A method for calculating the weights of factors affecting landslide surface deformation

Abstract

The present invention provides a method for calculating the weights of factors affecting landslide surface deformation, which belongs to the field of engineering geological technology and solves the technical problem that the landslide displacement prediction is performed due to the same contribution of the landslide influencing factors to the displacement, thereby causing inaccurate results. An attention-based LSTM model is constructed, including: two LSTM models, the time series data of these external factors are input into the attention-based LSTM model for model training, and finally a landslide displacement prediction model is obtained; the attention-based LSTM model outputs the weights of different external factors and judges the influence of different external factors on the landslide displacement, thereby achieving effective prediction of the landslide displacement. The present invention can calculate the weights of various external factors in the landslide displacement prediction based on historical monitoring data, and then judge the influence of various factors on the landslide displacement prediction based on the weights, and finally effectively predict the landslide displacement.

Description

Technical Field

[0001] The invention belongs to the field of engineering geological research, and in particular relates to a method for calculating the weights of factors affecting landslide surface deformation. Background Art

[0002] Landslides are one of the most common natural geomorphic processes, causing huge casualties and property losses around the world every year. According to the National Geological Bulletin and related national geological disaster data released annually by the Ministry of Natural Resources, the total number of geological disasters in my country reached 129,408 between 2010 and 2020, of which landslides accounted for more than 70%. In addition, most provinces and cities in my country are threatened by landslides to varying degrees. It can be seen that my country's landslide disaster prevention and control still faces huge challenges.

[0003] In the past few decades, with the rapid development of intelligent methods, nonlinear models driven by geological big data have received increasing attention in landslide displacement prediction. These models apply nonlinear methods to construct the relationship between landslide displacement and external triggering factors (reservoir water level, reservoir water level fluctuation, rainfall), such as support vector machine (SVM), velocity inverse method and long short-term memory (LSTM) neural network. However, in these models, there are few quantitative and comparative studies on the role of external triggering factors in landslide deformation. For example, in the article "Study on the Stability of Reservoir Bank Slopes under Long-term Reservoir Water Level Fluctuation", only the effect of reservoir water level fluctuation on landslide deformation was studied, ignoring the effects of rainfall and reservoir water level; in "Study on Displacement Prediction of Baishui River Landslide Based on Time Series and PSO-SVR Coupling Model", the authors only used machine learning models to predict the displacement of Outang landslide in the Three Gorges Reservoir area, but did not analyze the weights of reservoir water level and rainfall in surface deformation. In order to solve the problem of quantitative and comparative study of external factors in landslide deformation and improve the physical interpretability of nonlinear models, this study introduced the temporal attention module. The temporal attention module determines the influence of each factor on landslide deformation by assigning different weights to external triggering factors.

[0004] In summary, although many models have been developed for the study of landslide deformation, many research methods of such topics lack the quantification and comparison of external triggering factors in landslide deformation. In order to solve this technical problem, the present invention provides a method for calculating the weights of factors affecting landslide surface deformation. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a method for calculating the weights of factors affecting landslide surface deformation, which solves the problem of quantifying and comparing external triggering factors in landslide deformation, and calculates the weights of various influencing factors in landslide surface deformation by introducing an attention module.

[0006] The technical solution adopted by the present invention is specifically as follows:

[0007] A method for calculating the weights of factors affecting landslide surface deformation, characterized in that the calculation method comprises the following steps:

[0008] S1. Collect historical data of current landslide displacement and external factors affecting landslide deformation at the same sampling time and sampling interval; the sampling interval is the time step, set the time segmentation step, and divide the historical data according to the set time segmentation step to generate a sample data set for landslide displacement prediction. The samples in the sample data set are represented by X, and the labels are the corresponding landslide displacements. A sample contains S small samples corresponding to S moments. The data of each small sample is an external factor. The sample data set is standardized and divided into a training set and a validation set.

