Mountain landslide early warning system and method based on modal decomposition and machine learning

Abstract

The invention discloses a landslide early warning system and method based on modal decomposition and machine learning, and particularly relates to the field of artificial intelligence big data driving in geotechnical engineering and geological disaster prediction and early warning, a data transmission device transmits data monitored by a data monitoring and acquisition system to a monitoring station and a cloud storage server; the data monitoring and acquisition system comprises a displacement sensor, an underground water level sensor, a humidity sensor and a data transmission device; the displacement sensors are arranged on the surface and inside a mountain slope to monitor the displacement of the slope, the underground water level sensors are arranged at the slope top, the slope body and the slope toe to monitor the water level of the slope, and the humidity sensors are arranged on the surface of the slope to monitor weather conditions; the landslide displacement is predicted through a data analysis system, and the data analysis system comprises an SSA-VMD model and a CNN-BiLSTM-Attention model.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence big data driven in geotechnical engineering and geological disaster prediction and early warning, and in particular to a landslide early warning system and method based on modal decomposition and machine learning. Background Art

[0002] The alpine canyon area has rich water resources. In some areas where water energy is concentrated, a large number of reservoirs and dams have been built. At the same time, these areas have strong recent and modern tectonic activities, deep river valleys, and complex and harsh geological environments. In such an environment, the slopes of the reservoir are prone to instability and landslides. Among them, changes in reservoir water levels and seasonal rainfall are recognized as the most common factors that induce instability and deformation of reservoir slopes. Once the slope of the reservoir becomes unstable and forms a landslide, a dam will form in the valley to block the river, causing surges and dam collapses, etc. The disaster chain seriously affects the operational safety and environmental safety of the project. Therefore, making full use of monitoring data to realize an accurate and reliable prediction and early warning system has important theoretical significance and engineering application value for ensuring the safe operation of the project and reducing disaster losses.

[0003] At present, there are many studies and achievements on landslide early warning systems and methods, including monitoring equipment, drone data collection, Beidou satellite data transmission, real-time simulation, and the use of various machine learning and optimization algorithms. In landslide applications, the nonlinearity of displacement monitoring data makes high-precision prediction challenging. Traditional displacement prediction models are insufficient in accuracy and require technicians to operate and analyze them manually to a certain extent, resulting in large errors. Therefore, a landslide early warning system and method based on the SSA-VMD-CNN-BiLSTM-Attention model is developed, and more algorithms are introduced to quickly and accurately analyze and predict data, so as to improve the rapid analysis and prediction of landslide monitoring big data, and improve the accuracy and reliability of landslide early warning, to meet the urgent needs of ensuring the safety of major projects and personnel. Summary of the invention

[0004] To this end, the present invention provides a landslide early warning system and method based on modal decomposition and machine learning to solve the problems raised in the background technology.

[0005] In order to achieve the above-mentioned purpose, the present invention provides the following technical solutions: a landslide early warning method based on modal decomposition and machine learning, wherein a data transmission device transmits data monitored by a data monitoring and acquisition system to a cloud storage server, and the cloud storage server is responsible for storing original and processed predicted data; the data monitoring and acquisition system comprises a displacement sensor, a groundwater level sensor, a humidity sensor and a data transmission device; the displacement sensor is arranged on the surface and inside of a mountain slope to monitor the displacement of the slope, the groundwater level sensor is arranged at the top, the body and the foot of the slope to monitor the water level of the slope, and the humidity sensor is arranged on the surface of the slope to monitor the weather conditions;

[0006] The landslide displacement is predicted by the data analysis system set up at the monitoring station. The data analysis system includes the SSA-VMD model and the CNN-BiLSTM-Attention model.

[0007] (1) The original landslide displacement time series is divided into three subsequences using the SSA-VMD model. The optimal decomposition subsequences are selected as the trend term displacement sequence, the periodic term displacement sequence, and the random term displacement sequence according to the sample entropy value of the periodic term.

