A landslide early warning method based on data mining

Abstract

The application discloses a landslide early warning method based on data mining, comprising the following steps: obtaining historical monitoring data of a target position, and storing the obtained data in a database; performing variational mode decomposition on obtained ground surface absolute displacement data to obtain a limited number of intrinsic mode components; performing wavelet threshold denoising on each intrinsic mode component obtained through decomposition, and finally reconstructing the denoised components to obtain denoised and reconstructed displacement data; predicting the displacement of the denoised displacement data by using a long short-term memory neural network method; constructing a deep belief neural network model, and using landslide displacement prediction data to perform early warning on whether a landslide occurs. The method solves the problem of poor prediction accuracy in landslide displacement early warning in the prior art, thereby realizing a more comprehensive and scientific early warning method.

Description

TECHNICAL FIELD

[0001] The present application relates to geological disaster monitoring and early warning technology, in particular to a landslide early warning method based on data mining. BACKGROUND

[0002] At present, the occurrence of landslide disasters can cause great harm to public infrastructure. Landslides usually have the characteristics of concealment, great destructive power and strong randomness, and the traditional investigation and supervision are mainly carried out in the form of manual investigation. This method has high precision, but it is difficult to achieve large-scale and rapid delineation of landslide range. With the development of science and technology, by means of computer system, the historical monitoring data of landslide body are learned, and then the displacement of landslide body after a period of time is inferred, and finally the position of future landslide occurrence is reasonably predicted and early warning is issued. Therefore, the response plan can be made as soon as possible to save the lives and property safety of the people.

[0003] From the aspect of data monitoring: in actual engineering, it is difficult to promote large-scale monitoring of multiple types of sensors due to the limitation of funds and manpower, and displacement monitoring is used as the most extensive monitoring method. In the complex and variable conditions in the field, the data monitored usually contains multiple interference signals, and signal denoising can improve the accuracy of people's identification of information and provide protection for subsequent signal processing. The denoising of the signal is mainly from the most classic Fourier transform to analyze the global information of the signal to the window Fourier transform, which overcomes the defects of Fourier transform in processing non-stationary and nonlinear signals. Then to a new time-frequency analysis method--wavelet transform, which can focus on any details of the signal for multi-resolution time-frequency domain analysis. The present application improves the wavelet denoising method and proposes a signal denoising method based on VMD decomposition and wavelet threshold denoising, which can effectively eliminate the interference of Gaussian white noise on the signal to a certain extent.

[0004] In landslide displacement time series prediction: Initially, most landslide deformation prediction models were developed using machine learning and deep learning, such as convolutional neural networks (CNNs) and backpropagation (BP) neural networks to predict landslide deformation trends, but the results were not ideal. Later, Recurrent Neural Networks (RNNs) were used to predict landslide deformation trends, as RNNs can handle arbitrarily long sequences. However, in practical applications, RNNs suffer from the vanishing gradient problem. To overcome the shortcomings of traditional RNNs, a special type of RNN called LSTM (Laser-Signal Memory) neural network was proposed. Compared to traditional RNNs, the basic unit of the hidden layer in LSTM is a storage block. A storage block contains memory cells and three "gates" (forget gate, input gate, and output gate). These three "gates" regulate the flow of information into and out of the storage cell. The input gate controls the input vector entering the storage cell. The forget gate controls whether information from the previous time step is remembered or forgotten. The forget gate filters information, retaining useful information and discarding useless information. By introducing these three "gates," LSTM can handle time series signals very well.

[0005] In landslide early warning, intelligent early warning methods have higher accuracy rates compared to traditional methods, but they still have their own problems and limitations. For example, the expert experience knowledge required by expert systems is difficult to acquire; artificial neural networks are prone to getting trapped in local optima; and support vector machines, being binary classification algorithms, suffer from overlapping or inseparable classifications when applied to multi-class problems, and their classification efficiency is not high. Deep learning methods, such as deep belief networks (DBN) and convolutional neural networks (CNN), have stronger feature extraction capabilities and fault tolerance than shallow machine learning methods, achieving better classification results and showing great promise for application in the field of landslide early warning. Summary of the Invention

[0006] The main objective of this invention is to provide a landslide early warning method based on data mining, which solves the problem of poor prediction accuracy in landslide displacement early warning in the prior art. It utilizes a deep belief network combined with historical monitoring data of landslides and real-time prediction data of landslide displacement for early warning, thereby achieving a more comprehensive and scientific early warning method.

[0007] The technical solution adopted in this invention is: a landslide early warning method based on data mining, comprising:

[0008] Step 1: Obtain historical monitoring data for the target location and store the obtained data in the database;

[0009] Step 2: Perform variational mode decomposition (VMD) on the acquired absolute surface displacement data to obtain a finite number of intrinsic mode components (IMF).

[0010] Step 3: Wavelet threshold denoising is performed on each intrinsic mode component (IMF) obtained by decomposition, and finally the denoised components are reconstructed to obtain denoised and reconstructed displacement data;

[0011] Step 4: The denoised displacement data is subjected to displacement prediction by using a long short-term memory neural network (LSTM) method.

