A landslide monitoring and early warning method, equipment, medium and product

Abstract

The present application discloses a landslide monitoring and early warning method, equipment, medium and product, which relate to the field of landslide monitoring and early warning. The method comprises: combining the evolution law of the landslide itself and the mechanism of rainfall on landslide deformation, extracting feature engineering from landslide monitoring data, determining the sample set features of a rainfall-type red layer residual slope accumulation landslide cumulative displacement prediction model, generating a predicted displacement sequence according to the rainfall-type red layer residual slope accumulation landslide cumulative displacement prediction model; determining the predicted landslide time by using the predicted displacement sequence and an improved inverse velocity method; determining the predicted landslide time by using the improved inverse velocity method for the same landslide at different deformation stages; a and A The value of is different. a and A The value of changes with the evolution of the deformation stage; the difference between the predicted landslide time and the current time is used as the early warning indicator, and a rainfall-type red layer residual slope accumulation landslide monitoring and early warning model based on the early warning indicator and the landslide displacement rate is constructed to monitor and warn the landslide. This application can achieve timely early warning of landslides.

Description

Technical Field

[0001] The present application relates to the field of landslide monitoring and early warning, and in particular to a landslide monitoring and early warning method, equipment, medium and product. Background Art

[0002] Rainfall-induced red bed residual slope landslides frequently occur in the northeastern Sichuan Basin. Under the disturbance of rainfall, the time from the appearance of obvious deformation to the occurrence of landslides is generally relatively short. These landslides are unconventional, sudden landslides induced by rainfall disturbance, and the deformation is more sudden. Therefore, it is difficult to predict the time of such landslides, and it is difficult to achieve advance warning and occurrence time forecast of such landslides through conventional methods. Summary of the invention

[0003] The purpose of this application is to provide a landslide monitoring method, equipment, medium and product to solve the problem of difficulty in timely warning of landslides.

[0004] To achieve the above objectives, this application provides the following solutions.

[0005] In a first aspect, the present application provides a landslide monitoring and early warning method, comprising the following steps.

[0006] Combined with the evolution law of the landslide itself and the mechanism of rainfall on landslide deformation, feature engineering is performed on the landslide monitoring data to determine the sample set characteristics of the rainfall-type red layer residual slope accumulation landslide cumulative displacement prediction model; the landslide monitoring data include displacement data, landslide slope, rainfall, and physical and mechanical parameters of the residual slope accumulation layer during rainfall infiltration; the sample set characteristics include historical displacement data characteristics, rainfall factor characteristics, stability coefficient characteristics, rainfall infiltration depth characteristics and time factor characteristics.

[0007] Based on the sample set characteristics, a predicted displacement sequence is generated according to the rainfall-type red layer residual slope accumulation landslide cumulative displacement prediction model.

[0008] The predicted landslide time is determined by using the predicted displacement sequence and the improved velocity inverse method, which specifically includes: determining the velocity inverse-time relationship according to the velocity inverse-time curve of the landslide; the velocity inverse-time relationship is: ;in, V for speed; t f is the time when the landslide occurred, t For time; A and a All are empirical constants.

[0009] The speed inverse-time relationship is processed logarithmically, and the logarithm-processed speed inverse-time relationship is converted into a linear equation; wherein the speed inverse-time relationship after logarithm processing is: .

[0010] Based on the linear equation, combined with the velocity inverse-time data, the a and A The power relationship of .

[0011] Determine according to the power relationship a and A The change curve over time determines the a and A The empirical value of ; ; m 1. n 1. m 2. n 2. p is a constant.

[0012] According to the a and A The time-varying change curve identifies the deformation acceleration points in the accelerated deformation phase.

[0013] Based on the deformation acceleration point, the predicted displacement sequence is used and the improved velocity inverse method-time relationship is applied to determine the predicted landslide time; the same landslide at different deformation stages in the improved velocity inverse method a and A The value of is different. a and A The value of changes with the evolution of the deformation stage.

[0014] The difference between the predicted landslide time and the current time is used as an early warning indicator, and a rainfall-type red layer residual slope accumulation landslide monitoring and early warning model based on the early warning indicator and the landslide displacement rate is constructed.

[0015] Landslides are monitored and warned based on the rainfall-type red layer residual slope accumulation landslide monitoring and early warning model.

[0016] In a second aspect, the present application provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any of the landslide monitoring and early warning methods described above.

[0017] In a third aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the landslide monitoring and early warning methods described above.

[0018] In a fourth aspect, the present application provides a computer program product, including a computer program, which, when executed by a processor, implements any of the landslide monitoring and early warning methods described above.

[0019] According to the specific embodiments provided by the present application, the present application discloses the following technical effects: the present application combines the evolution law of the landslide itself and the mechanism of rainfall on landslide deformation, performs feature engineering extraction on the landslide monitoring data, and determines the sample set characteristics of the rainfall-type red layer residual slope accumulation landslide cumulative displacement prediction model. The sample set characteristics include historical displacement data characteristics, 10 rainfall factor characteristics, stability coefficient characteristics, rainfall infiltration depth characteristics and time factor characteristics, a total of 14 specific characteristics, so as to effectively utilize the correlation with the cumulative displacement of the landslide, so that the model prediction effect is better; and based on the sample set characteristics, a predicted displacement sequence is generated according to the rainfall-type red layer residual slope accumulation landslide cumulative displacement prediction model. The accurate and reliable rainfall-type red layer residual slope accumulation landslide cumulative displacement prediction model established according to multiple characteristics can provide data reference value for landslide early warning; and the velocity inverse method is improved to change the same landslide at different deformation stages. a and A The value of a and A The value of changes with the evolution of the deformation stage. Compared with the landslide imminent sliding time prediction method based on the traditional velocity inverse method, the landslide monitoring and early warning method based on the improved velocity inverse method to predict the landslide imminent sliding time is more suitable for sudden landslides. The difference between the predicted landslide time and the current time is used as an early warning indicator to construct a rainfall-type red layer residual slope accumulation landslide monitoring and early warning model based on the early warning indicator and the landslide displacement rate. The rainfall-type red layer residual slope accumulation landslide cumulative displacement prediction model is established by coupling a bidirectional long short-term memory network (Bidirectional Long Short-Term Memory, Bi-LSTM) and an information extraction network (Informer). The constructed rainfall-type red layer residual slope accumulation landslide cumulative displacement prediction model has higher prediction accuracy, thereby monitoring and early warning of landslides, accurately predicting the future deformation development trend of landslides, and realizing timely early warning of landslides. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0021] Figure 1 A flow chart of a landslide monitoring and early warning method provided in one embodiment of the present application.

