Landslide displacement prediction method based on deep learning, storage medium and equipment
Through deep learning methods combined with frequency analysis and BiLSTM-reduction model, the problem of insufficient frequency domain analysis and time series complexity processing of landslide displacement prediction is solved, and the accuracy of landslide displacement prediction is improved.
Patent Information
- Application Number
- CN202411419436.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-12
- Publication Date
- 2025-01-17
AI Technical Summary
The existing landslide displacement prediction methods lack frequency domain analysis and traditional models to deal with time series complexity and long-term dependence, making it difficult to improve prediction accuracy.
Using a deep learning-based method, the landslide displacement signal and pregnancy factor signal are decomposed into intrinsic mode signals through variational modal decomposition, quantifying the correlation degree between periodic term displacement and pregnancy factor signal, and using BiLSTM-reduction The model introduces a simple gate mechanism to build a landslide trend term and periodic term displacement prediction model.
By combining frequency analysis and maximum information coefficient, the accuracy of feature screening is improved, the two-way timing characteristics are effectively captured, noise and redundant information are eliminated, and the accuracy of landslide displacement prediction is improved.
Smart Images
Figure CN119322924A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of landslide displacement prediction, and in particular to a landslide displacement prediction method, storage medium, and device based on deep learning. Background Art
[0002] Landslides are a common and frequent geological disaster, often posing a major threat to the life and property safety of residents in the surrounding areas of landslides. Landslide disasters are extremely destructive, especially in mountainous areas, hills, and around reservoirs. Studies have shown that there is a close correlation between the occurrence of landslides and their displacement, and landslide displacement is an important indicator reflecting the occurrence of landslides. Generally speaking, the change in landslide displacement is positively correlated with the probability of landslide disasters, that is, the greater the displacement, the higher the possibility of landslides. Therefore, accurate monitoring and prediction of landslide displacement has become a core link in the landslide disaster early warning and forecasting system. Through the effective prediction of landslide displacement, potential landslide risks can be identified in advance, and timely response measures can be provided to decision makers to avoid or reduce the damage to life and property caused by disasters.
[0003] Currently, the prediction of landslide displacement is usually based on the analysis of multiple influencing factors such as landslide displacement, reservoir water level and rainfall. However, most existing research and prediction methods focus on the analysis of the time domain, and there is a lack of research on the frequency domain characteristics of signals such as landslide displacement, reservoir water level and rainfall, which limits the improvement of prediction accuracy to a certain extent. In addition, traditional landslide prediction models are insufficient in dealing with the complexity and long-term dependence of time series signals, making it difficult to capture the complex change patterns in landslide displacement. Summary of the invention
[0004] The purpose of the present invention is to solve the problems of lack of frequency domain analysis in existing landslide displacement prediction and insufficient time series complexity and long-term dependence in traditional models, and propose a landslide displacement prediction method based on deep learning, comprising the following steps:
[0005] S1. Obtain landslide displacement signal and disaster-pregnant factor signal;
[0006] S2. Decomposing the landslide displacement signal and the disaster-pregnant factor signal by using variational mode decomposition to obtain the intrinsic mode signals of the landslide displacement signal and the disaster-pregnant factor signal, dividing the intrinsic mode signal of the landslide displacement signal into the intrinsic mode signal corresponding to the periodic term displacement and the intrinsic mode signal corresponding to the trend term displacement; dividing the intrinsic mode signal corresponding to the periodic term displacement, the intrinsic mode signal corresponding to the trend term displacement, and the intrinsic mode signal of the disaster-pregnant factor signal into a first data set and a second data set;
[0007] S3, the degree of correlation between the intrinsic mode signal corresponding to the displacement of the quantized periodic term and the intrinsic mode signal of the disaster-pregnant factor signal;
[0008] S4. Based on BiLSTM, introduce a reduction gate in its output part to generate a BiLSTM-reduction model. Input the first dataset into the BiLSTM-reduction model to construct a landslide trend-term displacement prediction model based on the BiLSTM-reduction model; select the eigenmode signals of the disaster-causing factor signals with a high degree of correlation with the eigenmode signals corresponding to the periodic-term displacement in the second dataset and input them into the BiLSTM-reduction model to construct a landslide periodic-term landslide displacement prediction model based on the BiLSTM-reduction model.
