CN114358192B - Multi-source heterogeneous landslide data monitoring and fusing method - Google Patents

Abstract

The invention discloses a multi-source heterogeneous landslide data monitoring and fusing method, which comprises the following steps: obtaining a weighted correlation degree by combining a maximum mutual information coefficient method MIC and a gray correlation analysis method GRA to reflect the influence of multisource heterogeneous landslide monitoring variables on landslide deformation displacement and a common change trend, further carrying out optimization on characteristic factors according to the weighted correlation degree, carrying out stepwise regression fitting analysis on the optimally obtained characteristic factors to obtain corresponding regression coefficients, further calculating to obtain a final regression equation and a fusion result, and carrying out reliability and effectiveness evaluation on the data fusion result by adopting a landslide stage judgment and trend prediction method. The method adopts a multi-source heterogeneous data fusion means to carry out multivariate data effective analysis and processing so as to obtain a more reliable and accurate data fusion result and provide valuable reference information for landslide prediction, thereby effectively improving the precision of landslide prediction.

Description

A Multi-source Heterogeneous Landslide Data Monitoring Fusion Method

technical field

The invention relates to the technical field of data fusion, in particular to a multi-source heterogeneous landslide data monitoring fusion method.

Background technique

As a new and multi-field interdisciplinary subject, multi-source data fusion technology has been widely used in landslide deformation monitoring after more than ten years of exploration and development. With the emergence of many sensors, how to extract information from multi-source heterogeneous sensors and carry out effective fusion processing is the current research difficulty. In landslide monitoring, a single sensor information cannot fully reflect the deformation characteristics of landslides, and the reliability of the prediction results obtained is not high, it is necessary to combine multi-source heterogeneous sensor information for effective feature extraction and comprehensive analysis and processing to eliminate The redundancy and mutual exclusivity among the data can obtain more reliable and accurate prediction results.

At present, the effective feature extraction and comprehensive analysis and processing of source heterogeneous sensor information have the problem of one-sidedness and unreliability of single landslide monitoring information, which makes the prediction inaccurate.

Contents of the invention

An embodiment of the present invention provides a multi-source heterogeneous landslide data monitoring and fusion method, including:

Obtain multi-source heterogeneous monitoring variable data;

Divide multi-source heterogeneous monitoring variables into dependent variables and characteristic variables;

Calculate the maximum mutual information coefficient MIC of every two multi-source heterogeneous landslide monitoring variables, and screen out the characteristic variables that have the greatest impact on landslide deformation;

Determine the single-point displacement sequence that reflects the deformation characteristics of the landslide as the reference column, and the data sequence composed of factors that affect the landslide deformation is the comparison column;

Calculate the gray correlation coefficient and gray correlation degree of the reference sequence and the comparison sequence;

Calculate the weighted correlation degree according to the maximum mutual information coefficient MIC and the gray correlation degree;

Feature selection is performed according to the weighted correlation degree, and the final feature variable is screened out;

Construct feature selection based on weighted correlation degree-stepwise regression feature-level data fusion model;

Using the weighted correlation degree-based feature selection-stepwise regression feature-level data fusion model to carry out multi-source heterogeneous information fusion to provide effective auxiliary information for landslide prediction and forecasting.

One step further, it also includes data preprocessing for multi-source heterogeneous monitoring variables:

Outlier elimination, missing value completion and data smoothing and denoising.

Further, the step of calculating the maximum mutual information coefficient MIC of each two multi-source heterogeneous landslide monitoring variables includes:

Given the variables i and j, grid the scatter diagram composed of the two variables in column i and row j, and find the maximum mutual information value;

Normalize the maximum mutual information value;

Select the maximum value of mutual information at different scales as the MIC value;

Get the feature variable with the highest degree of correlation with the dependent variable.

One step closer, the gray correlation coefficient, the calculation formula includes:

与

分别表示两级最小差和两级最大差。Among them, ρ is the resolution coefficient, 0<ρ<1, if the smaller ρ, the greater the difference between the correlation coefficients, the stronger the ability to distinguish, usually ρ is 0.5, |x ₀ (k) _-xi (k)| The absolute difference between the comparison sequence and the corresponding elements of the reference sequence,

and

Respectively represent the two-level minimum difference and the two-level maximum difference.

One step further, the weighted correlation degree, the calculation formula includes:

式中n为待选择特征变量的总数，MIC(A，B_i)表示特征变量A和特征变量B_i的最大互信息系数MIC。

In the formula, n is the total number of feature variables to be selected, and MIC(A, B _i ) represents the maximum mutual information coefficient MIC between feature variable A and feature variable B _i .

