GCN-LSTM-based tailing dam multi-point settlement deformation prediction method

Abstract

The invention discloses a GCN-LSTM-based tailing dam multi-point settlement deformation prediction method. The method comprises the steps of collecting and preprocessing original settlement monitoring data of monitoring points in a research area; constructing a graph structure for settlement data between all monitoring point pairs, setting a threshold value, connecting node pairs with significant correlation coefficient relationships by using edges, constructing a weighted undirected graph, and converting the weighted undirected graph into a normalized adjacent matrix as the input of a graph convolutional network GCN; extracting the spatial topology of each monitoring point by using a GCN; capturing long-term time dependence in the settlement process by using a gating mechanism of the LSTM network; and fusing the GCN and the LSTM network, and finally outputting a predicted value through a full connection layer. According to the method, the spatial topological graph among the monitoring points of the tailing dam is constructed, and the GCN and LSTM networks are fused, so that the spatial correlation among the monitoring points and the time dynamic characteristics of the settlement data are effectively captured, and high-precision tailing dam multipoint settlement prediction is realized.

Description

Technical Field

[0001] This invention relates to a method for predicting settlement and deformation of tailings dams, and more particularly to a method for predicting settlement and deformation of tailings dams at multiple points based on GCN-LSTM. Background Technology

[0002] Tailings dams are crucial infrastructure in the mining and beneficiation industry, used to store tailings waste generated during mining operations. The stability of these tailings dams has a critical impact on the ecological environment and production operations; abnormal settlement or dam failure often leads to severe environmental damage and loss of life and property. The settlement process of tailings dams is influenced by various factors, including the ore body structure, material properties, loads, and hydrological conditions, exhibiting complex nonlinear spatiotemporal coupling characteristics. Therefore, accurate prediction of tailings dam settlement deformation is essential for ensuring the safe operation of the dam and providing timely early warnings.

[0003] Traditional methods for predicting tailings dam settlement mainly include physical and mechanical models and time-series statistical models. Physical models (such as finite element analysis) usually require accurate geological parameters and complex numerical calculations, making them difficult to adapt to actual working conditions. Time-series analysis methods (such as ARIMA and grey models) perform well in handling linear relationships, but they are difficult to characterize complex nonlinearities and spatial correlations between multiple monitoring points.

[0004] In recent years, machine learning and deep learning methods have begun to be applied to the field of tailings dam deformation prediction. Recurrent neural networks such as LSTM can process long-term temporal information, but their utilization of spatial topology is limited. Graph neural networks (such as GCN) can extract spatial structural information, but in single applications, they often lack the capture of temporal dynamic features. Existing research mostly focuses on settlement prediction at single points or in single regions, neglecting the spatial correlation between monitoring points within the dam body. Furthermore, the parameter sensitivity and computational efficiency of deep models also limit further improvements in prediction performance. Therefore, there is an urgent need for a high-precision settlement deformation prediction method that can fully integrate the spatial correlation and temporal dynamic information of multiple monitoring points in tailings dams. Summary of the Invention

[0005] To address the shortcomings of the aforementioned technologies, this invention provides a method for predicting settlement and deformation of tailings dams at multiple points based on GCN-LSTM.

[0006] To solve the above technical problems, the technical solution adopted by this invention is: a method for predicting settlement and deformation of tailings dams at multiple points based on GCN-LSTM, comprising the following steps: Step 1: Data Acquisition and Preprocessing: Collect raw settlement monitoring data from monitoring points in the study area and preprocess the raw settlement monitoring data; Step 2: Construct a graph structure for the settlement data of all monitoring points: Quantize the settlement time series data of each monitoring point pairwise, set a threshold, connect the nodes with significant correlation coefficients with edges, construct a weighted undirected graph, and convert it into a normalized adjacency matrix as the input of the graph convolutional network GCN. Step 3: Spatial Feature Extraction: The spatial topology of each monitoring point is extracted using a graph convolutional network (GCN), and neighbor node information is aggregated through a multi-layer GCN. Step 4: Temporal Feature Extraction: Input the spatial topological features extracted by the Graph Convolutional Network (GCN) and the time series data into the LSTM network, and use the gating mechanism of the LSTM network to capture the long-term temporal dependence in the settlement process. Step 5: Spatiotemporal feature fusion and prediction. The settlement data of each time step of the spatial features output by GCN are fused and input into LSTM for joint modeling. Finally, the predicted value is output through a fully connected layer.

