Slope early warning method based on mutation theory and recurrent neural network

Abstract

The invention discloses a slope early warning method based on a mutation theory and a recurrent neural network, and the method comprises the steps: collecting corresponding slope displacement data through a geological collection device, calculating the difference value of adjacent nodes to obtain slope displacement change data, and predicting a mutation point when a slope is unstable, so as to achieve the purpose of slope early warning. Slope displacement change data is divided into overall trend data and fluctuation data by adopting a decomposition method, the overall trend data is input into an RNN for training, the fluctuation data is input into an RNN + Attention network for training, the predicted trend data and fluctuation data are added to obtain final displacement change prediction data, and the final displacement change prediction data serve as input of a mutation model. And predicting to obtain a final slope instability abrupt change point. According to the method, the principle of the mutation theory is combined with the RNN model, the defects of a traditional method are overcome, and the accuracy and timeliness of slope instability early warning are improved.

Description

Technical Field

[0001] The present application relates to the technical field of geological disaster monitoring and early warning, and in particular to a slope early warning method based on mutation theory and recurrent neural network. Background Art

[0002] Slope warning is a crucial direction in the geological disaster warning system. It mainly targets disasters such as landslides, mudslides, collapses, etc. that are highly random and destructive. It provides accurate warning information before the disaster occurs, so as to formulate response measures, which can effectively avoid casualties and property losses.

[0003] Traditional slope stability early warning methods usually rely on physical models and statistical analysis, mainly predicting displacement curves through a single model, such as linear regression, and then calculating the range of the tangent angle to achieve the effect of early warning. However, this method has poor prediction accuracy for slope instability, and the early warning often requires manual analysis, with a low degree of automation. In addition, in order to improve the model's predicted displacement, some researchers have proposed the use of multi-source data sets, including soil moisture, deep displacement, crack displacement, slope displacement and other data. However, this method also has obvious limitations, such as data is difficult to obtain, and the multi-source feature training model may gradually forget the distribution of real data, resulting in irreversible defects.

[0004] In view of this, there is an urgent need for a slope early warning method based on catastrophe theory and recurrent neural network. The catastrophe theory and RNN are combined to realize slope early warning. RNN performs well in processing time series data and can be used to predict the dynamic changes of slope displacement. Catastrophe theory, as a theory for dealing with sudden changes in system behavior, can effectively describe the sudden change state of slope instability. Summary of the invention

[0005] In view of the shortcomings of the prior art mentioned above, the present application proposes a slope early warning method based on mutation theory and recurrent neural network, which solves the defect of low automation in the prior art. The mutation theory is used to directly process the predicted displacement changes to obtain the mutation point, that is, the time when the slope becomes unstable; at the same time, the single displacement curve is modally decomposed into trend data and fluctuation data to improve the prediction ability of the model, and also enhances the model's ability to process complex data, which can better adapt to various geological conditions and environmental changes, and provide better assistance for slope disaster early warning.

[0006] In order to achieve the above-mentioned purpose, this application specifically adopts the following technical solutions:

[0007] A slope early warning method based on catastrophe theory and recurrent neural network includes the following steps:

[0008] Step 1: Preprocess the collected slope displacement data, and then calculate the difference between adjacent nodes to obtain the slope displacement change data;

[0009] Step 2: Decompose the slope displacement change data to obtain overall trend data and periodic fluctuation data;

[0010] Step 3: Build the RNN+Attention network and define the mutation model;

[0011] Step 4: Use the RNN network to train the overall trend data to obtain the first prediction data, and use the RNN+Attention network to train the periodic fluctuation data to obtain the second prediction data;

[0012] Step 5: Add the first prediction data and the second prediction data to obtain the final displacement change prediction data, use the displacement change prediction data as input, use the mutation model to predict the mutation point, and complete the slope warning.

[0013] As an optional technical solution, the slope displacement data in step 1 is collected by geological acquisition equipment.

[0014] As an optional technical solution, the data preprocessing includes denoising, and a density-based clustering algorithm is used to denoise the data.