[0022] Figure 2 This is the result diagram of the importance of a landslide feature.

[0023] Figure 3 Schematic diagram of rainfall infiltration calculation model.

[0024] Figure 4 A simplified diagram of the improved infinite slope calculation.

[0025] Figure 5 This is the relationship diagram between the accumulated displacement and stability coefficient characteristics of the landslide and the rainfall infiltration depth.

[0026] Figure 6 This is a graph of average monthly rainfall for each month.

[0027] Figure 7 This is the Bi-LSTM structure diagram.

[0028] Figure 8 This is the Informer structure diagram.

[0029] Fig. 9 Schematic diagram of the landslide cumulative displacement prediction model.

[0030] Fig.10 for a With time t Relationship graph.

[0031] Fig.11 for A With time t Relationship curve diagram; t ≥0.211.

[0032] Fig.12 It is a graph of speed short-term moving average (SMA) and speed long-term moving average (LMA).

[0033] Fig.13 Comparative analysis diagram of landslide time prediction between the improved velocity inverse method and the traditional velocity inverse method.

[0034] Fig.14 This is a comparative analysis chart of landslide time prediction using displacement prediction data and original data. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0036] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0037] Taking rainfall-type red layer residual slope accumulation landslide as the research object, the inventors explored the monitoring and early warning method suitable for this type of landslide. a. A Fixed, but in the actual landslide time prediction calculation, different landslide a. A Different values ​​indicate that the same landslide has different deformation stages. a. A Therefore, the traditional velocity inverse method has low accuracy and poor timeliness in predicting the imminent landslide time. However, this application identifies the accelerated deformation point and intends to establish different velocity inverse method according to different deformation stages of the landslide. a. A The value can improve the accuracy and timeliness of the prediction of the imminent sliding time of red layer residual slope landslide.

[0038] The present application embodiment provides a landslide monitoring and early warning method, such as Figure 1 As shown, the method is executed by a computer device, and specifically can be executed by a computer device such as a terminal or a server alone, or can be executed by a terminal and a server together, and includes the following steps 101 to 105.

[0039] Step 101: In combination with the evolution law of the landslide itself and the mechanism of rainfall on landslide deformation, feature engineering is performed on the landslide monitoring data to determine the sample set characteristics of the rainfall-type red layer residual slope accumulation landslide cumulative displacement prediction model; the landslide monitoring data includes displacement data, landslide slope, rainfall, and physical and mechanical parameters of the residual slope accumulation layer during rainfall infiltration; the sample set characteristics include historical displacement data characteristics, rainfall factor characteristics, stability coefficient characteristics, rainfall infiltration depth characteristics, and time factor characteristics.

[0040] Step 102: Based on the sample set characteristics, a predicted displacement sequence is generated according to the rainfall-induced red layer residual slope accumulation landslide cumulative displacement prediction model.

[0041] Step 103: Determine the predicted landslide time using the predicted displacement sequence and the improved velocity inverse method.

[0042] Furthermore, according to the curve between the inverse of the velocity and time of the landslide, the inverse of the velocity and time relationship is determined; the inverse of the velocity and time relationship is: ;in, V for speed; t f is the time when the landslide occurred, t For time; A and a All are empirical constants.

[0043] The speed inverse-time relationship is processed logarithmically, and the logarithm-processed speed inverse-time relationship is converted into a linear equation; wherein the speed inverse-time relationship after logarithm processing is: .

[0044] Based on the linear equation, combined with the velocity inverse-time data, the a and A The power relationship of .

[0045] Determine according to the power relationship a and A The change curve over time determines the a and A The empirical value of ; m 1. n 1. m 2. n 2. p is a constant.

[0046] According to the a and A The time-varying change curve identifies the deformation acceleration points in the accelerated deformation phase.

[0047] Based on the deformation acceleration point, the predicted displacement sequence is used and the improved velocity inverse method-time relationship is applied to determine the predicted landslide time; the same landslide at different deformation stages in the improved velocity inverse method a and A The value of is different. a and A The value of changes with the evolution of the deformation stage.

[0048] Step 104: taking the difference between the predicted landslide time and the current time as an early warning indicator, and constructing a rainfall-induced red layer residual slope accumulation landslide monitoring and early warning model based on the early warning indicator and the landslide displacement rate.

[0049] Step 105: monitoring and warning the landslide according to the rainfall-type red layer residual slope accumulation landslide monitoring and early warning model.

[0050] In an exemplary embodiment, before step 101, the monitoring data needs to be preprocessed, and the preprocessing steps are as follows.

[0051] The main contents of data preprocessing usually include outlier processing, missing value filling and data normalization. In this application, since the original monitoring data of the landslide studied in this application is relatively complete, there is no need to fill in missing values, and no further explanation of missing value filling is given.

[0052] (1) Outlier processing: The interval prediction model is constructed for the accumulated displacement of the landslide through the differential autoregressive moving average model, and the sliding window algorithm is used to predict its segmented subsequences. The confidence interval is corrected using the predicted value to propose an outlier factor. β to identify outliers.