One step closer, the steps of feature optimization include:

Sort the calculated weighted correlation degrees from large to small;

The characteristic variables are sorted and screened according to the weighted correlation degree;

Calculate the weight of each preferred feature after sorting;

时，筛选停止，得到最终的特征变量；When the optimal feature weight

When , the screening stops and the final feature variable is obtained;

Among them, J _S is the sum of the weighted correlation degrees of each feature variable, J _j is the weighted correlation degree of the jth feature variable to be screened, ω _j is the jth optimal feature weight, and α is a given threshold.

In a further step, it also includes: analyzing and comparing the fusion results of feature optimization based on weighted correlation degree-stepwise regression and the fusion results of BP neural network;

Establish a BP neural network fusion model, take the independent variable as the system input variable, and the dependent variable as the system output variable;

Establish a multi-input single-output BP neural network fusion model with two hidden layers;

Using the two indicators of improved tangent angle and deformation rate, the feature selection based on weighted correlation degree-stepwise regression fusion and BP neural network data fusion model are evaluated in stages;

The long-short-term memory network artificial neural network LSTM is used to perform prediction comparison analysis based on feature selection-stepwise regression fusion data and BP neural network fusion data based on weighted correlation degree respectively.

The embodiment of the present invention provides a multi-source heterogeneous landslide data monitoring fusion method, compared with the prior art, its beneficial effects are as follows:

1. MIC is used to measure the degree of correlation between two variables. Compared with other correlation analysis methods, MIC is not only suitable for linear and nonlinear data, but also has universality, fairness and symmetry. high accuracy.

2. Combine mutual information weight and gray relational degree, use weighted relational degree to measure the importance of characteristic factors to landslide deformation, and calculate optimal characteristic weight, filter out characteristic factors according to threshold, combine mutual information and gray relational characteristics to Perform feature optimization to make the result of feature optimization more reliable.

Description of drawings

Fig. 1 is a flow chart of feature optimization based on weighted correlation degree in the fusion method of the present invention;

Fig. 2 is the RNN model figure in the evaluation analysis of the present invention;

Fig. 3 is the RNN model hidden layer cell structure in the evaluation analysis of the present invention;

Fig. 4 is the cell structure of the hidden layer of the LSTM model in the evaluation analysis of the present invention;

Fig. 5 is the distribution diagram of monitoring points in the experimental research area of the present invention;

Fig. 6 is the feature selection-stepwise regression fusion result based on the weighted correlation degree in the experimental part of the present invention;

Fig. 7 is the fusion result of the BP neural network model of the experiment part of the present invention.

Detailed ways

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

Referring to Figures 1-7, an embodiment of the present invention provides a multi-source heterogeneous landslide data monitoring and fusion method, the method comprising:

S1. Maximum mutual information calculation. This patent adopts the minepy class library in the python program software to realize the MIC algorithm. It is mainly divided into three steps: 1) Given i and j, grid the scatter diagram composed of two variables in column i and row j, and calculate the maximum mutual information value; Normalization processing; 3) Select the maximum value of mutual information at different scales as the MIC value; 4) Get the feature variable with the highest degree of correlation with the dependent variable for subsequent regression prediction. MIC is used to measure the degree of association between two variables, and compared with other association analysis methods, MIC is not only suitable for linear and nonlinear data, but also has universality, fairness and symmetry, and has a high Accuracy.

S2. Calculation of gray relational degree. The single-point displacement sequence that can reflect the deformation characteristics of the landslide is determined as the reference sequence, and the data sequence composed of factors that affect the landslide deformation is the comparison sequence, and the reference sequence and the comparison sequence are dimensionless, and then the reference sequence and the comparison sequence are obtained. Gray correlation coefficient and gray correlation degree.

S3. Optimizing weighted correlation features. Mutual information measures the influence of features on landslide deformation, its weight reflects the effectiveness of features, and the gray correlation degree quantifies the consistency between features and landslide deformation. Combining mutual information weight and gray relational degree, the weighted relational degree is used to measure the importance of characteristic factors to landslide deformation, and the optimal characteristic weight is calculated, and the characteristic factors are screened out according to the threshold. Combining the characteristics of mutual information and gray correlation for feature selection makes the result of feature selection more reliable.

