A tunnel floor drum deformation monitoring method based on time series data

Abstract

This invention relates to the field of geotechnical engineering safety monitoring technology, specifically disclosing a method for monitoring tunnel floor heave deformation based on time-series data. The method deploys an integrated sensor network at the tunnel floor heave monitoring section, simultaneously collecting time-series data on multiple parameters such as floor displacement, surrounding rock stress, and groundwater level. This data is then cleaned and feature extracted to generate a multi-dimensional feature time-series sequence. A combined deep learning model integrating a graph convolutional network and gated recurrent units is constructed. The graph convolutional network captures the spatial propagation relationship between different monitoring points, while the gated recurrent units uncover the temporal evolution pattern to achieve multi-step rolling prediction of floor displacement at future moments. The prediction results are input into a pre-constructed multi-level risk matrix for verification. Based on threshold classification, four levels of early warning signals (safety, caution, warning, or alarm) are automatically output and released in real time via a terminal, providing intelligent technical support for tunnel floor heave disaster prevention.

Description

Technical Field

[0001] This invention belongs to the field of geotechnical engineering safety monitoring technology, specifically, it relates to a method for monitoring tunnel floor heave deformation based on time-series data. Background Technology

[0002] During long-term operation, the tunnel is affected by a combination of factors, including complex geological conditions, groundwater activity, and changes in surrounding rock stress. The heave and deformation of the bottom slab has gradually become a prominent problem affecting structural safety and operational stability.

[0003] Existing tunnel floor heave deformation monitoring technologies typically rely on periodic or manual measurements of single physical quantities, such as focusing only on the absolute change in floor displacement or monitoring surrounding rock stress and groundwater level in isolation. They lack a systematic consideration of multi-parameter spatiotemporal coupling effects. This single-point, single-parameter monitoring method struggles to capture the complex nonlinear dynamic characteristics of multi-field coupling during floor heave deformation. For example, the feedback relationship between displacement, stress, and seepage is often analyzed in isolation, resulting in monitoring data that fails to reflect the true mechanical evolution mechanism. Furthermore, existing methods lack effective means to identify and process the anomalous fluctuations, chaotic characteristics, and long-term memory effects implicit in time-series data, such as sudden increases in groundwater level or other factors. Early warning information such as stress relaxation is easily misjudged as noise and directly rejected, thus losing key early warning features. At the predictive model level, existing technologies mostly rely on empirical formulas, regression analysis, or single recurrent neural networks, which cannot handle the spatial propagation relationship between monitoring points. Bottom heave deformation often propagates along the longitudinal direction of the tunnel between different sections. Traditional models treat this as an independent time series and ignore the spatial topology of the sensor network of adjacent sections, resulting in a lag in the prediction of deformation trends and insufficient accuracy. This restricts the ability to identify bottom heave disasters in actual tunnel engineering, and warnings are often only issued after significant deformation has occurred, which cannot meet the needs of preventive maintenance.

[0004] To address the aforementioned problems, this invention proposes a method for monitoring tunnel floor deformation based on time-series data. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for monitoring tunnel floor deformation based on time-series data, solving the problem of the lack of multi-parameter spatiotemporal coupling prediction and nonlinear feature fusion early warning in existing technologies.

[0006] The objective of this invention can be achieved through the following technical solutions: A method for monitoring tunnel floor heave deformation based on time-series data, the method comprising: Step 1: Deploy an integrated sensor network at the target point of the bottom heave monitoring section of the target tunnel. Extract the bottom plate displacement, surrounding rock stress and groundwater level of the bottom heave monitoring section in real time at the synchronous sampling frequency from the beginning of the monitoring cycle to construct multi-parameter time series data. Step 2: After performing data cleaning on the multi-parameter time series data, data sequences associated with various parameters are generated, and multi-dimensional feature time series sequences are generated based on the numerical performance of each data sequence. Step 3: Construct a combined deep learning model based on gated recurrent units (GRU) and graph convolutional networks (GCN); GCN is used to capture the spatial propagation relationship between different target points, while GRU is used to capture the evolution law of multidimensional feature time series in the time dimension, and output the predicted value of the base plate displacement at future time. Step four: Verify the predicted displacement value of the base plate with the pre-constructed multi-level risk matrix, and output the corresponding level of early warning signal.

[0007] As a further aspect of the present invention, in step one, the bottom heave monitoring section of the target tunnel and the target point at the bottom heave monitoring section are determined by the operator after on-site survey. The target points at the same bottom drum monitoring section include at least one first monitoring point, one second monitoring point, and one third monitoring point.

[0008] As a further aspect of the present invention, in step one, the integrated sensor network includes a hydrostatic level, a vibrating wire pressure cell, and a pore water pressure gauge. The hydrostatic level is installed at the first monitoring point, the vibrating wire pressure cell is installed at the second monitoring point, and the pore water pressure gauge is installed at the third monitoring point. All sensors are connected to the system terminal via the system bus to perform data transmission.

[0009] As a further aspect of the present invention, the specific method for constructing multi-parameter time-series data in step one is as follows: Sensors deployed at the first, second, and third monitoring points of the target location at any monitoring section of the bottom drum; The current time is taken as the start time t1 of the preset monitoring cycle for data collection; Starting from time t1, the displacement of the bottom plate, the stress of the surrounding rock, and the groundwater level collected by the hydrostatic level, vibrating wire pressure cell, and pore water pressure gauge at each sampling time are extracted and summarized in chronological order, and recorded as multi-parameter time series data.

[0010] As a further aspect of the present invention, the specific method for generating the data sequences associated with various parameters in step two is as follows: All base plate displacements are extracted from the multi-parameter time series data and rearranged in chronological order, which is recorded as the initial base plate displacement sequence. Calculate the standard deviation σ and mean μ of the initial base plate displacement sequence, and use the isolated forest algorithm to filter out abnormal base plate displacements in the initial base plate displacement sequence; The abnormal base plate displacement is checked against [μ-3σ,μ+3σ]. If the abnormal base plate displacement is outside the range of [μ-3σ,μ+3σ], it is considered that the abnormal base plate displacement has failed the check and is marked as candidate data to be removed. If there are N consecutive abnormal base plate displacements in the initial base plate displacement sequence, including candidate rejection data, and none of them pass the verification, then the N base plate displacements are considered normal data. Conversely, all candidate data to be removed will be removed, and linear interpolation will be used to fill in the removed data. Extract all base plate displacements from the initial base plate displacement sequence after the above operations, and denot them as base plate displacement sequences D1, D2, ..., Dj, where D1 corresponds to t1, and j represents the total number of sampling times within a monitoring period; Similarly, determine the surrounding rock stress sequence F1, F2, ..., Fj and the groundwater level W1, W2, ..., Wj.

