning information is a graded early warning instruction that integrates prediction bias and prediction uncertainty, including warning level, risk location, confidence probability, and emergency response prompts.

[0029] Based on this, historical groundwater level observation data and water inrush disaster data from multiple monitoring points in the tunnel project were acquired, and their geographical locations or tunnel mileage coordinates were recorded to construct a multi-point time series dataset. Specifically, continuous hydrological monitoring data from multiple points along the entire tunnel were collected, and a spatiotemporal two-dimensional time series dataset was constructed by aligning the data with time. The three-dimensional coordinates of the monitoring points were simultaneously entered, and the data were standardized to eliminate dimensional differences. In engineering practice, auxiliary features such as rainfall and surrounding rock permeability coefficients can be added to enhance the richness of the dataset. It should be noted that during the construction of the multi-point time series dataset, outlier detection and removal can be performed on the original observation data of each monitoring point. Short-term missing segments can be filled using linear interpolation or the mean of neighboring points, and Z-score normalization can be performed on the entire time series to eliminate dimensional differences between different points. A preliminary seasonal division is established within the dataset based on the dry season, normal water season, and wet season.

[0030] Subsequently, for missing values ​​in the time series data, the seasonal-trend decomposition (STL) based on LOESS and discrete wavelet transform (DWT) were used to reconstruct the data at multiple scales. Specifically, the time series was first decomposed using LOESS local regression to separate the annual / monthly periodic seasonal fluctuations of groundwater, the long-term slow seepage trend, and the noisy residuals. Then, discrete wavelets, such as the db4 wavelet, were used to decompose the residuals at multiple scales. Soft thresholding was used to filter out high-frequency electromagnetic noise from the sensors. The clean residuals, trend, and seasonal terms were reconstructed to obtain the target time series. Compared with single wavelet or STL decomposition, this method takes into account both periodic extraction and separation of abrupt noise, and is suitable for nonlinear water level fluctuations in karst tunnels.

[0031] Subsequently, in the process of constructing the graph adjacency matrix, in addition to using the coordinates of monitoring points corresponding to historical groundwater hydrological data, a graph structure can also be constructed based on the mileage location in the tunnel project. The monitoring points are used as nodes, and the reciprocal of the distance between points or the Gaussian kernel function value is used as the edge weight to construct a graph adjacency matrix that represents the hydraulic connection between points. Specifically, the pairwise Euclidean distance is calculated based on the three-dimensional coordinates of the measuring points, and the scale parameter σ is automatically solved using kernel density estimation. The association weight of the measuring points is calculated through the Gaussian kernel exponential function to generate the adjacency matrix. This replaces the traditional 0-1 binary adjacency matrix, realizing that the closer the distance and the stronger the hydraulic coupling, the higher the weight, which conforms to the physical law of groundwater seepage. The adaptive σ of kernel density estimation avoids the modeling deviation caused by manually fixed parameters.

[0032] Subsequently, a Residual Graph Convolutional Network (ResGCN) is constructed. This network performs graph convolution operations on the time series data using a graph adjacency matrix and preserves node feature differences through residual connections to extract spatial dependency features. Then, a Long Short-Term Memory (LSTM) network is constructed. The time series data containing spatial features output from the ResGCN is input into the LSTM network to extract dependencies in the temporal dimension. Specifically, a two-layer graph convolution is superimposed with residual connections. The input adjacency matrix is ​​used to perform graph convolution operations on the target time series data. The input features of each layer are added to the convolution output to preserve shallow hydrological spatial features, thus addressing the gradient vanishing problem in deep networks. The spatial features output by the GCN are fed into the LSTM, and a three-gating mechanism is used to capture cross-temporal dependencies, outputting the hidden state sequences at each time step. Existing technologies often employ a single-layer GCN without residual structures, which easily loses multi-point coupling features of fault zones. This scheme, with its residual structure, significantly improves the stability of deep spatial feature extraction.

[0033] Subsequently, the hidden state sequences output by the Long Short-Term Memory Network are weighted and aggregated through an attention mechanism to enhance the representation ability of long-term time dependencies. Specifically, an adaptive attention mechanism is introduced to calculate the weights of the hidden states at each time step, assigning high weights to high-risk precursor time series and low weights to stable baseline time series, and then weighting and summing to fuse global time series features. The predicted water levels at multiple measurement points are then mapped and output through a fully connected layer, thus overcoming the shortcomings of traditional LSTM's equal weighting of time series and weakening of short-term abrupt changes in sudden water inrush signals.

[0034] Finally, the output of the attention mechanism is mapped to the final groundwater level prediction result through a fully connected layer. The water level prediction result is compared with the measured water level value, the prediction residual time series is calculated, and the emergency water inrush disaster warning level is divided. Specifically, the prediction residual is calculated by comparing the predicted water level with the real-time measured water level, and the emergency water inrush warning level is initially divided into four levels according to the residual amplitude. The warning output is corrected by combining the prediction uncertainty.