[0012] Step 5: A deep belief neural network (DBN) model is constructed, and whether a landslide occurs is predicted by using the landslide displacement prediction data.

[0013] Further, in step 1, the historical monitoring data of the target position is obtained, and the obtained data is stored in a database, including the following aspects: slope, slope shape, 1-hour rainfall and 24-hour rainfall, and 6 disaster-causing factors such as absolute ground displacement monitoring data and crack instrument monitoring displacement as research data; and the collected information data is transmitted to a remote server through wireless network technology and stored in the database.

[0014] Further, in step 2, the time series landslide displacement data is subjected to variational mode decomposition (VMD) to obtain a finite number of intrinsic mode components (IMF); the decomposition of the displacement signal is mainly to determine the size of the mode decomposition number K value, and the process is as follows:

[0015] Step 21: Residual error analysis

[0016] Assuming different k values, input the landslide monitoring data as the original input sequence, and perform VMD decomposition and reconstruction; by solving the root mean square error (RMSE) of each reconstructed sequence and the original input sequence, the deviation between the reconstructed sequence and the original sequence is measured:

[0017] ;

[0018] Step 22: ADF stability analysis, according to the ADF value of the analyzed sequence, select the K value that minimizes the possibility of component or order non-stationarity;

[0019] Step 23: Analyze the correlation coefficient between the intrinsic mode components and the residual error (Re) under different K values and the original input sequence, and finally determine the size of the K value; input the displacement data into the VMD model to obtain K intrinsic mode component data.

[0020] Further, in step 3, wavelet threshold denoising is performed on each IMF obtained by decomposition, and finally the denoised components are reconstructed to obtain reconstructed displacement data, and the specific process is as follows:

[0021] Step 31: For landslide displacement monitoring data, the compactly supported biorthogonal wavelet (dbN) is selected as the wavelet basis function, and the processing effect is best when the decomposition level is 3, and then the wavelet soft threshold method is selected for noise reduction processing, and the threshold rule is selected as the wavelet soft threshold Rigorous SURE;

[0022] Step 32: Use db3 wavelet to perform wavelet transform on the IMF component to obtain a group of wavelet coefficients corresponding to the decomposition level;

[0023] Step 33: Compare the decomposed wavelet coefficients with the selected threshold rule. If the wavelet coefficient is greater than the threshold, it is considered that the wavelet coefficient is mainly composed of useful signals, and the wavelet coefficient is retained. Otherwise, it is considered that the wavelet coefficient is mainly composed of noise signals, and the wavelet coefficient is discarded;

[0024] Step 34: Reconstruct the retained signals to obtain the noise-reduced IMF component;

[0025] Step 35: Perform steps 2-4 on each IMF component until all IMF components are subjected to wavelet threshold noise reduction operation;

[0026] Step 36: Finally, reconstruct the noise-removed IMF component to obtain the noise-reduced reconstructed landslide displacement data signal.

[0027] Further, in step 4, the long short-term memory neural network (LSTM) method is used to predict the displacement of the displacement data after noise reduction, and the specific process is as follows:

[0028] Step 41: Establish an LSTM landslide displacement time series prediction model, which consists of an input layer, two hidden layers and an output layer;

[0029] Step 42: Data segmentation: divide the noise-reduced reconstructed displacement data into 80% training set and 20% validation set; import the data into the LSTM prediction model;

[0030] Step 43: Determine the hyperparameters: determine the number of hidden layer nodes of the neural network through the empirical equation:

[0031] ;

[0032] wherein, is the number of input layer neurons; is the number of output layer neurons, is the number of training set neurons, is an integer between 2 and 10;

[0033] Step 44: The number of neurons obtained by the empirical formula also needs to be obtained by continuous experiments to obtain more accurate results; in order to intuitively compare the pros and cons of the model prediction results under different parameters, the mean error (MAE), mean error percentage (MAPE), and root mean square error (RMSE) are introduced for different results comparison:

[0034] ;

[0035] In the formula The i-th group of predicted values is represented by The true value at time m+1 in the i-th group of samples is represented by L, and the number of samples used in one iteration is represented by L;

[0036] Step 45: Step 5: continuously adjust the parameters to obtain the best LSTM model parameters.

[0037] Further, in step 5, a deep belief network (DBN) model is constructed using historical monitoring data, and landslide displacement prediction data is input into the model to perform early warning on whether a landslide occurs, and the model structure is:

[0038] The structure of the DBN model is: 3 RBM networks connected in sequence, 1 BP network, and 1 Softmax classifier;

[0039] The input is: monitored 1-hour rainfall and 24-hour rainfall, slope gradient and slope shape, crack instrument monitoring displacement, absolute ground displacement, and corresponding warning categories;

[0040] The output is the warning category of the Softmax classifier: first-level warning, second-level warning, third-level warning, and no warning in 4 states;

[0041] In the unsupervised pre-training phase of the DBN, the Dropout algorithm is introduced, and the Dropout technology is used for regularization processing: in the pre-training phase, under the premise that the input and output of the neural network remain unchanged, a certain probability is used to randomly adopt the weights of the hidden layer nodes, and each time the adjustment is made, a part of the neurons do not participate in the forward propagation training process; in particular, the probability of Dropout is set to 50%.