[0009] S2. Building an attention-based LSTM model

[0010] The attention-based LSTM model includes a first LSTM model, a second LSTM model, and a fully connected layer. The specific model architecture is:

[0011] The first LSTM model has S LSTM units, a factor attention module is connected between two adjacent LSTM units, the output of the last LSTM unit is the output of the first LSTM model, the output of the first LSTM model is used as the input of the second LSTM model, the hidden layer of the second LSTM model is connected to the time attention module, and the state of the hidden layer of the second LSTM model is updated after being processed by the time attention module, thereby obtaining the output of the second LSTM model, and the output of the second LSTM model is connected to a fully connected layer, the output of the fully connected layer is the landslide displacement data, and at the same time, the weight α of the factor at time s calculated by each factor attention module in the first LSTM model is s As output;

[0012] The factor attention module includes three fully connected layers and two activation functions, namely, the first fully connected layer, the second fully connected layer, the third fully connected layer, the tanh activation function, and the Softmax activation function. The calculation method of the factor attention module at any time s in the first LSTM model is: the hidden layer state h output by the LSTM unit at the previous time (s-1) s-1 and cell state c s-1 Input to the second fully connected layer in the factor attention module, M external factor data at the sth moment Input to the first fully connected layer, and finally the output of the first and second fully connected layers is processed by the tanh activation function as the input of the third fully connected layer. The output of the third fully connected layer is processed by the softmax activation function as the factor weight of the factor attention module at the sth moment, and is combined with x s After multiplication, it is used as the input of the LSTM unit at the sth moment in the first LSTM model;

[0013] S3, using the data in the training set to train the parameters of the attention-based LSTM model in step S2 to obtain a landslide displacement prediction model;

[0014] S4. After processing the external factors affecting the deformation of the landslide in the landslide area to be calculated in the manner of step S1, a number of samples divided according to the set time segmentation step are obtained, and the several samples are input into the landslide displacement prediction model obtained in step S3, and the displacement prediction result of the landslide is output, and at the same time, the weights of the influence of different external factors in the area on the landslide deformation can be output.

[0015] The external factors include rainfall, reservoir water level, and reservoir water level fluctuation.

[0016] The temporal attention module includes a fully connected layer and a softmax activation function. The second LSTM model contains K LSTM units, each of which has its own hidden state h k , there are K hidden states in total. The hidden states of all LSTM units in the second LSTM model are combined as the input H of the temporal attention module. After a fully connected layer and a softmax activation function, they are multiplied with the input H of the temporal attention module to obtain the new hidden state of the second LSTM model. The second LSTM model is updated with the new hidden state to obtain the output of the second LSTM model.

[0017] Weighted input data of the LSTM unit at time s in the first LSTM model Hidden state h s and the cell state c s The updates are:

[0018]

[0019]

[0020] where σ s is the softmax activation function; f is an LSTM unit, which can be calculated according to formulas (8)-(13); M is the number of features, which refers to the number of types of external factors, that is, the number of factors; α s is the weight of the factor at the sth moment, x s is the M external factor data at the sth moment.

[0021] The calculation formula of the training loss function RMSE of the landslide displacement prediction model in step S3 is:

[0022]

[0023] where s n and are the monitored values ​​and model predicted values ​​in the training set, respectively, and s l and are the monitored values ​​and model predicted values ​​in the test set, respectively; N and L are the number of samples in the training set and the test set, respectively.

[0024] The factor attention module focuses on the key factors affecting the surface deformation of the landslide by assigning different weights to different external factors; the larger the factor weight assignment value, the greater the influence of the factor on the landslide displacement prediction; the smaller the factor weight assignment value, the smaller the influence of the factor on the landslide displacement prediction.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] (1) The present invention uses the attention module and LSTM model to capture important environmental factor information from external factors that affect landslide deformation, and realizes the calculation of key factors that affect landslide surface deformation based on historical monitoring data. The significance of determining the key factors that affect landslide deformation is that it can help people prevent landslides. The present invention aims to standardize the evaluation of key factors that affect landslide deformation, improve the effectiveness and rationality of landslide warning and prevention, and play a certain guiding role in landslide disaster prevention and mitigation.

[0027] (2) The present invention comprehensively considers the interrelationships between different factors that affect landslide deformation, calculates the weights of various factors in landslide surface deformation, and determines the degree of influence of a factor on landslide deformation according to the weight of the factor. The larger the weight value, the greater the influence of the factor on landslide deformation.

