Slope deformation space-time prediction method based on LSTM-GCN-Attention

Abstract

The invention discloses a side slope deformation space-time prediction method based on LSTM-GCN-Attention, and belongs to the technical field of side slope prediction.The method starts with monitoring of side slope deformation data through a total station deploying and controlling triangular observation network, the monitoring data are expanded through a two-way average interpolation method of a self-adaptive sliding window, and the side slope deformation space-time prediction result is obtained. The method combines the advantages of a long-short-term memory neural network (LSTM) and a graph convolutional neural network (GCN) in the aspect of capturing time sequence and spatial structure features, introduces an attention mechanism to enhance the attention capability of a model on important spatial and temporal features, and dynamically adjusts the node and feature weight, so that the prediction result is more explanatory and robust. And a refined analysis approach is provided for similar side slope deformation time-space synchronization prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of slope prediction, and in particular relates to a spatiotemporal prediction method for slope deformation based on LSTM-GCN-Attention. Background Art

[0002] With the rapid development of engineering construction and transportation infrastructure, slope stability has become a research hotspot in the field of geotechnical engineering. Slope deformation prediction is a key link in disaster prevention and mitigation, and is of great significance for ensuring engineering safety and reducing disaster risks. Traditional slope prediction methods mainly rely on empirical formulas and mathematical statistical models, such as regression analysis based on time series, differential equations, or geomechanical models. Although these methods can reflect the basic laws of slope deformation to a certain extent, they are difficult to effectively handle complex nonlinear dynamic processes. In addition, these methods often ignore the interaction between spatial information and multidimensional data, resulting in limited prediction accuracy.

[0003] In recent years, with the development of machine learning and deep learning technologies, more and more intelligent algorithms have been applied to slope deformation prediction. For example, machine learning models such as support vector machine (SVM) and random forest (RF) can capture the nonlinear characteristics of data to a certain extent, but their ability to mine time series information is insufficient; deep learning models such as recurrent neural network (RNN) and long short-term memory network (LSTM) have significant advantages in processing time series data, but their ability to model complex spatial relationships in slope deformation is limited. At the same time, graph convolutional network (GCN) has attracted much attention due to its excellent performance in processing spatial graph structure data, but existing GCN models usually ignore the temporal dynamic characteristics, the correlation of information between nodes, and slope characteristics. In addition, some existing methods often use simple weighting or superposition methods when processing spatiotemporal data, which makes it difficult to fully mine the complex correlation in spatiotemporal data, resulting in a large room for improvement in the stability and accuracy of the prediction results. Summary of the invention

[0004] The present invention proposes a spatiotemporal prediction method for slope deformation based on LSTM-GCN-Attention. By combining the advantages of long short-term memory neural network (LSTM) and graph convolutional neural network (GCN), and introducing the Attention mechanism, the accuracy and reliability of slope deformation prediction are comprehensively improved. Specifically, this method uses the LSTM model to deeply explore the time-dependent characteristics of slope deformation data, and combines the spatial topological relationship between slope monitoring points with GCN modeling to solve the spatiotemporal complexity that traditional methods cannot handle. The Attention mechanism further enhances the model's ability to pay attention to important spatiotemporal features, dynamically adjusts the node and feature weights, and makes the prediction results more interpretable and robust. Through this multi-level spatiotemporal feature extraction and fusion, this method can handle complex nonlinear dynamic processes, has high precision, strong generalization ability and good engineering applicability, and provides an efficient and reliable solution for slope deformation prediction.