[0042] Further, in step 5, the deep belief network (DBN) is used to integrate landslide historical monitoring data, slope support information data, and landslide displacement prediction data to perform early warning on whether a landslide occurs, and the specific implementation process of the deep belief network (DBN) is:

[0043] Step 51: The certainty factor (CF) is used to calculate the coefficients of the landslide disaster factor influencing factors, and the Mapminmax normalization method is used to normalize the landslide data to the [0, 1] interval.

[0044] Step 52: the collected historical data after processing is divided into training samples and test samples;

[0045] Step 53: inputting data into the model, using an unsupervised greedy algorithm, initializing the link weight between layers of the restricted Boltzmann machine unit network model And the bias value Pre-training the deep belief network; in order to prevent dependence between neural nodes and overfitting, 50% dropout is introduced in the visible layer;

[0046] Step 54: selecting an activation function for the hidden layer, performing Gibbs sampling and contrast divergence algorithm for multiple iterations of pre-training each layer to update the model parameters, and obtaining a deep belief network model;

[0047] Step 55: using the BP neural network algorithm and the gradient descent method to further fine-tune and optimize the model parameters of the pre-trained deep belief network from top to bottom;

[0048] Step 56: the output of the DBN is judged by the Softmax classifier to determine the type of early warning; the test sample is brought into the training model for verification, and finally the optimal early warning model is obtained;

[0049] Step 57: inputting the landslide displacement prediction data into the optimal model, and obtaining the grade of landslide early warning through the Softmax classifier.

[0050] Advantages of the present application:

[0051] The landslide early warning method based on data mining of the present application solves the problem of poor prediction accuracy in landslide displacement early warning in the prior art, and uses the deep belief network to combine the historical monitoring data of landslides and the real-time prediction data of landslide displacement for early warning, thereby realizing a more comprehensive and scientific early warning method.

[0052] In addition to the purposes, features and advantages described above, the present application has other purposes, features and advantages. The present application will be further described below with reference to the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0053] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application, illustrate the preferred embodiments of the application, and assist in the explanation of the application. The drawings are not intended to limit the application in any way.

[0054] Figure 1 is the flowchart of the variational mode decomposition and wavelet threshold denoising of the present application;

[0055] Figure 2is a flow chart of the long short-term memory neural network of the present application running;

[0056] Figure 3 is a flow chart of the deep belief network of the present application running;

[0057] Figure 4 is a displacement time curve after VMD-wavelet denoising of the present application;

[0058] Figure 5 is a prediction graph of the LSTM model of the present application;

[0059] Figure 6 is an error rate curve graph of the DBN test set of the present application. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application, and are not used to limit the present application.

[0061] Reference Figures 1 to 6 A landslide early warning method based on data mining, comprising:

[0062] Step 1: obtaining the historical monitoring data of the target position, and storing the obtained data in the database;

[0063] Step 2: performing variational mode decomposition (VMD) on the obtained ground surface absolute displacement data to obtain a limited number of intrinsic mode components (IMF);

[0064] Step 3: performing wavelet threshold denoising on each intrinsic mode component (IMF) obtained by decomposition, and finally reconstructing the denoised components to obtain denoised reconstructed displacement data;

[0065] Step 4: using the long short-term memory neural network (LSTM) method to predict the displacement of the displacement data after denoising processing;

[0066] Step 5: constructing a deep belief neural network (DBN) model, and using the landslide displacement prediction data to predict whether a landslide will occur.

[0067] In step 1, obtaining the historical monitoring data of the target position and storing the obtained data in the database includes the following aspects: 6 disaster-causing factors such as slope, slope shape, 1-hour rainfall, 24-hour rainfall, ground surface absolute displacement monitoring data, and crack instrument monitoring displacement are used as research data; and the collected information data is transmitted to a remote server through wireless network technology and stored in the database.

[0068] The step 2 is to perform a variational mode decomposition (VMD) on the time series landslide displacement data to obtain a finite number of intrinsic mode components (IMFs); the decomposition of the displacement signal is mainly to determine the size of the mode decomposition number K value, and the process is as follows:

[0069] Step 21: residual error analysis

[0070] Assuming different k values, input the landslide monitoring data as the original input sequence, perform VMD decomposition and reconstruction; by solving the root mean square error (RMSE) of each reconstruction sequence and the original input sequence, the deviation between the reconstruction sequence and the original sequence is measured:

[0071] ;

[0072] Step 22: ADF stability analysis, according to the ADF value of the analyzed sequence, select the K value that minimizes the possibility of component or order non-stationarity;

[0073] Step 23: analyze the correlation coefficient between the intrinsic mode components and the residual error (Re) under different K values and the original input sequence, and finally determine the size of K value; input the displacement data into the VMD model to obtain K intrinsic mode component data.