S4. Stepwise regression analysis. The characteristic factors are introduced into the model one by one, the F test is carried out after each explanatory variable is introduced, and the t test is carried out for the explanatory variables that have been selected one by one. When the introduction of the explanatory variables originally introduced becomes no longer significant, the delete. To ensure that the regression equation before each new variable introduced contains only significant variables. Repeat until no significant explanatory variables are selected into the regression equation, and no insignificant explanatory variables are removed from the regression equation.

S5. Evaluation and analysis. The first step is to compare the fusion results. In order to evaluate the reliability of feature selection based on weighted correlation degree-stepwise regression feature-level data fusion, the fusion results of feature optimization based on weighted correlation degree-stepwise regression and BP neural network fusion results were used to conduct stage discriminant analysis and comparison. First, a BP neural network fusion model is established, with the independent variable as the system input variable and the dependent variable as the system output variable, and a multi-input single-output BP neural network fusion model with two hidden layers is established. And use the two indicators of improved tangent angle and deformation rate to carry out stage evaluation on the feature selection based on weighted correlation degree-stepwise regression fusion and BP neural network data fusion model, to prove the effectiveness of the fusion results of the present invention. The second step is to predict and compare the fusion results. LSTM (long-short-term memory network artificial neural network) is used to predict and compare the single-point data of the dependent variable of GNSS monitoring points, feature selection based on weighted correlation degree-stepwise regression fusion data and BP neural network fusion data. The Keras in the python program is used to build the LSTM model, and two indicators of MRE (mean relative error) and MAE (mean absolute error) are used to evaluate the prediction results to prove the reliability of the fusion results of the present invention.

specifically:

(其中n_x,y为落在第(x,y)格子中的数据点数，n为总的数据点数)，同理获得P(x)、P(y)的估计。然后计算随机变量间的互信息，因为m乘以n的网格划分数据点的方式不止一种，所以要获得使互信息最大的网格划分，并使用归一化因子，将互信息的值转化到(0,1)区间内。最后，找到能使归一化互信息最大的网格分辨率，作为MIC的度量值。其中网格的分辨率限制为m×n<B,B＝f(data_size)＝n^0.6，MIC计算公式为

S1. Maximum mutual information calculation. The multi-source heterogeneous sensors on the landslide are correlated, and one sensor will be affected by multiple other sensors comprehensively. Therefore, it is necessary to calculate the correlation degree of the multi-source heterogeneous landslide monitoring variables and screen out the ones that most affect the landslide deformation. The eigenfactors are used for subsequent fusion predictions. Before performing MIC (Maximum Mutual Information Coefficient) calculation, it is first necessary to understand information entropy and mutual information. Mutual information refers to the degree of correlation between two random variables, and the average amount of information after the redundancy is excluded from the information is called "information entropy", then the method of MI (mutual information) can be used to screen out the influence of landslide deformation. eigenfactors with less redundant information. Compared with MI, MIC has higher accuracy, and is not limited to a specific function type, so the degree of correlation between variables can be obtained. If there is an association between two variables, the sets of their corresponding data points are distributed in a two-dimensional space, and the data space is divided by a grid of m times n, so that the frequency of data points falling in the (x, y) grid As an estimate of P(x,y),

(where n _{x, y} are the number of data points falling in the (x, y) grid, and n is the total number of data points), similarly obtain the estimates of P(x) and P(y). Then calculate the mutual information between random variables, because there are more than one way to divide the data points into the grid of m times n, so it is necessary to obtain the grid division that maximizes the mutual information, and use the normalization factor to convert the value of the mutual information Convert to (0,1) interval. Finally, find the grid resolution that maximizes the normalized mutual information as a measure of MIC. The resolution of the grid is limited to m×n<B, B=f(data_size)=n ^0.6 , and the MIC calculation formula is

The specific calculation steps are as follows: 1) Calculate the maximum mutual information value. Given i and j, grid the scatter diagram composed of two variables X and Y with i columns and j rows, and find the maximum mutual information value. However, given i and j, multiple different gridding schemes can be obtained, and the mutual information value corresponding to each scheme needs to be calculated to find the gridding scheme that maximizes the mutual information. 2) Normalize the largest mutual information value. Dividing the obtained maximum mutual information by log(min(X,Y)) is normalization. 3) Select the maximum value of mutual information at different scales as the MIC value. Then select the features that have a greater impact on the landslide deformation, and eliminate the features with less information, so that the variables used for modeling are more representative.