[0007] Preferably, in step one, a denoising preprocessing method based on db4 wavelet and SURE threshold is performed; this includes the following three stages: 1.1) Use db4 wavelet for three-level multi-resolution decomposition to peel the signal layer by layer into low-frequency trends and high-frequency details; 1.2) Perform SURE thresholding on the high-frequency coefficients of each layer to remove high-frequency noise and retain the detailed features of the settlement signal; 1.3) Wavelet reconstruction: The signal is recombined into a denoised signal by wavelet reconstruction, removing noise components and restoring it to the original signal form.

[0008] Preferably, step two, the graph structure construction, specifically includes the following steps: 2.1) Pearson correlation coefficient is introduced for pairwise quantitative analysis: The Pearson correlation coefficient is calculated for each pair of settlement monitoring data samples from monitoring points. The formula is as follows: In the formula, r is the Pearson correlation coefficient; n is the total number of samples; i is the i-th sample; Xi and Yi are the two variables involved in the calculation. , These are the average values ​​of the two variables involved in the calculation; 2.2) Based on a preset threshold, if the value is greater than the threshold, an edge is established between node i and node j, and rij is used as the weight of the edge; then, the constructed weighted adjacency matrix A is subjected to symmetric normalization to obtain the normalized adjacency matrix. : In the formula: A is a matrix of |V|×|V|, representing the connection relationship between nodes in the graph, and D is the degree matrix.

[0009] Preferably, the SURE threshold is 0.6.

[0010] Preferably, in step three, the specific process of spatial feature extraction is as follows: using a normalized adjacency matrix, GCN fuses the features of each monitoring point and its neighbors layer by layer, and realizes the extraction and aggregation of spatial features of monitoring points through multi-layer convolution operations. The calculation of each layer of graph convolution can be expressed as: In the formula, =A+I, yes The degree matrix is ​​given by H(l), where H(l) represents the node features of the l-th layer, W(l) represents the trainable weights, and σ is the activation function. By stacking multiple layers of GCN structure, the spatial dependency relationship between tailings dam monitoring points can be comprehensively characterized.

[0011] Preferably, the LSTM network introduces three gating mechanisms: a forget gate, an input gate, and an output gate, as shown in the following formula: Forgotten Gate: ; Input Gate: ; ; Cell state: ; Output gate: ; ( ).

[0012] in, yes t Input variables at time, It is the information output by the previous LSTM unit. It is an activation function. W It is a weight matrix; b It is a bias term. , , These are the forgotten value, the input gate's output value, and the output gate's output value, respectively. yes t Candidate value variable at time, yes t The state of neurons at any given time, This is the output value of the current LSTM unit.

[0013] Preferably, the Adam optimizer is used to tune the hyperparameters of the GCN-LSTM model obtained by joint modeling.

[0014] To address the problems of low accuracy and insufficient utilization of spatiotemporal features in existing tailings dam multi-point settlement prediction methods, this invention discloses a GCN-LSTM-based method for predicting settlement deformation at multiple points in tailings dams. By constructing a spatial topology map between monitoring points of the tailings dam and integrating GCN and LSTM networks, the method effectively captures the spatial correlation between monitoring points and the temporal dynamic characteristics of settlement data, thereby achieving high-precision multi-point settlement prediction for tailings dams. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the framework of the prediction method of the present invention.

[0016] Figure 2 This is a time series data graph of the initial dam monitoring points in an application example of the present invention.

[0017] Figure 3 This is a time series data graph of the monitoring point of the first-level sub-dam in an application example of the present invention.

[0018] Figure 4 This is a time series data graph of the monitoring point of the secondary sub-dam in an application example of the present invention.

[0019] Figure 5 This is a comparison chart of data before and after X1 denoising in an application example of the present invention.

[0020] Figure 6 This is a comparison chart of data before and after D2 denoising in an application example of the present invention.

[0021] Figure 7 This is a graph showing the results of Pearson correlation analysis in an application example of the present invention.

[0022] Figure 8 This is a spatial correlation diagram of monitoring points in an example of the application of the present invention.