[0008] (2) Through the SSA-VMD model, each influencing factor sequence is decomposed into two subsequences, which represent the low-frequency and high-frequency parts of the influencing factor sequence respectively. The optimal decomposition subsequence is obtained according to the sample entropy value of the low-frequency part, and the maximum mutual information coefficient MIC and gray correlation GRG value of the decomposition subsequence of each factor and each displacement subsequence are calculated, and the correlation between the influencing factor subsequence and the displacement subsequence is comprehensively judged;

[0009] (3) Based on the two subsequence input sequences of the influencing factors decomposed in (2), a single-factor CNN-BiLSTM-Attention model is constructed for the trend item displacement sequence, and a multi-factor CNN-BiLSTM-Attention model is constructed for the period item displacement sequence and the random item displacement sequence. The single-factor and multi-factor CNN-BiLSTM-Attention models are trained using the training data set and the validation data set is predicted;

[0010] (4) The final cumulative landslide displacement prediction result is the sum of the predicted values ​​of the trend term displacement, the period term displacement and the random term displacement. The prediction performance of the method is verified and evaluated by comparing the cumulative landslide displacement monitoring sequence.

[0011] Preferably, the landslide displacement time series is expressed as follows according to the additive model of the time series:

[0012] X(t)=T(t)+P(t)+R(t)

[0013] Where: X(t) is the time series displacement value; T(t) is the trend term function; P(t) is the period term function; R(t) is the uncertainty random term function;

[0014] The intrinsic mode function IMF of VMD in the SSA-VMD model is an amplitude modulation-frequency modulation signal expression as follows:

[0015] u k (t) = A k (t)cos(φ k (t))

[0016] Where: u k (t) is the modal signal; phase φ k (t) is a non-decreasing function and is related to the instantaneous amplitude A k (t) are all non-negative;

[0017] In the SSA-VMD model, VMD uses the input signal sequence and the sum of each mode as the constrained variation, and realizes adaptive decomposition by finding the optimal solution of the constrained variation model. Then its constrained variation expression is written as:

[0018]

[0019] Where: K is the number of required modal components, k is an integer between 1 and K; {u k )={u1,…,u k} is the modal component obtained by the final decomposition; {ω k )={ω1,…,ω k} is the actual center frequency of each modal component; is the symbol of partial derivative; δ(t) is the Dirac function; * is the convolution operator;

[0020] To solve the above equation, the Lagrange operator λ is introduced to transform the constrained variational problem into an unconstrained variational problem, and the expanded Lagrange expression is obtained as follows:

[0021]

[0022] Where: ε is the decomposition parameter to reduce the interference of Gaussian noise;

[0023] The alternating direction multiplier iteration algorithm is used to optimize the modal components and center frequencies, and the saddle point of the unconstrained model is searched to obtain the optimal solution of the constrained model, and then K modal components are obtained.

[0024] Preferably, the number of modal components K=2, and the penalty factor α and the rise time step τ in the VMD model are optimized using the sparrow optimization algorithm SSA, which can effectively avoid the influence of subjective factors.

[0025] fitness=RMSE(m,M)·SampEn(IMF2)

[0026] As the fitness value of the SSA algorithm;

[0027] Where RMSE(m,M) is the root mean square error between the reconstructed sequence and the original sequence; SampEn(IMF2) is the sample entropy value of the low-frequency part of the periodic term displacement sequence or influencing factors after decomposition.

[0028]

[0029] Where m is the embedding dimension, which is 1 or 2, and r is the similarity tolerance, which is 0.1 to 0.25σ. x , σ x is the standard deviation of the sequence, is the probability that two sequences match points at similar capacity, and is the probability that two sequences match points;

[0030]

[0031] Where: x t is the value of the original sequence at time t, is the value of the reconstructed sequence at time t, and n is the length of the sequence.