无量纲化后的数据序列形成如下矩阵：

3)逐个计算每个被评价对象指标序列(比较序列)与参考序列对应元素的绝对差值，即|x₀(k)-x_i(k)|(k＝1,…,m；i＝1,…,n,)n为被评价对象的个数。4)确定

与

5)计算关联系数。关联系数计算公式为：

其中ρ为分辨系数，0<ρ<1，若ρ越小，关联系数间差异越大，区分能力越强，通常ρ取0.5。6)计算关联度。对各评价对象(比较序列)分别计算其个指标与参考序列对应元素的关联系数的均值，以反映各评价对象与参考序列的关联关系，并称为关联度，记为：

S2. Calculation of gray relational degree. It is not convincing to use only MIC results for feature selection. The MIC analysis representing the degree of influence and the gray correlation analysis representing the degree of consistency can be combined for comprehensive analysis to obtain feature factors that are more suitable for data fusion. Through the non-dimensional processing of the multi-source heterogeneous landslide monitoring feature sequence, the correlation coefficient and correlation degree are calculated. The specific process is: 1) Determine the reference column sequence and compare the column sequence. Let the reference column sequence be Y={Y(k)|k=1,2,…n}; the comparison column sequence be X _i ={X _i (k)|k=1,2,…,n}, i= 1,2,...,m. 2) Dimensionalization of variables. Since the different dimensions of the characteristic factors are not convenient for comparison, it is necessary to perform dimensionless processing.

The dimensionless data sequence forms the following matrix:

3) Calculate the absolute difference between each index sequence (comparison sequence) of the evaluated object and the corresponding element of the reference sequence one by one, namely |x ₀ (k) _-xi (k)|(k=1,...,m; i= 1,...,n,) n is the number of evaluated objects. 4) OK

and

5) Calculate the correlation coefficient. The formula for calculating the correlation coefficient is:

Among them, ρ is the resolution coefficient, 0<ρ<1. If ρ is smaller, the difference between the correlation coefficients is greater and the discrimination ability is stronger. Usually, ρ is set to 0.5. 6) Calculate the correlation degree. For each evaluation object (comparison sequence), calculate the mean value of the correlation coefficient between each index and the corresponding element of the reference sequence to reflect the correlation between each evaluation object and the reference sequence, and it is called the correlation degree, which is recorded as:

式中n为待选择特征的总数。S3. Calculating the weighted correlation degree. Combining the maximum mutual information weight with the gray correlation degree, the weighted correlation degree of the corresponding feature is obtained to reflect the feature. The greater the weighted correlation degree, the more important the feature is, and the calculation formula is:

where n is the total number of features to be selected.

式中J_S为优选集合中各特征的加权关联度之和，J_j为第j个待筛选特征的加权关联度，ω_j为第j个优选特征权重，该值小于α时，认为特征筛选完毕。S4. Feature optimization. The calculated weighted correlations are sorted from large to small, and the feature with the largest weighted correlation is selected to be added to the optimized set, and removed from the to-be-optimized set. Filter in order from large to small, and calculate the weight of the preferred feature:

In the formula, J _S is the sum of the weighted correlation degrees of each feature in the optimal set, J _j is the weighted correlation degree of the jth feature to be screened, and ω _j is the weight of the jth optimal feature. When the value is less than α, it is considered that the feature screening complete.

S5. Stepwise regression analysis. The basic idea of stepwise regression is to introduce variables into the model one by one, perform F-test after each explanatory variable is introduced, and perform t-test on the explanatory variables that have been selected one by one, when the introduction of the original explanatory variables becomes no longer significant , delete it. To ensure that the regression equation before each new variable introduced contains only significant variables. This is an iterative process until neither significant explanatory variables are selected into the regression equation nor insignificant explanatory variables are removed from the regression equation. The specific steps of stepwise regression are as follows:

Step 1: Build an augmented matrix

其中

即可得到扩充了的增广矩阵

其中R＝(r_ij)_m×m，r_yy＝1，r_y＝(r_1y,r_2y,…,r_my)'。Calculate l _ij , l _iy , l _yy and r _ij , r _iy , the formulas are:

in

The extended augmented matrix can be obtained

Where R=(r _ij ) _m×m , r _yy =1, r _y =(r _1y , r _2y , . . . , r _my )'.

其中Step 2: Perform elimination transformation on step s, the result is

in

The third step: factor elimination.

