MV-BP neural network landslide prediction method integrating multi-source data
Through the MV-BP neural network method of integrating multi-source data, the problems of large manpower and material consumption and accuracy deviation in landslide monitoring and early warning are solved, and more accurate landslide prediction is achieved and prediction accuracy is improved.
Patent Information
- Application Number
- CN202410959608.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-17
- Publication Date
- 2024-11-15
AI Technical Summary
Existing landslide monitoring and early warning methods consume a lot of manpower and material resources, the measurement accuracy is biased, and due to the complexity and uncertainty of landslides, it is difficult to accurately predict various influencing factors.
The MV-BP neural network method with integrated multi-source data is adopted to obtain the full-factor input data of landslide deformation displacement, perform intelligent preprocessing, establish an MV-BP neural network model, and predict landslides to improve the accuracy of prediction.
More accurate surface displacement change data is achieved, the accuracy of landslide prediction is improved, manpower and material investment is reduced, and the reliability of prediction is enhanced.
Smart Images

Figure CN118964973A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of geological disaster monitoring and early warning, and in particular relates to an MV-BP neural network landslide prediction method integrating multi-source data. Background Art
[0002] In the geological disaster-prone areas of Sichuan-Tibet and Xinjiang-Tibet in my country, landslides, mud-rock flows, collapses and other disasters occur frequently and cause great harm. The monitoring and early warning of landslides have always been a concern of experts and scholars at home and abroad. At present, the monitoring and early warning methods for landslides are mainly geodetic precision instrument measurement methods, including longitude and latitude measurement, surface displacement measurement, etc. However, these instruments not only consume a lot of manpower and material resources during the measurement process, but also have certain deviations in measurement accuracy. At the same time, due to the complexity and uncertainty of landslides, there are many factors affecting the occurrence of landslides. Therefore, comprehensive consideration of various factors is a necessary prerequisite for accurate prediction. Summary of the invention
[0003] The purpose of the present invention is to provide a MV-BP (Moving average Backpropagation) neural network landslide prediction method based on integrated multi-source data to achieve accurate prediction of landslide disasters.
[0004] The technical solution to achieve the purpose of the present invention is:
[0005] A MV-BP neural network landslide prediction method based on multi-source data includes the following steps:
[0006] Step 1: Obtain all input data of landslide deformation and displacement to construct a training data set;
[0007] Step 2: Intelligently preprocess the multi-source monitoring data to obtain input values for model training;
[0008] Step 3: Establish an MV-BP neural network detection and early warning model, input data, and obtain prediction results;
[0009] Step 4: Use the evaluation index model to calculate the accuracy of the MV-BP neural network prediction results.
[0010] Compared with the prior art, the significant advantages of the technical solution of the present invention are:
[0011] The MV-BP neural network is used to predict landslides. Through multi-source data collection, the input value of landslide deformation displacement can be obtained, thereby realizing the construction of a training data set. Using the MV-BP-based neural network model to predict landslides can obtain more accurate surface displacement change data, thereby improving the accuracy of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0012] Figure 1 This is a flow chart of landslide prediction using MV-BP neural network integrating multi-source data.
[0013] Figure 2 Flowchart for intelligent preprocessing of multi-source data.
[0014] Figure 3 This is the processing flow chart of MV-BP neural network.
[0015] Figure 4 Establish a flow chart for the MV-BP neural network model. DETAILED DESCRIPTION
[0016] The following is combined with Figure 1-4 The present invention is further described with reference to the accompanying drawings and specific embodiments.
[0017] Combination Figure 1 This embodiment provides a MV-BP neural network landslide prediction method integrating multi-source data, including the following steps:
[0018] Step 1: Obtain all input data of landslide deformation and displacement and construct a training data set, which specifically includes the following steps:
[0019] The full-factor input data of landslide deformation and displacement include three types: deformation monitoring data, stress-strain monitoring data, and inducing factor monitoring data. Deformation monitoring data is generally obtained through Beidou monitoring stations, total stations, displacement meters, three-dimensional laser scanning, InSAR, ground-based SAR (synthetic aperture radar), etc. This method obtains deformation monitoring data through Beidou monitoring stations and InSAR (interferometric radar). Stress-strain monitoring data includes monitoring the changes in stress-strain data of anchor cables after landslide anchoring. This method obtains stress monitoring data through NPR anchor cable mechanics collection. Inducing factor monitoring data includes meteorological information such as rainfall, temperature and humidity, and natural activities such as earthquakes. This method uses a rain gauge to obtain rainfall data.
[0020] Step 2: Intelligently preprocess the multi-source monitoring data to obtain the input value for model training. Figure 2 , the specific process is as follows:
[0021] The multi-source data obtained in this method mainly include Beidou monitoring data, InSAR data, NPR anchor mechanics data, and rainfall data. Among them, Beidou monitoring data and InSAR data require intelligent preprocessing before they can be used.