[0011] As a further aspect of the present invention, the specific method for generating the multidimensional feature time series sequence in step two is as follows: Align the bottom plate displacement sequence D1, D2, ..., Dj, the surrounding rock stress sequence F1, F2, ..., Fj, and the groundwater level sequence W1, W2, ..., Wj according to the time sequence to form a triplet sequence {(Di,Fi,Wi)}, where i is the counting index, 1≤i≤j; For all triples, a chaotic phase space reconstruction is performed with embedding dimension m=3 and time delay τ; the historical m time values ​​of each type of parameter constitute a phase point, thereby mapping the one-dimensional time series data to the three-dimensional phase space, where the time delay τ is determined by the first minimum value of the mutual information method; Starting from time t1, continuously lock the sliding window T, where T = j / 10; Calculate the maximum Lyapunov exponent λ of the base plate displacement sequence in three-dimensional phase space across all sliding windows. max And the rescaled range Hurst exponent (HST); The joint mutual information X(D;F) between the bottom plate displacement and the surrounding rock stress is calculated using X(D;F)=H(D)+H(F)−H(D,F), where the information entropy H is estimated using the equidistant histogram method. Similarly, calculate the joint mutual information X(D;W) and the joint mutual information X(F;W); The maximum Lyapunov exponent λ within any sliding window T. max The rescaled range Hurst exponent (HST), combined with the joint mutual information X(D;F), X(D;W), and X(F;W), generates a multidimensional feature time series RT={λ} associated with the corresponding sliding window T.max ,HST,X(D;F),X(D;W),X(F;W)}.

[0012] As a further aspect of the present invention, the specific method for constructing the combined deep learning model based on the gated recurrent unit (GRU) and the graph convolutional network (GCN) in step three is as follows: Using the first, second, and third monitoring points within the bottom drum monitoring section as graph nodes, and the corresponding first, second, and third monitoring points in adjacent bottom drum monitoring sections as associated graph nodes, a graph structure is constructed to generate a normalized adjacency matrix. Obtain a combined deep learning model consisting of gated recurrent units (GRUs) and graph convolutional networks (GCNs), including fully connected layers; The graph convolutional network GCN uses two layers of graph convolution, with a multidimensional feature temporal sequence as input and a spatial feature matrix as output. The gated recurrent unit (GRU) uses the flattened spatial feature matrix output by the GCN as the input for each time step. It employs a two-layer stacked GRU to capture the temporal evolution pattern and outputs the final hidden state. The final hidden state is connected to the fully connected layer, and the predicted value of the base plate displacement at the future time Δt is output. The future time Δt is the time after the current monitoring cycle ends, which is preset by the operator and is an integer multiple of the sampling frequency interval. The combined deep learning model is trained using the mean squared error loss function and the Adam optimizer. The training sliding window length is 1 / 10 of the monitoring period, and the training sliding window step size is half of the training sliding window length.

[0013] As a further aspect of the present invention, the specific method for outputting the predicted value of the base plate displacement at future times in step three is as follows: Extracting the multidimensional feature time series RT={λ from an arbitrary sliding window T max After forward propagation through GCN and GRU, the fully connected layer outputs a scalar value d_Δt, which is the predicted value of the bottom plate displacement at the future time Δt. Adjust the output dimension of the fully connected layer to the integer M preset by the operator; Obtain M different future moments input by the operator: Δt, Δt2, ..., ΔtM; The fully connected layer outputs M different predicted values ​​of the base plate displacement at future times. Each predicted value of the base plate displacement is used as part of the prediction input for the next future time, and the prediction results for subsequent future times are generated in a rolling manner and then output.

[0014] As a further aspect of the present invention, in step four, the specific method for verifying the predicted value against the pre-constructed multi-level risk matrix and outputting a warning signal of the corresponding level is as follows: A multi-level risk matrix is ​​constructed, which takes the predicted value of the base plate displacement as the input variable and divides it into four consecutive risk level intervals, with each risk level interval corresponding to a warning level. Input the predicted future base plate displacement values ​​output from step three into the multi-level risk matrix; If the predicted displacement of the base plate is lower than the first threshold, a safety signal is output. If the predicted displacement of the base plate is greater than or equal to the first threshold and lower than the second threshold, an attention signal is output. If the predicted displacement of the base plate is greater than or equal to the second threshold but lower than the third threshold, an alarm signal will be output. If the predicted displacement of the base plate is greater than or equal to the third threshold, an alarm signal will be output. The first, second, and third thresholds are set by the operators according to the actual needs of the tunnel, and the warning signals are released in real time through the system terminal via sound and light, SMS and platform push.

[0015] The beneficial effects of this invention are as follows.

[0016] (1) This invention overcomes the limitations of single-parameter monitoring by integrating sensor networks to synchronously collect time-series data of displacement, stress and water level, providing a more comprehensive data foundation for the analysis of tunnel floor heave deformation. Based on data cleaning and statistical feature extraction, the reliability and availability of the original data are improved. The deep learning model combining gated recurrent units and graph convolutional networks can simultaneously capture the spatial propagation relationship between different monitoring points and the long-term evolution law of the time-series data itself, realizing the prediction of future floor displacement. The prediction results are dynamically verified with the multi-level risk matrix, and graded early warning signals are automatically output, improving the intelligence level and early warning timeliness of tunnel floor heave monitoring.

[0017] (2) This invention determines the monitoring section and target point of the tunnel floor heave through on-site survey, and integrates a static level, a vibrating wire pressure cell and a pore water pressure gauge to correspond to the three key parameters of floor displacement, surrounding rock stress and groundwater level, respectively, to realize the synchronous monitoring of the multi-physical field coupling effect of the tunnel floor heave, and overcome the limitations of single index judgment; at the same time, the system bus connects the sensor and the terminal, automatically performs periodic data collection and summarizes it into multi-parameter time series data in chronological order, reducing the error and cost caused by manual intervention, and providing a continuous and reliable data foundation for subsequent analysis of the evolution law of the floor heave and the establishment of an early warning model.

[0018] (3) This invention improves the reliability of base plate displacement monitoring through multi-level data cleaning and nonlinear feature extraction; combined with the anomaly detection strategy of isolated forest and 3σ criterion, it introduces a continuous anomaly retention rule to effectively identify extreme outliers and avoids the accidental deletion of physically meaningful continuous anomaly data due to short-term fluctuations, thus ensuring the integrity of the data sequence; it adopts chaotic phase space reconstruction and nonlinear dynamic indicators such as maximum Lyapunov exponent and Hurst exponent to reveal the chaotic characteristics and long-term memory of the base plate displacement system, providing sensitive criteria for predicting instability trends; based on joint mutual information, it quantifies the coupling correlation between multiple parameters, enabling the multi-dimensional feature time series to comprehensively reflect the synergistic influence of surrounding rock stress and groundwater level on base plate deformation, thereby enhancing the ability to identify precursors of displacement abrupt changes under complex geological conditions.