[0015] As an optional technical solution, the data preprocessing includes processing missing values ​​using the adjacent mean filling method, that is, looping through the entire sequence, and if the current node has a missing value, the mean of the previous node and the next node is used as the value of the current node.

[0016] As an optional technical solution, the data preprocessing includes taking differences between adjacent distances and taking differences between adjacent nodes of the denoised data until the traversal is completed.

[0017] As an optional technical solution, the decomposition of the slope displacement change data in step 2 adopts an empirical mode decomposition method, and multiple intrinsic mode functions are obtained by decomposition, each of which represents a different frequency component of the signal;

[0018] A threshold frequency component is preset, and all inherent mode functions greater than the threshold frequency component are added together to obtain the overall trend data. In order to retain the characteristics of the original data, the overall trend data obtained by subtracting the original data is used as the periodic fluctuation data.

[0019] As an optional technical solution, the specific process of building the RNN+Attention network in step 3 is as follows:

[0020] S1: Define the input structure of the network. The input size is the time step and dimension. The input shape is defined as (30, 1), which means that the predicted input is one-dimensional data for 30 days.

[0021] S2: Construct the GRU of the recurrent neural network, set the number of neurons to 32, and the activation function to tanh;

[0022] S3: Build the LSTM layer, set the number of neurons to 96, and the activation functions to tanh and sigmoid;

[0023] S4: Construct the Attention mechanism, set trainable parameters, multiply the input features by the parameters, obtain the score of each node through the activation function tanh, and then obtain the weight of each node through the activation function softmax, and finally obtain the features after the attention mechanism;

[0024] S5: Define additional layers, including two fully connected Dense layers with 128 neurons, and a Dropout regularization network structure to prevent overfitting, with the coefficient set to 0.25;

[0025] S6: Define the output layer, which is a fully connected layer and has a predefined prediction length and size (5, 1).

[0026] As an optional technical solution, the specific definition process of defining the mutation model in step 3 is:

[0027] A1: Determine state variables and control parameters

[0028] State variables: describe the internal state of the system;

[0029] Control parameters: describe external conditions or factors that affect changes in system states;

[0030] A2: Constructing the potential function

[0031] The landslide prediction-inverse velocity formula is used as the potential function, and its formula is as follows:

[0032]

[0033] In the formula, t is used to represent the state variable, which represents the time change, A and α are the parameters of the external factors, T is the predicted time when the landslide becomes unstable, and V is the moving speed of the landslide;

[0034] A3: Analyze the equilibrium state, input the slope displacement change data predicted by the RNN model, and perform linear regression analysis on Vt. According to the regression analysis results, predict the time T1 when the slope occurs. At this time, T1 represents the predicted mutation point.

[0035] As an optional technical solution, in step 4, the overall trend data of the RNN network training is used to obtain the first prediction data, and the periodic fluctuation data of the RNN+Attention training is used to obtain the second prediction data, wherein the RNN network training and the RNN+Attention training both include the following steps:

[0036] All data are split into 80% training data + 20% test data. The training data is first normalized separately, and then the sliding window method is used to preprocess the data.

[0037] The beneficial effects of this application include:

[0038] By applying the mutation theory, the potential mutation points of the slope can be accurately identified, avoiding the misjudgment and missed judgment caused by the large data volatility of the traditional method. The RNN model can accurately predict the future landslide trend by learning the features in the historical data, thereby providing early warning signals before the slope occurs. Combining the advantages of the mutation theory and the RNN model, the early warning system can not only identify sudden landslide precursors, but also provide stable early warning results through the time series prediction model. This comprehensive model shows higher stability and robustness in the face of various slope state changes, has real-time early warning capabilities and good adaptability, and provides strong technical support for the prevention and mitigation of slope disasters.