[0022] S21: Beidou monitoring data preprocessing: Beidou monitoring data preprocessing includes establishing ADEM (semi-supervised learning model) environmental model, RTK (real-time dynamic carrier phase difference technology) position solution, and data domain denoising based on EEMD (empirical mode decomposition), as follows:
[0023] (1) Establish an ADEM environmental model to optimize the recognition capability of Beidou gross errors and improve the utilization rate of Beidou data. The specific process is as follows:
[0024] Subtract the signal-to-noise ratio data of base station b and monitoring station r of the common-view satellite k, and get the sequence
[0025]
[0026] in, is the observed value of baseline b, is the observation value of monitoring station r, is the difference between the observations at the two stations.
[0027] By setting the satellite altitude Signal-to-Noise Ratio Threshold to filter The set of stationary sequences:
[0028]
[0029] Among them, Select represents the selected stable sequence; E′ is the empirical threshold of the satellite altitude angle, is the satellite elevation angle, S0 is the empirical threshold of the signal-to-noise ratio, and in this embodiment, E′=30° and S0=45dB.
[0030] calculate The standard deviation of the satellite k is calculated with [-3σ, 3σ] as the interval. Traverse all common-view satellites one by one to determine the overall The interval range is [S1, S2], and the specific formula is as follows:
[0031]
[0032] Among them, floor means rounding down, and ceil means rounding up.
[0033] The rejection criteria for observation data based on altitude angle and signal-to-noise ratio are:
[0034]
[0035] E0 is the minimum threshold of satellite elevation angle.
[0036] Satellites that meet the above criteria are marked as gross errors and removed. The above is the ADEM environmental model process.
[0037] (2) RTK solution: By utilizing the highly similar characteristics of the observation values of the base station and the monitoring station, such as tropospheric delay, ionospheric delay, satellite clock error, and ephemeris error, these errors can be fully eliminated through differential methods to obtain a centimeter-level displacement monitoring sequence of the landslide body.
[0038] (3) Data domain denoising based on EEMD. In actual environments, there are complex non-modeling errors and random noises in landslide monitoring data, which are difficult to separate effectively and difficult to eliminate using differential methods. The EEMD method can effectively reduce the occurrence of modal aliasing and is suitable for complex monitoring environments.
[0039] S22: InSAR data preprocessing. InSAR data preprocessing includes atmospheric delay error correction. The specific process is as follows:
[0040] The troposphere and ionosphere are corrected by using a comprehensive atmospheric correction model, and the SAR image is bandpass filtered in the range direction to separate two forward-looking and backward-looking SAR images with different frequencies. After forming interference patterns, the forward-looking and backward-looking interference patterns are linearly combined to separate the ionospheric phase with dispersive characteristics and the non-dispersive phase combination. Among them, φ1 and φ2 represent the center frequencies of the forward-looking and backward-looking SAR images after frequency division, and the formula is as follows:
[0041]
[0042] Where φ1 and φ2 are the forward-looking and backward-looking differential interference unwrapping phases after spectral frequency division, respectively; λ1 and λ2 are the forward-looking and backward-looking wavelength lengths; Δr nd is the change in the line of sight of the master and slave image radars, including terrain phase error, deformation phase, orbit phase, etc.; k and c are constant terms; ΔTEC is the ionospheric change of the master and slave images. The ionospheric change of the master and slave images can be obtained by linearly combining the above formulas:
[0043]
[0044] The InSAR observation results can be corrected according to the ionospheric changes obtained from the above formula. The above is the InSAR data preprocessing process.
[0045] S23: The pre-processed Beidou monitoring data and InSAR data, together with the NPR anchor mechanics data and rainfall data, are used as the input of the MV-BP neural network model.
[0046] Step 3: Establish the MV-BP neural network detection and early warning model, input data, and obtain the prediction results; Figure 3, the specific process is as follows:
[0047] S31: Normalize the input data of the neural network model. Since the input source data belongs to different dimensions, the input data needs to be normalized. The formula is as follows:
[0048]
[0049] Where x' is the normalized data, x is the original input data, min(x) is the minimum value in this type of data, and max(x) is the maximum value in this type of data.
[0050] S32: Obtain historical surface deformation data, and decompose the surface deformation data.
[0051] Decompose the surface deformation displacement data and convert the historical surface deformation displacement data X t Decompose into internal factors I t and predisposing factors t , I t and P t The formula is as follows:
[0052]
[0053] P t =X t -I t ;
[0054] In the formula, t is the time before the surface deformation occurs, n is the cumulative number of times, and X t is the surface deformation displacement data, I t For internal factors, P t As the predisposing factor.