[0009] S5. Predict the displacement of the landslide based on the constructed landslide trend-term displacement prediction model and landslide periodic-term landslide displacement prediction model.
[0010] Furthermore, the disaster-causing factor signals include rainfall signals and reservoir water level change signals.
[0011] Furthermore, the acquisition period of the disaster-causing factor signals is one day.
[0012] Furthermore, according to the central frequency of the eigenmode signals of the landslide displacement signal, the eigenmode signals of the landslide displacement signal are divided into eigenmode signals corresponding to periodic-term displacement and eigenmode signals corresponding to trend-term displacement.
[0013] Furthermore, the degree of correlation between the eigenmode signals corresponding to the periodic-term displacement and the eigenmode signals of the disaster-causing factor signals is quantified as:
[0014]
[0015] where MIC(x, y) represents the degree of correlation between x and y, x represents the eigenmode signal of the disaster-causing factor signal, y represents the eigenmode signal corresponding to the periodic-term displacement, I(x, y) represents the mutual information between x and y, a and b respectively represent the number of samples of x and y, and a×b < B represents the constraint on the total number of grids, and B represents a set threshold.
[0016] Furthermore, the input and output of the BiLSTM-reduction model are expressed as:
[0017] z i =σ(W zi [Y i-1 ,S i +b zi )
[0018] r i =σ(W ri [Y i-1 ,S i]+b ri )
[0019]
[0020] Among them, z i Represents the weight matrix W z The i-th index of , σ represents the activation function, W zi , W ri , W Yi Represent the weight matrix of the simplified layer, b zi 、b ri 、b Yi They represent the bias of the reduced layer, Y i-1 Indicates the output value at the previous moment, S i Represents the input of the simplified layer at the current moment, Y i represents the output of the current simplified layer, z ti represents the output of the restoration gate, Indicates the intermediate state of output information, ri Represents the output of the hidden state.
[0021] The present invention also proposes a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the above-mentioned landslide displacement prediction method based on deep learning is implemented.
[0022] The present invention also proposes an electronic device, comprising a processor and a memory, wherein the processor and the memory are interconnected, wherein the memory is used to store a computer program, the computer program comprises computer-readable instructions, and the processor is configured to call the computer-readable instructions to execute the above-mentioned landslide displacement prediction method based on deep learning.
[0023] The beneficial effects brought by the technical solution provided by the present invention are:
[0024] The present invention quantifies the correlation between disaster-prone factors such as reservoir water level and rainfall and landslide displacement by combining frequency analysis and maximum information coefficient, thus improving the accuracy of feature screening. In addition, the improved BiLSTM-reduction model effectively captures bidirectional time series features and eliminates noise and redundant information by introducing a simple gate mechanism, thereby improving the accuracy of landslide displacement prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1 is a flow chart of a landslide displacement prediction method based on deep learning according to an embodiment of the present invention;
[0026] Figure 2 is a block diagram of an electronic device in an exemplary embodiment of the present invention;
[0027] Figure 3 is a relevant data monitoring signal for landslide prediction in an embodiment of the present invention;
[0028] Figure 4 is the landslide displacement decomposition result based on VMD in an embodiment of the present invention;
[0029] Figure 5 is the rainfall reservoir water level rainfall decomposition result based on VMD in the embodiment of the present invention, Figure 5 (a) The rainfall decomposition result, Figure 5 (b) Reservoir water level decomposition results;
[0030] Figure 6 It is a screening and decomposition scheme for the displacement characteristics of the landslide periodic term in the embodiment of the present invention;
[0031] Figure 7 It is the overall displacement prediction result of the landslide in the embodiment of the present invention. DETAILED DESCRIPTION
[0032] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.