[0028] (3) The present invention further improves the interpretability of the landslide prediction model by calculating the weights of various factors in landslide deformation, and is helpful for analyzing and revealing the deformation mechanism of reservoir-induced landslides. Compared with the traditional support vector machine (SVM) model, the model proposed in this application is used to predict landslide displacement, and the results obtained have a smaller average root mean square error. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0030] Figure 1 A flow chart of a method for calculating weights of factors affecting landslide surface deformation according to the present invention.

[0031] Figure 2 Schematic diagram of the structure of the factor attention module in the present invention.

[0032] Figure 3 Schematic diagram of the structure of the temporal attention module in the present invention.

[0033] Figure 4 Paulownia Bay landslide monitoring network.

[0034] Figure 5 Distribution of factor attention weights on the Paulownia Bay landslide. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. Of course, the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0036] A method for calculating the weights of factors affecting landslide surface deformation comprises the following steps:

[0037] S1. Collect historical data of landslide displacement and external factors affecting landslide deformation (rainfall, reservoir water level, reservoir water level fluctuation) in the region at the same sampling time and sampling interval; the sampling interval is the time step, set the time segmentation step, and segment the historical data according to the set time segmentation step to generate a sample data set for landslide displacement prediction. The samples in the sample data set are represented by X, and the labels are the corresponding landslide displacements. A sample contains S small samples, and the data of each small sample is an external factor. The sample data set is standardized and divided into a training set and a validation set. The displacement data comes from the GPS instrument installed on the landslide; the rainfall data comes from the rain gauge installed on the landslide; the reservoir water level data comes from the daily public record data of the Yangtze River Navigation Bureau; the reservoir water level fluctuation is calculated using the above-mentioned reservoir water level data, and these data are recorded in a table for subsequent use.

[0038] S2. Build an attention-based LSTM model, including the first LSTM model, the second LSTM model and the fully connected layer;

[0039] The first LSTM model has S LSTM units, and a factor attention module is connected between two adjacent LSTM units. The output of the last LSTM unit is the output of the first LSTM model and serves as the input of the second LSTM model. The hidden layer of the second LSTM model is connected to the time attention module. After being processed by the time attention module, the state of the hidden layer of the second LSTM model is updated to obtain the output of the second LSTM model. The output of the second LSTM model is connected to a fully connected layer, and the output of the fully connected layer is the landslide displacement data. At the same time, the weight αs of the factor at time s calculated by each factor attention module in the first LSTM model is used as the output.

[0040] The factor attention module includes three fully connected layers and two activation functions, namely, the first fully connected layer, the second fully connected layer, the third fully connected layer, the tanh activation function, and the Softmax activation function. The calculation method of the factor attention module at any time s in the first LSTM model is as follows: the hidden layer state h output by the LSTM unit at the previous time (s-1) s-1 and cell state c s-1 Input to the second fully connected layer in the factor attention module, M external factor data at the sth moment Input to the first fully connected layer, and finally the output of the first and second fully connected layers is processed by the tanh activation function as the input of the third fully connected layer. The output of the third fully connected layer is processed by the softmax activation function as the factor weight of the factor attention module at the sth moment, and is combined with x s After multiplication, it serves as the input of the LSTM unit at the sth moment in the first LSTM model.

[0041] The temporal attention module includes a fully connected layer and a softmax activation function. The second LSTM model contains K LSTM units, each of which has its own hidden state h k , there are K hidden states in total. The hidden states of all LSTM units are combined as the input H of the temporal attention module. After a fully connected layer and a softmax activation function, they are multiplied with the input H of the temporal attention module to obtain the new hidden state of the second LSTM model. The second LSTM model is updated with the new hidden state to obtain the output of the second LSTM model.

[0042] S3. Use the external factor time series data in the training set to train the parameters of the attention-based LSTM model in step S2 to obtain a landslide displacement prediction model.