[0074] The step 3 is to perform wavelet threshold denoising on each IMF obtained by decomposition, and finally reconstruct the denoised components to obtain the reconstructed displacement data, and the specific process is as follows:

[0075] Step 31: Preferably, a compactly supported biorthogonal wavelet (dbN) is selected as the wavelet basis function for landslide displacement monitoring data, and the processing effect is best when the decomposition level is 3, and then a wavelet soft threshold method is selected for denoising processing, and the threshold rule is selected as wavelet soft threshold Rigorous SURE;

[0076] Step 32: Use db3 wavelet to perform wavelet transform on the IMF component to obtain a group of wavelet coefficients corresponding to the decomposition level;

[0077] Step 33: Compare the decomposed wavelet coefficients with the selected threshold rule, if the wavelet coefficient is greater than the threshold, it is considered that the wavelet coefficient is mainly composed of useful signals, and the wavelet coefficient is retained, otherwise, it is considered that the wavelet coefficient is mainly composed of noise signals, and the wavelet coefficient is discarded;

[0078] Step 34: reconstruct the retained signals to obtain denoised IMF components;

[0079] Step 35: perform steps 2-4 on each IMF component until all IMF components are subjected to wavelet threshold denoising operation;

[0080] Step 36: Finally, the denoised IMF components are reconstructed to obtain the denoised reconstructed landslide displacement data signal.

[0081] In step 4, the displacement data after noise reduction is predicted by using a long short-term memory neural network (LSTM) method, and the specific process is as follows:

[0082] Step 41: An LSTM landslide displacement time series prediction model is established, which consists of an input layer, two hidden layers and an output layer;

[0083] Step 42: Data segmentation: the denoised reconstructed displacement data is divided into 80% training set and 20% validation set; the data is imported into the LSTM prediction model;

[0084] Step 43: Determination of hyperparameters: the number of hidden layer nodes of neural network is determined by empirical equation:

[0085] ;

[0086] In the formula, is the number of input layer neurons; is the number of output layer neurons, is the number of training set neurons, is an integer between 2 and 10;

[0087] Step 44: The number of neurons obtained by empirical formula still needs to be obtained by continuous test to get more accurate effect; in order to intuitively compare the pros and cons of model prediction results under different parameters, mean error (MAE), mean error percentage (MAPE) and root mean square error (RMSE) are introduced for different result comparison:

[0088] ;

[0089] In the formula represents the i-th group of predicted values, represents the true value of the m+1 time in the i-th group of samples, and L is the number of samples used in one iteration;

[0090] Step 45: Step 5: the parameters are continuously adjusted to obtain the best LSTM model parameters.

[0091] In step 5, the historical monitoring data is used to construct a deep belief network (DBN) model, and then the landslide displacement prediction data is input into the model to predict whether the landslide will occur or not, and the model structure is:

[0092] The structure of DBN model is: 3 RBM networks, 1 BP network and 1 Softmax classifier are connected in turn;

[0093] The input is: monitored 1-hour rainfall and 24-hour rainfall, slope of the slope body and slope shape, crack instrument monitored displacement, absolute displacement of the ground surface, and corresponding warning categories;

[0094] The output is the warning category of the Softmax classifier: first-level warning, second-level warning, third-level warning, and no warning;

[0095] To solve the overfitting in network training, the Dropout algorithm is introduced in the unsupervised pre-training stage of DBN, and the Dropout technology is used for regularization processing: in the pre-training stage, under the premise that the input and output of the neural network remain unchanged, a certain probability is used to randomly adopt the weight of the hidden layer node, and a part of neurons does not participate in the forward propagation training process each time; in particular, the probability of Dropout is set to 50%.

[0096] In step 5, the deep belief network (DBN) is used to integrate the landslide historical monitoring data, slope support information data, and landslide displacement prediction data to perform early warning on whether the landslide occurs or not, and the specific implementation process of the deep belief network (DBN) is as follows:

[0097] Step 51: The certainty factor (CF) is used to calculate the coefficients of each influencing factor of the landslide disaster factor, and the Mapminmax normalization method is used to normalize the landslide data to the [0, 1] interval;

[0098] Step 52: The collected and processed multiple sets of historical data are divided into training samples and test samples;

[0099] Step 53: The data is input into the model, and the unsupervised greedy algorithm is used to initialize the link weight between each layer of the restricted Boltzmann machine unit network model and the bias value The deep belief network is pre-trained; in order to prevent dependence between neural nodes and overfitting, 50% dropout is introduced in the visible layer;

[0100] Step 54: Select an activation function for the hidden layer, perform Gibbs sampling and contrast divergence algorithm for multiple iterations of pre-training on each layer to update the model parameters, and obtain the deep belief network model;

[0101] Step 55: The pre-trained deep belief network is further fine-tuned and optimized by using the BP neural network algorithm and the gradient descent method from top to bottom;

[0102] Step 56: The output of the DBN is judged by the Softmax classifier to determine the type of warning; the test sample is brought into the training model for verification, and the optimal warning model is finally obtained;

[0103] Step 57: input landslide displacement prediction data in the optimal model, and get the grade of landslide warning through the Softmax classifier.