[0033] The flowchart of the landslide displacement prediction method based on deep learning in an embodiment of the present invention is as follows: Figure 1 , specifically including the following steps:
[0034] S1. Obtain landslide displacement signals and disaster-pregnant factor signals. Disaster-pregnant factor signals include: rainfall signals and reservoir water level change signals.
[0035] Specifically, the collection period of landslide displacement signal is one month, and the collection period of disaster-pregnant factors is one day. The collected landslide displacement signal Y=[Y1,Y2,…,Y n ], the disaster factors mainly include rainfall and reservoir water level, and the rainfall signal M=[M1,M2,…,M n ] and reservoir water level signal N=[N1,N2,…,N n ], where n is the total number of months collected, that is, the number of samples of landslide displacement, Y i is the landslide displacement corresponding to the ith month, M i is the daily rainfall signal collected in the ith month [M i1 ,M i2 ,…,M im ],N i is the daily reservoir water level signal collected in the ith month [N i1 ,N i2 ,…,N im ],M ij and N ijare the rainfall and reservoir water level signals collected on the jth day in the ith month respectively. m is the specific day of each month and is a variable value related to i.
[0036] The collected rainfall signal and reservoir water level signal are processed to make them correspond to the acquisition period of landslide displacement, and the cumulative rainfall of each month is obtained. 11 ,x 12 ,...,X 1n ] Then the monthly reservoir water level change X2 = [X 21 ,X 22 ,...,X 2N ], where X 1j and X 2j are the cumulative rainfall in the jth month and the change in reservoir water level in that month: X 1j =M i1 +M i2 +...+M im ,X 2j =M im -M i1 .
[0037] S2. Decompose the landslide displacement signal and the disaster-pregnant factor signal using variational mode decomposition to obtain the intrinsic mode signals of the landslide displacement signal and the disaster-pregnant factor signal, and divide the intrinsic mode signal of the landslide displacement signal into the intrinsic mode signal corresponding to the periodic term displacement and the intrinsic mode signal corresponding to the trend term displacement; divide the intrinsic mode signal corresponding to the periodic term displacement, the intrinsic mode signal corresponding to the trend term displacement, and the intrinsic mode signal of the disaster-pregnant factor signal into a first data set and a second data set.
[0038] Specifically, VMD is used to perform multi-layer decomposition on the landslide displacement signal Y, the monthly cumulative rainfall signal X1, and the monthly reservoir water level change signal X2, where the landslide displacement signal is decomposed into K1 groups of intrinsic mode signals [IMF Y1 ,IMF Y2 ,...,IMF YK1 ], the monthly cumulative rainfall signal is decomposed into K2 groups of intrinsic mode signals [IMF X11 ,IMF X12 ,...,IMF X1K2 ], the monthly reservoir water level change signal is decomposed into K3 groups of intrinsic mode signals [IMF X21 ,IMF X22 ,...,IMF X2K3 ]; According to the signal after landslide displacement decomposition [IMF Y1 ,IMF Y2 ,...,IMF YK1, divide the eigenmode signal group Y corresponding to the periodic term displacement according to the central frequency of each eigenmode signal p and the trend term displacement Y t , where Y p =IMF Yt , [Y t ,Y p =[IMF Y1 ,IMF Y2 ,...,IMF YK1 .