[0023] Figure 9 This is a diagram showing the initial dam settlement prediction results in an example of the application of this invention.

[0024] Figure 10 This is a diagram showing the settlement prediction results of a dam in an example of the application of this invention.

[0025] Figure 11 This is a diagram showing the settlement prediction results of the secondary dam in an embodiment of the present invention. Detailed Implementation

[0026] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0027] To address the problems of low accuracy and insufficient utilization of spatiotemporal features in existing tailings dam multi-point settlement prediction methods, this invention proposes a GCN-LSTM-based method for predicting multi-point settlement and deformation of tailings dams. Figure 1 The framework flowchart, as shown, includes the following steps: Step 1: Data Acquisition and Preprocessing; Raw settlement monitoring data from monitoring points in the study area were collected, and the raw settlement monitoring data were cleaned to remove errors and duplicate values.

[0028] A wavelet thresholding method based on db4 wavelet and SURE threshold is used for denoising to eliminate high-frequency noise and preserve the true settlement trend.

[0029] In step 1 above, wavelet denoising processing includes the following three stages: 1.1) Using db4 wavelet for 3-level multi-resolution decomposition: db4 wavelet has good time-frequency localization characteristics, is suitable for complex nonlinear and non-stationary signal processing, and can effectively distinguish high-frequency noise and low-frequency settlement trend in settlement signal, thereby peeling the signal layer by layer into low-frequency trend and high-frequency details.

[0030] 1.2) Perform SURE (Stein's Unbiased Risk Estimate) thresholding on each layer of high-frequency coefficients: This method is based on unbiased risk estimation and can automatically select the optimal threshold so that the denoised signal is closest to the original signal in terms of root mean square error (RMSE), thereby effectively removing high-frequency noise and preserving the detailed features of the settlement signal.

[0031] 1.3) Wavelet Reconstruction: The processed coefficients are recombined into a denoised signal through wavelet reconstruction, removing noise components and restoring the original signal form. This denoising method can effectively remove noise from tailings dam settlement data while preserving the main signal characteristics, and is suitable for processing tailings dam settlement monitoring data.

[0032] Step 2: Construct a graph structure for the settlement data of all monitoring points: Quantize the settlement time series data of each monitoring point pairwise, set a threshold, connect the nodes with significant correlation coefficients with edges, construct a weighted undirected graph, and convert it into a normalized adjacency matrix as the input of the graph convolutional network (GCN). Step 2 above, the graph structure construction specifically includes the following steps: 2.1) Pearson correlation coefficient is introduced for pairwise quantitative analysis: The Pearson correlation coefficient is calculated for each pair of settlement monitoring data samples from monitoring points. The formula is as follows: In the formula, r Pearson correlation coefficient; n The total number of samples; i For the first i One sample; X i , Yi These are the two variables involved in the calculation; , These are the average values ​​of the two variables involved in the calculation.

[0033] 2.2) Based on a preset threshold, preferably 0.6, if the threshold is greater than this, an edge is established between node i and node j, and... r ij The weights are used as edge weights; then, the constructed weighted adjacency matrix A is subjected to symmetric normalization to obtain the normalized adjacency matrix. : In the formula: A It is a matrix |V|×|V|, representing the connection relationship between nodes in the graph, where D is the degree matrix.

[0034] Step 3: Spatial Feature Extraction: Extract the spatial topology of each monitoring point using a graph convolutional network (GCN): Aggregate neighbor node information through multi-layer GCN to enhance the ability to express the overall structure of the tailings dam.

[0035] In step three above, the Graph Convolutional Network (GCN), as a hierarchical implementation of the Graph Neural Network (GNN), realizes the transmission and aggregation of information between settlement monitoring points through graph convolution operations.

[0036] By utilizing a normalized adjacency matrix, the Graph Convolutional Neural Network (GCN) fuses features from each monitoring point and its neighbors layer by layer, effectively capturing local spatial structure. As the network depth increases, the receptive field of each node gradually expands, and the feature representations progressively integrate information from a wider neighborhood, thereby enhancing the ability to model the overall topology and spatial relationships of the tailings dam and improving the accuracy of settlement prediction. The GCN extracts and aggregates spatial features of monitoring points through multiple convolutional operations. The computation of each graph convolutional layer can be represented as: In the formula, = A + I , yes The degree matrix, H (l) It is the first l Layer node characteristics, W (l) These are trainable weights. σ It is an activation function (such as ReLU).