[0055] S33: Construct the MV-BP neural network model, and use the decomposed deformation data to train the model to obtain the intrinsic factor monitoring and early warning model and the induced factor monitoring and early warning model. Figure 4 , the specific process is as follows:
[0056] (1) Determine the number of neurons in each layer: According to the input characteristics of the training samples, determine the number of neurons in the input layer, hidden layer, and output layer of the BP neural network. Assuming the number of neurons in the input layer is m, the number of neurons in the hidden layer is L, and the number of neurons in the output layer is k, the empirical formula for determining the number of neurons in the hidden layer is as follows:
[0057]
[0058] Where a is the system parameter;
[0059] In the neural network model in this embodiment, two hidden layers are set, one of which has 50 neurons and the second has 10 neurons.
[0060] (2) Initialize the network:
[0061] Set the parameters of the MV-BP neural network model, the learning rate is 0.01, the number of input layers is 1, and the number of iterations is 100. At this time, the internal factor I t and predisposing factors t Input to the network.
[0062] (3) Calculate the hidden layer output:
[0063] Calculate the output value of each neuron in the hidden layer of the network. Suppose the output value of the jth neuron in the i-th hidden layer is I ij , the formula is as follows:
[0064]
[0065] Where i is the layer number, i = {1, 2, .., n}, n is the total number of output layers, ω tr is the connection weight between the rth neuron in the previous layer and the tth neuron in the current hidden layer, δ r is the threshold of the rth neuron in the previous layer. F(·) is the activation function, and the formula is as follows:
[0066] F(U) = max(0,U);
[0067] Where U is the neuron output parameter;
[0068] (4) Calculate the output of the output layer:
[0069] Suppose the output value of the pth neuron in the output layer is R P , the activation function is F(U), then:
[0070]
[0071] Among them, ω tr is the connection weight between the rth neuron in the previous layer and the tth neuron in the current output layer, I r represents the output value of the rth neuron in the previous layer, δ r The connection threshold of the rth neuron in the previous layer.
[0072] (5) Calculate the prediction error of the current network:
[0073] According to the output value R of the current network P The true value Y of the sample data P Calculate the prediction error E of the current network P,as follows:
[0074]
[0075] (6) Determine the calculated error E P Whether the requirements are met, if so, end the training; if not, proceed to step (7).
[0076] (7) Update weights;
[0077] The weight ω of the rth neuron in each layer r as follows:
[0078]
[0079] Wherein, r represents the learning rate, which is 0.01 in this embodiment; ω r is the weight of the rth neuron in each layer, α is the influence coefficient, x r is the neuron transmission coefficient, ω mr is the weight from neuron m to neuron r, E p Warning error.
[0080] (8) Update threshold;
[0081] Update the threshold of each hidden layer neuron and output layer neuron in the BP neural network. The updating method of the threshold of the rth neuron in each hidden layer is:
[0082]
[0083] (9) After updating the new network weights and thresholds, re-execute steps (4) to (6) until the error reaches the required level and the training is terminated.
[0084] (10) The MV-BP neural network prediction model is established based on the connection weights and thresholds of each layer of neurons obtained after training. The MV-BP neural network prediction model includes the relationship between the intrinsic factor detection and early warning model and the inducing factor detection and early warning model.
[0085] S34: Input Beidou monitoring data and InSAR data into the intrinsic factor monitoring and early warning model, and input NPR anchor cable mechanical data and rainfall data into the induced factor monitoring and early warning model. Adding the results of the two models together, the final result of the MV-BP neural network prediction can be obtained.
[0086] Step 4: Use the evaluation index model to calculate the accuracy of the MV-BP neural network prediction results. The evaluation index model includes root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and determination coefficient (R 2 )4 indicators are used to evaluate the accuracy of the model. The formula is as follows:
[0087]
[0088] Where, MAE is the mean absolute error;
[0089]
[0090] Where RMSE is the root mean square error;
[0091]
[0092] Where, MAPE is the mean absolute percentage error;
[0093]
[0094] In the formula, R 2 is the coefficient of determination, represents the average measured value, represents the calculated value, y={y1,y2,...,y n} indicates the measured value.
[0095] In this embodiment, the method of the present invention adopts MV-BP neural network to predict landslides. Based on multi-source data of landslides, prediction is performed through MV-BP neural network model to obtain more accurate prediction results, thereby improving the accuracy of prediction.
[0096] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A MV-BP neural network landslide prediction method based on multi-source data, characterized in that: The following steps are involved: Step 1: Obtain all input data of landslide deformation and displacement to construct a training data set; Step 2: Intelligently preprocess the multi-source monitoring data to obtain input values for model training; Step 3: Establish an MV-BP neural network detection and early warning model, input data, and obtain prediction results; Step 4: Use the evaluation index model to calculate the accuracy of the MV-BP neural network prediction results.