[0039] S3. Quantify the correlation degree between the eigenmode signal corresponding to the periodic term displacement and the eigenmode signal of the disaster-causing factor signal
[0040] Combined with the concept of the maximum information coefficient, quantify the correlation degree between the eigenmode functions with different frequencies contained in the monthly cumulative rainfall signal and the eigenmode functions with different frequencies of the periodic term displacement respectively
[0041] Among them, the value range of y is the value range of x is [IMF X21 ,IMF X22 ,...,IMF X2K3 , I(x,y) represents the mutual information between x and y, a×b<B is the constraint on the total number of grids, a and b represent the number of samples of x and u respectively, and B is a set threshold, approximately the 0.6th power of the data volume. Finally, the maximum information coefficient between the eigenmode functions with different frequencies contained in the monthly reservoir water level change signal and the eigenmode functions with different frequencies of the periodic term displacement is obtained
[0042] S4. Based on BiLSTM, introduce a reduction gate in its output part to generate a BiLSTM-reduction model. Input the first dataset into the BiLSTM-reduction model to construct a landslide trend term displacement prediction model based on the BiLSTM-reduction model; select the eigenmode signal of the disaster-causing factor signal with a high correlation degree corresponding to the periodic term displacement in the second dataset and input it into the BiLSTM-reduction model to construct a landslide periodic term landslide displacement prediction model based on the BiLSTM-reduction model
[0043] Specifically, first, based on BiLSTM, mine the features of each input data in the forward and reverse time series spaces to obtain the forward correlation feature S fi and the reverse correlation feature S bi , integrate the forward and reverse correlation features, and finally the output of BiLSTM is S i , where Sfi , S bi and S i It corresponds to the forward correlation feature, reverse correlation feature and BiLSTM output result of the i-th input sample in the time series space.
[0044] Based on this, a reduction gate is introduced to process the features extracted by BiLSTM, and a deeper analysis is performed on the currently extracted features, so as to further filter out the effective features in the data and save them, and eliminate the redundant information in the data:
[0045] z i =σ(W zi [Y i-1 ,S i ]+b zi )
[0046] r i =σ(W ri [Y i-1 ,S i ]+b ri )
[0047]
[0048] Among them, z i Represents the weight matrix W z The i-th index of , σ represents the activation function, W zi , W ri , W Yi Represent the weight matrix of the simplified layer, b zi , b ri , b Yi They represent the bias of the reduced layer, Y i-1 Indicates the output value at the previous moment, S i Represents the input of the simplified layer at the current moment, Y i represents the output of the current minimalist layer, z ti represents the output of the restoration gate, Indicates the intermediate state of output information, ri Represents the output of the hidden state.
[0049] S5. Predict the displacement of the landslide based on the constructed landslide trend term displacement prediction model and landslide period term landslide displacement prediction model.
[0050] First, the sorted landslide trend term displacement Y t , periodic displacement Y p, monthly cumulative rainfall X1 and monthly reservoir water level change signal X2 are decomposed to obtain their respective intrinsic mode function groups, and the data is divided into training set and test set. Then, based on the training set data in the landslide trend term displacement, a landslide trend term displacement prediction model based on the BiLSTM-reduction model is constructed to predict the long-term trend of the landslide. The prediction of the landslide period term displacement needs to be carried out on the test set. In order to more accurately predict the intrinsic mode signals of each frequency in the periodic term, the intrinsic mode signals of each different frequency are When making predictions, the rainfall eigenmode signals with different center frequencies that are highly correlated are selected. and reservoir water level intrinsic mode signal As input features. In this way, a landslide period term displacement prediction model is established based on the BiLSTM-reduction model. Finally, the established landslide displacement prediction model is verified using the test set data, and the landslide trend term displacement prediction results are used as input features. t And the periodic term displacement prediction results of each different center frequency O pa The overall displacement prediction result O of the landslide is finally obtained by superposition, where O pa It is the prediction result corresponding to the ath intrinsic mode signal in the periodic term.
[0051] In an exemplary embodiment, a computer-readable storage medium is included, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned landslide displacement prediction method based on deep learning is implemented.
[0052] See also Figure 2 In an exemplary embodiment, an electronic device is also included, including at least one processor, at least one memory, and at least one communication bus.
[0053] Among them, a computer program is stored in the memory, and the computer program includes computer-readable instructions. The processor calls the computer-readable instructions stored in the memory through a communication bus to execute the above-mentioned landslide displacement prediction method based on deep learning.