[0019] (4) This invention constructs a combined deep learning model of gated recurrent units and graph convolutional networks. The graph convolutional network uses adjacent cross-section monitoring points as graph nodes to effectively capture spatial structural dependencies, while the gated recurrent units further extract the evolutionary patterns of long-term sequences. The combination of the two achieves efficient modeling of complex spatiotemporal features. Secondly, the use of sliding window training, mean squared error loss function and Adam optimizer ensures fast model convergence and strong generalization ability. It supports rolling prediction of multiple future moments, flexibly adapts to different early warning needs, and uses a hierarchical early warning mechanism based on a multi-level risk matrix to map the predicted values ​​to four levels: safety, attention, warning, and alarm. The predicted values ​​are released in real time through sound and light, SMS and platform, enabling on-site personnel to respond accurately according to the risk level, thereby improving the safety management level of tunnel construction and operation. Attached Figure Description

[0020] The invention will now be further described with reference to the accompanying drawings.

[0021] Figure 1 This is a flowchart illustrating the method described in this invention.

[0022] Figure 2 This is a schematic diagram of the integrated sensor network described in Embodiment 2 of the present invention.

[0023] Figure 3 This is a schematic diagram of the deployment of the integrated sensor network described in Embodiment 2 of the present invention.

[0024] Figure 4 This is a schematic diagram of the structure of the combined deep learning model described in this invention. Detailed Implementation

[0025] 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, and 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.

[0026] Example 1: This embodiment provides a method for monitoring tunnel floor heave deformation based on time-series data. The main process is as follows: Figure 1 As shown. The method specifically includes the following steps: Step 1: Deploy an integrated sensor network at the target point of the bottom heave monitoring section of the target tunnel. Extract the bottom plate displacement, surrounding rock stress and groundwater level of the bottom heave monitoring section in real time at the synchronous sampling frequency from the beginning of the monitoring cycle to construct multi-parameter time series data. Step 2: After performing data cleaning on the multi-parameter time series data, data sequences associated with various parameters are generated, and multi-dimensional feature time series sequences are generated based on the numerical performance of each data sequence. Step 3: Construct a combined deep learning model based on gated recurrent unit (GRU) and graph convolutional network (GCN); use GCN to capture the spatial propagation relationship between different target points, and use GRU to capture the evolution law of multidimensional feature time series in the time dimension, and output the predicted value of the base plate displacement at future time. Step four: Verify the predicted displacement value of the base plate with the pre-constructed multi-level risk matrix, and output the corresponding level of early warning signal.

[0027] Example 2: Based on Example 1, this embodiment uses real-time acquisition of multi-parameter time-series data and deep learning models to achieve future prediction and graded early warning of floor heave displacement. This makes the time-series data-based tunnel floor heave deformation monitoring method applicable to sections of railway or highway tunnels with high floor heave risk during operation or construction, such as tunnel sections crossing weak surrounding rock or water-rich strata.

[0028] In specific implementation, firstly, step one as described in Example 1 is executed, and an integrated sensor network is deployed at the target point of the bottom drum monitoring section of the target tunnel. After on-site survey, the operator determines the location of the bottom drum monitoring section. The distance between adjacent bottom drum monitoring sections is set to 20m. This data can be adjusted according to actual needs. Only an example value is given here.

[0029] At each bottom heave monitoring section, the operators determine the target points, which include three monitoring points: the first monitoring point is located at the center of the invert arch, the second monitoring point is located in the surrounding rock at the junction of the invert arch and the sidewall, and the third monitoring point is located in the aquifer below the invert arch.

[0030] The integrated sensor network includes a hydrostatic level with a range of 0–50 mm and an accuracy of 0.1 mm, a vibrating wire pressure cell with a range of 0–2 MPa and an accuracy of 0.01 MPa, and a pore water pressure meter with a range of 0–0.5 MPa and an accuracy of 0.001 MPa.

[0031] A static level is installed at the first monitoring point, and its storage tank is connected to the static level at the benchmark point through a connecting pipe to eliminate the influence of benchmark settlement; a vibrating wire pressure cell is buried inside the surrounding rock at the second monitoring point, with the pressure cell's force-bearing surface facing the invert arch; a pore water pressure gauge is installed in the borehole at the third monitoring point, with the borehole depth penetrating the tunnel floor slab and entering the aquifer, and the borehole is filled with permeable sand and gravel.

[0032] All sensors described above are connected to the system terminal via an RS-485 system bus. The system terminal has a built-in real-time clock for synchronous sampling. See details... Figure 2 and Figure 3 .

[0033] Starting from the beginning of the monitoring cycle at time t1, the system terminal simultaneously triggers the hydrostatic level, vibrating wire pressure cell, and pore water pressure gauge to collect data at a synchronous sampling frequency of 1 time / hour. Each time, the system collects the bottom plate displacement, surrounding rock stress, and groundwater level. The collected data is then summarized in chronological order to form multi-parameter time-series data. This parameter time-series data is stored in the system terminal in structured text format and uploaded to the cloud server via wireless network every 24 hours for persistent storage and real-time retrieval.

[0034] Next, step two is performed to clean the multi-parameter time series data obtained above and generate data sequences associated with various parameters. Based on the numerical performance of each data sequence, a multi-dimensional feature time series sequence is generated.

[0035] In this embodiment, data cleaning employs a combination of the 3-standard-deviation criterion ([μ-3σ,μ+3σ]) and linear interpolation; specifically as follows: Data sequences of base plate displacement, surrounding rock stress, and groundwater level, each aligned with time, were extracted separately. Outliers exceeding the range of [μ-3σ, μ+3σ] were removed, and missing points were filled in using linear interpolation. The displacement sequence of the bottom plate after cleaning is denoted as D1, D2, ..., Dj, the stress sequence of the surrounding rock is F1, F2, ..., Fj, and the groundwater level sequence is W1, W2, ..., Wj; where j is the total number of sampling times in a monitoring cycle. For example, if a cycle is 30 days and the sampling frequency is 1 time / hour, then j = 720. Subsequently, the triplet sequence {(Di,Fi,Wi)} is formed by aligning the triplets in time order; where i is the counting index, 1≤i≤j; Statistical features within each sliding window are calculated, including the mean and variance of the base plate displacement, the maximum Lyapunov exponent, and the Pearson correlation coefficient between the base plate displacement and the surrounding rock stress. These features are then combined into a multidimensional feature time series. The length of the sliding window is set to j / 10=72, and the maximum Lyapunov exponent is calculated using the Wolf algorithm.