[0104] Summary of the application and data sources

[0105] Research area

[0106] The research of the application relies on the landslide body of HP21 (K60+806~K61+449) section of the south side of Taoyuhe slope in LJ-15 section of the south section of Xi'an Outer Ring Expressway, and the length of the relying engineering section is 3.393km. The landslide occurs in the late middle period of Quaternary System. Rainwater infiltrates into the top surface of mudstone along the pores of loess under the condition of heavy rain or continuous rainfall. The structure of Tertiary System mudstone is dense, forming a water-resisting layer. The shear strength of soil body decays quickly under the condition of saturation, and the landslide is formed under the condition of continuous rainfall. According to the investigation, since the Wenchuan earthquake occurred in 2008, the secondary landslides HP21-1 and HP21-2 have produced different degrees of sliding, resulting in cracks and tilting of houses. It is analyzed that the secondary landslides HP21-1 and HP21-2 are currently in an unstable state. The shear outlet of the front edge of the main landslide is located 5-10m below the modern river channel, there is no free surface of sliding, and the whole landslide is in a stable state.

[0107] Data acquisition

[0108] The generation of landslide is caused by multiple harmful factors, so the present application selects slope, slope shape, 1-hour rainfall, 24-hour rainfall, and 6 disaster-causing factors such as surface absolute displacement monitoring data and crack instrument monitoring displacement as research data in combination with external geographical environment and the convenience of collecting data. The collected information data is transmitted to a remote server through wireless network technology and stored in a database.

[0109] Variational mode decomposition

[0110] In the landslide warning model based on deep belief network, first, the collected surface absolute displacement data is used for prediction, then the collected monitoring data is used to construct a warning model, and finally the predicted displacement data is imported into the optimal warning model to perform real-time warning on whether landslide occurs. Before using the monitoring data to predict displacement, the data needs to be processed to remove outliers and reduce noise: noise reduction is realized by variational mode decomposition plus wavelet threshold denoising method. The data of the present application is recorded once every 1 hour from August 1, 2022 to August 14, 2022, and there are 311 groups of data, and the monitoring data of the first two days are as shown in Table 1:

[0111] ;

[0112] Variational mode decomposition (VMD) is a non-recursive signal processing method, which can decompose time series data into a series of intrinsic mode functions (IMF) with limited bandwidth by iteratively searching for the optimal solution of variational mode. And VMD has good anti-noise ability and can overcome the problem of frequency aliasing. The obtained landslide displacement data is decomposed by VMD to obtain a limited number of IFM. In the process of variable mode decomposition, the input parameters of VMD include f, alpha, tau, K, DC, init, and tol. Among them, f represents the time series signal that needs to be decomposed; alpha represents the mode constraint strength, that is, when the time series signal is decomposed into different IMFs, the limited bandwidth of the frequency band; tau represents the tolerance to noise; K represents the number of modes that need to be decomposed; DC represents whether the mode is a direct current component; init represents the initialization parameter for determining the center frequency of each IMF; and tol represents the accuracy of function convergence. The selection of the value of K needs to be considered comprehensively, and the selection of the value will affect whether under-decomposition or overlap will occur in the process of variational mode decomposition. The best K value is determined by residual error analysis, ADF stability analysis and correlation analysis of different K values, and the specific process is as follows:

[0113] Step 1: Residual error analysis

[0114] First, assume different K values: K=2,3…20, input landslide displacement data as the original input sequence, and perform VMD decomposition and reconstruction. By solving the root mean square error (RMSE) of each reconstructed sequence and the original input sequence, the deviation between the reconstructed sequence and the original sequence is measured:

[0115] ;

[0116] Step 2: ADF test is a unit root test method to detect whether the time sequence is stable. Generally, the smaller the estimator of ADF test statistic, the more negative the original hypothesis, and the smaller the possibility of non-stationary sequence. Therefore, the K value that minimizes the possibility of component or sequence non-stationarity is selected by the ADF value of the analyzed sequence.

[0117] Step 3: Analyze the correlation coefficient between the intrinsic mode component (IMF) and the residual (Re) under different K values and the original input sequence, and finally determine the size of K value. The displacement data is input into the VMD model to obtain K intrinsic mode components, and Table 2 shows the experimental parameters.

[0118] ;

[0119] Each IMF obtained by decomposition is denoised by wavelet thresholding, and finally the denoised components are reconstructed to obtain the reconstructed surface absolute displacement data.

[0120] Step 1: For landslide displacement monitoring data, the compactly supported biorthogonal wavelet (dbN) is preferred as the wavelet basis function through multiple practical applications, and the processing effect is best when the decomposition level is 3, followed by the wavelet soft threshold method for noise reduction processing, and the threshold rule is selected as the wavelet soft threshold Rigorou step Step 2: Use db3 wavelet to perform wavelet transform on the intrinsic modal component (IFM) obtained by variational modal decomposition (VMD) to obtain a group of wavelet coefficients corresponding to the decomposition level;

[0121] Step 2: Use db3 wavelet to perform wavelet transform on the intrinsic modal component (IFM) obtained by variational modal decomposition (VMD) to obtain a group of wavelet coefficients corresponding to the decomposition level;

[0122] Step 3: Compare the decomposed wavelet coefficients with the selected threshold value. If the wavelet coefficient is greater than the threshold value, it is considered that the wavelet coefficient is mainly composed of useful signals, and the wavelet coefficient is retained. Otherwise, it is considered that the wavelet coefficient is mainly composed of noise signals, and the wavelet coefficient is discarded.