②计算

③若F>F_出，则执行第四步；反之，则进行s+1次消去变换，然后转入二、三两步进行计算。①Choose j ₀ so that

② calculation

③ If F>F _{is out} , then execute the fourth step; otherwise, perform s+1 times of elimination transformation, and then transfer to the second and third steps for calculation.

Step 4: Introduce regression factors. Suppose s, {j}, f are still defined in step 2.

②计算

其中

③若F<F_进，则执行第五步；反之，则进行s+1次消去变换，引入第k₀个回归因子，然后转入二、三、四步进行计算。①Choose k ₀ such that

② calculation

in

③ If F<F _advance , then execute the fifth step; otherwise, perform s+1 times of elimination transformation, introduce the k _0th regression factor, and then transfer to the second, third, and fourth steps for calculation.

In the fifth step, variables cannot be introduced or eliminated at this time. The final regression equation is ^y=^b ₀ +∑ _j∈{j} ^b _j x _j , where

S6. Evaluation and analysis.

式中

为第l层第i个节点的输出值；

第l层第i个节点的激活值；

为第l-1层第j个节点到第l层第i个节点的连接权值；

为第l层第i个节点的阈值；N_l为第l层节点数。为了提高输出层神经元的误差精度，采用梯度递降算法进行反向误差传播。通过梯度递降算法对每一层神经元之间的连接权重进行调整，使最终的总体误差会沿减少方向改变。其算法公式为：

(η为学习率)，权值调整公式为

The first step is to compare the fusion results. In order to evaluate the reliability of feature selection based on weighted correlation degree-stepwise regression feature-level data fusion, the fusion results of this model and the fusion results of BP neural network were used to conduct stage discriminant analysis and comparison. First, a BP neural network fusion model is established, with the independent variable as the system input variable and the dependent variable as the system output variable, and a multi-input single-output BP neural network fusion model with two hidden layers is established. The topology of BP neural network includes three parts: input layer, hidden layer and output layer, and also includes two processes: forward multi-layer feedforward stage and reverse error correction stage. The forward multi-layer feed-forward stage is a forward process to calculate the actual input and output of each node in each layer sequentially from the input layer, and the mathematical model is

In the formula

is the output value of the i-th node in layer l;

The activation value of the i-th node in layer l;

is the connection weight from the jth node of layer l-1 to the ith node of layer l;

is the threshold of the i-th node in layer l; N _l is the number of nodes in layer l. In order to improve the error accuracy of neurons in the output layer, a gradient descent algorithm is used for reverse error propagation. The connection weights between neurons in each layer are adjusted through the gradient descent algorithm, so that the final overall error will change in the direction of reduction. Its algorithmic formula is:

(η is the learning rate), the weight adjustment formula is

And using the improved tangent angle (refer to Xu Qiang, an improved tangent angle and the corresponding landslide early warning criterion) and deformation rate to optimize the feature based on the weighted joint degree-stepwise regression fusion model and BP neural network data fusion The model is evaluated in stages, as shown in the table based on the deformation rate threshold and the comprehensive early warning basis of the deformation process.

Table 2 Comprehensive early warning criteria based on deformation rate threshold and deformation process

The second step is to predict and compare the fusion results. Feature-level fusion is an effective analysis and processing of multi-source heterogeneous information obtained from landslide monitoring. This will improve the accuracy of the forecast. In order to discuss the effectiveness of feature-level fusion in improving the accuracy of landslide prediction, LSTM (long-term short-term memory network artificial neural network) is used to analyze the single-point data of GNSS monitoring point dependent variables, feature selection based on weighted joint degree-stepwise regression fusion The data is fused with the BP neural network for forecasting and comparative analysis. The LSTM model is built using Keras in the python program. The LSTM neural network model is a new type of neural network algorithm based on the improvement of the ordinary cyclic neural network. In the LSTM model, the RNN cells in the hidden layer are replaced by LSTM cells, which can effectively overcome The problem that gradients can quickly disappear during backpropagation gives it a long-term memory that can handle long-term series data. Compared with the RNN model, 3 gating switches are set inside the LSTM unit, as shown in Figure 3, where i is the input gate, f is the forgetting gate, c is the cell state, o is the output, and σ and tanh are Sigmoid respectively and hyperbolic tangent activation functions.