[0036] Then, proceed to step three to construct a combined deep learning model based on gated recurrent units (GRU) and graph convolutional networks (GCN).

[0037] refer to Figure 4 The deep learning model shown is based on a combination of Gated Recurrent Unit (GRU) and Graph Convolutional Network (GCN), including an input layer, a GCN layer, a flattening layer, a GRU layer, and an output layer.

[0038] In this embodiment, the inputs to the combined deep learning model are a multidimensional feature time series RT, graph nodes, edges, and a normalized adjacency matrix A. The multidimensional feature time series RT includes five features: the maximum Lyapunov exponent λmax, the Hurst exponent HST, and joint mutual information X(D;F), X(D;W), and X(F;W). Graph nodes correspond to monitoring points: the first, second, and third monitoring points within each bass drum monitoring section. Edges: fully connected between different monitoring points within the same section; connections between monitoring points of the same type in adjacent sections. The normalized adjacency matrix A is an edge weight matrix calculated using a Gaussian kernel function based on the spatial distance between monitoring points.

[0039] This embodiment employs a two-layer graph convolutional network (GCN): GCN layer 1 has 64 output channels and uses the ReLU activation function; GCN layer 2 has 32 output channels and uses the ReLU activation function. The input is the multidimensional feature time series obtained in step two, and the output is a spatial feature matrix. GCN is used to capture the spatial propagation relationship between different monitoring points and extract the spatial feature matrix.

[0040] In this embodiment, the flattening layer flattens the spatial feature matrix (number of nodes × 32) output by GCN along the node dimension.

[0041] This embodiment employs a two-layer stacked gated recurrent unit (GRU). GRU layer 1 contains 128 hidden units; GRU layer 2 contains 128 hidden units (stacked structure). The GRU uses the flattened vector of the spatial feature matrix output by the GCN as input for each time step to capture the temporal evolution. The fully connected layer maps the 128-dimensional final hidden state to a scalar output.

[0042] Finally, the fully connected layers of the combined deep learning model output the future time Δt, for example, Δt=12 hours, which is the predicted value of the base plate displacement after 12 sampling intervals.

[0043] The combined deep learning model was trained using the mean squared error loss function. The Adam optimizer's initial learning rate was set to 0.001, the training sliding window length was j / 10 = 72 time steps, the step size was 36 time steps, and the training epochs were 200. The combined deep learning model was trained on a cloud server using historical data from the current time over the past 6 months. The combined deep learning model captures the spatial topological features of the monitoring network through GCN and captures the temporal evolution patterns through GRU. The combination of these two methods effectively predicts the future trend of tunnel floor displacement.

[0044] Finally, in step four, the predicted displacement value of the base plate is verified against the pre-constructed multi-level risk matrix, and an early warning signal of the corresponding level is output. The multi-level risk matrix is ​​defined as follows: If the predicted displacement of the base plate is less than 10mm, a green safety signal is output. If the predicted displacement of the base plate is 10mm or less and the predicted displacement value is less than 25mm, output a blue warning signal. If the predicted displacement of the base plate is less than 50 mm and 25 mm is ≤, a yellow warning signal will be output. If the predicted displacement of the base plate is ≥50mm, a red alarm signal will be output.

[0045] After receiving the warning level returned by the cloud server, the system terminal sends the warning information to the maintenance personnel through on-site audible and visual alarms, such as red LEDs and buzzers, as well as SMS push modules.

[0046] This embodiment integrates a sensor network to synchronously collect three types of parameters: base plate displacement, surrounding rock stress, and groundwater level. Combined with a deep learning model, it achieves accurate prediction of base plate deformation. Compared with existing methods based solely on displacement monitoring, this improves the accuracy and lead time of early warning and avoids false alarms or missed alarms caused by sudden increases in surrounding rock stress or changes in groundwater level.

[0047] Example 3: This embodiment is based on embodiment 2, and further optimizes the data cleaning and multi-dimensional feature extraction methods to better suit scenarios where there are abnormal fluctuations in the monitoring data caused by temporary sensor failures or external interference.

[0048] The deployment of the sensor network integrated in step one of this embodiment is basically the same as that in embodiment 2. The specific difference is that the number of monitoring points in the target points of the bottom drum monitoring section is increased to five. There are two first monitoring points, which are symmetrically deployed on the left and right sides of the invert arch, two second monitoring points, which are deployed at the foot of the left and right side walls, and one third monitoring point, which is deployed in the deepest water layer below the centerline of the tunnel. All sensors are still connected to the system terminal through the CAN bus, and the sampling frequency is increased to 1 time / 30 minutes.

[0049] In step two, data cleaning employs a combined strategy of the Isolation Forest algorithm and the three-standard-deviation (3σ) criterion. Specifically, all base plate displacements are extracted from the multi-parameter time-series data and arranged in chronological order to obtain an initial base plate displacement sequence. The standard deviation σ and mean μ of this sequence are calculated, and the Isolation Forest algorithm with 100 isolated trees and a subsampling size of 256 is used to filter out abnormal base plate displacements in the initial base plate displacement sequence. For each base plate displacement value x marked as abnormal, it is checked whether it is within the interval [μ-3σ, μ+3σ]. If x is outside this interval, it is considered to have failed the check and is marked as a candidate for removal; if x is within the interval, the data is retained.

[0050] Furthermore, if N consecutive bottom plate displacement values ​​appear in the initial bottom plate displacement sequence, all of which are candidate rejection data (N=3 in this embodiment), then all N consecutive data are treated as normal data and retained to cope with instantaneous and continuous regular anomalies caused by tunnel blasting vibration or heavy vehicle passage; otherwise, all non-continuous candidate rejection data are rejected, and linear interpolation is used to fill in the data points at the rejection positions.

[0051] The linear interpolation formula is: D_k = D_{k-1} + (D_{k+1} - D_{k-1}) × (k - (k-1)) / ((k+1) - (k-1)); where D_k is the missing point to be filled, and D_{k-1} and D_{k+1} are the nearest valid values ​​before and after the missing point, respectively.

[0052] The bottom plate displacement sequence after the above processing is denoted as D1, D2, ..., Dj. Similarly, the surrounding rock stress sequence F1, F2, ..., Fj and the groundwater level sequence W1, W2, ..., Wj are obtained, where j = 30 days × 48 times / day = 1440.