[0054] The embodiment of the present invention takes the displacement data of a landslide monitored from January 2007 to December 2012 as the research object, and performs landslide displacement prediction based on the collected landslide displacement and disaster-pregnant factor data. The specific process is as follows:
[0055] (1) The collected disaster-prone factors are processed and organized into parameters corresponding to the landslide displacement sampling period. The monthly reservoir water level is organized into the monthly reservoir water level change, and the monthly rainfall is organized into the monthly cumulative rainfall. The organized data are as follows: Figure 3 shown.
[0056] (2) Based on VMD, the landslide displacement is decomposed into layers. The decomposition layer number of the landslide displacement is set to 4, and the decomposition layer number of the rainfall and reservoir water level signal is set to 3. The final decomposition result of the landslide displacement is as follows: Figure 4 As shown in Figure 2, the decomposition results of the reservoir water level and rainfall correlation signals are as follows: Figure 5 shown.
[0057] (3) Since the frequency of change of the intrinsic mode signal IMF1 in the landslide displacement decomposition result is relatively small, it is regarded as the displacement of the landslide trend term, and the rest are the intrinsic mode signals of different frequencies contained in the periodic term.
[0058] (4) The maximum information coefficient is used to quantify the degree of correlation between the intrinsic mode signals of different central frequencies contained in the disaster-pregnant factors and the landslide periodic displacement. The maximum information coefficients of the intrinsic mode signals of different frequencies in rainfall and landslide periodic displacement, and the maximum information coefficients of the intrinsic mode signals of different frequencies in reservoir water level and landslide periodic displacement are shown in Table 1. The intrinsic mode signals of different frequencies are tested and need to be combined with the disaster-pregnant factor signals of different frequencies with their maximum information coefficients greater than 0.3 as their input features. The specific scheme is as follows: Figure 6 shown.
[0059] Table 1 Maximum information coefficient between the periodic term and external disaster factors
[0060]
[0061] (5) A landslide displacement prediction model based on the BiLSTM-reduction model was established. The data from January 2007 to December 2011 were used as the training set to train the prediction model. The data collected from January 2012 to December 2012 were used as the test set to verify the prediction performance of the model. The landslide displacement prediction model was trained based on the training data.
[0062] (6) Based on the test set, the trained model is used to predict the displacement of the landslide trend term and the period term, and the prediction results are superimposed to obtain the displacement prediction result of the entire landslide. The final prediction result is as follows: Figure 6 As shown in Table 2, the performance indicators of the landslide displacement model in this paper compared with the comparison model are shown in Table 2. The indicators are: RMSE (Root Mean Square Error), MAE (Mean Absolute Error), R 2 (R-Square, coefficient of determination).
[0063] Table 2 Overall prediction performance indicators of different models
[0064] Model RMSE(mm) MAE(mm) <![CDATA[R 2 ]]> BP 11.0880 9.4697 0.979 LSTM 8.5689 6.9467 0.982 GRU 9.4399 7.7466 0.979 BiLSTM 7.5789 6.6072 0.984 BiLSTM-reduction 5.6875 4.8354 0.991
[0065] from Figure 7 It can be clearly observed from Table 2 that the accuracy of the BiLSTM-reduction model proposed in the present invention in the landslide displacement prediction task is much higher than that of the other comparison models, including BP neural network, LSTM (Longshort-term memory), GRU (Gated Recurrent Unit), BiLSTM (Long short-term memory), with Actual (real value) as a reference. It is further verified that the landslide displacement prediction model proposed in the present invention can initially remove redundant information to a large extent by virtue of its ability to mine bidirectional features in the time series space, and reduce the influence of noise on the final landslide displacement prediction accuracy, and then conduct a deeper mining of the features that have been generated based on the reduction gate, thereby ensuring the maximum utilization of effective information and improving the prediction performance of the high model in the landslide displacement time series task.