[0032] Optimize the penalty factor α and the rising time step τ to find the optimal fitness, and fuse the sample entropy and the root mean square error. This can reflect both the integrity of the decomposed sequence and the decomposition effect of the sequence. The specific steps are as follows:

[0033] (1) Initialize the parameters input to the sparrow optimization algorithm.

[0034] (2) Randomly generate a series of α and τ as the initial positions of the sparrow population.

[0035] (3) Perform VMD decomposition on the displacement sequence at the current sparrow position, and calculate the sample entropy SampEn(IMF2) of the decomposed periodic term sequence or the low-frequency part of the influencing factor.

[0036] (4) Obtain the new sparrow fitness value fitness, find the individual with the best fitness and the individual with the worst fitness, and update the positions of the discoverer, follower, and warning device.

[0037] (5) Repeat (3) and (4) until the maximum number of iterations is reached, and output the sparrow position and fitness value at this time as the optimal solution.

[0038] Preferably, the CNN-BiLSTM-Attention model uses a CNN framework composed of a double-layer one-dimensional convolution layer and a pooling layer to automatically extract internal features of the displacement sequence. The convolution layer performs effective nonlinear local feature extraction of the time series. The pooling layer selects the maximum pooling method to compress the extracted features and generate more critical feature information. The BiLSTM hidden layer modeling learns the internal dynamic change law of the local features extracted by CNN, iteratively extracts more complex global features from the local features, and uses the features generated by the BiLSTM hidden layer as the input of the Attention mechanism. The Attention mechanism is used to automatically distinguish the importance of the time information extracted by the BiLSTM hidden layer in a weighted manner, which can more effectively utilize the time series properties of the displacement time series itself and mine deep time correlations. The attention mechanism can effectively reduce the loss of historical information and highlight the information of key historical time points to weaken the influence of redundant information on the displacement prediction results. Then, the output of the Attention layer is used as the input of the fully connected layer, and the final displacement prediction result is output through the fully connected layer.

[0039] The bidirectional long short-term memory neural network BiLSTM is an improvement on the traditional LSTM and is composed of a forward and backward LSTM, both of which will affect the results.

[0040] A i =f1(ω1x i +ω2A i-1 )

[0041] B i =f2(ω3x i +ω5B i-1 )

[0042] Y i =f3(ω4A i +ω6B i )

[0043] Where: f1, f2, f3 are activation functions between different layers, x i Input word for time i, A i is the forward LSTM unit output value at time i, B i is the output value of the LSTM unit after time i, Y i is the final output result at time i, ω1 and ω2 are x i , A i-1 Input the weight of the forward LSTM unit at time i, ω3 and ω5 are x i , B i-1 Input the weights of the backward LSTM unit at time i, ω4 and ω6 are A i , B iInput the weight of the output layer at time i; the Attention mechanism assigns weights according to different features, and the attention state conversion process is as follows:

[0044] S ti =Vtanh(Wh t +Uh i +b),i=1,2,3…,t-1

[0045]

[0046] h′ t =f(F,h t ,y t )

[0047] Where: a ti Input value h to the BiLSTM hidden layer t The attention weights for the current input, y1,y2,y3,...,y t is the input sequence, h1,h2,h3,...,h t is the corresponding input sequence y1,y2,y3,...,y t The hidden layer state value, h' t is the final feature vector V, W, U, b is the learning parameter of the model, S ti To use the SoftMax function to process a ti Required input, F is h t As a result of weighted averaging, these parameters are continuously iterated during the model training process.

[0048] The present invention also discloses a landslide early warning system based on modal decomposition and machine learning, which is characterized in that: the early warning system executes the above method, and the early warning system also includes a discrimination early warning system, the discrimination early warning system includes a discrimination relationship and a warning notification system, the discrimination relationship is that the slope of the obtained slope displacement curve exceeds a preset value for a period of time, and the discrimination relationship is met, which will trigger the warning notification system, transmit the slope information with landslide risk to relevant units, release relevant information to the public, and take emergency measures to deal with it.