[0005] To this end, the present invention adopts the following technical solutions:

[0006] A spatiotemporal prediction method for slope deformation based on LSTM-GCN-Attention includes the following steps:

[0007] Step 1) Select the target slope work point, design the total station triangulation observation network layout plan, and continuously monitor the horizontal displacement and vertical settlement of the slope according to the design plan, and collect and organize the observation data;

[0008] Step 2) using a bidirectional average interpolation method with an adaptive sliding window to interpolate the observed data to generate a smooth and continuous data signal;

[0009] Step 3) Data normalization: using the range transformation method to normalize the interpolated data in step 2);

[0010] Step 4) Constructing a spatiotemporal data matrix: Using the Python open source numerical computing library Numpy, construct a spatiotemporal sequence data matrix D[T, V, F] with three dimensions of time, space, and node features, where T is the time dimension, V is the space dimension, and F is the feature dimension;

[0011] Step 5) Construct a spatial adjacency matrix: Use the ratio of the height difference between two points to the inclination distance as the spatial association weight coefficient to construct the spatial adjacency matrix A adj [V,V]:

[0012]

[0013] Among them, w ij represents the spatial association weight coefficient, H i , N i Respectively represent the height and inclination coordinates of node i, Hj , H j They represent the height and inclination coordinates of node j respectively;

[0014] Step 6) Divide the training set and the test set: Divide the sample data into a training set and a test set in proportion, and adjust the proportion through multiple pre-training;

[0015] Step 7) LSTM network modeling: Use the long short-term memory network LSTM to model the temporal dynamic characteristics of slope deformation;

[0016] Step 8) GCN network modeling: Use graph convolutional network GCN to capture the spatial dependency between monitoring points;

[0017] Step 9) Attention mechanism optimization and model fusion: Introduce the attention mechanism to dynamically adjust the weights of spatial nodes and time series features to improve the ability to focus on important features; fuse the LSTM time features and the GCN space features, weight them through the Attention mechanism, and generate the final spatiotemporal feature representation; then use the fully connected layer to output the deformation prediction value:

[0018]

[0019] Among them, FC is the fully connected layer, To predict the results;

[0020] Step 10) Model training and evaluation: Build a deep learning Pytoch framework based on Python language, establish a GPU-accelerated deep learning environment suitable for this model and carry out model training; the loss function is mean square error (MSE), and the adaptive optimizer Adam is used to accelerate convergence; the model is evaluated using the test set, and the evaluation indicators include mean square error (MAE), mean square error (MSE) and determination coefficient (R) 2 ;

[0021]

[0022] Among them, the predicted value is: True value: y={y1,y1,...,y n}.

[0023] Furthermore, in step 2), the initial sliding window size is selected according to the data characteristics; for each data point, the window size is adjusted according to the signal gradient or noise level; then the average value in the sliding window is calculated in the forward and reverse directions respectively and the results are fused, and the formula is expressed as:

[0024]

[0025] in and Indicates the results of forward and reverse sliding respectively.

[0026] Furthermore, in step 3), the normalization formula is:

[0027]

[0028] Where x is the original data, x′ is the normalized data, min(x) and max(x) refer to the minimum and maximum values ​​of the original data x, respectively;

[0029] Furthermore, in step 7), LSTM captures long-term dependencies through memory cells and gating mechanisms, and the formula is as follows:

[0030] Input gate: i t =σ(W i x t +U i h t-1 +b i )

[0031] Forget gate: f t =σ(W f x t +U f h t-1 +b f )

[0032] Output gate: o t =σ(W o x t +U o h t-1 +b o )

[0033] Status update: c t =f t ⊙c t-1 +i t ⊙tanh(W c x t +U c xh t-1 +b c )

[0034] h t =o t ⊙tanh(c t )

[0035] Among them, x t Input for the current time, h t is the hidden state at the current moment, c t is the memory unit state, W, U, b are model parameters, and σ is the activation function;

[0036] bi , b f , b o , b c The bias parameters b corresponding to the input gate, forget gate, output gate and state update respectively;

[0037] U i , U f , U o , U c Corresponding to the input gate, forget gate, output gate and state update respectively with the current input x t The associated weight matrix U;

[0038] W i , W f , W o , W c They correspond to the input gate, forget gate, output gate and the state update and the hidden state h at the previous moment. t-1 The associated weight matrix W.

[0039] Furthermore, in step 8), GCN performs convolution operation through the adjacency matrix and node features of the graph, and the formula is as follows:

[0040] H (l+1) =σ(D -1 / 2 AD -1 / 2 H (l) W (l) )

[0041] Among them, H (l) is the node feature of the lth layer, A is the adjacency matrix, D is the node degree matrix, W (l) is the trainable weight matrix and σ is the activation function.