[0066] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A landslide displacement prediction method based on deep learning, characterized in that: It includes the following steps: S1. Obtain landslide displacement signals and disaster-causing factor signals; S2. Use variational mode decomposition to decompose the landslide displacement signals and disaster-causing factor signals to obtain the intrinsic mode signals of the landslide displacement signals and disaster-causing factor signals. Divide the intrinsic mode signals of the landslide displacement signals into the intrinsic mode signals corresponding to the periodic displacement and the intrinsic mode signals corresponding to the trend displacement; Divide the intrinsic mode signals corresponding to the periodic displacement, the intrinsic mode signals corresponding to the trend displacement, and the intrinsic mode signals of the disaster-causing factor signals into a first data set and a second data set; S3. Quantify the correlation degree between the intrinsic mode signals corresponding to the periodic displacement and the intrinsic mode signals of the disaster-causing factor signals; S4. Based on BiLSTM, introduce a reduction gate in its output part to generate a BiLSTM-reduction model. Input the first data set into the BiLSTM-reduction model to construct a landslide trend displacement prediction model based on the BiLSTM-reduction model; Select the intrinsic mode signals of the disaster-causing factor signals with a high correlation degree with the intrinsic mode signals corresponding to the periodic displacement in the second data set and input them into the BiLSTM-reduction model to construct a landslide periodic landslide displacement prediction model based on the BiLSTM-reduction model; S5. Predict the displacement of the landslide based on the constructed landslide trend displacement prediction model and landslide periodic landslide displacement prediction model.
2. The landslide displacement prediction method based on deep learning according to claim 1 is characterized in that: The disaster-causing factor signals include: rainfall signals, reservoir water level change signals.
3. The landslide displacement prediction method based on deep learning according to claim 1 is characterized in that: According to the central frequency of the intrinsic mode signals of the landslide displacement signals, divide the intrinsic mode signals of the landslide displacement signals into the intrinsic mode signals corresponding to the periodic displacement and the intrinsic mode signals corresponding to the trend displacement.
4. The landslide displacement prediction method based on deep learning according to claim 1 is characterized in that: The acquisition period of the disaster-causing factor signals is one day.
5. The landslide displacement prediction method based on deep learning according to claim 1 is characterized in that: Quantifying the correlation degree between the intrinsic mode signals corresponding to the periodic displacement and the intrinsic mode signals of the disaster-causing factor signals is expressed as: Where MIC(x,y) represents the correlation degree between x and y, x represents the intrinsic mode signals of the disaster-causing factor signals, y represents the intrinsic mode signals corresponding to the periodic displacement, I(x,y) represents the mutual information between x and y, a and b respectively represent the number of samples of x and y, and a×b < B represents the constraint on the total number of grids, and B is a set threshold.
6. The method for predicting landslide displacement based on deep learning according to claim 1, characterized in that: The input and output of the BiLSTM-reduction model are expressed as: z i =σ(W zi [Y i-1 ,S i ]+b zi ) r i =σ(W ri [Y i-1 ,S i ]+b ri ) Y i ％ =tanh(W Yi ×[r i Y i-1 ,S i ]+b Yi ) AND i =(1-z ti )AND i-1 +z ti Hey i ％ Among them, z i Represents the weight matrix W z The i-th index of , σ represents the activation function, W zi , W ri , W Yi Represent the weight matrix of the simplified layer, b zi 、b ri 、b Yi They represent the bias of the reduced layer, Y i-1 Indicates the output value at the previous moment, S i Represents the input of the simplified layer at the current moment, Y i represents the output of the current simplified layer, z ti Represents the output of the restoration gate, Y i ％ Indicates the intermediate state of output information, ri Represents the output of the hidden state.
7. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the method according to any one of claims 1-6.
8. An electronic device, characterized in that: It includes a processor and a memory, the processor is connected to the memory. Among them, the memory is used to store a computer program, the computer program includes computer-readable instructions, and the processor is configured to call the computer-readable instructions to execute the method according to any one of claims 1-6.