分别为损失函数l()的一阶导数和二阶导数，最终得到的目标函数为：

W和b分别对应的权重系数矩阵和偏置项。(2)反向计算每个LSTM细胞的误差项，包括按时间和网络层级2个反向传播方向。(3)根据相应的误差项，计算每个权重的梯度。(4)应用基于梯度的优化算法更新权重。The forget gate uses the Sigmoid unit to output a vector between 0 and 1 by looking at the information of h _t-1 and x _t . The 0-1 value in the vector indicates which information in the cell state c _t-1 is retained or discarded . 0 means not reserved, 1 means reserved. f _t =σ(W _f ·[h _t-1 ,x _t ]+b _f ). The input gate is used to update the cell state. First input the information of the previous hidden state and the current input information into the Sigmoid function, adjust the output value between 0 and 1 to decide which information to update, 0 means unimportant, 1 means important. At the same time, the hidden state and current input are transmitted to the tanh function, and the value is compressed between -1 and 1 to adjust the network, and then the tanh output is multiplied by the Sigmoid output, and the Sigmoid output will determine which information in the tanh output is important and need to be reserved. i _t =σ[W _f ·[h _t-1 , x _t ]+b _i ],~C=tanh(W _C ·[h _t-1 ,x _t ]+b _C ). The output gate controls the value of the next hidden state, which can be used for prediction. First pass the previous hidden state and current input to the Sigmoid function, and at the same time pass the newly obtained unit state to the tanh function, then multiply the tanh output and the Sigmoid output to obtain new information about the hidden state as the output value of the current unit output; finally the new cell state and hidden state are synchronized to the next time step. o _t =σ[W _o ·[h _t−1 ,x _t ]+b _o ], h _t =o _t *tanh(C _t ). The LSTM model training process adopts the classic backpropagation algorithm, which is divided into four steps: (1) calculate the output value of the LSTM cell according to the forward calculation method,

are the first and second derivatives of the loss function l() respectively, and the final objective function is:

W and b correspond to the weight coefficient matrix and bias term respectively. (2) Reversely calculate the error term of each LSTM cell, including 2 backpropagation directions according to time and network level. (3) Calculate the gradient of each weight according to the corresponding error term. (4) Apply a gradient-based optimization algorithm to update the weights.

检测期望值与实际值之间的距离大小，对预测精度进行衡量。Two indicators, MRE (Mean Relative Error) and MAE (Mean Absolute Error), were used to evaluate the prediction accuracy.

Detect the distance between the expected value and the actual value to measure the prediction accuracy.

Example:

The experimental data of the present invention adopts the landslide monitoring data from March 28 to October 4, 2019 in Heifangtai Dangchuan 7# landslide, Yongjing County, Gansu Province, with the sampling rate of days as the sampling rate, a total of 191 days, including two sets of GNSS Monitoring data (HF06, HF07), three sets of displacement meter monitoring data (DCF11, DCF14, DCF15) and three meteorological data of humidity, temperature and rainfall. The landslide occurred at 4:00 on October 5, 2019. All five groups of monitoring equipment monitored the displacement change data of the landslide deformation, and multi-source heterogeneous sensors can be used for data fusion and accuracy determination. The distribution of monitoring points in the experimental area is shown in Figure 5.

The present invention first performs data preprocessing on multi-sensor variables and environmental factors. The preprocessing includes abnormal value elimination, missing value completion, and data smoothing and denoising. Firstly, MIC (maximum mutual information) calculation is performed on the preprocessed data to determine Mutual information weights are used to obtain the eigenfactors that have the greatest impact on landslide deformation. Then calculate the gray relational degree to obtain the gray relational degree value, and finally use the weighted relational degree formula to obtain the weighted relational degree value, and determine the final feature factor by calculating the feature optimization weight. Table 3 shows the MIC weight, gray correlation degree and weighted correlation degree results of GNSS monitoring point HF06 and other GNSS monitoring data, displacement meter monitoring data, rainfall, temperature, and humidity. The greater the weighted correlation degree, the higher the impact and closeness of the feature on landslide deformation.

Table 3 Weighted association calculation table

Sort the results obtained by the weighted correlation method that combines MIC and gray correlation: GNSS monitoring point HF07 > displacement meter DCF11 data > displacement meter DCF14 data > displacement meter DCF15 data > cumulative rainfall in the previous 48 hours > humidity > temperature > rainfall Quantities, that is, displacement sensor monitoring data, GNSS monitoring data, cumulative rainfall data in the first 48 hours, temperature, and humidity data have a greater impact on landslides. Table 4 shows the feature selection weight size calculated according to the weighted correlation degree, and the feature selection is carried out through this value.