[0049] The present invention has the following advantages:

[0050] The present invention provides a fast and accurate algorithm model for a landslide early warning system and method. The system and method introduce an SSA-VMD model to decompose slope displacement monitoring data into three subsequences, namely, a trend term, a periodic term and a random term, decompose each influencing factor sequence into two subsequences, respectively representing the low-frequency and high-frequency parts of the influencing factor sequence, judge the correlation, and finally use the trained CNN-BiLSTM-Attention model to predict the sub-displacement sequence data. The system and method can perform fast and accurate data processing and prediction. The empirical mode decomposition decomposes the data curve into a trend term and a periodic term, and uses a least squares vector machine for analysis and prediction respectively, which has the advantages of fast and high efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is the flow chart of the SSA-VMD model of the present invention.

[0052] Figure 2 It is the displacement prediction flow chart of the present invention.

[0053] Figure 3 This is a flow chart of the CNN-BiLSTM-Attention model of the present invention.

[0054] Figure 4 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0055] The following is a description of the implementation of the present invention by specific embodiments. People familiar with the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] like Figure 1-4 As shown, the landslide early warning system based on the SSA-VMD-CNN-BiLSTM-Attention model of the present invention includes a data monitoring and acquisition system, a data analysis system, and a discrimination and early warning system. The data monitoring and acquisition system realizes the monitoring, transmission and storage of the displacement and environmental information of the slope; the data analysis system realizes the analysis and prediction of the slope monitoring data; and the discrimination and early warning system realizes the judgment of the slope landslide and the release of early warning information.

[0057] The data monitoring and acquisition system includes displacement sensors, groundwater level sensors, humidity sensors and data transmission devices. Displacement sensors are arranged on the surface and inside of the slope to monitor slope displacement. Groundwater level sensors and humidity sensors monitor slope weather changes and water level changes caused by rainfall, providing a basis for inferring slope landslides. The data transmission device can transmit slope displacement monitoring data, water level monitoring data and humidity monitoring data to the monitoring station and cloud server for storage.

[0058] like Figure 2 As shown in the figure, the data analysis system includes the SSA-VMD model and the CNN-BiLSTM-Attention model. The main steps of the model to predict landslide displacement are as follows:

[0059] (1) The original landslide displacement time series is divided into three subsequences using the SSA-VMD model. The optimal decomposition subsequences are selected as the trend term displacement sequence, the periodic term displacement sequence, and the random term displacement sequence according to the sample entropy value of the periodic term.

[0060] (2) Decompose each influencing factor sequence into two subsequences through the SSA-VMD model, representing the low-frequency and high-frequency parts of the influencing factor sequence respectively, and obtain the optimal decomposition subsequence based on the sample entropy of the low-frequency part. Calculate the maximum mutual information coefficient (MIC) and gray relational grade (GRG) value of the decomposition subsequence of each factor and each displacement subsequence, and comprehensively judge the correlation between the influencing factor subsequence and the displacement subsequence;

[0061] (3) Based on the two subsequence input sequences of the influencing factors decomposed in (2), a single-factor CNN-BiLSTM-Attention model is constructed for the trend item displacement sequence, and a multi-factor CNN-BiLSTM-Attention model is constructed for the period item displacement sequence and the random item displacement sequence. The model is trained using the training data set and the validation data set is predicted;

[0062] (4) The final cumulative landslide displacement prediction result is the sum of the predicted values ​​of trend term displacement, period term displacement and random term displacement. The prediction performance of the model is verified and evaluated by comparing the cumulative landslide displacement monitoring sequence.