[0123] Step 4: Reconstruct the retained signal to obtain the IMF component after noise reduction;

[0124] Step 5: Perform steps 2-4 on each IMF component until all IMF components have been subjected to wavelet threshold noise reduction;

[0125] Step 6: Finally, reconstruct the IMF component after removing noise to obtain the landslide displacement data signal after noise reduction :

[0126] ;

[0127] The processed data after variational modal decomposition and wavelet threshold noise reduction of the original data Figure 4 It can be seen that the input monitoring displacement time series signal retains the sharp peaks and mutations of the original displacement time curve after the above processing, and solves the non-stationary phenomenon existing in the time series.

[0128] Optimize the parameters to establish a long short-term memory neural network (LSTM) model, then input the original monitoring data and VMD-wavelet denoised data into the LSTM model for prediction, and compare the prediction results.

[0129] Use the long short-term memory neural network (LSTM) model to predict the displacement data of the target area; the specific process is as follows:

[0130] Step 1: Establish an LSTM landslide displacement time series prediction model, which consists of an input layer, two hidden layers and an output layer;

[0131] Step 2: data segmentation: the denoising reconstructed displacement data is divided into 80% training set and 20% validation set; the data is imported into the LSTM prediction model;

[0132] Step 3: determination of hyperparameters: when constructing the LSTM network prediction model, several original parameters of the network need to be set in advance, and such original parameters are also called hyperparameters, which specifically include the number of hidden layers of the network, the number of nodes of each layer, the frequency of neuron inactivation, the loss function, the reactivation parameter, the number of iterations, etc. At present, there is still lack of a relatively perfect method to realize the selection of the hyperparameters of the LSTM neural network, and it is usually necessary to adjust through the test effect and the experience of parameter adjustment; the number of nodes of the hidden layer of the neural network is determined by combining the empirical equation:

[0133] ;

[0134] In the formula, is the number of input layer neurons; is the number of output layer neurons, is the number of neurons in the training set, is an integer between 2 and 10;

[0135] Step 4: the number of neurons obtained through the empirical formula still needs to be obtained through continuous experiments to obtain more accurate results; in order to intuitively compare the pros and cons of the prediction results of the model under different parameters, the mean error (MAE), the mean error percentage (MAPE) and the root mean square error (RMSE) are introduced for different result comparison:

[0136] ;

[0137] In the formula represents the i-th group of predicted values, represents the true value of the m+1 time in the i-th group of samples, and L is the number of samples used in one iteration.

[0138] Step 5: the best LSTM model is obtained by continuously adjusting the parameters.

[0139] In order to analyze the advantages and disadvantages of the model, the original displacement data after simple processing is input into the above model as model 1, and the data after VMD-wavelet denoising is input into the model as model 2 for displacement prediction, as shown in Figure 5 The prediction results obtained by the data processing method have obvious advantages and are more consistent with the actual monitoring data.

[0140] The historical monitoring data of landslide points are used for model training of a deep belief network (DBN), and then the prediction data obtained by using an LSTM model are input into the trained DBN model for landslide early warning, so as to construct a landslide early warning model. How to construct the DBN model will be described in detail below:

[0141] The structure of the constructed DBN model is: three RBM networks connected in sequence, one BP network and one Softmax classifier;

[0142] The input is: monitored 1-hour rainfall and 24-hour rainfall, slope gradient and slope shape of the slope body, crack instrument monitored displacement, absolute ground displacement and corresponding early warning categories;

[0143] The output is the early warning category of the Softmax classifier: 0-no early warning, 1-third-level early warning, 2-second-level early warning, 3-first-level early warning, four states are coded, and the output codes corresponding to each state are shown in Table 3:

[0144] ;

[0145] Data preprocessing

[0146] The factors affecting geological disasters are complex and diverse, and are described in a qualitative and non-uniform manner, which cannot be directly calculated mathematically. In order to fully evaluate the sensitivity between different disaster factors, the present application uses a certainty factor (CF) to calculate the coefficients of each influencing factor of the landslide disaster factor.

[0147] Firstly, according to the commonly used classification method, each disaster factor is divided into different subsets, and then the sensitivity of each disaster factor subset is calculated by using a formula. The specific expression of the CF function is:

[0148] ;

[0149] In the formula: is the possibility of landslide disaster in data category a, is the possibility of landslide disaster in the entire study area, and the change interval of CF is [-1, 1]. A positive value represents a high possibility of landslide occurrence and poor geological environmental conditions; a negative value represents a low possibility of landslide occurrence and good geological environmental conditions. When the CF value is 0, it represents uncertainty whether a landslide occurs. The sensitivity coefficients of the disaster factor subsets calculated by using the certainty factor method are shown in Table 4:

[0150] ;

[0151] Then, normalization processing is performed: the landslide data and the slope geometric parameters are normalized to the interval [0, 1] by using the Mapminmax normalization method:

[0152] ;

[0153] wherein, is the normalized value, and respectively represent the original data, the maximum and minimum values of the data set.