[0053] Next, a multidimensional feature time series sequence is generated. Specifically, the following method is used: the base plate displacement sequence D1, D2, ..., Dj, the surrounding rock stress sequence F1, F2, ..., Fj, and the groundwater level sequence W1, W2, ..., Wj are aligned according to time sequence to form a triplet sequence {(Di,Fi,Wi)}, where i=1,2,...,j; chaotic phase space reconstruction is performed on all triplets, with an embedding dimension m=3. The time delay τ is determined by calculating the first minimum value using the mutual information method. In this embodiment, τ is calculated to be 6 sampling intervals. Through phase space reconstruction, the parameter values ​​of each type of parameter (such as base plate displacement, surrounding rock stress, groundwater level, etc.) at m historical moments constitute a phase point, thereby mapping the one-dimensional time series data to a three-dimensional phase space. For example, the point of the base plate displacement sequence in the phase space can be represented as: P_D(i)=[D_i,D_{i-τ},D_{i-2τ}] (i≥2τ+1).

[0054] Then, starting from the initial time t1, the sliding window T is continuously locked, with a window length of T = j / 10 = 144 time points, corresponding to 3 days.

[0055] Within each sliding window T, calculate the following data: The maximum Lyapunov exponent λ of the base plate displacement sequence in three-dimensional phase space was calculated using the small data quantity method. max The maximum Lyapunov exponent λ max Used to characterize the sensitivity of the tunnel floor drum system to initial conditions and the intensity of chaos; The Hurst index (HST) of the rescaled range was calculated using the R / S analysis method. The long-term memory of the sequence, such as the persistence or anti-persistence of the deformation trend, was also calculated and evaluated using the R / S analysis method. The joint mutual information X(D;F) of the bottom plate displacement and the surrounding rock stress is calculated as follows: X(D;F) = H(D) + H(F) - H(D,F), where the information entropy H is estimated using the equidistant histogram method, and the number of histogram intervals is set as the number of data points within the window.

[0056] Similarly, calculate the joint mutual information X(D;W) and X(F;W).

[0057] Finally, the above five data points are combined into a multidimensional feature time series RT={λ} corresponding to the sliding window T. max ,HST,X(D;F),X(D;W),X(F;W)}, when the sliding window slides with a step size of 1 sampling time, a series of multidimensional feature time series sequences RT are obtained, thus forming a set of multidimensional feature time series sequences.

[0058] It should be noted that the algorithms involved in this embodiment, such as the R / S analysis method and the small data method, are all considered to be covered by the prior art, and therefore will not be elaborated on in this solution.

[0059] Steps three and four of this embodiment are basically the same as those of embodiment 2, except that the input dimension of the combined deep learning model is changed to 5, corresponding to five different parameters in the multidimensional feature time sequence RT, and the sliding window length during training of the combined deep learning model is fixed at 144 time points.

[0060] This embodiment, by introducing chaotic phase space reconstruction and nonlinear dynamic features, such as the Lyapunov exponent, Hurst exponent, and mutual information features, can more sensitively capture the critical transition signal of the bottom deformation from steady state to instability, and can improve the generalization ability of the combined deep learning model under more complex geological conditions.

[0061] The technical effect achieved by this embodiment is that, through the combined cleaning strategy of isolated forest and 3σ criterion, discontinuous outliers are effectively removed while the authenticity of continuous outliers is preserved. Secondly, through chaotic phase space reconstruction and nonlinear feature extraction, the combined deep learning model can understand the dynamic evolution law hidden in multi-parameter time series data. Compared with traditional statistical feature methods, it can improve the early warning capability before the bottom drum mutation.

[0062] Example 4: Based on Example 3, this embodiment focuses on the specific construction, training, and multi-step rolling prediction of the combined deep learning model, so as to better apply to scenarios that require predicting the displacement trend of the base plate at multiple future moments.

[0063] Steps one and two of this embodiment are exactly the same as those in embodiment 3, and will not be repeated here.

[0064] Step three of this embodiment involves constructing a combined deep learning model based on gated recurrent units (GRUs) and graph convolutional networks (GCNs).

[0065] First, a graph structure is constructed using the first, second, and third monitoring points within the bottom heave monitoring section as graph nodes. In this embodiment, a total of L bottom heave monitoring sections are arranged along the longitudinal direction of the tunnel, where L=10, and each bottom heave monitoring section contains 5 monitoring points, with the specific type being the same as in Embodiment 3.

[0066] The total number of nodes in the graph convolutional network GCN is N=10, the bottom drum monitoring section × 5 monitoring points / bottom drum monitoring section = 50, and the edge connection rules between nodes are as follows: Within the same bottom drum monitoring section, different monitoring points are fully connected to capture the spatial coupling relationship of different parameters on the same section. Connecting monitoring points of the same type in adjacent floor drum monitoring sections, such as the first monitoring point of the q-th floor drum monitoring section and the first monitoring point of the (q+1)-th section, to capture the propagation relationship of floor drum deformation along the longitudinal direction of the tunnel; thereby constructing a normalized adjacency matrix A, with a dimension of 50×50, where the element A_ij represents the edge weight between node i and node j, and is calculated using a Gaussian kernel function: A_ij = exp(-dist(i,j)). 2 / (2σ1 2 )), where dist(i,j) represents the spatial distance between two monitoring points, and σ1 is the standard deviation of the distance between all node pairs.

[0067] In this embodiment, the structural reference of the combined deep learning model Figure 4 The combined deep learning models specifically include the following: The input layer receives a multidimensional feature time sequence RT, with the dimension being the sliding window length × feature dimension, i.e., 144 × 5; The two-layer graph convolutional network (GCN) has 64 output channels in the first layer and 32 output channels in the second layer. The activation function for both layers is ReLU. The spatial feature matrix output by the second-layer GCN is flattened along the node dimension to a 1600-dimensional vector, which is the number of nodes × 32, i.e., 50 × 32 = 1600, and used as the input for each time step. The gated recurrent unit (GRU) adopts a two-layer stacked structure, with 128 hidden units in each layer and a time step equal to the length of the sliding window, i.e., 144 steps. The final hidden state is output by the last layer of the GRU, with a dimension of 128. The fully connected layer maps the 128-dimensional hidden state to a scalar output, which is the predicted value of the base plate displacement at the future time Δt.

[0068] It should be noted that the future time Δt is preset by the operator to be an integer multiple of the sampling frequency interval. In this embodiment, the sampling interval is 30 minutes, and Δt is 12 hours, that is, 24 sampling intervals.

[0069] The combined deep learning model was trained using the mean squared error loss function and the Adam optimizer. The initial learning rate was 0.001, which decayed to 0.9 times the original rate every 50 epochs. The training sliding window length was 144 time steps, the sliding step size was 72 time steps, the batch size was 32, and a total of 500 epochs were trained. The training data used the monitoring data from the previous 8 months, and the validation set used the data from the 9th month. The early stopping mechanism stopped training when the validation loss did not decrease for 20 consecutive epochs.