[0035] In summary, relying solely on multi-point observation data without requiring external variable input, this method is applicable to both groundwater level prediction and tunnel inrush water prediction scenarios. It achieves accurate predictions in data-scarce areas, reduces dependence on the density of external monitoring facilities, and expands the method's applicability. The multi-scale reconstruction method, coupled with STL decomposition and discrete wavelet transform, effectively repairs missing groundwater levels, improves data quality, and provides complete and reliable time-series data for subsequent modeling. The residual graph convolutional network effectively extracts spatial dependency features between monitoring points, and the residual connection mechanism alleviates the over-smoothing problem of the graph convolutional network, maintaining the feature differences between nodes at different monitoring points. Combining LSTM and attention mechanisms, short-term and long-term temporal dependencies are extracted respectively, and the attention mechanism effectively alleviates the information loss problem in long-sequence prediction. The collaborative working mechanism of residual graph convolution, LSTM, and attention mechanisms produces a cumulative performance improvement, outperforming a single LSTM baseline model. Uncertainty analysis can quantify the prediction mean and confidence interval, providing the probability of the warning level and increasing the reliability of the prediction results.

[0036] Furthermore, in step S104, the process of processing the multi-point time series dataset using the LOESS-based seasonal trend decomposition algorithm and discrete wavelet transform to obtain the target time series data is specifically implemented as follows in this embodiment: The multi-point time series dataset is decomposed into a trend term, a seasonal term, and a residual term using the LOESS-based seasonal trend decomposition algorithm. The residual term is processed by the discrete wavelet transform to remove high-frequency noise and obtain a denoised residual term. The trend term, the seasonal term, and the denoised residual term are then superimposed to obtain the target time series data.

[0037] The time series consists of three components: a trend term, representing the slow, long-term changes in groundwater, a smooth time series curve formed by long-term seepage from the tunnel surrounding rock and seasonal changes in groundwater recharge; a seasonal term, representing the periodic fluctuations in groundwater, a fixed-period water level fluctuation signal formed by monthly / seasonal rainfall and surface water recharge; a residual term, representing the remaining components after removing the trend and seasonal terms from the original time series, including sensor random noise and precursor signals of sudden water inrush; a discrete wavelet transform residual denoising process, which involves multi-scale wavelet decomposition of the residual term to distinguish between high-frequency equipment noise figures and low-frequency sudden water inrush mutation coefficients, filtering out noise figures to reconstruct a clean residual; a denoised residual term, representing the residual component that removes random acquisition noise and retains the instantaneous abnormal signal of sudden water inrush; and a process of superimposing the three terms to reconstruct the target time series, which involves linearly adding the trend term, seasonal term, and denoised residual to restore a complete, noise-free water level time series that retains all hydrological characteristics.

[0038] Based on this, STL decomposition is used to decompose time series data into trend, seasonal, and residual terms. Then, discrete wavelet transform is used to perform multi-scale decomposition on the residual term to remove high-frequency noise components. Finally, the reconstructed trend, seasonal, and denoised residual terms are superimposed to obtain the complete time series. Specifically, in the actual STL decomposition process, the trend window period is set to an odd number, and the seasonal window period is set according to the data sampling frequency. The discrete wavelet transform uses Daubechies or Symlet wavelet basis functions, and the number of decomposition levels is adaptively determined according to the data length. The specific calculation of STL decomposition is as follows: , Among them, Y t It is the original time series, corresponding to multi-point time series data, S t It is a seasonal component, corresponding to the seasonal term, T. t It is a trend component, corresponding to the trend term, R. t These are the residual components, corresponding to the residual terms.

[0039] The specific calculation of the discrete wavelet transform is as follows: , Where x(t) represents the original discrete-time sequence, t is the discrete-time index, T represents the total length of the signal x(t), W(m, n) are the wavelet coefficients at scale m and translation n, the parameter m represents the scale factor controlling the wavelet frequency resolution, and n represents the translation parameter determining the wavelet's position in the time domain. The function ψ is the mother wavelet function, which generates the wavelet basis through scaling and translation. The original discrete-time sequence is expressed as: , Where A n (t) represents the approximate component at level n, reflecting the low-frequency trend of the signal, while D i (t) represents capturing the detail components of high-frequency changes at different scales.

[0040] In addition, the MAD median absolute deviation method can be used to adaptively calculate the denoising threshold. The wavelet coefficients of high-frequency noise below the threshold are set to zero, and the denoising residual is obtained by inverse wavelet transform reconstruction. Unlike global Gaussian filtering, wavelet multi-scale decomposition can accurately retain the instantaneous surge water spike signal in the residual and only filter out high-frequency jitter noise.

[0041] Therefore, by setting a LOESS smoothing window adapted to the hydrological time series, the original time series at multiple locations is decomposed point by point, separating the long-term slow seepage trend term, the rainfall-driven periodic seasonal term, and the noisy residual term. Compared with the moving average STL, the LOESS nonparametric fitting does not require a preset periodic function and is adapted to the irregular periodic fluctuations of groundwater in karst tunnels. Discrete wavelet transform is used to decompose the residual term in multiple layers, and the noise-free residual, the original trend term, and the original seasonal term are superimposed step by step to generate a clean and complete target time series. The natural evolution law of groundwater is preserved throughout the process, and only the interference from equipment acquisition is removed, without losing any precursory abrupt changes of sudden water inrush.