[0039] Other beneficial effects or advantages of the present application will be described in detail in conjunction with the specific structure in the specific implementation manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative labor. In addition, it should be understood that the proportional relationship of the various components in the drawings of this specification does not represent the proportional relationship in the actual material selection and design, which is only a schematic diagram of the structure or position, where:

[0041] Figure 1 A flow chart of the method of this application;

[0042] Figure 2 This is a schematic diagram of data filtering in this application;

[0043] Figure 3 This is a schematic diagram of calculating data difference in this application;

[0044] Figure 4 This is a schematic diagram of modal decomposition data in this application;

[0045] Figure 5 This is a schematic diagram of the model prediction in this application. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations.

[0047] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application for which protection is sought, but merely represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.

[0048] In the description of this application, it should be noted that the terms used, such as "top" and "bottom", refer to the part of this application that is close to the top when in use as the top, and the part that is close to the bottom as the bottom; the terms used, such as "first" and "second", are only used to distinguish expressions, and do not indicate or imply that they have a difference in importance or order; the terms used, such as "inside" and "outside", refer to the inside and outside of a specific contour. The use of the above terms is only to facilitate a clear and simple description of the technical solution of this application, and cannot be understood as a limitation on this application.

[0049] Example:

[0050] This embodiment discloses a slope early warning method based on catastrophe theory and recurrent neural network, comprising the following steps:

[0051] Step 1: Preprocess the collected slope displacement data, and then calculate the difference between adjacent nodes to obtain the slope displacement change data;

[0052] Step 2: Decompose the slope displacement change data to obtain overall trend data and periodic fluctuation data;

[0053] Step 3: Build the RNN+Attention network and define the mutation model;

[0054] Step 4: Use the RNN network to train the overall trend data to obtain the first prediction data, and use the RNN+Attention network to train the periodic fluctuation data to obtain the second prediction data;

[0055] Step 5: Add the first prediction data and the second prediction data to obtain the final displacement change prediction data, use the displacement change prediction data as input, use the mutation model to predict the mutation point, and complete the slope warning.

[0056] The present invention is a method for slope early warning. Before training the prediction model, it only needs to collect long-term slope displacement data. First, pre-process the data, decompose the processed slope change data into trend data and periodic fluctuation data, input the periodic fluctuation data into the recurrent neural network for prediction, and then add it to the trend data to obtain the final actual prediction value. Finally, input the prediction value into the mutation model to obtain the mutation point when the slope is unstable. The flow chart of the slope early warning method based on mutation theory and recurrent neural network is as follows: Figure 1 shown.

[0057] The existing slope early warning method mainly predicts the displacement curve through a single model, such as linear regression, and then calculates the range of the tangent angle to achieve the early warning effect. However, this method has poor prediction accuracy for slope instability, and the early warning mostly requires manual analysis, with a low degree of automation. In addition, in order to improve the prediction displacement of the model, some researchers have proposed the use of multi-source data sets, including soil moisture, deep displacement, crack displacement, slope displacement and other data. However, this method also has obvious limitations, such as data is difficult to obtain, and the multi-source feature training model may gradually forget the distribution of real data, resulting in irreversible defects.

[0058] Therefore, a slope early warning method based on mutation theory and RNN is proposed. This method makes up for the defect of low automation. The mutation theory is used to directly process the predicted displacement changes to obtain the mutation point, that is, the time when the slope becomes unstable. At the same time, the single displacement curve is modally decomposed into trend data and fluctuation data to improve the prediction ability of the model. It also enhances the model's ability to process complex data, can better adapt to various geological conditions and environmental changes, and provide better assistance for slope disaster early warning.

[0059] More specifically:

[0060] In step 1, data preprocessing is performed on the collected (raw data) slope displacement data, including denoising, processing missing values, and calculating differences between adjacent nodes.

[0061] Denoising: When the instrument collects data, the data may have obvious noise due to external factors. It is necessary to remove the obvious noise before inputting the network training. Here, a density-based clustering algorithm DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is used to denoise the data to ensure the effective removal of outliers. The relevant steps are as follows:

[0062] For a point p, its ε neighborhood is defined as:

[0063] N(p)={q∈D|disp(p, q)≤ε}

[0064] Where disp(p, q) represents the Euclidean distance between points p and q.