[0053] This method is used to process data outliers. β After the abnormal point is identified by the size of the value, the abnormal point is replaced by the landslide monitoring data of the previous moment.

[0054] ;in, T i is the original data, u cl is the upper limit of the modified confidence interval of the predicted value, l cl is the lower limit of the revised confidence interval for the predicted value.

[0055] When the original data deviates to twice the interval length, β The value is 3, then according to the abnormal factor β The size of the abnormal level is divided into three levels, as shown in Table 1.

[0056] Table 1 Abnormality level table

[0057]

[0058] (2) Normalization: Data normalization makes the data more standardized by unifying the dimensions and range of sample set features to the same scale.

[0059] MaxAbsScaler (maximum value normalization) is used to normalize the data. This method normalizes the data range by scaling all eigenvalues ​​to a ratio between 0 and 1, and maintains the relative relationship between data points. After the calculation of the cumulative displacement prediction model of rainfall-type red layer residual slope accumulation landslide is completed, the data obtained by the cumulative displacement prediction model of rainfall-type red layer residual slope accumulation landslide can be obtained by denormalization to obtain the predicted data.

[0060] ;in, x It is the original data. x ′ is the data after standardization, | x max | is the maximum absolute value of the original data.

[0061] In an exemplary embodiment, step 101 may be replaced by the following steps 201 to 204 .

[0062] Step 201: In combination with the evolution law of the landslide itself and the mechanism of rainfall on landslide deformation, the rainfall factors are analyzed based on the historical average rainfall, the previous effective rainfall and the number of rainy days, and the importance of the decision tree algorithm is used to screen the rainfall factors and determine the characteristics of the rainfall factors.

[0063] Step 202: Determine the stability coefficient characteristics and rainfall infiltration depth characteristics according to the permeability coefficient, shear strength and landslide slope.

[0064] Step 203: Based on the rainfall data of different months, a time feature analysis is performed to determine the time factor features, and the time factor features are classified according to the similarity between the time factor features; the time factor features include 12 months; the classification results of the time factor features are as follows: November, December, January, February and March are in the first category, April, May, June and October are in the second category, and July, August and September are in the third category.

[0065] Step 204: Determine the sample set characteristics of the rainfall-induced red layer residual slope accumulation landslide cumulative displacement prediction model according to the historical displacement data characteristics, the rainfall factor characteristics, the stability coefficient characteristics, the rainfall infiltration depth characteristics and the time factor characteristics.

[0066] Furthermore, the sample set feature extraction specifically includes the following steps.

[0067] After the training set, validation set, and test set were divided, the rainfall factors were analyzed based on the historical average rainfall, previous effective rainfall, and number of rainy days. The two characteristics of rainfall infiltration depth and stability coefficient were calculated through the permeability coefficient, shear strength, and landslide slope. Finally, the time characteristic analysis was performed, and the sample set characteristics of the landslide displacement prediction model were extracted through feature combination.

[0068] First, the data after outlier processing is divided into training set, validation set and test set according to the chronological order, and the historical average rainfall, previous effective rainfall and rainfall days are considered to obtain 3h average rainfall, 6h average rainfall...14-day average rainfall; 1-day effective rainfall, 2-day effective rainfall...14-day effective rainfall; 2-day rainfall days, 4-day rainfall days...14-day rainfall days, a total of 32 rainfall factors. However, too many features will reduce the efficiency of model training, so feature screening is needed to remove redundant rainfall factors, and determine the features of the 10 rainfall factors after screening. Figure 2 The "d" in it stands for "day".

[0069] Furthermore, the importance of the decision tree algorithm can be used I ( F ) to filter.

[0070] ; In the formula, is the feature in the decision tree algorithm F Gini impurity of ; is the feature in the decision tree algorithm i Gini impurity of ; N is the total number of samples of the node before splitting.

[0071] After ranking the importance results of the features according to the decision tree algorithm, irrelevant features with lower importance are removed, and the final screening results are as follows: Figure 2 shown.

[0072] According to relevant research, the Van-Genuchten (VG) model of unsaturated soil and the improved Green-Ampt rainfall infiltration model are used to calculate the rainfall infiltration depth. The calculation model is as follows: Figure 3 As shown, Direction perpendicular to the slope The direction is parallel to the slope surface. is the soil moisture content, assuming that the volume moisture content in the conductive zone is uniformly distributed.

[0073] (1) When the permeability coefficient k s ≥ Table rainfall intensity q Time, rainfall time t The corresponding rainfall infiltration depth z w : ;in, represents the initial volumetric moisture content of the soil, It indicates the volume moisture content of soil in the conductive zone under rainfall infiltration.

[0074] (2) When k s < q hour, ;in, S is the matrix suction of the soil at the infiltration front; M It is the difference between the saturated volumetric moisture content of the soil and the initial volumetric moisture content of the soil; α is the slope; t For rainfall time; k s is the permeability coefficient of soil.

[0075] In addition, according to the unsaturated soil shear strength theory and limit equilibrium method proposed by Fredlund et al. in 1993, combined with the improved infinite slope calculation diagram, such as Figure 4 As shown, W is the weight of the sliding body, N 1 is the weight of the sliding body perpendicular to the landslide, F 1 is the sliding force of the sliding body, is the shear stress at the wetting peak, is the normal stress at the wetting peak, The pore water pressure is used to derive the calculation method of the landslide stability coefficient when the potential sliding surface is located in the following three situations.

[0076] (1) Under natural conditions, the landslide stability coefficient F s : .

[0077] (2) Under rainfall conditions and rainfall infiltration depth z w Less than soil thickness H hour, .

[0078] (3) Under rainfall conditions and rainfall infiltration depth z w Equal to the soil thickness H hour, .

[0079] in, c ′ is the effective cohesion; is the effective friction angle; is the suction friction angle of shear strength changing with soil matrix suction, and its value is approximately equal to the effective friction angle ; is the saturated density of soil; H is the soil layer thickness; is the saturated density of soil; z f It is the suction head at the wetting front.