[0042] Finally, for the missing values ​​in the time series data, STL and DWT were used to reconstruct the data at multiple scales, and the results are as follows: Figure 2 The diagram illustrates a time-series data reconstruction method for multi-point water level prediction and sudden water inrush early warning. It achieves hierarchical decomposition of the time series into three components, processing them separately to preserve the inherent cycle and long-term seepage patterns of groundwater while completely eliminating high-frequency noise from sensors. Wavelet threshold denoising accurately distinguishes noise from sudden water inrush signals, without erasing short-term, minute precursors to water level surges, thus improving the completeness of precursor identification. Preprocessing significantly improves the time-series signal-to-noise ratio, accelerating subsequent model training convergence, reducing training loss, and further minimizing the error in water level prediction at measurement points. The step-by-step standardized time-series processing workflow is adaptable to mixed monitoring data from different geological sections of tunnels and different sensor models, exhibiting stronger generalization capabilities.

[0043] Furthermore, in step S106, the process of constructing a graph adjacency matrix based on the coordinates of monitoring points corresponding to historical groundwater hydrological data is specifically implemented as follows in this embodiment: Calculate the Euclidean distance dᵢ between the coordinates of the i-th monitoring point and the coordinates of the j-th monitoring point; based on the Euclidean distance dᵢ and the scale parameter σ determined by the kernel density estimation, construct the graph adjacency matrix aᵢ=exp(−dᵢ² / σ²).

[0044] Wherein, the Euclidean distance dᵢ is the straight-line distance between the i-th and j-th monitoring points in the three-dimensional coordinate system, quantifying the spatial proximity of the measuring points; the closer the distance, the stronger the hydraulic connectivity. The kernel density estimation is a parameter-free probability density calculation method that automatically solves the Gaussian kernel scale parameter σ based on the distance distribution of all measuring points, without the need for manual experience-based assignment. The scale parameter σ is the Gaussian kernel bandwidth, which controls the decay rate of the distance-related weights of the measuring points. σ adapts to the distribution density of the measuring points; σ is smaller for dense measuring points and larger for sparse measuring points. The graph adjacency matrix is ​​an N×N square matrix, where N is the total number of monitoring points. The element in the i-th row and j-th column of the matrix is ​​the Gaussian kernel association weight of measuring points i and j, which serves as the input to the graph convolutional network to represent the spatial association of nodes. The Gaussian kernel association weight is an exponential decay function; the larger the Euclidean distance between two points, the closer the weight approaches 0, and the closer the distance approaches 0, the closer the weight approaches 1, accurately mapping the relationship between spatial distance and hydraulic coupling strength.

[0045] Based on this, the three-dimensional spatial coordinates of all water level sensors are extracted, and the Euclidean distance between each pair of spatial lines is calculated. The tunnel elevation difference is simultaneously incorporated, which conforms to the three-dimensional seepage law of groundwater flow and is superior to the two-dimensional distance calculation in the plane. Kernel density fitting is performed on the pairwise distances of all measuring points, and the optimal bandwidth σ is automatically solved based on the variance of the distance distribution. In densely populated sections of measuring points, σ ​​is automatically reduced and the weight difference is amplified. In sparse sections, σ is increased to retain weak hydraulic correlation, replacing the manual fixed parameter scheme and eliminating subjective modeling bias. The distance dᵢ and adaptive σ of each pair of measuring points are substituted into the exponential function to calculate the correlation weight, which is filled into the corresponding position of the square matrix to generate a continuous value adjacency matrix. The matrix weights are continuously distributed from 0 to 1, which accurately quantifies the degree of synchronous change of water level between different measuring points.

[0046] Therefore, the continuous Gaussian kernel adjacency matrix fully represents the hydraulic coupling gradient relationship of the measuring points, which is different from the simple partitioning of binary matrices. The accuracy of extracting spatial dependency features by graph convolution is greatly improved. The kernel density estimation automatically adapts to the density distribution of measuring points, and the adjacency matrix is ​​constructed according to a unified standard for the entire tunnel, eliminating the need for segmented manual parameter adjustment, thus automating and standardizing the modeling. The distance is calculated based on three-dimensional coordinates, which fully considers the obstruction effect of the tunnel elevation difference on groundwater connectivity. The spatial correlation modeling fits the real hydrogeological seepage mechanism. The physical meaning of the adjacency matrix weights is clear, which improves the interpretability of the model and makes it easier for engineers to understand the spatial hydraulic coupling relationship of the measuring points.

[0047] Furthermore, in step S108, the process of extracting the spatial dependency features of the target time series data through the graph adjacency matrix based on the residual graph convolutional network is specifically implemented as follows in this embodiment: The target time series data is graph convolved using the graph adjacency matrix through the residual graph convolutional network, which includes at least two graph convolutional layers and residual connections, wherein the residuals add the input features to the output features of the graph convolutional layers.

[0048] The graph convolutional layer is a network layer that aggregates node features spatially based on the graph adjacency matrix, weighting and fusing hydrological features from adjacent measuring points to extract spatial coupling features. The residual connection representation is a cross-layer direct connection channel that directly adds and fuses the original input features of the graph convolutional layer with the convolutional output features. The residual graph convolutional network is a network that stacks two or more graph convolutional layers and configures residual direct connection channels to extract nonlinear spatial dependency features from multi-point time series data. The input features are the target time series data fed into the graph convolutional layer, i.e., the water level time series vectors after noise reduction at each measuring point. The graph convolutional output features are the multi-layer spatial fusion features after weighted aggregation by the adjacency matrix.