[0154] Constructing a deep belief network model

[0155] In the model of machine learning, if the parameters of the model are too many and the training samples are too few, the trained model is prone to overfitting phenomenon. In order to prevent overfitting phenomenon, the Dropout algorithm is introduced in the unsupervised pre-training stage of DBN, and the Dropout technology is used for regularization processing: in the pre-training stage, under the premise that the input and output of the neural network remain unchanged, the weights of the hidden layer nodes are randomly used with a certain probability, and a part of neurons does not participate in the forward propagation training process each time; In particular, the probability of Dropout is set to 50%.

[0156] The specific implementation process of training the deep belief network (DBN) is as follows:

[0157] Step 1: divide the collected and processed multiple groups of historical data into training samples and test samples;

[0158] Step 2: input the data into the model, and use the unsupervised greedy algorithm to initialize the link weights between the layers of the restricted Boltzmann machine unit network model and the bias value Pre-train the deep belief network. In order to prevent dependence between neural nodes and overfitting phenomenon, 50% dropout (randomly freeze 50% of the nodes, and retain their weights in network training) is introduced in the visible layer, and the selected nodes restore the previously retained weights in the next training process, and then randomly select part of the nodes to repeat the process. The structure diagram of the deep belief network in the embodiment of the application is shown in Figure 3 .

[0159] Step 3: select an activation function for the hidden layer, perform Gibbs sampling method and contrast divergence algorithm to pre-train each layer multiple times, update the model parameters, and obtain the deep belief network model. The parameter update formula is as follows:

[0160] ;

[0161] In the formula, represents the updated value of the weight matrix, represents the updated value of the i-th input layer bias, represents the update value of the jth hidden layer bias, represents the bias between the visual layer (hidden layer) neurons, represents the bias reconstruction between the visual layer (hidden layer) neurons. represents the learning rate, and the value is 0.0001-0.5.

[0162] Step 4: For the pre-trained deep belief network, the BP neural network algorithm and the gradient descent method are used to further fine-tune and optimize the model parameters from top to bottom. In addition, due to the use of dropout technology in the training process, the error is set to zero with a probability of 50% when calculating the error of the network calculation node in the fine-tuning optimization process. The output error is used to estimate the error of the previous layer of the output layer, and after the error estimation of each layer is obtained through the backward propagation learning, the gradient descent method is used to calculate and update the weight of each node, and the reconstruction error is minimized layer by layer.

[0163] Step 5: The output of the DBN is judged by the Softmax classifier to determine the type of early warning; the test sample is brought into the training model for verification, and finally the optimal early warning model is obtained;

[0164] Step 6: The environmental monitoring data and the landslide displacement prediction data are input into the optimal model, and the grade of landslide early warning is obtained through the Softmax classifier.

[0165] Through the above six steps, a multi-factor landslide induced landslide prediction model of DBN can be constructed. The slope, slope shape, 1-hour rainfall and 24-hour rainfall, and six disaster factors such as surface absolute displacement monitoring data and crack instrument monitoring displacement are selected as the input of the neural network. The Dropout mechanism is used to randomly sample the hidden layer node weight with a certain probability to avoid the occurrence of overfitting phenomenon, and finally the BP algorithm is used to fine-tune and optimize the parameter value of each layer. Through Figure 6 As shown in the figure, the error rate of the early warning effect of the DBN model with the Dropout mechanism is lower, and the effect is better.