[0037] By stacking multiple layers of GCN structures, the receptive field of nodes can be gradually expanded, enabling the model to integrate neighborhood information at different spatial scales, thereby comprehensively depicting the spatial dependencies between tailings dam monitoring points.

[0038] Step 4: Temporal Feature Extraction: Input the spatial topological features extracted by the Graph Convolutional Network (GCN) and the time series data into the LSTM network, and use the gating mechanism of the LSTM network to capture the long-term temporal dependence in the settlement process.

[0039] In step four above, the Long Short-Term Memory (LSTM) network is a deep learning architecture, an improvement on the traditional recurrent neural network. LSTM effectively captures the temporal dependence of tailings dam settlement by processing historical settlement data sequences. Its input layer receives data processed by a multi-layer graph convolutional network (GCN). This data includes the settlement information of the monitoring point itself and the fused spatial information of neighboring monitoring points. Compared with traditional RNNs, LSTM introduces three gating mechanisms: a forget gate, an input gate, and an output gate. It solves the gradient vanishing and gradient exploding problems present in RNNs through memory units. Its core formula is as follows: Forgotten Gate: ; Input Gate: ; ; Cell state: ; Output gate: ; ( ).

[0040] in, yes t Input variables at time, It is the information output by the previous LSTM unit. It is an activation function. W It is a weight matrix; b It is a bias term. , , These are the forgotten value, the input gate's output value, and the output gate's output value, respectively. yes t Candidate value variable at time, yes t The state of neurons at any given time, This is the output value of the current LSTM unit.

[0041] Step 5: Spatiotemporal feature fusion and prediction. The spatial features output by GCN are fused with the settlement data at each time step and input into LSTM for joint modeling. Finally, the predicted value is output through a fully connected layer, and the Adam optimizer is used for hyperparameter tuning to prevent overfitting.

[0042] In step five above, the jointly modeled GCN-LSTM model integrates a Graph Convolutional Network (GCN) and a Long Short-Term Memory (LSTM) network to address the complex spatiotemporal graph structure in tailings dam settlement monitoring data. The GCN module focuses on capturing the spatial dependencies between monitoring points, effectively fusing the features of each monitoring point with the information of its neighboring monitoring points through graph convolution operations, forming a deep understanding of the spatial structure of the tailings dam. The LSTM module excels at modeling the temporal dynamics of settlement data, processing historical settlement data sequences to capture long-term dependencies. In the GCN-LSTM model, the data input to the LSTM module is a clever combination of spatial features processed by GCN and settlement data at the current time step. This fusion method ensures that the LSTM can fully consider spatial correlations when processing time-series data, achieving accurate prediction of tailings dam settlement.

[0043] At each time step, the spatial features extracted by the GCN module are combined with the settlement data for that time step to form an input vector. Subsequently, the LSTM module updates its hidden state at each time step, fusing the dynamic information of the time series with spatial correlations. Therefore, the final hidden state contains the spatiotemporal joint features of the tailings dam settlement. A fully connected layer transforms the final hidden state of the LSTM module into a prediction output, generating a settlement prediction value for each monitoring point.

[0044] The following detailed description of the GCN-LSTM-based multi-point settlement and deformation prediction method for tailings dams will be provided through specific application examples.

[0045] Step 1: Data acquisition and preprocessing.

[0046] The tailings dam of the Hongyangou copper mine in Lueyang County, Shaanxi Province, was selected as the research object. A total of 20 monitoring points were set up on the tailings dam body. The monitoring points were divided into three categories according to the dam structure: initial dam, accumulation dam (first-level sub-dam, second-level sub-dam) and secondary dam. A total of 20 settlement monitoring points were set up, namely: initial dam (J1, G1, X1, G2, J2), first-level sub-dam (D1, D2, D3, D4), second-level sub-dam (C1, C2, C3, C4), and secondary dam (G9, G10, G11, J9, J10, J11).