[0043] S4, after processing the external factors affecting the deformation of the landslide in the area to be calculated in the manner of step S1, a number of samples segmented according to the set time segmentation step are obtained, and the samples are input into the landslide displacement prediction model obtained in step S3, and the displacement prediction result of the landslide is output, and at the same time, the weights of the influence of different external factors (rainfall, reservoir water level, reservoir water level fluctuation) on the deformation of the landslide can be output. The weights of the influence on the deformation of the landslide at different times can be output through the first LSTM model.

[0044] The attention-based LSTM model takes external factors as input data and processes them through the attention-based LSTM model; specifically, the factor attention module dynamically assigns weights to different environmental factors in a single time step, and the temporal attention module assigns weights to the hidden state of each time step; by adopting these two attention mechanisms, the attention-based LSTM model can dynamically adjust the attention weights and make full use of environmental variables; the specific calculation method of the attention module includes the following steps:

[0045] S21, LSTM model

[0046] Long short-term memory (LSTM) neural network is a branch of recurrent neural network (RNN). Compared with artificial neural network (ANN), RNN adds a state body between hidden layers to store information, so RNN can learn relevant information from time series input data. In the structure of LSTM, the state body is replaced by a self-recurrent memory block, which consists of an input gate, a forget gate, an output gate, and a memory unit, providing long-term memory and forgetting functions for LSTM. The forget gate determines how much historical information can be stored in the memory unit, the input gate determines how much information can be saved in the memory unit at the current moment, and the output gate is used to convert the current unit state in LSTM to the next unit state. The gate operation and unit state of the LSTM model can be expressed as follows using equations 8 to 13:

[0047] i s =σ(w xi x s +w hi h s-1 +b i ) (8)

[0048] f s =σ(w x fx s +w h fh s-1 +b f ) (9)

[0049] o s =σ(w xo xs +w ho h s-1 +b o ) (10)

[0050] g s =tanh(w xc x s +w hc h s-1 +b c ) (11)

[0051] c s =f s ·c s-1 +i s ·g s (12)

[0052] h s =o s ·tanh(c s ) (13)

[0053] Among them, i s 、f s , o s and c s They are the input gate, forget gate, output gate and four outputs of the storage unit; b i , b f , b o and b c are their corresponding bias vectors. xi 、w xf 、w xo 、w xc and w hi 、w hf 、w ho 、w hc are the input matrices and hidden matrices of the three gates and storage units mentioned above. σ is the sigmoid function, and tanh is the hyperbolic tangent function.

[0054] This application introduces an attention mechanism based on the LSTM neural network to improve the performance of LSTM. At the same time, the introduction of the attention mechanism can assign weights to different input factors (factor attention) and hidden states (temporal attention) in the LSTM model.

[0055] S22, Factor Attention Module

[0056] Time series data X = (x1, x2, x s ,…,x S )=(x 1 ,x 2 ,x m ,…,xM ) S ∈R S×M , where X represents a sample input to the attention-based LSTM model, S represents the number of small samples in a sample, and M represents the number of features, which is the same as the number of external factors. represents the external factor data with a time step of s (including rainfall, reservoir water level, and all factors of reservoir water level change), Represents the historical sequence data of the external environmental factor m containing S time steps.

[0057] The factor attention module includes three fully connected layers and two activation functions. The calculation method of the factor attention module at any time s in the first LSTM model is as follows: the hidden layer state h output by the LSTM unit at the previous time (s-1) s-1 and cell state c s-1 Input to the second fully connected layer in the factor attention module, the M external factor data at the sth moment The output of the first and second fully connected layers is processed by the first activation function tanh function as the input of the third fully connected layer. The output of the third fully connected layer is e s After being processed by the second activation function softmax function, it is the weight α of the factor at the sth moment s , and with x s After multiplication, it is used as the input of the LSTM unit at the sth moment in the first LSTM model;

[0058] The weight of the factor at the sth moment is α s The calculation method is as follows:

[0059]

[0060]

[0061] Where V f ∈R S , W f ∈R S×2p , U f ∈R S×S are the weight matrices of the third, second, and first fully connected layers (MLP), respectively; b f ∈R S is the bias of the second fully connected layer; σ c is the tanh activation function, h s-1 ∈R p and c s-1 ∈R p is the hidden state and unit state at the previous moment, and p is the size of each hidden layer;

[0062] Weighted input data of the LSTM unit at time s in the first LSTM model Hidden state h s and the cell state c s The updates are as follows:

[0063]

[0064]

[0065] where σ s is the softmax activation function; f is an LSTM unit, which can be calculated according to formulas (8)-(13); M is the number of features, which refers to the number of types of external factors, that is, the number of factors.