[0166] The above only describes the preferred embodiments of the present application, and does not limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A landslide early warning method based on data mining, characterized in that, include: Step 1: Obtain historical monitoring data for the target location and store the obtained data in the database; Step 2: Perform variational mode decomposition on the obtained absolute surface displacement data to obtain a finite number of intrinsic mode components; Step 3: Perform wavelet threshold denoising on each intrinsic mode component obtained from the decomposition, and finally reconstruct the denoised components to obtain the denoised and reconstructed displacement data. Step 4: Use a long short-term memory neural network to predict the displacement of the noise-reduced shifted data; Step 5: Construct a deep belief neural network model and use landslide displacement prediction data to provide early warning of landslides; In step 1, historical monitoring data of the target location is acquired and stored in a database. This data includes the following aspects: slope, slope shape, 1-hour rainfall, 24-hour rainfall, absolute surface displacement monitoring data, and displacement monitored by crack meters, among other six disaster-causing factors, as research data. The collected information data is transmitted to a remote server via wireless network technology and stored in the database. In step 2, variational mode decomposition is performed on the obtained time-series landslide displacement data to obtain a finite number of intrinsic mode components. The most important aspect of decomposing the displacement signal is determining the value of the number of mode components, K. The process is as follows: Step 21: Residual Error Analysis Assuming different values ​​of k, landslide monitoring data are used as the original input sequence, and VMD decomposition and reconstruction are performed. The deviation between the reconstructed sequence and the original sequence is measured by solving the root mean square error between each reconstructed sequence and the original input sequence. ； Step 22: ADF stability analysis, select the K value that minimizes the possibility of component or sequence nonstationarity based on the ADF values ​​of the analyzed sequence; Step 23: Analyze the correlation coefficients between the intrinsic modal components and residuals and the original input sequence under different K values, and finally determine the value of K; input the displacement data into the VMD model to obtain K intrinsic modal component data; In step 3, wavelet threshold denoising is performed on each IMF obtained by decomposition, and finally the denoised components are reconstructed to obtain the reconstructed displacement data. The specific process is as follows: Step 31: For landslide displacement monitoring data, compactly supported bioorthogonal wavelets are preferred as wavelet basis functions. The processing effect is best when the number of decomposition layers is 3. Secondly, wavelet soft thresholding is used for noise reduction. The threshold rule is wavelet soft thresholding Rigorous SURE. Step 32: Use the db3 wavelet to perform wavelet transform on the IMF components to obtain a set of wavelet coefficients corresponding to the decomposition level; Step 33: Compare the decomposed wavelet coefficients with the selected threshold rule. If the wavelet coefficient is greater than the threshold, it is considered that the wavelet coefficient is mainly composed of useful signals and the wavelet coefficient is retained. Otherwise, it is considered that the wavelet coefficient is mainly composed of noise signals and the wavelet coefficient is discarded. Step 34: Reconstruct the retained signal to obtain the denoised IMF component; Step 35: Perform steps 2-4 on each IMF component until all IMF components have undergone wavelet thresholding noise reduction. Step 36: Finally, the noise-removed IMF components are reconstructed to obtain the noise-reconstructed landslide displacement data signal. In step 4, a long short-term memory neural network is used to predict the displacement of the noise-reduced shift data. The specific process is as follows: Step 41: Establish an LSTM landslide displacement time series prediction model, which consists of one input layer, two hidden layers, and one output layer; Step 42: Data Segmentation: Divide the denoised and reconstructed displacement data into an 80% training set and a 20% validation set; import the data into the LSTM prediction model; Step 43: Determining Hyperparameters: Determine the number of hidden layer nodes in the neural network using empirical equations. ； In the formula, It is the number of neurons in the input layer; It is the number of neurons in the output layer. It is the number of neurons in the training set. Integers between 2 and 10; Step 44: The number of neurons obtained through empirical formulas still needs to be tested continuously to achieve more accurate results; to intuitively compare the model's prediction results under different parameters, mean error, mean error percentage, and root mean square error are introduced for comparison of different results: ； ； ； In the formula This represents the predicted value of the i-th group. Let L represent the true value at time m+1 in the i-th sample group, and L be the number of samples used in one iteration. Step 45: Step 5: Continuously adjust the parameters to obtain the optimal LSTM model parameters; In step 5, a deep belief network model is constructed using historical monitoring data, and then landslide displacement prediction data is input into the model to provide an early warning of whether a landslide will occur. The model structure is as follows: The DBN model has the following structure: three RBM networks, one BP network, and one Softmax classifier connected in sequence. The inputs are: the monitored 1-hour and 24-hour rainfall, the slope and shape of the slope, the displacement monitored by the crack meter, the absolute displacement of the ground surface, and the corresponding warning category; The output is the warning category of the Softmax classifier: four states: Level 1 warning, Level 2 warning, Level 3 warning, and no warning. The Dropout algorithm is introduced into the unsupervised pre-training stage of DBN, and regularization is performed using Dropout technology: During the pre-training stage, while keeping the neural network input and output unchanged, the weights of the hidden layer nodes are randomly selected with a certain probability. Each adjustment excludes a portion of neurons from the forward propagation training process; specifically, the probability of Dropout is set to 50%. In step 5, a deep belief network is used to comprehensively analyze historical landslide monitoring data, slope support information data, and landslide displacement prediction data to provide an early warning of whether a landslide is likely to occur. The specific implementation process of using the deep belief network is as follows: Step 51: Use the deterministic coefficients to calculate the coefficients of each influencing factor of landslide disaster factors, and then use the Mapminmax normalization method to normalize the landslide data to the [0,1] interval; Step 52: Divide the collected and processed historical data into training samples and test samples; Step 53: Input the data into the model and use an unsupervised greedy algorithm to initialize the link weights between the layers of the restricted Boltzmann machine unit network model. and bias value The deep belief network is pre-trained; to prevent dependency between neurons and overfitting, 50% dropout is introduced in the visible layer. Step 54: Select activation functions for the hidden layers, perform Gibbs sampling and contrastive divergence algorithm to perform multiple iterations of pre-training for each layer to update the model parameters and obtain the deep belief network model; Step 55: For the pre-trained deep belief network, the BP neural network algorithm and gradient descent method are used to further fine-tune and optimize its model parameters from top to bottom; Step 56: Use the DBN output to determine the type of warning through the Softmax classifier; input the test samples into the training model for verification, and finally obtain the optimal warning model; Step 57: Input landslide displacement prediction data into the optimal model and obtain the landslide warning level through the Softmax classifier.

Images