[0049] Based on this, the residual connection adds the input features to the output features of the graph convolutional layer to alleviate the oversmoothing problem in graph convolutional networks and the information decay problem in long short-term memory networks. The residual graph convolutional network includes at least two graph convolutional layers and residual connections. The residual connection adds the input features to the output features of the graph convolutional layer to alleviate the oversmoothing problem in graph convolutional networks and the information decay problem in long short-term memory networks, and maintains the feature differences between different monitoring nodes.

[0050] For example, such as Figure 3 The provided multi-point water level prediction model structure diagram is shown in the multi-point water level prediction and sudden water inrush early warning method. In practical applications, the residual graph convolutional network includes two graph convolutional layers, each with 64 output channels, a ReLU activation function, and a dropout rate of 0.2. The residual connection directly adds the input features to the output features across the graph convolutional layers. The LSTM network has 128 hidden units, 2 layers, and a dropout rate of 0.2.

[0051] Furthermore, in step S110, the process of weighted aggregation of the hidden state sequence based on the attention mechanism is specifically implemented as follows in this embodiment: Attention weights are calculated for the hidden states at each time step of the Long Short-Term Memory network using an additive attention mechanism or a scaled dot product attention mechanism; and the entire sequence of hidden states is then summed based on these attention weights.

[0052] Among them, additive attention, namely Bahdanau attention, uses a two-layer neural network to calculate the similarity weight of the hidden state, which is suitable for scenarios with inconsistent query and key vector dimensions and has stronger nonlinear fitting ability; scaled dot product attention calculates similarity through vector dot product and scales the stable gradient by dividing by the square root of the dimension, which has high parallelism of matrix operations and faster computation efficiency; the time step hidden state is a high-dimensional time series feature vector output by LSTM at each acquisition moment, corresponding to the water level evolution information of different time periods; the attention weight uses a 0~1 normalized coefficient to represent the contribution of the corresponding time step hidden state to the final water level prediction, and the weight is automatically increased during the period of sudden water inrush; the weighted summation and aggregation process is to multiply all time step hidden state vectors by the corresponding attention weight and then sum them to obtain a global feature vector that integrates key time series information of the whole time period.

[0053] Based on this, the attention mechanism employs additive attention or scaled dot product attention to calculate attention weights for the hidden states at each time step of the Long Short-Term Memory network, and then performs a weighted summation of the hidden states to highlight the information contribution of key time steps. Specifically, such as... Figure 3 The multi-point water level prediction model structure diagram provided is shown in the multi-point water level prediction and sudden water inrush early warning method. The attention mechanism adopts an additive attention approach, calculating the attention score through a single hidden layer feedforward network, and then obtaining the attention weights through Softmax function normalization. The final groundwater level prediction result is as follows. Figure 4 The prediction results of the multi-point water level prediction model for a proposed method for multi-point water level prediction and sudden water inrush early warning are shown in the figure.

[0054] Therefore, it automatically distinguishes between stable baseline time series and time series of precursory sudden water inrushes, amplifies the feature weights of risk periods, and significantly improves the sensitivity of identifying small, gradual water level anomalies; it is compatible with both high and low dimensional LSTM hidden states, adaptively switches attention calculation methods, and the model's generalization ability adapts to tunnel engineering with different monitoring sampling frequencies; it integrates all historical time series information without losing early precursor signals of the incubation stage of sudden water inrushes, and compared with traditional schemes that only use the end hidden states, its ability to capture long-term time series dependencies is greatly enhanced; the attention weights are visually interpretable, allowing engineers to intuitively see which time periods water level changes induce the risk of sudden water inrushes, facilitating geological disaster source tracing analysis.

[0055] Furthermore, in step S112, the process of determining the early warning information for sudden water inrush disaster by combining the predicted groundwater level with the measured water level is specifically implemented as follows in this embodiment: By comparing the predicted groundwater level with the measured groundwater level, the predicted residual time series is calculated and the level of sudden water inrush disaster warning is classified. The predicted groundwater level is processed using the Monte Carlo Dropout method to calculate the prediction confidence interval. The prediction confidence interval is then correlated with the level of sudden water inrush disaster warning to obtain the sudden water inrush disaster warning information.

[0056] The predicted residual time series consists of the time-series differences between the predicted water level at the same measuring point and the real-time sensor measured value. A larger residual indicates a deviation from the normal evolution pattern of the water level, and a higher risk of sudden water inrush. The sudden water inrush disaster warning level is a multi-level risk label based on the residual amplitude, typically divided into four levels: safe, concerning, warning, and emergency sudden water inrush, matching the emergency response standards for tunnel engineering. Monte Carlo Dropout (MCDO) involves continuously enabling network dropout with random deactivation during the inference phase, generating multiple sets of prediction results through multiple forward propagations, and using a prediction variance quantification model to predict uncertainty. It belongs to the approximate Bayesian uncertainty estimation method; the prediction confidence interval is the upper and lower limits of the water level prediction value calculated from the mean and variance of multiple Monte Carlo sampling prediction values. The wider the interval, the higher the uncertainty of the model prediction and the lower the credibility of the warning; the correlation between the confidence interval and the warning level is that the final warning output is corrected by combining the risk level determined by the residual and the width of the confidence interval. Under high uncertainty, the warning level is lowered for the same residual to avoid blindly raising the level and triggering false alarms; the sudden flood disaster warning information is a standardized warning output message that integrates the risk level, prediction confidence probability, coordinates of high-risk points, and uncertainty coefficient.