[0047] The data collection period was from January 2020 to November 2021, with a sampling frequency of once daily, resulting in approximately 700 time-series data points. Figure 2-4 As shown; To address the problem of abrupt changes and high-frequency noise intertwining in the original settlement sequence, which easily obscures the true deformation signal, this invention employs the tightly supported and orthogonally excellent db4 (Daubechies 4) wavelet basis for three-level multi-resolution decomposition, peeling the signal layer by layer into low-frequency trends and high-frequency details. For the thresholding strategy, an adaptive thresholding criterion based on Stein's unbiased risk estimation (SURE) is introduced to perform soft thresholding denoising on the high-frequency coefficients of each layer. This avoids the defects of "over- or under-crimping" with fixed thresholds and dynamically adjusts with the local signal-to-noise ratio, maximizing the preservation of weak energy details related to deformation. After reconstruction by inverse wavelet transform, a smooth and physically meaningful settlement time-series curve is obtained: abrupt changes in amplitude are significantly weakened, high-frequency spikes are effectively eliminated, while key features such as the overall trend, stage-specific abrupt changes, and annual periodic fluctuations are fully preserved. Figures 5-6 A comparison of the data before and after processing at the same monitoring point is presented. It can be seen that the noise reduction curve and the original data are highly consistent at extreme points and inflection points, with a correlation coefficient of over 0.98. This fully verifies the reliability and superiority of the method in extracting weak deformation information of tailings dams, and provides a high signal-to-noise ratio data foundation for subsequent spatial correlation analysis and early warning modeling.

[0048] Step 2: Construct a spatial association map of monitoring points.

[0049] Step 2.1 Correlation analysis: To systematically characterize the spatial linkages of different monitoring points in a tailings dam during the settlement process, this invention introduces the Pearson correlation coefficient to quantify the pairwise settlement time-series data of each monitoring point, uncovering their potential spatial dependencies. Given the complex geological structure and nonlinear load transfer paths of tailings dams, not all interactions between monitoring points are significant. Simply retaining all connections would lead to an overly dense network and the burying of crucial information. Therefore, based on robustness pre-tests, the correlation threshold is set to 0.6, retaining only monitoring point pairs with |r| ≥ 0.6 and passing the 0.05 significance level test. This effectively eliminates weak and spurious correlation noise, highlighting spatial associations with engineering significance.

[0050] Furthermore, using the selected highly relevant detection point pairs as edges and monitoring points as nodes, a relevant network graph that combines representativeness and explanatory power is constructed (see...). Figure 7 The thickness of the edges in the diagram is directly proportional to the absolute value of the correlation coefficient, and the size of the nodes is positively correlated with their weighted degree centrality. This can intuitively reveal the settlement hotspots, potential deformation transmission paths, and key control nodes of the dam body, providing spatial topological basis for subsequent zonal modeling, anomaly tracing, and risk early warning.

[0051] Step 2.2 Construction of Spatial Association Map of Monitoring Points and Spatial Information Modeling: Tailings dam settlement safety monitoring relies on the collaborative work of multiple sensors, including GNSS, hydrostatic levels, and fiber optic gratings. Among these, the real-time settlement sequences collected by settlement monitoring sensors at minute-level frequencies are considered the most direct "barometer" of dam deformation. Since different sensors share the continuous medium of the dam foundation-dam body in space, their data naturally exhibit mechanical interaction. Furthermore, the same sensor is driven by external factors such as reservoir water level, rainfall, and ore discharge at different times, demonstrating significant dynamic non-stationary characteristics. To fully reproduce this complex coupling relationship of "spatial linkage and temporal evolution" in the model, this invention introduces a weighted undirected graph to uniformly model the monitoring field: each node in the graph corresponds to the settlement observation value of a monitoring point at a specific time, and the node attributes are the denoised settlement sequence segments; the edge weights directly quantify the spatial correlation strength between pairs of monitoring points, replacing the shortcomings of traditional Euclidean distance in reflecting the true stress transmission path of the soil and rock mass. In the GCN-LSTM hybrid architecture, the GCN module aggregates neighborhood information on the graph at each time step by using spectral domain convolution. By superimposing multiple receptive fields, it automatically captures multi-scale spatial structural features from local dam sections to the entire dam body, thereby extracting multi-granular embeddings with physical interpretation, providing node representations rich in spatial context for subsequent LSTM temporal evolution modeling.