[0080] The stability coefficient and rainfall infiltration depth can be calculated based on the above formula, such as Figure 5 shown.

[0081] Since the rainfall data in different months have significant differences in temporal characteristics, such as Figure 6 As shown in the figure, the average rainfall in different months varies significantly, and the rainy season is from July to September every year, which leads to frequent landslides. The month can be included in the sample set as a feature. However, considering that the time factor feature contains twelve categories, some of which have high similarity, it can be considered to be reclassified, among which November, December, January, February and March are in the first category, April, May, June and October are in the second category, and July, August and September are in the third category.

[0082] Finally, the sample set features of the landslide displacement prediction model are extracted through feature combination.

[0083] In an exemplary embodiment, in terms of model selection, the rainfall-type red layer residual slope accumulation landslide is affected by rainfall data factors, and the deep learning method with deep neural network is more suitable for the prediction of nonlinear landslide displacement time series data affected by rainfall characteristics. Therefore, the rainfall-type red layer residual slope accumulation landslide cumulative displacement prediction model is established by coupling a bidirectional long short-term memory network and an information extraction network.

[0084] In an exemplary embodiment, Bi-LSTM is a neural network improved and optimized by a long short-term memory neural network (LSTM). LSTM is a deep learning algorithm widely used in time series processing and prediction. Its original intention of designing it is to overcome the shortcomings of conventional recurrent neural networks (RNNs) in dealing with long-term dependency problems.

[0085] The core of the LSTM network is a basic unit consisting of a forget gate, an input gate, and an output gate. The forward propagation process of the three logic gates is shown in the following formula.

[0086] .

[0087] .

[0088] .

[0089] Where: f t , i t , o t They are t Always keep in mind the state of the forget gate, input gate, and output gate; W f , W i , W o They are the weight matrices of the forget gate, input gate, and output gate respectively; b f , b i , b o They are the bias items of the forget gate, input gate and output gate respectively; W s is the input unit weight matrix; b s is the input unit state bias term; s t is a temporary state quantity; h t-1 and x t They are t-1 The intermediate output at time and the input at time constitute t Input of each door at all times; c t-1 and c t They are t -1 moment and t State memory variables at each moment; h t for t The intermediate output at the moment; σ is the sigmoid activation function; ⊙ represents the multiplication of matrix elements; tanh is the hyperbolic tangent activation function.

[0090] Figure 7 It is a schematic diagram of the Bi-LSTM network, where x n For input data, n =1,2,..., the input of the model is processed by the LSTM network in both the sequential and reverse directions, and the output result is the concatenation of the outputs in these two directions. Therefore, the model output obtained by network training contains the information of the time-coded input sequence in both the sequential and reverse directions. Due to the use of weight reuse, this method does not increase the demand for data volume while enhancing the model's expressiveness, reducing the risk of underfitting.

[0091] Considering that the historical accumulated displacement data of landslides, rainfall factors, residual slope accumulation layer strength factors, and time factors are independent time series, this application couples these characteristic information that affect landslide displacement through a vector representation method.

[0092] Firstly, the accumulated displacement value of the landslide at a certain moment in the sample set and its related features are concatenated into a vector to form a new time series data; then, the sliding window method is used to generate the feature graphs of the input time series data one by one to facilitate the subsequent network calculation.

[0093] Furthermore, Informer is a new time series feature learning model that solves the problem of high computational complexity faced by the Transformer model when learning long time series while ensuring the effect of feature learning.

[0094] Transformer is a neural network model proposed by Google in 2017 that is completely composed of self-attention. It effectively solves the long-term dependency problems of traditional recurrent neural networks and long short-term memory networks, and has achieved great success in natural language processing, image recognition, time series processing and other fields.

[0095] The computational processing of the Informer model consists of two parts: the encoder and the decoder, both of which contain a position information embedding layer, a multi-head probabilistic sparse self-attention layer, and a self-attention distillation layer.

[0096] The basic logic is: first, determine the time-coded input sequence; second, extract features from the neural network composed of a multi-head probabilistic sparse self-attention layer and a self-attention distillation layer; finally, obtain the estimated value of the landslide displacement prediction based on the fully connected layer. The model structure is as follows: Figure 8 shown.

[0097] The linear transformation of sine and cosine functions is used for position encoding. The specific calculation method is: .

[0098] in, For location POS and dimensions j The coded value on POS for t The time instant is a fixed position in the original time-coded input sequence, j is the dimension of the position encoding, j =1, 2, …, d model , d model Dimensions encoding the position; L x is the length of the original time-coded input sequence. When representing the time-coded input sequence of the encoder and decoder, L x The values ​​of are the lengths of their time-coded input sequences, L en and L de ; i is the index of any dimension in the position encoding; Is a positive integer.

[0099] The attention mechanism of the multi-head probabilistic sparse self-attention layer is expressed as: ; is the attention weight, Q is the query vector (Query), K is the key vector (Key), V is the value vector (Value); is the representation of the query vector; d K is the key vector K Dimensions; K T is the key vector K The transposed matrix of .

[0100] Secondly, the result of the probabilistic sparse attention layer is input into the self-attention distillation layer, which consists of a one-dimensional convolutional layer with an activation function of ELU Function, the convolution is followed by a pooling layer with a stride of 2. d Layer to d The calculation formula for +1 layer is: ;in, The time step length in the time series is t , the number of layers is d +1 layer value; ELU is the activation function; It is a multi-head probabilistic sparse self-attention mechanism With self-attention mechanism B The key operation, Convld For a one-dimensional convolution operation on the time series, use ELU Function to activate; MaxPool For the max pooling layer operation, the temporal encoding input sequence is cropped.