[0066] S23, Temporal Attention Module

[0067] The temporal attention module includes a fully connected layer and a softmax activation function. The second LSTM model contains K LSTM units. The hidden states of all LSTM units are combined as the input H of the temporal attention module. After being processed by a fully connected layer and a softmax activation function, they are multiplied with the input H of the temporal attention module and then used as the output of the second LSTM model. Figure 2 The process of temporal attention is briefly shown. First, the input data is the hidden state of the LSTM unit at different time steps k, which is processed by a fully connected layer and outputs the temporal attention weight; at each time step, the temporal attention weight is assigned to the hidden state. The weight assignment method and final output of the temporal attention module are shown below:

[0068] H=[h1,h2,...,h k ] K×p (18)

[0069] β=softmax(A t (H))=[β1,β2,...,β k ] 1×K (19)

[0070]

[0071] where h k is the hidden state of the kth LSTM unit, and p is the size of each hidden state. t is a fully connected layer (fully connected neural network), H is a combination of K hidden layers, h at is the output of the final hidden state of the LSTM unit in the second LSTM model.

[0072] As a further optimization scheme for calculating the weights of factors affecting landslide surface deformation provided by the present invention, in step S3, the landslide displacement prediction model has a training loss function calculation formula as follows:

[0073]

[0074] where s n and are the monitored values ​​and model predicted values ​​in the training set, respectively, and s l and are the monitored values ​​and model predicted values ​​in the test set, respectively; N and L are the number of samples in the training set and the test set, respectively.

[0075] As a further optimization scheme for the calculation method of the weights of factors affecting landslide surface deformation provided by the present invention, the factor attention module focuses on the key factors affecting landslide surface deformation by assigning different weights to different external factors; the larger the factor weight assignment value, the greater the influence of the factor on the landslide displacement prediction; the smaller the factor weight assignment value, the smaller the influence of the factor on the landslide displacement prediction.

[0076] Example

[0077] (1) Collect external factors and deformation characteristics that affect landslide deformation in the area

[0078] The Paotongwan landslide is located in Wushan County, Chongqing, on the east side of Daxi River, a tributary of the Yangtze River. The average annual rainfall and temperature in the area are 1000-1400 mm and 16.4 degrees Celsius, respectively. Precipitation events are generally concentrated in the rainy season from May to September, and the cumulative rainfall during this period can reach 70% of the annual cumulative rainfall. In addition, for flood control and power generation, the reservoir water level shows periodic fluctuations within a hydrological year. From November to January of the following year, the water level remains at a maximum level of about 175 meters, and then from January to June, the reservoir water level begins to drop to 145m, which is maintained for about 3 months until September. Finally, the reservoir water level rises to a maximum level of 175m between September and November. In summary, the main external factors affecting landslide deformation are: rainfall, reservoir water level, and reservoir water level fluctuation.