[0070] In this embodiment, the specific method for outputting the predicted value of the base plate displacement at future times is as follows: Extract the multidimensional feature time sequence RT={λmax,HST,X(D;F),X(D;W),X(F;W)} of any sliding window T. After forward propagation through GCN and GRU, the fully connected layer outputs a scalar value d_Δt, which is the predicted value of the base plate displacement 12 hours later. Next, to achieve multi-step prediction, this embodiment adjusts the output dimension of the fully connected layer to an integer M preset by the operator.

[0071] In this embodiment, M=5, and five different future times are obtained from the operator's input: Δt1=12 hours, Δt2=24 hours, Δt3=36 hours, Δt4=48 hours, and Δt5=60 hours. The prediction adopts a rolling generation method. First, the displacement value d_Δt1 after 12 hours is predicted using the current sliding window RT. Then, d_Δt1 is used as a known value and concatenated with the original sequence to construct a new sliding window. The earliest sampling time is discarded, and the model is input again to predict the displacement value d_Δt2 after 24 hours. This process is repeated until the prediction results of all five future times are generated. Based on this rolling prediction, the problem of excessive cumulative error in the model by directly outputting multiple times can be avoided.

[0072] In this embodiment, step four is basically the same as in embodiment 2. The difference is that a multi-level risk matrix verification is performed independently for the predicted value at each future time, and warning signals for the next 12 hours, 24 hours, 36 hours, 48 ​​hours, and 60 hours are output. If the warning level at any future time reaches the warning or alarm level, the system terminal immediately issues the corresponding level of sound, light, and SMS warning and displays it on the screen.

[0073] The beneficial effect of this embodiment is that it captures the spatial propagation relationship between different monitoring sections and between different monitoring points on the same section through graph convolutional networks, enabling the combined deep learning model to use monitoring data from neighboring areas to assist in the prediction of the current bottom heave monitoring section. It is suitable for long tunnels with drastic changes in geological conditions. The multi-step rolling prediction mechanism provides a longer response time window for engineering maintenance, which facilitates the early organization of construction machinery and materials.

[0074] Example 5: Based on the above embodiments, this embodiment further specifies the construction details of the multi-level risk matrix and the specific output method of the early warning signal. Specifically, steps one, two, and three all adopt the implementation method described in embodiment four, but parameters such as sampling frequency and monitoring cycle can be adjusted according to different tunnel levels.

[0075] This embodiment uses a high-speed railway tunnel traversing a fault fracture zone as an example. The tunnel passes through a region of moderately weathered, thick-layered layered limestone and dolomite interbedded with marl, with bedding angles ranging from 15° to 25°. The tunnel is designed for a speed of 350 km / h. The invert arch structure is made of C40 reinforced concrete, and the surrounding rock at the arch foot is strongly weathered silty mudstone. The groundwater level is located 3 to 5 meters below the invert arch. Operators set up a monitoring section every 30 meters along the tunnel's longitudinal direction, for a total of 15 sections. Each section has 5 monitoring points. Similar to Embodiment 4, the sampling frequency is once every 15 minutes, and the monitoring period is 90 consecutive days.

[0076] For the aforementioned high-speed railway tunnel, the specific method for constructing a multi-level risk matrix in step four is as follows: First, data from 12 historical cases of floor heave under similar geological conditions in the area where the high-speed railway tunnel is located are collected, and the critical displacement values ​​at which the floor heave reaches the lining cracking, track irregularity exceeding limits, and invert arch structure failure are statistically analyzed. Then, combined with the sensitivity analysis of the strength parameters of the layered rock mass structure surface, the limit displacement corresponding to the floor heave reaching the critical value of lining cracking is 65mm. Considering a safety factor of 1.3, the alarm threshold is set to 50mm, the warning threshold is set to 30mm (corresponding to 46% of the limit displacement), and the attention threshold is set to 15mm (corresponding to 23% of the limit displacement). Based on this, the operators calibrate three thresholds according to the actual needs of the tunnel: the first threshold, specifically the safety threshold, is set to 15mm; the second threshold, specifically the attention threshold, is set to 30mm; and the third threshold, specifically the alarm threshold, is set to 50mm.

[0077] A multi-level risk matrix, using the predicted displacement of the base plate as input variable, divides the risk into four consecutive risk level intervals: Interval 1: Predicted displacement of the base plate < 15mm, corresponding to the safety level, outputs a green safety signal; Interval 2: 15mm ≤ predicted displacement of base plate < 30mm, corresponding to the attention level, output a blue attention signal; For interval 3: 30mm ≤ predicted displacement of the base plate < 50mm, the corresponding warning level is to output a yellow warning signal; Interval 4: If the predicted displacement of the base plate is ≥50mm, a red alarm signal will be output, corresponding to the alarm level.

[0078] In this embodiment, the predicted values ​​of the base plate displacement at three time points in the next 6 hours, 12 hours, and 24 hours are input into a multi-level risk matrix. If the predicted value at any future time point falls into interval 2, interval 3, or interval 4, the system terminal will issue a warning signal in real time through three channels, specifically including: on-site alarm, SMS push, and platform push.

[0079] On-site audible and visual alarms are installed at the tunnel entrance and every 200 meters inside the tunnel, with different sounding modes depending on the warning level.

[0080] SMS push notifications are sent to the pre-bound mobile phone numbers of six maintenance personnel via the 4G DTU module built into the system terminal.

[0081] The platform pushes early warning data packets to the railway maintenance section's monitoring and dispatching platform via the MQTT protocol. The data packets include the cross-section location, predicted displacement value, and early warning level. At the same time, the system terminal locally saves complete early warning logs, including timestamps, early warning levels, trigger thresholds, operator confirmation time, and processing results.

[0082] The beneficial effects of this embodiment are as follows: the multi-level risk matrix based on historical floor heave data has clear physical meaning and engineering operability; the four-level classification is consistent with the early warning levels in the current railway tunnel maintenance specifications; and the three release methods of sound and light, SMS and platform complement each other, ensuring that early warning information can reach maintenance personnel at different levels in a timely manner, thereby reducing the risk of train interruption caused by sudden floor heave.

[0083] Example 6: Based on Examples 1-5, this embodiment further optimizes the tunnel floor heave deformation monitoring method based on time-series data for the complex scenario in cold-region tunnels where floor heave deformation is affected by the coupling effect of both surrounding rock rheology and frost heave.

[0084] In seasonally or perennially frozen soil regions, tunnel floor heave deformation is composed of two superimposed components: one is the floor heave component D_heave caused by surrounding rock rheology, stress relaxation, and groundwater seepage; the other is the frost heave component D_frost caused by frost heave forces during the freeze-thaw cycle. These two components differ significantly in their temporal characteristics: the floor heave component exhibits a monotonically increasing or accelerating growth trend, while the frost heave component displays an approximately sinusoidal fluctuation characteristic with an annual cycle.