[0101] The decoder of the Informer model consists of a multi-head probabilistic sparse self-attention layer, a full attention layer, and a fully connected layer. Its time-coded input sequence consists of the sequence to be predicted and a known sequence before the time point of the sequence to be predicted, and the final landslide cumulative displacement prediction result is output through the fully connected layer of the decoder. The input calculation formula of the decoder is: .

[0102] in, X feed_decoder is the input sequence to be predicted; X token is a known sequence between the sequences to be predicted, which is used as a label; X placeholder is the sequence to be predicted, filled with 0; L token is the length of the historically known sequence; L y is the length of the sequence to be predicted.

[0103] This application combines the Bi-LSTM network and the Informer network to construct a rainfall-induced red layer residual slope landslide cumulative displacement prediction model. Fig. 9 shown.

[0104] For the landslide displacement prediction model, we first need to determine its input sequence and output sequence. By analyzing the factors affecting the landslide cumulative displacement data, the input data of the landslide displacement prediction model is XIt includes the following 14 features: historical displacement data features, 10 rainfall factor features, stability coefficient features, and time factor features, denoted as X =[ x 1 , x 2 , x 3 ,…, x 14 ], where the stability coefficient features include two residual slope layer strength factor features. In order to combine the advantages and disadvantages of the two deep learning models, the average value of their prediction results is taken for superposition and reconstruction to obtain the final predicted displacement sequence.

[0105] The landslide displacement prediction model is built using machine learning and deep learning libraries such as Pandas, Matplotlib, Numpy, Scikit-learn, and TensorFlow. The parameters of the Bi-LSTM model and the Informer model are shown in Table 2.

[0106] Table 2 Parameters of landslide displacement prediction model

[0107]

[0108] In order to quantify the prediction effect of landslide cumulative displacement, the root mean square error RMSE, mean absolute error MAE, mean absolute percentage error MAPE and goodness of fit R are used. 2 Four indicators are used to measure the displacement prediction error. The smaller the RMSE, MAE and MAPE, the better the prediction performance of the model. 2 The closer it is to 1, the better the model fitting effect is. The definitions of various indicators are as follows.

[0109] .

[0110] .

[0111] .

[0112] .

[0113] in, is the total number of predicted samples; y i for i The actual value of displacement at the moment; for i The predicted displacement value at the moment; is the average value of the actual displacement.

[0114] In an exemplary embodiment, step 103 is further described.

[0115] Generally speaking, gravity-type landslides can be effectively predicted using the inverse velocity method based on creep theory. However, rainfall-type landslides, especially the rainfall-type red-layer residual slope accumulation landslides mentioned in this application, are difficult to predict in a timely manner using the traditional inverse velocity method due to their sudden deformation and destruction trends.

[0116] The specific data of future displacement can be predicted by the landslide displacement prediction method. The present application can accurately simulate the influence of rainfall on landslide displacement data by considering multiple features including rainfall factors in the landslide displacement prediction model.

[0117] At present, landslide displacement prediction is mostly used for deformation trend analysis, and relatively less for the prediction of the imminent landslide time. The present application can avoid dependence on future data by performing analysis based on displacement prediction data, and predict the imminent landslide time more timely.

[0118] Since the formula of the traditional speed inverse method is basically based on assumptions a = 2. However, in actual situations, after the landslide enters the accelerated deformation stage, the data points of velocity inverse-time are not necessarily distributed on a straight line.

[0119] Different landslides in the prediction calculation of actual landslide imminent sliding time based on the velocity inverse method a , A Different values ​​indicate that the same landslide has different deformation stages. a , A The values ​​are also different. a , A The values ​​increase with the evolution of the deformation stage. After identifying the accelerated deformation point, this application intends to set up different a , A Get the value.

[0120] However, how to analyze a , A The value of is another difficult problem, because the formula is affected by two parameters at the same time, making it difficult to conduct quantitative analysis.

[0121] The velocity inverse-time curve of the rainfall-type red layer residual slope landslide studied in this application is analyzed below.

[0122] Taking the logarithm of both sides of the above equation, we get: .

[0123] In the above formula, replace the unknown terms and constant terms, and let: .

[0124] Then the original formula becomes: .

[0125] Can be determineda , A The values ​​are: ,in, y t For the moment t of y value, x t For the moment t of x value.

[0126] After extensive analysis of the velocity inverse-time data of other local red bed residual slope landslides, it was found that a The value increases with time, showing a trend of first increasing rapidly and then increasing slowly. A The value increases with time approximately linearly, then a , A The growth relationship over time can be summarized as a power relationship ( y = mx n ), according to the above formula and curve fitting, the empirical a , A Get the value.

[0127] .

[0128] Where: m 1= e 0.3396 , n 1=0.2178, m 2= e -5.9525 , n 2=0.9010, p is a constant, and its value is approximately 0.211. t The unit is hour (h). t Less than p hour, a and A All of them make the formula meaningful ( a >1).

[0129] According to the above formula, a and A Over time t The change curve of Fig.10 , Fig.11 As shown, a and A Over time t increases with the increase of .

[0130] Furthermore, the accelerated deformation stage is crucial to the successful early warning of landslides. Therefore, after the deformation acceleration point is identified, the displacement monitoring data of the accelerated deformation stage is used to predict the short-term landslide time. For rainfall-type red layer residual slope accumulation landslides, the speed of the landslide changes significantly from the constant speed deformation stage to the accelerated deformation stage. The deformation acceleration point can be identified by the intersection of the short-term moving average (SMA) of the speed and the long-term moving average (LMA) of the speed. Taking a certain landslide as an example, the identification of its deformation acceleration point is as follows: Fig.12 shown.