[0057] Based on this, the output of the attention mechanism is mapped to the final groundwater level prediction result through a fully connected layer. The predicted water level result is compared with the measured water level value to calculate the prediction residual time series and classify the emergency warning level for sudden water inrush disaster. Specifically, a prediction residual abnormality threshold is set. When the prediction residual of any monitoring point exceeds the prediction residual abnormality threshold within a continuous time window, it is determined that the prediction residual of the monitoring point has abruptly changed, and the monitoring point is marked as a suspected abrupt change point. For suspected abrupt change points, the rate of change of their measured water level value within a sliding time window is calculated, and multi-level thresholds are set. According to the set multi-level thresholds for prediction residual abnormality, the warning level is divided into green (safe) for low risk, yellow (attention) for medium risk, orange (warning) for high risk, and red (emergency sudden water inrush) for extremely high risk.

[0058] Subsequently, the Monte Carlo dropout method was used to quantify the uncertainty of the prediction results. The uncertainty of the water level prediction results was quantified by coverage and width, and the prediction mean and confidence interval were estimated, which were then converted into the probability of a risk warning level. Specifically, during the testing phase, the dropout layer was randomly deactivated, multiple forward propagations were performed, the prediction mean and confidence interval were calculated, the prediction uncertainty was estimated, and converted into the probability of a risk warning level. The uncertainty analysis used the following formulas to calculate the coverage confidence, average confidence interval width, and average relative deviation: , , , Where L t and U t Let I represent the lower and upper bounds of the prediction interval for the t-th test sample, respectively; m represents the number of test samples; I{·} represents the index function; ε is a small constant introduced to avoid division by zero; ŷ t Let y represent the predicted value of the t-th test sample. t This represents the corresponding observed value.

[0059] Therefore, as Figure 5 The diagram illustrates a feature importance analysis based on SHAP for a multi-point water level prediction and sudden water inrush early warning method. It identifies progressive water level anomalies based on prediction residuals, overcoming the limitation of fixed thresholds in capturing slow, imperceptible water inrush precursors and reducing missed alarms for concealed inrushes. Monte Carlo Dropout quantifies prediction confidence intervals, distinguishing between high-confidence and low-confidence prediction scenarios and dynamically adjusting early warning levels, significantly reducing false alarms in geologically complex sections. Early warning information simultaneously includes risk level and prediction confidence, providing differentiated emergency response guidelines for on-site construction. High-confidence emergency warnings result in immediate work stoppage, while low-confidence warnings require intensive manual retesting, optimizing emergency resource allocation. The method outputs synchronous, graded early warnings from multiple monitoring points along the entire line, enabling differentiated risk management across tunnel segments and adapting to comprehensive safety monitoring of long-distance, multi-fault tunnels.

[0060] Furthermore, in step S112, after determining the early warning information for sudden water inrush disaster by combining the predicted groundwater level results and the measured water level values, the sensor layout can be adjusted according to the prediction results. In this embodiment, the adjustment process is specifically implemented as follows: The SHAP method is used to quantify the contribution and interaction of the characteristics of each monitoring point to the groundwater level prediction results, screen the key observation points that affect the early warning of sudden water inrush, and optimize the sensor deployment based on the screening results.

[0061] The SHAP method is an algorithm based on game theory to quantify the marginal contribution of each input feature to the model's predicted output. A positive SHAP value indicates that an increase in the feature will exacerbate the risk of sudden water inrush, while a negative value indicates that it will suppress the risk. The absolute value of the value is the contribution intensity. The feature contribution degree is the weight of the influence of the hydrological time series features of each monitoring point on water level prediction and sudden water inrush risk assessment. The larger the absolute value of SHAP, the higher the contribution degree. The feature interaction relationship is the synergistic influence of the hydrological features of two or more monitoring points on the risk of sudden water inrush. For example, the synchronous rise of water levels at monitoring points on both sides of a fault indicates a positive synergistic risk interaction. The key observation points are the sensor points with the highest absolute value of SHAP contribution and strong risk interaction with other monitoring points. These are the main monitoring points for controlling the risk of sudden water inrush. During the sensor deployment optimization process, the deployment of tunnel sensors is adjusted based on the key point selection results. The sampling density of high-contribution points is strengthened, low-contribution redundant monitoring points are removed, and missing monitoring points in the fault interaction zone are added.

[0062] Based on this, the SHAP method is used to perform interpretability analysis on the prediction results, quantify the contribution of each input feature to the prediction results and their respective interactions, and select observation points with greater impact on the early warning of sudden water inrush based on the contribution, thereby optimizing the sensor point layout. Specifically, using the time-series characteristics of all monitoring points as input, the global SHAP contribution value of each monitoring point is calculated to quantify the independent impact of a single monitoring point on the overall water level prediction and the risk of sudden inrush. Simultaneously, the SHAP interaction values ​​of pairs of monitoring points are calculated to identify the synergistic risk effects of monitoring points in hydraulically connected sections, distinguishing between single-point independent risk and multi-point coupled risk. A SHAP contribution threshold is set, retaining monitoring points with contributions higher than the threshold and significant interactions as core key observation points, while eliminating redundant monitoring points with extremely low contributions and no risk interactions. The spatial distribution of key points and a list of risk interaction combinations are output. Based on the screening results, deployment optimization strategies are output, such as increasing the density of sensors and sampling frequency at high SHAP contribution key points; removing or reducing the sampling frequency of low-contribution redundant monitoring points; supplementing intermediate transitional monitoring points in fault sections with strong point interactions to improve the hydraulically coupled monitoring network. The final determined importance of the observation point characteristics is then assessed. Figure 6 The importance of observation point features in a multi-point water level prediction and sudden water inrush early warning method is illustrated in the schematic diagram for point optimization.