[0063] like Figure 3As shown in the figure, the CNN-BiLSTM-Attention model uses a CNN framework consisting of a double-layer one-dimensional convolution layer and a pooling layer to automatically extract the internal features of the displacement sequence. The convolution layer extracts effective nonlinear local features of the time series, and the pooling layer selects the maximum pooling method to compress the extracted features and generate more critical feature information. The BiLSTM hidden layer modeling learns the internal dynamic change law of the local features extracted by CNN, and iteratively extracts more complex global features from the local features. The features generated by the BiLSTM hidden layer are used as the input of the Attention mechanism. The attention mechanism is used to automatically distinguish the importance of the time information extracted by the BiLSTM hidden layer in a weighted manner, which can more effectively utilize the time series properties of the displacement time series itself and mine deep time correlations. The attention mechanism can effectively reduce the loss of historical information and highlight the information of key historical time points to weaken the impact of redundant information on the displacement prediction results. Then the output of the Attention layer is used as the input of the fully connected layer, and the final displacement prediction result is output through the fully connected layer.

[0064] The discrimination warning system includes a discrimination relationship and a warning notification system. The discrimination relationship is that if the slope of the slope displacement curve exceeds the preset value for a period of time, the warning notification system will be triggered. The warning notification system will transmit the information of the slope with landslide risk to the relevant units, release relevant information to the public, and take emergency measures.

[0065] Although the present invention has been described in detail above by general description and specific embodiments, it is obvious to those skilled in the art that some modifications or improvements can be made to the present invention. Therefore, these modifications or improvements made without departing from the spirit of the present invention all belong to the scope of protection claimed by the present invention.

Claims

1. A landslide early warning method based on modal decomposition and machine learning, characterized by: The data transmission device transmits the data monitored by the data monitoring and acquisition system to the cloud storage server; the data monitoring and acquisition system includes a displacement sensor, a groundwater level sensor, a humidity sensor and a data transmission device; the displacement sensor is arranged on the surface and inside of the mountain slope to monitor the displacement of the slope, the groundwater level sensor is arranged at the top, the body and the foot of the slope to monitor the water level of the slope, and the humidity sensor is arranged on the surface of the slope to monitor the weather conditions; The landslide displacement is predicted by the data analysis system set up at the monitoring station. The data analysis system includes the SSA-VMD model and the CNN-BiLSTM-Attention model. (1) The original landslide displacement time series is divided into three subsequences using the SSA-VMD model. The optimal decomposition subsequences are selected as the trend term displacement sequence, the periodic term displacement sequence, and the random term displacement sequence according to the sample entropy value of the periodic term. (2) Through the SSA-VMD model, each influencing factor sequence is decomposed into two subsequences, which represent the low-frequency and high-frequency parts of the influencing factor sequence respectively. The optimal decomposition subsequence is obtained according to the sample entropy value of the low-frequency part, and the maximum mutual information coefficient MIC and gray correlation GRG value of the decomposition subsequence of each factor and each displacement subsequence are calculated, and the correlation between the influencing factor subsequence and the displacement subsequence is comprehensively judged; (3) Based on the two subsequence input sequences of the influencing factors decomposed in (2), a single-factor CNN-BiLSTM-Attention model is constructed for the trend item displacement sequence, and a multi-factor CNN-BiLSTM-Attention model is constructed for the period item displacement sequence and the random item displacement sequence. The single-factor and multi-factor CNN-BiLSTM-Attention models are trained using the training data set and the validation data set is predicted; (4) The final cumulative landslide displacement prediction result is the sum of the predicted values ​​of the trend term displacement, the periodic term displacement and the random term displacement. The prediction performance of the method is verified and evaluated by comparing the cumulative landslide displacement monitoring sequence.