[0052] The specific construction process follows three steps: "mapping-association-integration". First, all monitoring points in the field are mapped to graph nodes one by one, and the settlement data after wavelet denoising at each point is used as the initial node attribute vector. Second, the Pearson correlation coefficient is calculated for any pair of nodes. When the correlation coefficient is greater than 0.6 and passes the significance test, an undirected edge is established with the edge weight set to |ρ|, which preserves strong correlation connections and gives interpretable numerical differences for different coupling strengths. Finally, all nodes and weighted edges are summarized to generate a sparse, symmetric weighted undirected graph, which is then transformed into a normalized adjacency matrix and used as the input to the GCN.

[0053] This method uses a data-driven approach to reveal and capture the potential spatial correlations within the dam body. The resulting graph topology is highly consistent with the spatial distribution of the on-site monitoring network. The main dam axis region exhibits high-density clustering due to deformation synergy, while edge nodes on both banks form weak connections through ridge monitoring chains. Radial bridging occurs around the flood discharge tunnel, fully validating the rationality of the graph structure. This graph provides a robust, sparse, and physically meaningful structural prior for training subsequent graph neural network models, enabling end-to-end joint optimization of spatial and temporal features within a unified framework. This lays a reliable topological foundation for tailings dam settlement prediction and anomaly tracing. The construction results are shown below. Figure 8 As shown.

[0054] Step 3: Model training and optimization.

[0055] The predictive performance of the tailings dam monitoring model is limited by the model parameter settings. Through multiple rounds of experimental verification and optimization, the optimal parameter configuration was determined, as shown in Table 1. Table 1 Model parameter setting results Table 1 Model parameter setting results

[0056] The optimizer used is Adam. Due to the limited number of tailings dam monitoring data samples and the sparse information, overfitting is prone to occur during training. Therefore, a regularization term needs to be introduced, and its coefficient needs to be analyzed to determine its impact on the model's generalization ability. Furthermore, to more effectively capture the time-dependent characteristics of the settling process, longer time series lengths need to be further adjusted and explored to improve the model's time-series modeling capabilities.

[0057] Step 4: Results Analysis.

[0058] The GCN-LSTM model can simultaneously model and predict multi-dimensional input features, helping managers to grasp the evolution trend of factors related to tailings dam stability in real time. To comprehensively evaluate the model's predictive performance under different dam structures, modeling and analysis were performed on 20 settlement monitoring points for the initial dam, the overflow dam, and the secondary dam, and the prediction results were compared with measured data. Experimental results show that the model exhibits high fitting accuracy and predictive stability in terms of overall trend, consistency, and local deformation response, verifying its practicality and reliability in tailings dam settlement prediction.

[0059] Step 4.1 The initial dam, as the earliest dam body built in the tailings dam, mainly undertakes the functions of initial sand retention and structural support. Its settlement process exhibits large amplitude and fluctuation. Figure 9 The results show that the predicted curves at monitoring points J1, G1, X1, G2, and J2 generally fit the measured curves well, especially in the range where the settlement rate changes significantly, the model can accurately capture the trend inflection points and their dynamic adjustment processes. Among them, point G2 has the smallest prediction error and is the most sensitive to settlement fluctuations; only the prediction at point J1 shows a certain lag when the settlement increases suddenly.

[0060] Step 4.2 As the main area for tailings dam settling, the settling process of the tailings dam is affected by the coupling of multiple factors such as tailings deposition thickness, particle structure, water content and consolidation rate. Figure 10The results show that the predicted data for monitoring points at the first-level sub-dams (D1–D5) and the second-level sub-dams (C1–C4) are highly consistent with the measured data. Especially during the stable, abrupt, and gradual settlement phases, the predicted curves closely follow the measured trends, demonstrating the model's strong ability to identify spatiotemporal features. Specifically, point D3 in the first-level sub-dam experienced the largest settlement amplitude, yet the GCN-LSTM model maintained a low prediction error, indicating its good adaptability to large-amplitude settlements. In contrast, the second-level sub-dam experienced smaller settlement amplitudes, and the predicted curves were more stable, further validating the model's ability to fit low-fluctuation time series. Overall, the results demonstrate that the GCN-LSTM model can effectively adapt to different settlement rates and patterns, exhibiting strong generalization ability and robustness.