[0131] Identifying deformation acceleration points based solely on the intersection of the above two lines may result in the identification of redundant deformation acceleration points. This application limits the number of acceleration points by setting a value range for the velocity magnitude of the deformation acceleration points. Studies have found that when the velocity magnitude of the deformation acceleration points is set to be greater than 0.06 mm / h, the intersection of the two lines can effectively identify the deformation acceleration points. When the velocity magnitude of a certain point is less than 1 mm / h, the deformation trend of the landslide changes significantly, and a new curve fitting should be performed, that is, the deformation acceleration points should be re-determined.

[0132] The predicted displacement sequence of the cumulative displacement of the landslide is obtained through the constructed thermal landslide cumulative displacement prediction model, where the predicted displacement sequence includes multiple predicted displacement data, and the time interval of the predicted displacement data is the sampling interval of the most recent monitoring time of the original data. A relative time is set up to compare and analyze the accuracy of the predicted landslide time, where the relative time is the difference between a certain time point and the deformation acceleration point, in hours, and the deformation acceleration point is time 0. After identifying the deformation acceleration point through the intersection of the short-term moving average (SMA) of velocity and the long-term moving average (LMA) of velocity, the predicted landslide time is obtained by using the predicted displacement data and applying the improved inverse velocity method.

[0133] The study of the above four landslides found that their common feature was that the acceleration continued to decrease before the landslide occurred. After the deformation acceleration point, the landslide velocity first increased and then slowly decreased. This phenomenon is inconsistent with the classical creep theory. Due to the continuous decrease in acceleration, the velocity inverse-time scatter points are not distributed on a straight line, but gradually approach the time axis in the form of a concave curve. a =2 is difficult to accurately predict landslide time. a =2, A =0.02) to compare the landslide time prediction with the actual landslide occurrence. It can be found that although the prediction accuracy of the traditional velocity inverse method is constantly improving with the increase of landslide monitoring data, the traditional velocity inverse method can only give a relatively accurate prediction when the landslide is approaching, such as Fig.13 shown.

[0134] When the rainfall-type red layer residual slope landslide enters the accelerated deformation stage, its displacement changes are somewhat sudden, and it is difficult to use the velocity inverse method to timely predict the landslide occurrence time. Analysis based on displacement prediction data can avoid dependence on future data and make landslide time predictions more timely. Its accuracy is comparable to Fig.14 shown.

[0135] In an exemplary embodiment, the process of establishing a rainfall-induced red layer residual slope accumulation landslide monitoring and early warning model is as follows.

[0136] In the early stage, the overall deformation amplitude of the rainfall-type red layer residual slope accumulation landslide is very small. After entering the accelerated deformation stage, its changes will increase rapidly in just a few minutes or even tens of seconds, showing strong suddenness and instantaneity. Therefore, it is likely that effective warning information cannot be issued in time using deformation indicators alone as early warning criteria. Comprehensively predicting the time of landslide, the difference between the predicted landslide time and the current time is used as the early warning indicator.

[0137] For emergencies, the warning levels of rainfall-type red layer residual slope accumulation landslide are divided into attention level, warning level, alert level and alarm level according to the danger, and the warning signals are represented by blue, yellow, orange and red respectively. The four warning levels are divided and determined as follows.

[0138] Warning level (blue): The deformation and destruction rate is relatively low at this stage, and the possibility of landslide damage in the short term is not high, so it is classified as a blue warning (long-term forecast).

[0139] Warning level (yellow): The rainfall-induced red layer residual slope landslide disaster deformation enters the early stage of accelerated deformation. At this time, the intersection of the short-term moving average of velocity (SMA) and the long-term moving average of velocity (LMA) can be used as an indicator to identify entering this stage, and it is set as a yellow warning (medium-term forecast).

[0140] Warning level (orange): indicates that the deformation has entered the middle stage of accelerated deformation; there is a high possibility of landslide damage within a few days or hours, and it is classified as orange warning (short-term forecast).

[0141] Alert level (red): indicates that the landslide deformation is about to enter the pre-slip stage. At this time, the precursor characteristics are already very obvious, and there is a high possibility of landslide damage in a very short period of time. It is classified as a red warning (pre-slip forecast).

[0142] The difference between the predicted landslide time and the current time Δ t and deformation rate v As early warning indicators, a landslide early warning model is established, as shown in Table 3.

[0143] Table 3 Landslide monitoring and early warning model based on early warning indicators

[0144]

[0145] The deformation rate of the red layer residual slope landslide is relatively small in the whole stage before the deformation acceleration, so the deformation rate is generally set to be relatively small before the deformation acceleration point is identified; after the deformation acceleration point is identified, because the deformation acceleration stage lasts for a short time, using days (d) as the warning unit may lead to a too large time scale, so consider using hours (h) as the warning unit. And through a large amount of analysis of the difference between the predicted landslide time and the current time and the landslide deformation rate data, the basis for the division of the warning indicators is obtained.

[0146] The warning logic of the early warning model is: when the deformation acceleration point is not identified, the deformation rate v Give warning when v <0.1mm / h, it is a blue warning, and when v ≥0.1mm / h, a yellow warning is issued; after the deformation acceleration point is identified, the difference between the predicted landslide time and the current time Δ t (Unit: hours) to give an early warning. t When the values ​​are between infinity, 72, 12, 6, and infinitesimal, respectively, blue, yellow, orange, and red warnings are issued. n monitoring time points (in this model n =5) When the warning level reaches other levels, the warning level will be upgraded or downgraded at that moment.

[0147] Different warning level results can respond to the cumulative displacement changes of the landslide, and can issue warnings more timely for points with obvious deformation. The warning results are compared and analyzed with the conventional warning results obtained by a geological disaster monitoring and early warning system. The warning method proposed in this application can achieve early warning of rainfall-type red layer residual slope accumulation landslides, and can issue warnings more timely than conventional warnings.