Table 4 Feature Optimization Results

Sort the weighted correlation degree and calculate the optimal feature weight, select the threshold α as 0.1, that is, when the feature weight is less than 0.1, it is considered that the impact of the feature on landslide deformation can be ignored, and the selection of feature factors is completed. The factors affecting landslide deformation obtained by optimizing the weighted correlation degree were subjected to stepwise regression fitting analysis. According to the analysis results, the GNSS monitoring point HF07 data, displacement meter monitoring data, accumulated rainfall data in the previous 48 hours, temperature, and humidity were used as independent variables. The GNSS monitoring point HF06 data is used as the dependent variable for stepwise regression analysis to obtain the corresponding regression coefficients and then calculate the final features and fusion results. And in the stepwise regression analysis, the optimal result of the model is obtained by comparing and analyzing the correlation coefficient, residual variance, F value, and significance of different models. The obtained regression coefficients are shown in Table 5.

Table 5 regression coefficient table

The stepwise regression model expression of the surface displacement of the landslide body at this place is obtained: surface displacement of the landslide body=(GNSS monitoring point HF07 data×0.587)+(displacement meter DCF11 data×0.036)+(displacement meter DCF14 data×0.519)-(displacement meter DCF15 data × 0.159) + (temperature × 0.028) + (humidity × 0.026) - (cumulative rainfall in the first 48 hours × 0.010), and then get the feature selection based on the weighted correlation degree - stepwise regression feature level fusion results, as shown in Figure 6 Show.

The stage discriminant analysis was carried out to compare the fusion results of feature selection based on weighted correlation degree-stepwise regression and the fusion results of BP neural network. Firstly, the BP neural network fusion model is established, and the GNSS monitoring point HF07 data, the displacement meter monitoring data, the accumulated rainfall in the previous 48 hours, temperature, humidity, etc. are used as the input data of the BP neural network model, and the GNSS monitoring point HF06 data is used as the expected For the output data, refer to the MIC analysis results to build a BP neural network fusion model with two hidden layers of multiple inputs and single output. After experimental analysis, the feature-level fusion results of BP neural network are obtained, as shown in Figure 7. It can be seen from the literature that when the tangent angle is greater than 80°, the landslide is already in the medium-acceleration stage. In this experiment, we only compare the two fusion results for the tangent angle before sliding. The comparison results of the tangent angle analysis are shown in Table 6. , the deformation rate analysis and comparison results are shown in Table 7.

Table 6 Two kinds of fusion results improve the tangent angle analysis results

Table 7 Deformation rate analysis results of two fusion results

It can be seen from the comparison of the two stage indicators that the improved tangent angle obtained by feature selection-stepwise regression based on the weighted correlation degree is closer to the moment of landslide instability, and it is more suitable for the real development state of the landslide when used for landslide stage discrimination. Accordingly, it can be shown that the fusion result of feature selection based on weighted correlation degree-stepwise regression has good reliability and accuracy in landslide stage discriminant analysis, and the fusion result is better.

Then, the LSTM network algorithm is used to conduct a comparative analysis of the landslide trend prediction on the GNSS monitoring point HF06 data, the feature selection based on the weighted correlation degree-stepwise regression fusion data and the BP neural network fusion data, and use MRE and MAE two accuracy evaluation indicators to carry out the analysis. Comparison of prediction accuracy.

Table 8 Comparison of prediction accuracy of two fusion results

It can be concluded from Table 8 that the MAE and MRE predicted by the feature selection based on the weighted correlation degree-stepwise regression fusion results are 9.9mm and 3.46%, respectively, and the MAE and MRE predicted by the BP neural network fusion results are 15.1mm and 4.33%, respectively. The MAE and MRE predicted by GNSS monitoring point HF06 data are 19.9mm and 4.51%, respectively. That is, the prediction accuracy of the feature selection-stepwise regression fusion result based on the weighted correlation degree is higher, which also proves that the feature-level fusion result is more accurate and reliable.

The above disclosures are only a few specific embodiments of the present invention. Those skilled in the art can make various changes and modifications to the embodiments of the present invention without departing from the spirit and scope of the present invention. However, the embodiments of the present invention are not limited thereto , any changes conceivable by those skilled in the art should fall within the protection scope of the present invention.