[0061] Step 4.3 The secondary dam, as an auxiliary structure of the tailings dam, has a small settlement amplitude and stable changes. Figure 11 The results show that the predicted values ​​at the six monitoring points (G9, G10, G11, J9, J10, J11) of the secondary dam highly overlap with the actual values, with almost no significant deviation, and accurately reflect the slow trend of settlement growth. The model's performance in the secondary dam area further demonstrates that the GCN-LSTM model still possesses good identification and prediction capabilities in low-dynamic regions, unaffected by settlement amplitude limitations. Because the results for the secondary dam are more complete and less affected by external disturbances, its prediction results are superior to those in other areas in both accuracy and stability.

[0062] Step 4.4 The combined prediction results for the three types of dams demonstrate that the GCN-LSTM model exhibits excellent global adaptability in tailings dam settlement prediction. Whether it's the initial dam and first-stage sub-dam experiencing severe settlement, or the stable secondary dam area, the model maintains high prediction accuracy and demonstrates good robustness to complex nonlinear deformation processes, different settlement amplitudes, and different time scales. This is mainly due to the ability of Graph Convolutional Networks (GCN) to extract spatial topological information and the efficient modeling of time-series dependencies by Long Short-Term Memory Networks (LSTM). This model can provide highly reliable technical support for stability assessment and risk warning of different structural regions of tailings dams.

[0063] In summary, this invention innovatively combines Graph Convolutional Neural Network (GCN) and Long Short-Term Memory Network (LSTM) for multi-point settlement and deformation prediction of tailings dams. GCN effectively captures the spatial correlation between monitoring points, while LSTM excels at processing time-series data and capturing temporal dynamic features. This combination overcomes the shortcomings of single models in processing spatiotemporally coupled data, providing a new solution for tailings dam settlement prediction.

[0064] This invention proposes a method for deep fusion of spatial and temporal features. At each time step, the spatial features extracted by GCN are combined with the settlement data of the current time step and used as input to the LSTM module. This ensures that spatial correlation is fully considered in the time series modeling process, and realizes comprehensive modeling of the tailings dam settlement process.

[0065] A wavelet thresholding method based on db4 wavelet and SURE threshold was used to denoise the raw settlement monitoring data. This method can effectively remove high-frequency noise while preserving the true trend of the settlement signal, providing a high-quality data foundation for subsequent feature extraction and model training.

[0066] By calculating the Pearson correlation coefficients between all pairs of monitoring points, a spatial correlation map of the monitoring points is constructed based on a preset threshold. This correlation-based graph structure accurately reflects the spatial relationships between monitoring points, providing a reasonable input data structure for GCN spatial feature extraction and enhancing the model's understanding of the spatial structure of tailings dams.

[0067] The GCN-LSTM model can adapt to the settlement characteristics of different structural regions of tailings dams, including the initial dam, the accumulation dam, and the secondary dam. Whether in the initial dam and accumulation dam where settlement is severe, or in the secondary dam area where changes are stable, the model can maintain high prediction accuracy, demonstrating good global adaptability and robustness, and can cope with deformation processes with different settlement amplitudes and time scales.

[0068] Experimental results show that the predictions of the GCN-LSTM model under different dam structures are highly consistent with the measured data. Especially in the range where settlement rates change significantly, the model can accurately capture trend inflection points and their dynamic adjustment processes, providing highly reliable technical support for the safety monitoring and risk early warning of tailings dams.

[0069] Furthermore, this invention provides a reliable technical means for the safety monitoring and risk early warning of tailings dams. Through high-precision settlement prediction, potential safety hazards can be detected in a timely manner, providing a scientific basis for the safe operation and management of tailings dams, and has significant practical application value.

[0070] The technical solution of this invention is not only applicable to the prediction of settlement and deformation of tailings dams, but also has certain scalability. Its model architecture and method, which integrates spatiotemporal features, can provide reference for other prediction problems involving spatiotemporally coupled data, such as earthquake monitoring and geological disaster early warning.

[0071] The above embodiments are not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the technical solution of the present invention are also within the protection scope of the present invention.