[0079] The deformation of the Paulownia Bay landslide began in the rainy season of 1998, and a GPS monitoring system was implemented in September 2006 to monitor the dynamic evolution of the landslide. Figure 4As shown in Figure 1, six GPS monitoring stations are located on the sliding body. GPS monitoring points in different areas of the landslide can provide detailed deformation data. Specifically, the cumulative displacement data measured by the GPS stations around the Ⅰ-Ⅰ' section in the southern part of the landslide show the largest deformation, with an average displacement of 251.5 mm for two points (WS01 and WS02). WS03 and WS04 have similar cumulative displacements in the middle of the landslide, with a cumulative displacement of about 200 mm and an average rate of about 2.38 mm / month. The recorded data show that the deformation is the smallest in the northern part of the landslide, with the cumulative displacements of two points (WS05 and WS06) both less than 170 mm. In summary, the deformation rate of the landslide accelerates from north to south. For the same section, the deformation of WS02, WS04 and WS06 (elevation 180-195 m) at the front edge of the landslide near Daxi River is greater than that of the points near the rear edge (WS01, WS03 and WS05). This can be explained by the erosion of the front edge of the landslide by the river, the weakening of the slope foot, and the accelerated deformation of the landslide. In general, the deformation of the Paulownia Bay landslide showed the characteristics of slow creep. The total displacement of the monitoring points ranged from 157 to 280 mm, with an average speed of 1.92 mm / month to 3.43 mm / month. The landslide also showed alternating slow deformation and accelerated deformation. The accelerated events occurred from May to September in 2007, 2008, 2009 and 2012, respectively. This period is the annual rainy season flood period. Among them, the completion time of the four accelerated events was 12 months (the entire monitoring period was selected as 75 months), but the increased values ​​accounted for about 60-80% of the total monitored displacement. Therefore, the reservoir water level and rainfall have a great influence on the temporal and spatial characteristics of landslide deformation.

[0080] (2) Input model

[0081] The monitored historical rainfall, reservoir water level and reservoir water level fluctuation data are segmented according to the set time step to generate sample data for landslide displacement prediction. One sample contains S small samples. The sample data is standardized and the landslide displacement prediction model trained by the constructed attention-based LSTM model is input into the landslide displacement prediction model.

[0082] (3) Calculate the weight of each influencing factor

[0083] Figure 4 The distribution of factor attention weights extracted from different time steps of the Paulownia Bay landslide is shown. For the Paulownia Bay landslide, the weights of the three external factors range from 0.1 to 0.55 (the sum of the weights is 1), indicating that the proposed model does not ignore any input factors that cause landslide deformation. Figure 5It can be seen from the figure that the influence of external factor weights on displacement prediction is not constant, but presents dynamic periodic changes in a hydrological year. The dynamic process of factor weights under different external conditions shows that the factor attention module can assign different factor weights to different input factors under different external conditions. Figure 4 As shown in the figure, for the Paulownia Bay landslide, the reservoir water level was the main influencing factor in the early monitoring period, while rainfall made a significant contribution to the landslide deformation at the end of the monitoring period. In addition, during the monitoring period, the influence of the reservoir water level decreased from 0.474 to 0.334, while the influence of rainfall increased from 0.376 to 0.478.

[0084] Furthermore, the factor weights exhibited similar fluctuations within a hydrological year. Figure 4 The figure shows the changes in the weights of the input factors during the monitoring period. The weight of the reservoir water level reaches a maximum of about 0.5 in October and then drops to about 0.3 in April of the following year as shown in the red box in the figure. From July to January of the following year, the weight of rainfall increases from 0.1 to 0.4. In the rainy season, the reservoir water level is the main influencing factor of deformation, but at the end of the monitoring, the reservoir water level and rainfall show similar weights (0.4) in the short term. In summary, for the Paulownia Bay landslide, rainfall is the main inducing factor of deformation, but in the rainy season, the influence of the reservoir water level increases and the weight is similar to the rainfall weight. It can be obtained that the model output is basically consistent with the actual collected landslide deformation characteristics, indicating that the model can accurately calculate the weight values ​​of various factors in the landslide surface deformation.

[0085] The present invention can calculate the weights of various external factors in landslide displacement prediction based on historical monitoring data, and then judge the influence of various factors on landslide displacement prediction according to the weights, and finally can effectively predict landslide displacement.