[0063] Therefore, this study addresses the black-box problem of spatiotemporal prediction models, quantifies the independent contribution and collaborative risk interaction of each monitoring point, enables the tracing of the source of sudden water inrush risk, and facilitates the location of key water-bearing structures by geologists. Data-driven optimization of sensor deployment is achieved by precisely reducing redundant monitoring points to lower the costs of monitoring hardware, computing power, and storage, while strengthening the monitoring density in high-risk sections and improving the ability to detect precursors. The list of key points is dynamically updated as tunnel excavation and geological exposure progress, continuously iterating and optimizing the entire monitoring network. The SHAP quantification results can generate standardized monitoring deployment reports, providing quantitative data support for early-stage sensor planning and mid-stage monitoring network upgrades in tunnel engineering, replacing traditional manual experience-based qualitative deployment schemes.

[0064] In summary, this study acquires groundwater level time-series data from multiple monitoring points in the study area; employs a multi-scale reconstruction method based on LOESS seasonal-trend decomposition and discrete wavelet transform to reconstruct missing data; constructs a graph structure based on the spatial location of monitoring points and extracts spatial dependency features using a residual graph convolutional network; combines a long short-term memory network and an attention mechanism to capture temporal dependencies; and classifies the early warning level of sudden water inrush disasters based on water level prediction results and residual thresholds. This method and system, which requires no additional variable input and relies solely on multi-point observation data, can achieve data reconstruction, prediction of water levels in sparse monitoring networks and tunnel sudden water inrushes, and optimization of sensor locations. It can predict sudden water inrush disasters under conditions of sparse monitoring networks, partially missing observation data, and multi-point monitoring, and exhibits good generalization ability and transferability.

[0065] Corresponding to the above method embodiments, this application also provides an embodiment of a multi-point water level prediction and sudden water inrush early warning device. Figure 7 A schematic diagram of a multi-point water level prediction and sudden water inrush early warning device according to an embodiment of this application is shown. Figure 7 As shown, the device includes: The acquisition module 702 is configured to acquire historical groundwater data from multiple monitoring points in the tunnel project, and to construct a multi-point time series dataset based on the historical groundwater data. The processing module 704 is configured to process the multi-point time series dataset using a LOESS-based seasonal trend decomposition algorithm and discrete wavelet transform to obtain target time series data. Module 706 is configured to construct a graph adjacency matrix based on the coordinates of the monitoring points corresponding to the historical groundwater data. The feature extraction module 708 is configured to extract the spatial dependency features of the target time series data through the graph adjacency matrix based on the residual graph convolutional network, and process the output of the residual graph convolutional network based on the long short-term memory network to extract the hidden state sequence with time dependency. The mapping module 710 is configured to perform weighted aggregation on the hidden state sequence based on the attention mechanism, and map the processing result of the attention mechanism to the groundwater level prediction result through a fully connected layer; The integrated module 712 is configured to integrate the groundwater level prediction results with the measured water level values ​​to determine early warning information for sudden water inrush disasters.

[0066] In an optional embodiment, the processing module 704 is further configured to: The multi-point time series dataset is decomposed into a trend term, a seasonal term, and a residual term using the LOESS-based seasonal trend decomposition algorithm. The residual term is processed by the discrete wavelet transform to remove high-frequency noise and obtain a denoised residual term. The trend term, the seasonal term, and the denoised residual term are then superimposed to obtain the target time series data.

[0067] In an optional embodiment, the building module 706 is further configured to: Calculate the Euclidean distance dᵢ between the coordinates of the i-th monitoring point and the coordinates of the j-th monitoring point; based on the Euclidean distance dᵢ and the scale parameter σ determined by the kernel density estimation, construct the graph adjacency matrix aᵢ = exp(−dᵢ² / σ²).

[0068] In an optional embodiment, the feature extraction module 708 is further configured to: The target time series data is graph convolved using the graph adjacency matrix through the residual graph convolutional network, which includes at least two graph convolutional layers and residual connections, wherein the residuals add the input features to the output features of the graph convolutional layers.

[0069] In an optional embodiment, the mapping module 710 is further configured to: Attention weights are calculated for the hidden states at each time step of the Long Short-Term Memory network using an additive attention mechanism or a scaled dot product attention mechanism; and the entire sequence of hidden states is then summed based on these attention weights.

[0070] In an optional embodiment, the integration module 712 is further configured to: By comparing the predicted groundwater level with the measured groundwater level, the predicted residual time series is calculated and the level of sudden water inrush disaster warning is classified. The predicted groundwater level is processed using the Monte Carlo Dropout method to calculate the prediction confidence interval. The prediction confidence interval is then correlated with the level of sudden water inrush disaster warning to obtain the sudden water inrush disaster warning information.