[0065] If a point p contains at least "Minpts" points in its ε neighborhood, then p is considered a core point;

[0066] If a point p is not a core point, but it is in the ε neighborhood of a core point q, then p is a boundary point;

[0067] If a point p is neither a core point nor a boundary point, it is considered a noise point;

[0068] If p is a core point, a new cluster is generated. For each core point in the current cluster, all points in its ε neighborhood are found and added to the current cluster. The expansion continues until all points are processed.

[0069] As mentioned above, ε is set to 3 and Minpts is set to 2. The effect of slope displacement data denoising is shown in the figure below: Figure 2 As shown, this step is crucial to remove data noise due to instrument errors or environmental interference.

[0070] Processing missing values: In order to better train the sequence features of the model, missing values ​​need to be processed. Here, the adjacent mean filling method is used, that is, the entire sequence is looped through. If the current node has a missing value, the mean of the previous node and the next node is used as the value of the current node.

[0071] Difference processing of adjacent nodes: In order to obtain the displacement change data of the slope, it is necessary to calculate the difference between the adjacent nodes of the denoised data until the traversal is completed. The effect diagram of the slope displacement change data is as follows Figure 3 As shown, the sudden change in the final displacement indicates slope instability.

[0072] Furthermore, the decomposition of the slope displacement change data in step 2 adopts the empirical mode decomposition method, and multiple intrinsic mode functions are obtained by decomposition, each intrinsic mode function represents a different frequency component of the signal; a threshold frequency component is preset, and all intrinsic mode functions greater than the threshold frequency component are added together to obtain the overall trend data. In order to retain the characteristics of the original data, the overall trend data obtained by subtracting the original data is used as periodic fluctuation data.

[0073] Specifically, the present invention adopts the empirical mode decomposition (EMD) method, the principle of which is based on the local characteristics of the signal, and a plurality of intrinsic mode functions (IMFs) are obtained by decomposition, each IMF representing a different frequency component of the signal.

[0074] like Figure 4 As shown in the figure, it is a sequence obtained by modal decomposition of slope displacement change data. Since the trends of sequences IM6 and IM7 are gentle, IM6+IM7 are added as trend item data. In order to retain the characteristics of the original data, the data after the original data-trend item is identified as periodic fluctuation item data.

[0075] Furthermore, the specific process of building the RNN+Attention network in step 3 is as follows:

[0076] S1: Define the input structure of the network. The input size is the time step and dimension. The input shape is defined as (30, 1), which means that the predicted input is one-dimensional data for 30 days.

[0077] S2: Construct the GRU of the recurrent neural network, set the number of neurons to 32, and the activation function to tanh;

[0078] S3: Build the LSTM layer, set the number of neurons to 96, and the activation functions to tanh and sigmoid;

[0079] S4: Construct the Attention mechanism, set trainable parameters, multiply the input features by the parameters, obtain the score of each node through the activation function tanh, and then obtain the weight of each node through the activation function softmax, and finally obtain the features after the attention mechanism;

[0080] S5: Define additional layers, including two fully connected Dense layers with 128 neurons, and a Dropout regularization network structure to prevent overfitting, with the coefficient set to 0.25;

[0081] S6: Define the output layer, which is a fully connected layer and has a predefined prediction length and size (5, 1).

[0082] As an optional technical solution, the specific definition process of defining the mutation model in step 3 is:

[0083] A1: Determine state variables and control parameters

[0084] State variables: describe the internal state of the system;

[0085] Control parameters: describe external conditions or factors that affect changes in system states;

[0086] A2: Constructing the potential function

[0087] The landslide prediction-inverse velocity formula is used as the potential function, and its formula is as follows:

[0088]

[0089] In the formula, t is used to represent the state variable, which represents the time change, A and α are the parameters of the external factors, T is the predicted time when the landslide becomes unstable, and V is the moving speed of the landslide;

[0090] A3: Analyze the equilibrium state, input the slope displacement change data predicted by the RNN model, and perform linear regression analysis on Vt. According to the regression analysis results, predict the time T1 when the slope occurs. At this time, T1 represents the predicted mutation point.