[0148] Based on the landslide monitoring data such as displacement and rainfall, this application establishes a rainfall-type red layer residual slope accumulation landslide displacement prediction model based on deep learning, which comprehensively considers rainfall factors, historical displacement data, landslide slope, and changes in physical and mechanical parameters of residual slope accumulation layers during rainfall infiltration; finally, based on the displacement prediction data of the landslide, according to a , A The power relationship between velocity and time is used to predict the landslide time by using the improved inverse velocity method, and a method based on the difference Δ between the predicted landslide time and the current time is constructed. t and landslide displacement rate v The rainfall-type red layer residual slope accumulation landslide monitoring and early warning model. This application has the following advantages.

[0149] (1) A feature extraction project for the displacement prediction model sample set was established. The project comprehensively considered 10 rainfall factor features based on historical average rainfall, previous effective rainfall, and number of rainy days, rainfall infiltration depth features based on landslide slope, landslide thickness, and test results, stability coefficient features, historical displacement data features, and time factor features, a total of 14 features; among which, the test results include permeability and shear strength. The results show that the 14 sample set features can effectively utilize the correlation with the accumulated displacement of the landslide, making the model prediction effect better.

[0150] (2) A landslide displacement prediction model based on Bi-LSTM and Informer deep learning methods considering feature engineering was constructed. In the landslide displacement prediction model based on Bi-LSTM and Informer, the displacement prediction data considering features such as rainfall are consistent with the original data, while the displacement prediction data without considering features such as rainfall are lower than the original data. This shows that the prediction model can learn the mechanism of rainfall on cumulative displacement to a certain extent, and considering features such as rainfall can obtain displacement prediction data more accurately. The prediction accuracy of the prediction model is high. By establishing a landslide displacement prediction model, future landslide monitoring data can be analyzed in advance. This method can be used to achieve advance warning and occurrence time prediction of such landslides.

[0151] (3) The landslide prediction method based on the traditional velocity inverse method was improved. By improving the velocity inverse method and statistically analyzing the velocity inverse-time data of the red layer residual slope landslide that has occurred, it was found that the velocity inverse method a The value increases with time, showing a trend of first increasing rapidly and then increasing slowly. A The increase of value over time is approximately linear, so a , A The growth relationship over time can be summarized as a power relationship ( y = mx n ), and based on this, we have gained experience in this type of landslide a , A Value: When t (h) is not less than 0.211, a = e 0.3396 × t 0.2178 , A =e -5.9525 × t 0.9010 , the speed inverse method can be improved according to this relationship.

[0152] (4) Improved the accuracy and timeliness of the prediction of the imminent sliding time of red layer residual slope landslide. The improved velocity inverse method proposed in this application can accurately predict the landslide time. With the increase of landslide monitoring data, the predicted landslide time is basically consistent with the actual landslide time. Similarly, by simulating rainfall disturbances through analysis based on displacement prediction data, the precursor information of landslide displacement changes can be obtained to a certain extent, and the prediction data can be used to predict the imminent sliding time more timely; when the traditional velocity inverse method is used to predict the landslide time, a relatively accurate prediction can only be given when the landslide is about to occur. The use of the improved velocity inverse method can effectively solve this problem, improving the timeliness and accuracy of the prediction of the imminent sliding time of rainfall-type red layer residual slope landslide.

[0153] (5) Improved the timeliness of early warning of red layer residual slope landslides. Through the landslide early warning model established in this application, the analysis of the example data of four typical landslides shows that the landslide monitoring and early warning method proposed in this application can achieve early warning of rainfall-type red layer residual slope landslides, and can issue early warnings more timely than conventional warnings.

[0154] In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the above method is implemented when the processor executes the computer program.

[0155] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, which implements the above method when executed by a processor.

[0156] In an exemplary embodiment, a computer program product is provided, including a computer program, which implements the above method when executed by a processor.

Claims

1. A landslide monitoring and early warning method, characterized in that: The landslide monitoring and early warning method comprises: Combined with the evolution law of landslide itself and the mechanism of rainfall on landslide deformation, feature engineering is performed on landslide monitoring data to determine the sample set characteristics of the rainfall-type red layer residual slope accumulation landslide cumulative displacement prediction model; the landslide monitoring data includes displacement data, landslide slope, rainfall, and physical and mechanical parameters of the residual slope accumulation layer during rainfall infiltration; the sample set characteristics include historical displacement data characteristics, rainfall factor characteristics, stability coefficient characteristics, rainfall infiltration depth characteristics, and time factor characteristics; Based on the sample set characteristics, a predicted displacement sequence is generated according to the rainfall-type red layer residual slope accumulation landslide cumulative displacement prediction model; The predicted landslide time is determined by using the predicted displacement sequence and the improved velocity inverse method, including: According to the curve between the inverse velocity and time of the landslide, the inverse velocity and time relationship is determined; the inverse velocity and time relationship is: ;in, V for speed; t f is the time when the landslide occurred, t For time; A and a All are empirical constants; The speed inverse-time relationship is processed logarithmically, and the logarithm-processed speed inverse-time relationship is converted into a linear equation; wherein the logarithm-processed speed inverse-time relationship is: ; Based on the linear equation, combined with the velocity inverse-time data, the a and A The power relationship of Determine according to the power relationship a and A The change curve over time determines the a and A The empirical value of ; m 1. n 1. m 2. n 2. p is a constant; According to the a and A The time-varying curve identifies the deformation acceleration point in the accelerated deformation stage; Based on the deformation acceleration point, the predicted displacement sequence is used and the improved velocity inverse method-time relationship is applied to determine the predicted landslide time; the same landslide at different deformation stages in the improved velocity inverse method a and A The value of is different. a and A The value of varies with the evolution of the deformation stage; The difference between the predicted landslide time and the current time is used as an early warning indicator, and a rainfall-type red layer residual slope accumulation landslide monitoring and early warning model based on the early warning indicator and the landslide displacement rate is constructed; Landslides are monitored and warned based on the rainfall-type red layer residual slope accumulation landslide monitoring and early warning model.