[0071] In an optional embodiment, the multi-point water level prediction and sudden water inrush early warning device further includes: The sensor optimization module is configured to quantify the contribution and feature interaction of each monitoring point's characteristics to the groundwater level prediction results using the SHAP method, screen key observation points that affect the early warning of sudden water inrush, and optimize sensor deployment based on the screening results.

[0072] The multi-point water level prediction and sudden water inrush early warning device provided in this application acquires historical groundwater hydrological data from multiple monitoring points in a tunnel project and constructs a multi-point time series dataset based on the historical groundwater hydrological data. The multi-point time series dataset is processed using a LOESS-based seasonal trend decomposition algorithm and discrete wavelet transform to obtain target time series data. A graph adjacency matrix is ​​constructed based on the coordinates of the monitoring points corresponding to the historical groundwater hydrological data. Spatial dependency features of the target time series data are extracted using the graph adjacency matrix based on a residual graph convolutional network, and the output of the residual graph convolutional network is processed using a long short-term memory network to extract hidden state sequences with time dependencies. The hidden state sequences are weighted and aggregated based on an attention mechanism, and the processing result of the attention mechanism is mapped to a groundwater level prediction result through a fully connected layer. The groundwater level prediction result is combined with the measured water level value to determine the early warning information for sudden water inrush disasters. Relying solely on multi-point observation data without requiring external variable input, this method is applicable to both groundwater level prediction and tunnel inrush water prediction scenarios. It achieves accurate predictions in data-scarce areas, reducing reliance on the density of external monitoring facilities and expanding its applicability. Through a multi-scale reconstruction method coupled with STL decomposition and discrete wavelet transform, missing groundwater level data is effectively restored, improving data quality and providing complete and reliable time-series data for subsequent modeling. The residual graph convolutional network effectively extracts spatial dependency features between monitoring points, and the residual connection mechanism alleviates the over-smoothing problem of the graph convolutional network, preserving the spatial relationships between different monitoring points. The system identifies the differences in features between nodes; it combines LSTM and attention mechanisms to extract short-term and long-term temporal dependencies, with the attention mechanism effectively mitigating information loss in long-sequence prediction; the collaborative working mechanism of residual graph convolution, LSTM, and attention mechanisms produces cumulative performance improvements, outperforming the single LSTM baseline model; uncertainty analysis quantifies the prediction mean and confidence interval, providing the probability of the warning level and increasing the reliability of the prediction results; the SHAP method performs interpretability analysis on the prediction results, quantifying the contribution of different observation points to the prediction results, thereby optimizing the sensor point layout.

[0073] The above is a schematic scheme of a multi-point water level prediction and sudden water inrush early warning device according to this embodiment. It should be noted that the technical solution of this multi-point water level prediction and sudden water inrush early warning device belongs to the same concept as the technical solution of the aforementioned multi-point water level prediction and sudden water inrush early warning method. Details not described in detail in the technical solution of the multi-point water level prediction and sudden water inrush early warning device can be found in the description of the technical solution of the aforementioned multi-point water level prediction and sudden water inrush early warning method. Furthermore, the components in the device embodiment should be understood as functional modules necessary to implement each step of the program flow or each step of the method; these functional modules are not actual functional divisions or separations. The device claims defined by such a set of functional modules should be understood as a functional module architecture that primarily implements the solution through the computer program described in the specification, and not as a physical device that primarily implements the solution through hardware.

[0074] Figure 8 A structural block diagram of a computing device 800 according to an embodiment of this application is shown. The components of the computing device 800 include, but are not limited to, a memory 810 and a processor 820. The processor 820 is connected to the memory 810 via a bus 830, and a database 850 is used to store data.

[0075] The computing device 800 also includes an access device 840, which enables the computing device 800 to communicate via one or more networks 860. Examples of these networks include a Public Switched Telephone Network (PSTN), a Local Area Network (LAN), a Wide Area Network (WAN), a Personal Area Network (PAN), or a combination of communication networks such as the Internet. The access device 840 may include one or more of any type of wired or wireless network interface (e.g., a Network Interface Card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) interface, a Wi-MAX interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, a Near Field Communication (NFC) interface, and so on.

[0076] In one embodiment of this application, the aforementioned components of the computing device 800 and Figure 8 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 8 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this application. Those skilled in the art can add or replace other components as needed.

[0077] The computing device 800 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or PCs. The computing device 800 can also be a mobile or stationary server.

[0078] The processor 820 is used to execute computer-executable instructions for each step of the multi-point water level prediction and sudden water inrush early warning method.

[0079] The above is a schematic representation of a computing device according to this embodiment. It should be noted that the technical solution of this computing device belongs to the same concept as the technical solution of the aforementioned multi-point water level prediction and sudden water inrush early warning method. Details not described in detail in the technical solution of the computing device can be found in the description of the technical solution of the aforementioned multi-point water level prediction and sudden water inrush early warning method.

[0080] An embodiment of this application also provides a computer-readable storage medium storing computer instructions that, when executed by a processor, are used to implement the steps of the multi-point water level prediction and sudden water inrush early warning method.

[0081] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium belongs to the same concept as the technical solution of the above-described multi-point water level prediction and sudden water inrush early warning method. For details not described in detail in the technical solution of the storage medium, please refer to the description of the technical solution of the above-described multi-point water level prediction and sudden water inrush early warning method.