[0085] At each monitoring section of the inverted arch, in addition to the two first monitoring points (symmetrically arranged on the left and right sides of the invert), two second monitoring points (arranged at the toes of the left and right side walls), and one third monitoring point (arranged in the deepest aquifer below the tunnel centerline) described in Example 3, two fourth monitoring points are added, symmetrically arranged on both sides of the roadbed centerline within the frost depth influence range below the inverted arch. A frost heave displacement gauge with a range of 0–30 mm and an accuracy of 0.01 mm is installed at the fourth monitoring point to collect the frost heave displacement caused by the freeze-thaw cycle in real time.

[0086] The hydrostatic level, vibrating wire pressure cell, pore water pressure gauge, and frost heave displacement gauge are all connected to the system terminal via a CAN bus, with a uniform sampling frequency of once every 30 minutes. The system terminal has a built-in real-time clock for synchronous sampling, ensuring that the four types of sensors trigger data acquisition at the same sampling time.

[0087] Starting from the beginning of the monitoring cycle at time t1, the system terminal simultaneously triggers the hydrostatic level, vibrating wire pressure cell, pore water pressure gauge and frost heave displacement gauge to collect data at the synchronous sampling frequency. The total displacement of the bottom plate, the surrounding rock stress, the groundwater level and the frost heave displacement are collected each time. The collected data are summarized in chronological order to form four-parameter time series data.

[0088] Specifically, the arithmetic mean of the readings of the two frost heave displacement gauges at the fourth monitoring point at the same sampling time is taken and recorded as the frost heave displacement Fz at that time. The parameter time series data is stored in the system terminal in a structured text format and uploaded to the cloud server for persistent storage every 24 hours via wireless network.

[0089] In this embodiment, the data cleaning part in step two is the same as in embodiment 3. A composite strategy combining the isolated forest algorithm and the 3σ criterion is used to identify and process outliers. The specific details will not be elaborated here.

[0090] After cleaning, the following four data sequences were obtained: Total displacement sequence of the base plate: DT1, DT2, ..., DTj, collected by the static level instrument at the first monitoring point, which is the total displacement including the frost heave component; Surrounding rock stress sequence: F1, F2, ..., Fj; Groundwater level sequence: W1, W2, ..., Wj; Frost heave displacement sequence: FZ1, FZ2, ..., FZj, collected by the frost heave displacement gauge at the fourth monitoring point.

[0091] The differential algorithm is used to separate the frost heave component D_frost and the bottom drum component D_heave: D_total=D_frost+D_heave; the frost heave component is deducted from the total displacement through periodic identification (annual cycle), and the remaining bottom drum component is fed into a combined deep learning model composed of a gated recurrent unit (GRU) and a graph convolutional network (GCN).

[0092] Based on the multi-level risk matrix constructed as in Example 5, a frost heave-specific early warning matrix is ​​then constructed, using the frost heave periodic intensity index P_frost and the frost heave residual ε as input variables: If P_frost < 1.2 and |ε i If the diameter is less than 5mm, output a normal frost heave signal; If 1.2 ≤ P_frost < 1.5 or 5 mm ≤ |ε i If the diameter is less than 10mm, output a frost heave warning signal. If 1.5 ≤ P_frost < 2.0 or 10 mm ≤ |ε i If the diameter is less than 15mm, a frost heave warning signal will be output. If P_frost ≥ 2.0 or |ε i If the diameter is ≥15mm, output a frost heave alarm signal.

[0093] The frost heave warning system operates independently of the bottom drum risk matrix, and both output warnings in parallel. If the bottom drum risk matrix outputs a warning or alarm signal, or if the frost heave warning system outputs a warning or alarm signal, the system terminal will immediately issue corresponding warning information through the audible and visual alarm, SMS push module, and monitoring and dispatch platform, and will mark the warning type on the display screen as: bottom drum type, frost heave type, or coupled type.

[0094] Then, if the following conditions are met simultaneously, a coupled alarm signal will be output: Condition 1: The output of the bass drum risk matrix is ​​greater than or equal to the warning level; Condition 2: The special early warning output for frost heave is ≥ the warning level; Condition 3, Pure bottom drum displacement rate v_D = (D j - D j-1 ) / Δt > 0.5mm / day.

[0095] A coupled alarm indicates that both frost heave and heave are aggravating simultaneously, requiring immediate activation of emergency response procedures.

[0096] This embodiment aims to achieve decoupled monitoring and independent early warning of frost heave component and bottom bulge component by adding frost heave monitoring points and introducing a differential separation algorithm.

[0097] All data in the formulas described above are numerical calculations after removing their dimensions. Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0098] The above description is merely an example and illustration of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0099] It should be stated that all user data collected in this application was collected with the user's consent and authorization, the use of user data is legal and compliant, and the use and processing of user data comply with the relevant laws, regulations and standards of the region.

Claims

1. A method for monitoring tunnel floor heave deformation based on time-series data, characterized in that, The method includes: Step 1: Deploy an integrated sensor network at the target point of the bottom heave monitoring section of the target tunnel. Extract the bottom plate displacement, surrounding rock stress and groundwater level of the bottom heave monitoring section in real time at the synchronous sampling frequency from the beginning of the monitoring cycle to construct multi-parameter time series data. Step 2: After performing data cleaning on the multi-parameter time series data, data sequences associated with various parameters are generated, and multi-dimensional feature time series sequences are generated based on the numerical performance of each data sequence. Step 3: Construct a combined deep learning model based on gated recurrent units (GRU) and graph convolutional networks (GCN); GCN is used to capture the spatial propagation relationship between different target points, while GRU is used to capture the evolution law of multidimensional feature time series in the time dimension, and output the predicted value of the base plate displacement at future time. Step four: Verify the predicted displacement value of the base plate with the pre-constructed multi-level risk matrix, and output the corresponding level of early warning signal.

2. The method for monitoring tunnel floor heave deformation based on time-series data according to claim 1, characterized in that, In step one, the bottom heave monitoring section of the target tunnel and the target point at the bottom heave monitoring section are determined by the operators after on-site survey. The target points at the same bottom drum monitoring section include at least one first monitoring point, one second monitoring point, and one third monitoring point.

3. The method for monitoring tunnel floor heave deformation based on time-series data according to claim 2, characterized in that, In step one, the integrated sensor network includes a hydrostatic level, a vibrating wire pressure cell, and a pore water pressure gauge. The hydrostatic level is installed at the first monitoring point, the vibrating wire pressure cell is installed at the second monitoring point, and the pore water pressure gauge is installed at the third monitoring point. All sensors are connected to the system terminal via the system bus to perform data transmission.