Claims

1. A GCN-LSTM-based tailings dam multi-point settlement deformation prediction method, characterized by: Comprising the following steps: Step one, data acquisition and pretreatment: collecting the original settlement monitoring data of the monitoring points in the study area, and pretreating the original settlement monitoring data; Step two, constructing a graph structure for the settlement data between all monitoring points: quantifying the settlement time series data of each monitoring point, setting a threshold, connecting the node pairs with significant correlation coefficient relationship with edges, constructing a weighted undirected graph, and converting it into a normalized adjacency matrix as the input of the graph convolution network (GCN); Step three, spatial feature extraction: using the graph convolution network (GCN) to extract the spatial topology of each monitoring point, and aggregating the neighbor node information through multiple layers of GCN; Step four, time feature extraction: inputting the spatial topology features extracted by the graph convolution network (GCN) and the time series data into the LSTM network, and using the gating mechanism of the LSTM network to capture the long-term temporal dependence in the settlement process; Step five, spatio-temporal feature fusion and prediction: fusing the settlement data of each time step of the spatial features output by the GCN, inputting it into the LSTM for joint modeling, and finally outputting the prediction value through the fully connected layer.

2. The GCN-LSTM-based tailings dam multi-point settlement deformation prediction method according to claim 1, characterized in that: In step one, the wavelet threshold method based on db4 wavelet and SURE threshold is used for denoising pretreatment; including the following three stages: 1.1) Use db4 wavelet for three-layer multi-resolution decomposition to strip the signal into low-frequency trend and high-frequency detail layer by layer; 1.2) Perform SURE threshold processing on each layer of high-frequency coefficients to remove high-frequency noise and retain the settlement signal detail features; 1.3) Wavelet reconstruction: recombine the denoised signal by wavelet reconstruction to remove noise components and restore it to the original signal form.

3. The GCN-LSTM-based tailings dam multi-point settlement deformation prediction method according to claim 2, characterized in that: In step two, the graph structure construction specifically includes the following steps: 2.1) Introduce Pearson correlation coefficient for pairwise quantitative analysis: calculate the Pearson correlation coefficient of each pair of monitoring point settlement monitoring data samples, the formula is: In the formula, r is a Pearson correlation coefficient; n is the total number of samples; i is the ith sample; Xi and Yi are the average values of the two variables participating in the calculation, respectively; and , ​ 2.2) According to the preset threshold, if it is greater than the threshold, an edge is established between node i and node j, and rij is taken as the weight of the edge; The constructed weighted adjacency matrix A is subjected to symmetric normalization to obtain a normalized adjacency matrix : Where: A is a matrix of ∣V∣x∣V∣, representing the connection relationship between nodes in the graph, and D is the degree matrix.

4. The GCN-LSTM-based tailings dam multi-point settlement deformation prediction method according to claim 3, characterized in that: The SURE threshold is preferably 0.

6.

5. The GCN-LSTM-based tailings dam multi-point settlement deformation prediction method according to claim 4, characterized in that: In step three, the specific process of spatial feature extraction is as follows: using the normalized adjacency matrix, GCN fuses the features of each monitoring point and its neighbors layer by layer, and realizes the extraction and aggregation of monitoring point spatial features through multi-layer convolution operation. The calculation of each layer of graph convolution can be represented as: wherein = A + I, is the degree matrix of the l-th layer, H(l) is the node features of the l-th layer, W(l) is the trainable weights, and σ is the activation function. By stacking multiple layers of GCN structure, the spatial dependence relationship between the monitoring points of the tailings dam is comprehensively described.

6. The GCN-LSTM-based tailings dam multi-point settlement deformation prediction method according to claim 5, characterized in that: The LSTM network introduces three gating mechanisms: forget gate, input gate and output gate, the formulas are as follows: forget gate: ; Input gate: ; ； Cell state: ; Output gate: ; ( )。 wherein, is t the input variable at time t, is the information output by the previous LSTM unit, is an activation function, W is a weight matrix; b is a bias term, , , are the output values of the forget value, input gate, and output gate, respectively, is t the candidate value variable at time t, is t the state of the neuron at time t, is the output value of the current LSTM unit.

7. The GCN-LSTM-based tailings dam multi-point settlement deformation prediction method according to claim 6, characterized in that: The GCN-LSTM model obtained by joint modeling adopts Adam optimizer for hyperparameter tuning.

Images