[0082] An embodiment of this application also provides a chip that stores a computer program, which, when executed by the chip, implements the steps of the multi-point water level prediction and sudden water inrush early warning method.

[0083] The foregoing has described specific embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0084] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0085] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0086] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0087] The preferred embodiments disclosed above are merely illustrative of this application. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this application. These embodiments are selected and specifically described in this application to better explain the principles and practical applications of this application, thereby enabling those skilled in the art to better understand and utilize this application. This application is limited only by the claims and their full scope and equivalents.

Claims

1. A method for multi-point water level prediction and sudden water inrush early warning, characterized in that, include: Historical groundwater data from multiple monitoring points in the tunnel project were acquired, and a multi-point time series dataset was constructed based on the historical groundwater data. The multi-point time series dataset is processed using the LOESS-based seasonal trend decomposition algorithm and discrete wavelet transform to obtain the target time series data. A graph adjacency matrix is ​​constructed based on the coordinates of the monitoring points corresponding to the historical groundwater hydrological data. Based on the residual graph convolutional network, the spatial dependency features of the target time series data are extracted through the graph adjacency matrix, and the output of the residual graph convolutional network is processed based on the long short-term memory network to extract the hidden state sequence with time dependency. The hidden state sequence is weighted and aggregated based on the attention mechanism, and the processing result of the attention mechanism is mapped to the groundwater level prediction result through a fully connected layer. Based on the combined results of the groundwater level prediction and the actual measured water level, early warning information for sudden water inrush disasters is determined.

2. The method according to claim 1, characterized in that, The multi-point time series dataset is processed using a LOESS-based seasonal trend decomposition algorithm and discrete wavelet transform to obtain the target time series data, including: The LOESS-based seasonal trend decomposition algorithm is used to decompose the multi-point time series dataset into trend, seasonal, and residual terms. The residual term is processed based on the discrete wavelet transform to remove high-frequency noise and obtain a denoised residual term. The target time series data is obtained by superimposing the trend term, the seasonal term, and the denoised residual term.

3. The method according to claim 1, characterized in that, The construction of a graph adjacency matrix based on the coordinates of the monitoring points corresponding to the historical groundwater hydrological data includes: Calculate the Euclidean distance dᵢ between the coordinates of the i-th monitoring point and the coordinates of the j-th monitoring point; Based on the Euclidean distance dᵢ and the scale parameter σ determined by the kernel density estimation, the graph adjacency matrix is ​​constructed as aᵢ=exp(−dᵢ² / σ²).

4. The method according to claim 1, characterized in that, The residual graph convolutional network extracts spatial dependency features of the target time series data through the graph adjacency matrix, including: The target time series data is graph convolved using the graph adjacency matrix through the residual graph convolutional network, which includes at least two graph convolutional layers and residual connections, wherein the residuals add the input features to the output features of the graph convolutional layers.

5. The method according to claim 1, characterized in that, The weighted aggregation of the hidden state sequence based on the attention mechanism includes: Attention weights are calculated for the hidden states of the Long Short-Term Memory network at each time step using an additive attention mechanism or a scaled dot product attention mechanism. The entire sequence of hidden states is weighted and summed according to the attention weights.

6. The method according to claim 1, characterized in that, The method of combining the predicted groundwater level with the measured groundwater level to determine early warning information for sudden water inrush disasters includes: By comparing the predicted groundwater level with the measured groundwater level, the predicted residual time series is calculated and the early warning level of sudden water inrush disaster is classified. The groundwater level prediction results are processed using the Monte Carlo Dropout method to calculate the prediction confidence interval. The prediction confidence interval is then correlated with the sudden water inrush disaster warning level to obtain the sudden water inrush disaster warning information.

7. The method according to claim 1, characterized in that, After combining the predicted groundwater level results with the measured water level values ​​to determine the early warning information for sudden water inrush disasters, the method further includes: The SHAP method is used to quantify the contribution and interaction of the characteristics of each monitoring point to the groundwater level prediction results, screen the key observation points that affect the early warning of sudden water inrush, and optimize the sensor deployment based on the screening results.

8. A multi-point water level prediction and sudden water inrush early warning device, characterized in that, include: The acquisition module is configured to acquire historical groundwater data from multiple monitoring points in the tunnel project, and to construct a multi-point time series dataset based on the historical groundwater data. The processing module is configured to process the multi-point time series dataset using a LOESS-based seasonal trend decomposition algorithm and discrete wavelet transform to obtain the target time series data. The construction module is configured to construct a graph adjacency matrix based on the coordinates of the monitoring points corresponding to the historical groundwater data. The feature extraction module is configured to extract the spatial dependency features of the target time series data based on the residual graph convolutional network through the graph adjacency matrix, and process the output of the residual graph convolutional network based on the long short-term memory network to extract the hidden state sequence with time dependency. The mapping module is configured to perform weighted aggregation of the hidden state sequence based on the attention mechanism, and map the processing result of the attention mechanism to the groundwater level prediction result through a fully connected layer; The integrated module is configured to combine the predicted groundwater level with the measured groundwater level to determine early warning information for sudden water inrush disasters.

9. A computing device, characterized in that, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the steps of the multi-point water level prediction and sudden water inrush early warning method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing computer instructions, characterized in that, When executed by the processor, this instruction implements the steps of the multi-point water level prediction and sudden water inrush early warning method according to any one of claims 1 to 7.

Images