[0086] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

[0087] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A method for calculating the weights of factors affecting landslide surface deformation, characterized in that: The calculation method includes the following steps: S1. Collect historical data of current landslide displacement and external factors affecting landslide deformation at the same sampling time and sampling interval; the sampling interval is the time step, set the time segmentation step, and segment the historical data according to the set time segmentation step to generate a sample data set for landslide displacement prediction. The samples in the sample data set are represented by X, and the labels are the corresponding landslide displacements. A sample contains S small samples corresponding to S moments. The data of each small sample is an external factor. The sample data set is standardized and divided into a training set and a validation set. S2. Building an attention-based LSTM model The attention-based LSTM model includes a first LSTM model, a second LSTM model, and a fully connected layer. The specific model architecture is: The first LSTM model has S LSTM units, a factor attention module is connected between two adjacent LSTM units, the output of the last LSTM unit is the output of the first LSTM model, the output of the first LSTM model is used as the input of the second LSTM model, the hidden layer of the second LSTM model is connected to the time attention module, and the state of the hidden layer of the second LSTM model is updated after being processed by the time attention module, thereby obtaining the output of the second LSTM model, and the output of the second LSTM model is connected to a fully connected layer, the output of the fully connected layer is the landslide displacement data, and at the same time, the weight α of the factor at time s calculated by each factor attention module in the first LSTM model is s As output; The factor attention module includes three fully connected layers and two activation functions, namely, the first fully connected layer, the second fully connected layer, the third fully connected layer, the tanh activation function, and the Softmax activation function. The calculation method of the factor attention module at any time s in the first LSTM model is: the hidden layer state h output by the LSTM unit at the previous time (s-1) s-1 and cell state c s-1 Input to the second fully connected layer in the factor attention module, M external factor data at the sth moment Input to the first fully connected layer, and finally the output of the first and second fully connected layers is processed by the tanh activation function as the input of the third fully connected layer. The output of the third fully connected layer is processed by the softmax activation function as the factor weight of the factor attention module at the sth moment, and is combined with x s After multiplication, it is used as the input of the LSTM unit at the sth moment in the first LSTM model; S3, using the data in the training set to train the parameters of the attention-based LSTM model in step S2 to obtain a landslide displacement prediction model; S4. After processing the external factors affecting the deformation of the landslide in the landslide area to be calculated in the manner of step S1, a number of samples divided according to the set time segmentation step are obtained, and the several samples are input into the landslide displacement prediction model obtained in step S3, and the displacement prediction result of the landslide is output, and at the same time, the weights of the influence of different external factors in the area on the landslide deformation can be output.

2. The method for calculating the weights of factors affecting landslide surface deformation according to claim 1, characterized in that: The external factors include rainfall, reservoir water level, and reservoir water level fluctuation.

3. The method for calculating the weights of factors affecting landslide surface deformation according to claim 1, characterized in that: The temporal attention module includes a fully connected layer and a softmax activation function. The second LSTM model contains K LSTM units, each of which has its own hidden state h k , there are K hidden states in total. The hidden states of all LSTM units in the second LSTM model are combined as the input H of the temporal attention module. After a fully connected layer and a softmax activation function, they are multiplied with the input H of the temporal attention module to obtain the new hidden state of the second LSTM model. The second LSTM model is updated with the new hidden state to obtain the output of the second LSTM model.

4. The method for calculating the weights of factors affecting landslide surface deformation according to claim 1, characterized in that: Weighted input data of the LSTM unit at time s in the first LSTM model Hidden state h s and the cell state c s The updates are: where σ s is the softmax activation function; f is an LSTM unit, which can be calculated according to formulas (8)-(13); M is the number of features, which refers to the number of types of external factors, that is, the number of factors; α s is the weight of the factor at the sth moment, x s is the M external factor data at the sth moment.

5. The method for calculating the weights of factors affecting landslide surface deformation according to claim 1, characterized in that: The calculation formula of the training loss function RMSE of the landslide displacement prediction model in step S3 is: where s n and are the monitored values ​​and model predicted values ​​in the training set, respectively, and s l and are the monitored values ​​and model predicted values ​​in the test set, respectively; N and L are the number of samples in the training set and the test set, respectively.

6. The method for calculating the weights of factors affecting landslide surface deformation according to claim 1, characterized in that: The factor attention module focuses on the key factors affecting the surface deformation of the landslide by assigning different weights to different external factors; the larger the factor weight assignment value, the greater the influence of the factor on the landslide displacement prediction; the smaller the factor weight assignment value, the smaller the influence of the factor on the landslide displacement prediction.

Images