4. The method for monitoring tunnel floor heave deformation based on time-series data according to claim 3, characterized in that, In step one, the specific method for constructing multi-parameter time-series data is as follows: Sensors deployed at the first, second, and third monitoring points of the target location at any monitoring section of the bottom drum; The current time is taken as the start time t1 of the preset monitoring cycle for data collection; Starting from time t1, the displacement of the bottom plate, the stress of the surrounding rock, and the groundwater level collected by the hydrostatic level, vibrating wire pressure cell, and pore water pressure gauge at each sampling time are extracted and summarized in chronological order, and recorded as multi-parameter time series data.

5. The method for monitoring tunnel floor heave deformation based on time-series data according to claim 1, characterized in that, In step two, the specific method for generating the data sequences associated with various parameters is as follows: All base plate displacements are extracted from the multi-parameter time series data and rearranged in chronological order, which is recorded as the initial base plate displacement sequence. Calculate the standard deviation σ and mean μ of the initial base plate displacement sequence, and use the isolated forest algorithm to filter out abnormal base plate displacements in the initial base plate displacement sequence; The abnormal base plate displacement is checked against [μ-3σ,μ+3σ]. If the abnormal base plate displacement is outside the range of [μ-3σ,μ+3σ], it is considered that the abnormal base plate displacement has failed the check and is marked as candidate data to be removed. If there are N consecutive abnormal base plate displacements in the initial base plate displacement sequence, including candidate rejection data, and none of them pass the verification, then the N base plate displacements are considered normal data. Conversely, all candidate data to be removed will be removed, and linear interpolation will be used to fill in the removed data. Extract all base plate displacements from the initial base plate displacement sequence after the above operations, and denot them as base plate displacement sequences D1, D2, ..., Dj, where D1 corresponds to t1, and j represents the total number of sampling times within a monitoring period; Similarly, determine the surrounding rock stress sequence F1, F2, ..., Fj and the groundwater level W1, W2, ..., Wj.

6. The method for monitoring tunnel floor heave deformation based on time-series data according to claim 5, characterized in that, In step two, the specific method for generating the multidimensional feature time series sequence is as follows: Align the bottom plate displacement sequence D1, D2, ..., Dj, the surrounding rock stress sequence F1, F2, ..., Fj, and the groundwater level sequence W1, W2, ..., Wj according to the time sequence to form a triplet sequence {(Di,Fi,Wi)}, where i is the counting index, 1≤i≤j; For all triples, a chaotic phase space reconstruction with embedding dimension m=3 and time delay τ is adopted. The historical m time values ​​of each type of parameter constitute a phase point, thereby mapping the one-dimensional time series data to the three-dimensional phase space. The time delay τ is determined by the first minimum value of the mutual information method. Starting from time t1, continuously lock the sliding window T, where T = j / 10; calculating the maximum Lyapunov exponent λ of the floorboard displacement sequence in all sliding windows in a three-dimensional phase space max and the rescaled range Hurst exponent HST; The joint mutual information X(D;F) between the bottom plate displacement and the surrounding rock stress is calculated using X(D;F)=H(D)+H(F)−H(D,F), where the information entropy H is estimated using the equidistant histogram method. Similarly, calculate the joint mutual information X(D;W) and the joint mutual information X(F;W); The maximum Lyapunov exponent λ within any sliding window T. max The rescaled range Hurst exponent (HST), combined with the joint mutual information X(D;F), X(D;W), and X(F;W), generates a multidimensional feature time series RT={λ} associated with the corresponding sliding window T. max ,HST,X(D;F),X(D;W),X(F;W)}.

7. The method for monitoring tunnel floor heave deformation based on time-series data according to claim 6, characterized in that, In step three, the specific method for constructing a combined deep learning model based on gated recurrent units (GRU) and graph convolutional networks (GCN) is as follows: Using the first, second, and third monitoring points within the bottom drum monitoring section as graph nodes, and the corresponding first, second, and third monitoring points in adjacent bottom drum monitoring sections as associated graph nodes, a graph structure is constructed to generate a normalized adjacency matrix. Obtain a combined deep learning model consisting of gated recurrent units (GRUs) and graph convolutional networks (GCNs), including fully connected layers; The graph convolutional network GCN uses two layers of graph convolution, with a multidimensional feature temporal sequence as input and a spatial feature matrix as output. The gated recurrent unit (GRU) uses the flattened spatial feature matrix output by the GCN as the input for each time step. It employs a two-layer stacked GRU to capture the temporal evolution pattern and outputs the final hidden state. The final hidden state is connected to the fully connected layer, and the predicted value of the base plate displacement at the future time Δt is output. The future time Δt is the time after the current monitoring cycle ends, which is preset by the operator and is an integer multiple of the sampling frequency interval. The combined deep learning model is trained using the mean squared error loss function and the Adam optimizer. The training sliding window length is 1 / 10 of the monitoring period, and the training sliding window step size is half of the training sliding window length.

8. The method for monitoring tunnel floor heave deformation based on time-series data according to claim 7, characterized in that, In step three, the specific method for outputting the predicted value of the base plate displacement at future times is as follows: Extracting the multidimensional feature time series RT={λ from an arbitrary sliding window T max After forward propagation through GCN and GRU, the fully connected layer outputs a scalar value d_Δt, which is the predicted value of the bottom plate displacement at the future time Δt. Adjust the output dimension of the fully connected layer to the integer M preset by the operator; Obtain M distinct future moments input by the operator: Δt, Δt2, ..., ΔtM; The fully connected layer outputs M different predicted values ​​of the base plate displacement at future times. Each predicted value of the base plate displacement is used as part of the prediction input for the next future time, and the prediction results for subsequent future times are generated in a rolling manner and then output.

9. The method for monitoring tunnel floor heave deformation based on time-series data according to claim 8, characterized in that, In step four, the specific method for verifying the predicted value against the pre-constructed multi-level risk matrix and outputting the corresponding level of early warning signal is as follows: A multi-level risk matrix is ​​constructed, which takes the predicted value of the base plate displacement as the input variable and divides it into four consecutive risk level intervals, with each risk level interval corresponding to a warning level. Input the predicted future base plate displacement values ​​output from step three into the multi-level risk matrix; If the predicted displacement of the base plate is lower than the first threshold, a safety signal is output. If the predicted displacement of the base plate is greater than or equal to the first threshold and lower than the second threshold, an attention signal is output. If the predicted displacement of the base plate is greater than or equal to the second threshold but lower than the third threshold, an alarm signal will be output. If the predicted displacement of the base plate is greater than or equal to the third threshold, an alarm signal will be output. The first, second, and third thresholds are set by the operators according to the actual needs of the tunnel, and the warning signals are released in real time through the system terminal via sound and light, SMS and platform push.

Images