[0042] The improvement of the present invention is:

[0043] 1. Make full use of the advantages of long short-term memory neural network (LSTM) and graph convolutional neural network (GCN) in capturing time series and spatial structure characteristics.

[0044] 2. The attention mechanism is introduced to enhance the model's ability to focus on important spatiotemporal features, dynamically adjust node and feature weights, and make the prediction results more interpretable and robust.

[0045] 3. Taking into full account the characteristics of the slope as a continuum, the deformation of different spatial nodes evolves in a coordinated manner rather than independently. This invention constructs a more scientific spatial adjacency matrix by introducing an association weight coefficient based on height difference and slope, that is, using the ratio of the height difference between two points to the inclination distance as the spatial association weight, accurately characterizing the degree of spatial association of monitoring points on the slope body.

[0046] The beneficial effects of the present invention are:

[0047] The present invention fully combines the advantages of time series modeling, spatial feature capture and dynamic weight optimization, and has the following significant advantages: the nonlinear time dynamic characteristics of slope deformation are captured by LSTM network, which effectively solves the problem that traditional empirical formulas and statistical methods are insufficient in modeling complex time dependencies; the spatial topological relationship of monitoring points is modeled by GCN, which fully considers the characteristics of the slope as a continuum, and the deformation of different spatial nodes is co-evolved rather than independent. The present invention constructs a more scientific spatial adjacency matrix by introducing an association weight coefficient based on height difference and slope, that is, the ratio of the height difference between two points to the inclination distance is used as the spatial association weight to accurately characterize the spatial association degree of monitoring points on the slope body. In addition, the Attention mechanism is introduced to dynamically allocate spatiotemporal feature weights, which further enhances the ability to focus on key features and improves the prediction accuracy and robustness of the model. The overall method realizes the deep fusion of spatiotemporal features, the accurate capture of nonlinear dynamic laws and the strong adaptability to complex geological environments, and provides reliable and advanced technical support for slope stability analysis and disaster warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a flowchart of the algorithm of the spatiotemporal prediction method of the present invention;

[0049] Figure 2 is a curve diagram before interpolation in step 2 of the embodiment of the present invention;

[0050] Figure 3 is a curve graph after interpolation in step 2 of the embodiment of the present invention;

[0051] Figure 4 is a spatial adjacency relationship distribution diagram of the measuring points in step 5 of the embodiment of the present invention;

[0052] Figure 5 It is the spatial adjacency weight matrix distribution diagram of step 5 of the embodiment of the present invention;

[0053] Figure 6 It is a horizontal displacement curve diagram of the spatial node JC21-0 in step 7 of the embodiment of the present invention;

[0054] Figure 7 It is a vertical displacement curve diagram of the spatial node JC21-0 in step 7 of the embodiment of the present invention;

[0055] Figure 8 It is a horizontal displacement prediction curve diagram of measuring point JC21-0 in step 10 of an embodiment of the present invention;

[0056] Fig. 9 It is a vertical displacement prediction curve diagram of measuring point JC21-0 in step 10 of an embodiment of the present invention;

[0057] Fig.10 It is a horizontal displacement prediction curve diagram of measuring point JC22-1 in step 10 of an embodiment of the present invention;

[0058] Fig.11 It is a vertical displacement prediction curve diagram of measuring point JC22-1 in step 10 of an embodiment of the present invention;

[0059] Fig.12 It is a horizontal displacement prediction curve diagram of measuring point JC22-7 in step 10 of an embodiment of the present invention;

[0060] Fig.13 It is a vertical displacement prediction curve diagram of measuring point JC22-7 in step 10 of an embodiment of the present invention;

[0061] Fig.14 It is a horizontal displacement prediction curve diagram of measuring point JC23-3 in step 10 of an embodiment of the present invention;