2. The MV-BP neural network landslide prediction method based on comprehensive multi-source data according to claim 1 is characterized in that: In step S1, all elements of landslide deformation and displacement are input, including deformation monitoring data, stress and strain monitoring data, and inducing factor monitoring data.
3. The MV-BP neural network landslide prediction method based on comprehensive multi-source data according to claim 1 is characterized in that: The intelligent preprocessing process in step 2 specifically includes: S21: Beidou monitoring data preprocessing: including the establishment of ADEM environmental model, RTK position solution, and data domain denoising based on EEMD; S22: InSAR data preprocessing: InSAR data preprocessing is performed by correcting the atmospheric delay error. S23: The pre-processed Beidou monitoring data and InSAR data, together with the NPR anchor mechanics data and rainfall data, are used as the input of the MV-BP neural network model.
4. The MV-BP neural network landslide prediction method based on integrated multi-source data according to claim 3 is characterized in that: The specific process of step S21 is as follows: (1) Establish an ADEM environmental model to optimize the recognition capability of BeiDou gross errors and improve the utilization rate of BeiDou data; (2) RTK solution, using the highly similar characteristics of the tropospheric delay, ionospheric delay, satellite clock error, and ephemeris error contained in the observation values of the base station and the monitoring station, these errors are fully eliminated through the difference method to obtain the second-level and millimeter-level displacement monitoring sequence of the landslide body; (3) EEMD-based data domain denoising is used to effectively reduce the occurrence of modal aliasing.
5. The MV-BP neural network landslide prediction method based on integrated multi-source data according to claim 4 is characterized in that: In step S21, the ADEM environmental model is established as: Subtract the signal-to-noise ratio data of base station b and monitoring station r of the common-view satellite k, and get the sequence in, is the observed value of baseline b, is the observation value of monitoring station r, is the difference between the observations at the two stations; By setting the satellite altitude Signal-to-Noise Ratio Threshold to filter The set of stationary sequences: Among them, Select represents the selected stable sequence; E′ is the empirical threshold of satellite altitude angle, and S0 is the empirical threshold of signal-to-noise ratio; The rejection criteria for observation data based on altitude angle and signal-to-noise ratio are: Where E0 is the minimum threshold of satellite elevation angle, and the sequence The value range of is the interval range [S1,S2]; Satellites that meet the above criteria are marked as gross errors and removed.
6. The MV-BP neural network landslide monitoring and early warning method based on integrated multi-source data according to claim 1 is characterized in that: Step S3 specifically includes: S31: normalizing the input data of the neural network model; S32: Acquire historical surface deformation data and decompose the surface deformation data; S33: constructing an MV-BP neural network model, and using the decomposed deformation data to train the model, to obtain an intrinsic factor monitoring and early warning model and an inducing factor monitoring and early warning model; S34: Input Beidou monitoring data and InSAR data into the intrinsic factor monitoring and early warning model, and input NPR anchor cable mechanical data and rainfall data into the induced factor monitoring and early warning model; add the results of the two parts of the model to obtain the final result of the MV-BP neural network prediction.
7. The MV-BP neural network landslide monitoring and early warning method based on integrated multi-source data according to claim 6 is characterized in that: In S32, the historical surface deformation displacement data X t Decompose into internal factors I t and predisposing factors t , I t and P t The formula is as follows: Where t is the time before the surface deformation occurs, n is the cumulative number of times, Xt is the displacement data of the surface deformation, It is the intrinsic factor, and Pt is the inducing factor.
8. The MV-BP neural network landslide monitoring and early warning method based on integrated multi-source data according to claim 6 is characterized in that: The specific steps in S33 are as follows: (1) Determine the number of neurons in each layer; (2) Initialize the network; (3) Calculate the hidden layer output; (4) Calculate the output of the output layer; (5) Calculate the prediction error of the current network; (6) Determine the calculated error E P Whether the requirements are met, if so, the training ends; if not, proceed to step (7); (7) Update weights; (8) Update threshold; (9) After updating the new network weights and thresholds, re-execute steps (4) to (6) until the error reaches the required level and the training is terminated; (10) The MV-BP neural network prediction model is established based on the connection weights and thresholds of each layer of neurons obtained after training.
9. The MV-BP neural network landslide monitoring and early warning method based on integrated multi-source data according to claim 1 is characterized in that: The evaluation index model in step 4 includes root mean square error, mean absolute error, mean absolute percentage error and determination coefficient.
Citation Information
Patent Citations
Landslide deformation displacement prediction method based on improved BP neural network
CN113283161A