[0091] In step 4, the overall trend data of the RNN network training is used to obtain the first prediction data, and the periodic fluctuation data of the RNN+Attention training is used to obtain the second prediction data, wherein the RNN network training and the RNN+Attention training both include the following steps:

[0092] All data are split into 80% training data + 20% test data. The training data is first normalized separately, and then the data is preprocessed using the sliding window method. The sliding window method defines the window size as 1, which can maximize the division of the training set. The division method is as follows:

[0093]

[0094] First, divide the first 30 data in the data set as input, and the 5 data from the 31st to the 35th as output test data. Then iterate forward one data at a time until all the data are divided.

[0095] During training, the model parameters need to be continuously tuned until the final prediction effect is obtained, and the currently trained model is the final result.

[0096] The final model test results are as follows: Figure 5As shown in the figure, according to the single prediction of 5 data, a total of 800 data are predicted, and the average error and root mean square error of each node are calculated to be 0.092 and 0.102 respectively. The predicted mutation point T is 745, which is 5 times different from the actual instability time of the landslide 750, and it is an early prediction. It can be seen that this method has a significant effect on the slope early warning capability.

[0097] In an optional embodiment, the present solution also includes one or more processors, a memory and one or more computer programs, wherein the one or more computer programs are stored in the memory, and the one or more computer programs include instructions, which, when executed by the electronic device, enable the one or more processors to execute the method proposed in the present application.

[0098] It can be known from the above description of the implementation method that those skilled in the art can clearly understand that all or part of the steps in the above implementation method can be implemented by means of software plus a necessary general hardware platform. Based on this understanding, the above computer program product may include but is not limited to APP;

[0099] In an optional embodiment, the present solution further includes a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method proposed in the present application is implemented.

[0100] In several embodiments provided by the present invention, if any function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, some technical solutions of the present invention are essentially or part of the technical solution that contributes to the prior art or part of the technical solution can be embodied in the form of the software product described below.

[0101] Continuing from the above, the above-mentioned device / terminal may be a computer device, and the hardware structure of the computer device may also specifically include: at least one processor, at least one communication interface, at least one memory and at least one communication bus; the processor, the communication interface and the memory may communicate with each other through the communication bus. Among them, the processor may be a central processing unit CPU, DSP, microcontroller or digital signal processor, and may also include GPU, embedded neural network processor (Neural-network Process Units; hereinafter referred to as: NPU) and image signal processor (Image Signal Processing; hereinafter referred to as: ISP), the processor may also include a specific integrated circuit ASIC, or one or more integrated circuits configured to implement the embodiment of the present invention, etc. In addition, the processor may have the function of operating one or more software programs, and the software programs may be stored in a storage medium such as a memory; and the aforementioned memory / storage medium may include: non-volatile memory (non-volatile memory), such as a non-removable disk, a U disk, a mobile hard disk, an optical disk, etc., as well as read-only memory (Read-Only Memory; hereinafter referred to as: ROM), random access memory (Random Access Memory; hereinafter referred to as: RAM), etc.

[0102] Finally, it should be understood that the embodiments described in this specification are only used to illustrate the principles of the embodiments of this specification. Other variations may also fall within the scope of this specification. Therefore, as an example and not a limitation, alternative configurations of the embodiments of this specification may be considered consistent with the teachings of this specification. Accordingly, the embodiments of this specification are not limited to the embodiments explicitly introduced and described in this specification.

Claims

1. A slope early warning method based on catastrophe theory and recurrent neural network, characterized in that: The following steps are involved: Step 1: Preprocess the collected slope displacement data, and then calculate the difference between adjacent nodes to obtain the slope displacement change data; Step 2: Decompose the slope displacement change data to obtain overall trend data and periodic fluctuation data; Step 3: Build the RNN+Attention network and define the mutation model; Step 4: Use the RNN network to train the overall trend data to obtain the first prediction data, and use the RNN+Attention network to train the periodic fluctuation data to obtain the second prediction data; Step 5: Add the first prediction data and the second prediction data to obtain the final displacement change prediction data, use the displacement change prediction data as input, use the mutation model to predict the mutation point, and complete the slope warning.