[0062] Fig.15 It is a vertical displacement prediction curve diagram of measuring point JC23-3 in step 10 of an embodiment of the present invention. DETAILED DESCRIPTION

[0063] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments:

[0064] (1) In this case, the deformation characteristics of the 9# slope of Panzhihua Airport were monitored by the method of deploying a triangular observation network with a total station. The surface displacement observation was achieved by using a geodetic system and a triangular observation network. The surface displacement observation of the landslide was carried out using a total station {Swiss Leica TS11 total station; accuracy: 1″(1mm+1.5ppm*D)}. The accuracy of the observation and the layout accuracy of the observation network were not lower than the national third-class technical standards required by the measurement specifications. The coordinates of the monitoring points are shown in Table 1. The monitoring points at the rear edge of the slope : JC23-0, JC23-1, JC22-0, JC22-1, JC21-0, JC21-1. Rear monitoring points: JC23-2, JC22-2, JC21-2, JC22-3. Middle monitoring points: JC21-4, JC23-4. Front monitoring points of the slope: JC23-3, JC21-3, JC22-4, JC22-5, JC21-5, JC23-5, JC22-6, JC22-7.

[0065] Table 1 Statistics of surface displacement and settlement observation points in slope deformation area

[0066]

[0067]

[0068] (3) Perform initial data processing, mainly including sliding average data interpolation and data normalization. The interpolation process is implemented by writing a bidirectional average interpolation program of an adaptive sliding window in Python, and the data normalization is implemented by writing a range transformation algorithm in Python. Taking the JC21-0 measuring point as an example, the time history curve of the measuring point after interpolation is as follows: Figure 2 and 3 shown.

[0069] (4) Use the Numpy module in Python to construct the slope deformation space-time matrix D[T, V, F], where the time dimension T = 372d, the space dimension V = 20, and the feature dimension F = 2, which are the horizontal displacement and vertical displacement, respectively. According to step 5, the spatial node adjacency matrix is ​​constructed, as follows: Figure 4 and 5 shown.

[0070] (5) Build a deep learning framework based on Pytoch = 2.4.1 and Cuda = 11.8, and establish a GPU-accelerated deep learning environment suitable for this model. Divide the constructed spatiotemporal data matrix into training set and test set in a ratio of 7:3, input the training set data and spatial adjacency matrix into LSTM-GCN-Attention for model training, and use the test set data to evaluate the model performance.

[0071] (6) Extract the prediction results of each spatial node. Taking spatial node JC21-0 as an example, the prediction results of LSTM-GCN-Attention and the traditional time series model LSTM are compared as follows: Figure 6 and 7 As shown in Table 2, it can be seen that the LSTM-GCN-Attention model is significantly better than the LSTM model in terms of the consistency with the actual value.

[0072] Table 2 Model evaluation indicators

[0073] Model LSTM LSTM-GCN-Attention MAE 0.403 0.494 MSE 0.555 0.523 <![CDATA[R 2 ]]> 0.923 0.962

[0074] (7) Use the trained optimal model to predict the horizontal displacement and vertical settlement of different spatial nodes. Taking the measuring points JC21-0, JC22-1, JC22-7, and JC23-3 as examples, the results are as follows: Figure 8-15 It can be seen that the improved spatiotemporal prediction method of slopes in the present invention can accurately predict the development trend of horizontal displacement and vertical displacement of slopes at different spatial positions.