2. The landslide monitoring and early warning method according to claim 1 is characterized in that: The rainfall-type red layer residual slope accumulation landslide cumulative displacement prediction model is established by coupling a bidirectional long-short term memory network and an information extraction network.

3. The landslide monitoring and early warning method according to claim 2 is characterized in that: The bidirectional long short-term memory network includes three logic gates; the logic gates include a forget gate, an input gate, and an output gate; the forward propagation process of the three logic gates is: ； ； ； in, f t , i t , o t They are t Always keep in mind the state of the forget gate, input gate, and output gate; W f , W i , W o They are the weight matrices of the forget gate, input gate, and output gate respectively; b f , b i , b o They are the bias items of the forget gate, input gate and output gate respectively; W s is the input unit weight matrix; b s is the input unit state bias term; s t is a temporary state quantity; h t-1 and x t They are t -1 The intermediate output at time and the input at time constitute t Input of each door at all times; c t-1 and c t They are t -1 moment and t State memory variables at each moment; h t for t The intermediate output at the moment; ⊙ is the matrix element multiplication; tanh is the hyperbolic tangent activation function.

4. The landslide monitoring and early warning method according to claim 2 is characterized in that: The information extraction network includes an encoder, a decoder and a fully connected layer; the encoder and the decoder both include a position information embedding layer, a multi-head probabilistic sparse self-attention layer and a self-attention distillation layer; The encoder uses the linear transformation of sine and cosine functions to perform position encoding, and the encoding method of the encoder is: ； in, PE ( pos,j ) t For location POS and dimensions j The coded value on POS for t The time instant is a fixed position in the original time-coded input sequence, j is the dimension of the position encoding, j =1, 2, …, d model , d model Dimensions encoding positions; L x The length of the original time-encoded input sequence; i is the index of any dimension in the position encoding; is a positive integer; The attention mechanism of the multi-head probabilistic sparse self-attention layer is: ； in, is the attention weight, Q is the query vector, K is the key vector, V is a value vector; is the representation of the query vector; d K is the key vector K Dimensions; K T is the key vector K The transposed matrix of The self-attention distillation layer d Layer to d The calculation formula for +1 layer is: ； in, The time step length in the time series is t , the number of layers is d +1 layer value; ELU is the activation function; It is a multi-head probabilistic sparse self-attention mechanism With self-attention mechanism B Key operations: Convld It is a one-dimensional convolution operation on the time series; MaxPool It is the maximum pooling layer operation; The input calculation formula of the decoder is: ； in, X feed_decoder is the input sequence to be predicted; X token is the known sequence between the sequences to be predicted; X placeholder is the sequence to be predicted, filled with 0; L token is the length of the historically known sequence; L y is the length of the sequence to be predicted; R is the set of real numbers.

5. The landslide monitoring and early warning method according to claim 1 is characterized in that: Combined with the evolution law of landslide itself and the mechanism of rainfall on landslide deformation, feature engineering is performed on landslide monitoring data to determine the sample set characteristics of the cumulative displacement prediction model of rainfall-type red layer residual slope landslide, including: Collect historical displacement data features; Combined with the evolution law of landslides and the mechanism of rainfall on landslide deformation, rainfall factors were analyzed based on historical average rainfall, previous effective rainfall and number of rainy days, and the importance of decision tree algorithm was used to screen rainfall factors and determine the characteristics of rainfall factors. Determine the stability coefficient characteristics and rainfall infiltration depth characteristics based on the permeability coefficient, shear strength and landslide slope; Based on the rainfall data of different months, time characteristic analysis is performed to determine the time factor characteristics, and the time factor characteristics are classified according to the similarity between the time factor characteristics; the time factor characteristics include 12 months; the classification result of the time factor characteristics is that November, December, January, February and March are in the first category, April, May, June and October are in the second category, and July, August and September are in the third category; The sample set characteristics of the rainfall-type red layer residual slope landslide cumulative displacement prediction model are determined according to the historical displacement data characteristics, the rainfall factor characteristics, the stability coefficient characteristics, the rainfall infiltration depth characteristics and the time factor characteristics.

6. The landslide monitoring and early warning method according to claim 5 is characterized in that: Combined with the evolution law of landslide itself and the mechanism of rainfall on landslide deformation, rainfall factors are analyzed based on historical average rainfall, previous effective rainfall and number of rainy days, and the importance of decision tree algorithm is used to screen rainfall factors and determine the characteristics of rainfall factors, including: Combined with the evolution law of landslide itself and the mechanism of rainfall on landslide deformation, 32 rainfall factors are determined based on the historical average rainfall, the previous effective rainfall and the number of rainy days; the historical average rainfall includes 3h average rainfall, 6h average rainfall, 12h average rainfall, 1 day average rainfall, 2 days average rainfall, 4 days average rainfall, 6 days average rainfall, 8 days average rainfall, 10 days average rainfall, 12 days average rainfall and 14 days average rainfall, where h is hours; the previous effective rainfall includes the effective rainfall every day within 14 days; the number of rainy days includes the number of rainy days in every two days of the 7 previous periods; The importance of the decision tree algorithm was used to screen 32 rainfall factors, and the characteristics of the 10 rainfall factors after screening were determined.

7. The landslide monitoring and early warning method according to claim 6 is characterized in that: The importance of the decision tree algorithm was used to screen 32 rainfall factors, and the characteristics of the 10 rainfall factors after screening were determined, including: Importance of using decision tree algorithm 32 rainfall factors were screened and the characteristics of 10 rainfall factors were determined; among them, is the feature in the decision tree algorithm F Gini impurity of ; is the feature in the decision tree algorithm i Gini impurity of ; N is the total number of samples of the node before splitting.

8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the landslide monitoring and early warning method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the landslide monitoring and early warning method described in any one of claims 1 to 7 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the landslide monitoring and early warning method described in any one of claims 1 to 7 is implemented.

Images