Claims (5)

1. A multi-source heterogeneous landslide data monitoring fusion method is characterized in that, comprising: Obtain multi-source heterogeneous monitoring variable data; Divide multi-source heterogeneous monitoring variables into dependent variables and characteristic variables; Calculate the maximum mutual information coefficient MIC of the dependent variable and the characteristic variable, and screen out the characteristic variable that most affects the landslide deformation; Determine the single-point displacement sequence that reflects the deformation characteristics of the landslide as the reference column, and the data sequence composed of factors that affect the landslide deformation is the comparison column; Calculate the gray correlation coefficient and gray correlation degree between the reference sequence and the comparative sequence; Calculate the weighted correlation degree according to the maximum mutual information coefficient MIC and the gray correlation degree; Feature selection is performed according to the weighted correlation degree, and the final feature variable is screened out; Perform stepwise regression fitting analysis on the characteristic variables obtained by optimization; Construct feature selection based on weighted correlation degree-stepwise regression feature-level data fusion model; Using the feature selection-stepwise regression feature-level data fusion model based on weighted correlation degree to carry out multi-source heterogeneous information fusion to provide effective auxiliary information for landslide prediction and forecasting; Described gray correlation coefficient, computing formula comprises:

Among them, ρ is the resolution coefficient, 0<ρ<1, if the smaller ρ, the greater the difference between the correlation coefficients, the stronger the ability to distinguish, usually ρ is 0.5, |x ₀ (k) _-xi (k)| The absolute difference between the comparison sequence and the corresponding elements of the reference sequence,

and

respectively represent the two-level minimum difference and the two-level maximum difference; n is the number of evaluated objects; Calculation of correlation degree: For each evaluation object, calculate the average value of the correlation coefficient between each index and the corresponding element of the reference sequence to reflect the correlation relationship between each evaluation object and the reference sequence, and it is called the correlation degree, which is recorded as:

The calculation formula of the weighted correlation degree includes:

where n is the total number of feature variables to be selected, and MIC(A, B _i ) represents the maximum mutual information coefficient MIC between feature variable A and feature variable B _i ; The preferred steps of the features include: Sort the calculated weighted correlation degrees from large to small; The characteristic variables are sorted and screened according to the weighted correlation degree; Calculate the weight of each preferred feature after sorting; When the optimal feature weight

When , the screening stops and the final feature variable is obtained; Among them, J _S is the sum of the weighted correlation degrees of each feature variable, J _j is the weighted correlation degree of the jth feature variable to be screened, ω _j is the jth optimal feature weight, and α is a given threshold; The stepwise regression fitting analysis includes: introducing the characteristic factors into the model one by one, carrying out the F test after each explanatory variable is introduced, and carrying out the t test to the explanatory variables which have been selected one by one, when the introduction of the explanatory variables originally introduced changes When it is no longer significant, delete it to ensure that the regression equation before each new variable introduced only contains significant variables; iteratively execute until no significant explanatory variables are selected into the regression equation, and there are no insignificant explanations Variables are removed from the regression equation. 2. a kind of multi-source heterogeneous landslide data monitoring fusion method as claimed in claim 1, is characterized in that, also comprises to multi-source heterogeneous monitoring variable data preprocessing: Outlier elimination, missing value completion and data smoothing and denoising. 3. a kind of multi-source heterogeneous landslide data monitoring fusion method as claimed in claim 1, is characterized in that, the step of the maximum mutual information coefficient MIC of described calculation dependent variable and characteristic variable, comprises: Given the variables i and j, grid the scatter diagram composed of the two variables in column i and row j, and find the maximum mutual information value; Normalize the maximum mutual information value; Select the maximum value of mutual information at different scales as the MIC value; Get the feature variable with the highest degree of correlation with the dependent variable. 4. a kind of multi-source heterogeneous landslide data monitoring fusion method as claimed in claim 1, is characterized in that, also comprises: To the feature optimization-stepwise regression fusion result based on weighted correlation degree and BP neural network fusion result are analyzed and compared , which includes: Establish a BP neural network fusion model, take the independent variable as the system input variable, and the dependent variable as the system output variable; Establish a multi-input single-output BP neural network fusion model with two hidden layers; The two indicators of improved tangent angle and deformation rate are used to evaluate the feature selection based on weighted correlation degree-stepwise regression fusion and BP neural network data fusion model. 5. a kind of multi-source heterogeneous landslide data monitoring fusion method as claimed in claim 4, is characterized in that, also comprises: The long-short-term memory network artificial neural network LSTM is used to perform prediction comparison analysis based on feature selection-stepwise regression fusion data and BP neural network fusion data based on weighted correlation degree respectively.

Images