Claims

1. A spatiotemporal prediction method for slope deformation based on LSTM-GCN-Attention, characterized in that: The following steps are involved: Step 1) Select the target slope work point, design the total station triangulation observation network layout plan, and continuously monitor the horizontal displacement and vertical settlement of the slope according to the design plan, and collect and organize the observation data; Step 2) using a bidirectional average interpolation method with an adaptive sliding window to interpolate the observed data to generate a smooth and continuous data signal; Step 3) Data normalization: using the range transformation method to normalize the interpolated data in step 2); Step 4) Constructing a spatiotemporal data matrix: Using the Python open source numerical computing library Numpy, construct a spatiotemporal sequence data matrix D[T, V, F] with three dimensions of time, space, and node features, where T is the time dimension, V is the space dimension, and F is the feature dimension; Step 5) Construct a spatial adjacency matrix: Use the ratio of the height difference between two points to the inclination distance as the spatial association weight coefficient to construct the spatial adjacency matrix A adj [V,V]: Among them, w ij represents the spatial association weight coefficient, H i , N i Respectively represent the height and inclination coordinates of node i, H j , N j They represent the height and inclination coordinates of node j respectively; Step 6) Divide the training set and the test set: Divide the sample data into a training set and a test set in proportion, and adjust the proportion through multiple pre-training; Step 7) LSTM network modeling: Use the long short-term memory network LSTM to model the temporal dynamic characteristics of slope deformation; Step 8) GCN network modeling: Use graph convolutional network GCN to capture the spatial dependency between monitoring points; Step 9) Attention mechanism optimization and model fusion: Introduce the attention mechanism to dynamically adjust the weights of spatial nodes and time series features to improve the ability to focus on important features; fuse the LSTM time features and the GCN space features, weight them through the Attention mechanism, and generate the final spatiotemporal feature representation; then use the fully connected layer to output the deformation prediction value: Among them, FC is the fully connected layer, To predict the results; Step 10) Model training and evaluation: Build a deep learning Pytoch framework based on Python language, establish a GPU-accelerated deep learning environment suitable for this model and carry out model training; the loss function is mean square error (MSE), and the adaptive optimizer Adam is used to accelerate convergence; the model is evaluated using the test set, and the evaluation indicators include mean square error (MAE), mean square error (MSE) and determination coefficient (R) 2 ; Among them, the predicted value is: True value: y={y1,y1,…,y n }.

2. The spatiotemporal prediction method for slope deformation based on LSTM-GCN-Attention according to claim 1 is characterized in that: In step 2), the initial sliding window size is selected according to the data characteristics; for each data point, the window size is adjusted according to the signal gradient or noise level; then the average value in the sliding window is calculated in the forward and reverse directions respectively and the results are fused, and the formula is expressed as: in and Indicates the results of forward and reverse sliding respectively.

3. The spatiotemporal prediction method for slope deformation based on LSTM-GCN-Attention according to claim 1 is characterized in that: In the step 3), the normalization formula is: Where x is the original data, x′ is the normalized data, min(x) and max(x) refer to the minimum and maximum values ​​of the original data x, respectively.

4. The method for spatiotemporal prediction of slope deformation based on LSTM-GCN-Attention according to claim 1 is characterized in that: In step 7), LSTM captures long-term dependencies through memory units and gating mechanisms, and the formula is as follows: Input gate: i t =σ(W i x t +U i h t-1 +b i ) Forget gate: f t =σ(W f x t +U f h t-1 +b f ) Output gate: o t =σ(W o x t +U o h t-1 +b o ) Status update: c t =f t ⊙c t-1 +i t ⊙tanh(W c x t +U c x h t -1 +b c ) h t =o t ⊙tanh(c t ) Among them, x t Input for the current time, h t is the hidden state at the current moment, c t is the memory unit state, W, U, b are model parameters, and σ is the activation function; b i , b f , b o , b c The bias parameters b corresponding to the input gate, forget gate, output gate and state update respectively; U i , U f , U o , U c Corresponding to the input gate, forget gate, output gate and state update respectively with the current input x t The associated weight matrix U; W i , W f , W o , W c They correspond to the input gate, forget gate, output gate and the state update and the hidden state h at the previous moment. t-1 The associated weight matrix W.

5. The method for spatiotemporal prediction of slope deformation based on LSTM-GCN-Attention according to claim 1 is characterized in that: In step 8), GCN performs convolution operation through the adjacency matrix and node features of the graph, and the formula is as follows: H (l+1) =σ(D -1 / 2 AD -1 / 2 H (l) W (l) ) Among them, H (l) is the node feature of the lth layer, A is the adjacency matrix, D is the node degree matrix, W (l) is the trainable weight matrix and σ is the activation function.

Images