2. A slope early warning method based on catastrophe theory and recurrent neural network as claimed in claim 1, characterized in that: The slope displacement data in step 1 is collected by geological acquisition equipment.

3. A slope early warning method based on catastrophe theory and recurrent neural network as claimed in claim 1, characterized in that: The data preprocessing includes denoising, and a density-based clustering algorithm is used to denoise the data.

4. A slope early warning method based on catastrophe theory and recurrent neural network as claimed in claim 1, characterized in that: The data preprocessing includes processing missing values, using the adjacent mean filling method, that is, looping through the entire sequence, if the current node has a missing value, the mean of the previous node and the next node is used as the value of the current node.

5. The slope early warning method based on catastrophe theory and recurrent neural network as claimed in claim 1 is characterized in that: The data preprocessing includes finding differences between adjacent distances and finding differences between adjacent nodes of the denoised data until the traversal is completed.

6. A slope early warning method based on catastrophe theory and recurrent neural network as claimed in claim 1, characterized in that: The decomposition of the slope displacement change data in step 2 adopts the empirical mode decomposition method, and a plurality of intrinsic mode functions are obtained by decomposition, each intrinsic mode function represents a different frequency component of the signal; A threshold frequency component is preset, and all inherent mode functions greater than the threshold frequency component are added together to obtain the overall trend data. In order to retain the characteristics of the original data, the overall trend data obtained by subtracting the original data is used as the periodic fluctuation data.

7. A slope early warning method based on catastrophe theory and recurrent neural network as claimed in claim 1, characterized in that: The specific process of building the RNN+Attention network in step 3 is as follows: S1: Define the input structure of the network. The input size is the time step and dimension. The input shape is defined as (30, 1), which means that the predicted input is one-dimensional data for 30 days. S2: Construct the GRU of the recurrent neural network, set the number of neurons to 32, and the activation function to tanh; S3: Build the LSTM layer, set the number of neurons to 96, and the activation functions to tanh and sigmoid; S4: Construct the Attention mechanism, set trainable parameters, multiply the input features by the parameters, obtain the score of each node through the activation function tanh, and then obtain the weight of each node through the activation function softmax, and finally obtain the features after the attention mechanism; S5: Define additional layers, including two fully connected Dense layers with 128 neurons, and a Dropout regularization network structure to prevent overfitting, with the coefficient set to 0.25; S6: Define the output layer, which is a fully connected layer and has a predefined prediction length and size (5, 1).

8. The slope early warning method based on catastrophe theory and recurrent neural network as claimed in claim 1 is characterized in that: The specific definition process of defining the mutation model in step 3 is: A1: Determine state variables and control parameters State variables: describe the internal state of the system; Control parameters: describe external conditions or factors that affect changes in system states; A2: Constructing the potential function The landslide prediction-inverse velocity formula is used as the potential function, and its formula is as follows: In the formula, t is used to represent the state variable, which represents the time change, A and α are the parameters of the external factors, T is the predicted time when the landslide becomes unstable, and V is the moving speed of the landslide; A3: Analyze the equilibrium state, input the slope displacement change data predicted by the RNN model, and perform linear regression analysis on Vt. According to the regression analysis results, predict the time T1 when the slope occurs. At this time, T1 represents the predicted mutation point.

9. The slope early warning method based on catastrophe theory and recurrent neural network as claimed in claim 1, characterized in that: In step 4, the overall trend data of the RNN network training is used to obtain the first prediction data, and the periodic fluctuation data of the RNN+Attention training is used to obtain the second prediction data, wherein the RNN network training and the RNN+Attention training both include the following steps: All data are split into 80% training data + 20% test data. The training data is first normalized separately, and then the sliding window method is used to preprocess the data.

Images