2. The landslide early warning method based on modal decomposition and machine learning according to claim 1 is characterized in that: The landslide displacement time series is expressed as follows according to the additive model of time series: X(t)=T(t)+P(t)+R(t) Where: X(t) is the time series displacement value; T(t) is the trend term function; P(t) is the period term function; R(t) is the random term function of uncertainty; The intrinsic mode function IMF of VMD in the SSA-VMD model is an amplitude modulation-frequency modulation signal expression as follows: u k (t)=A k (t)cos(φ k (t)) Where: u k (t) is the modal signal; phase φ k (t) is a non-decreasing function and is related to the instantaneous amplitude A k (t) are all non-negative; In the SSA-VMD model, VMD uses the input signal sequence and the sum of each mode as the constrained variation, and realizes adaptive decomposition by finding the optimal solution of the constrained variation model. Then its constrained variation expression is written as: Where: K is the number of required modal components, k is an integer between 1 and K; {u k }={u1,…,u k } is the modal component obtained by the final decomposition; {ω k }={ω1,…,ω k } is the actual center frequency of each modal component; is the symbol of partial derivative; δ(t) is the Dirac function; * is the convolution operator; To solve the above equation, the Lagrange operator λ is introduced to transform the constrained variational problem into an unconstrained variational problem, and the expanded Lagrange expression is obtained as follows: Where: ε is the decomposition parameter to reduce the interference of Gaussian noise; The alternating direction multiplier iteration algorithm is used to optimize the modal components and center frequencies, and the saddle point of the unconstrained model is searched to obtain the optimal solution of the constrained model, and then K modal components are obtained.

3. The landslide early warning method based on modal decomposition and machine learning according to claim 2 is characterized in that: The number of modal components K=2, and the penalty factor α and the rise time step τ in the VMD model are optimized using the sparrow optimization algorithm SSA, according to the formula fitness=RMSE(m,M)·SampEn(IMF2) As the fitness value of the SSA algorithm; Where RMSE(m, M) is the root mean square error between the reconstructed sequence and the original sequence; SampEn( / MF2) is the sample entropy value of the low-frequency part of the decomposed periodic term displacement sequence or influencing factors; in Where m is the embedding dimension, which is 1 or 2, and r is the similarity tolerance, which is 0.1 to 0.25σ. x , σ x is the standard deviation of the sequence, B m (r) is the probability that two sequences match m points under similar capacity r, B m+1 (r) is the probability of two sequences matching m+1 points; Where: x t is the value of the original sequence at time t, is the value of the reconstructed sequence at time t, and n is the length of the sequence; Optimize the penalty factor α and the rise time step τ to find the optimal fitness. The specific steps are as follows: (1) Initialize the parameters input to the sparrow optimization algorithm; (2) Randomly generate a series of penalty factors α and ramp-up time steps τ as the initial positions of the sparrow population; (3) Perform VMD decomposition on the displacement sequence at the current sparrow position, and calculate the sample entropy SampEn(IMF2) of the decomposed periodic term sequence or the low-frequency part of the influencing factor; (4) Obtain the new sparrow fitness value fitness, find the best fitness individual and the worst fitness individual, and update the positions of the discoverer, follower, and warning person; (5) Repeat (3) and (4) until the maximum number of iterations is reached, and output the sparrow position and fitness value at this time as the optimal solution.

4. The landslide early warning method based on modal decomposition and machine learning according to claim 1 is characterized in that: The CNN-BiLSTM-Attention model uses a CNN framework consisting of a double-layer one-dimensional convolution layer and a pooling layer to automatically extract the internal features of the displacement sequence. The convolution layer performs effective nonlinear local feature extraction of the time series. The pooling layer selects the maximum pooling method to compress the extracted features and generate more critical feature information. The BiLSTM hidden layer modeling learns the internal dynamic change law of the local features extracted by CNN, iteratively extracts more complex global features from the local features, and uses the features generated by the BiLSTM hidden layer as the input of the Attention mechanism. The Attention mechanism is used to automatically distinguish the importance of the time information extracted by the BiLSTM hidden layer in a weighted manner. The bidirectional long short-term memory neural network BiLSTM is an improvement on the traditional LSTM and consists of a forward and backward LSTM, both of which affect the results; A i =f1(ω1x i +ω2A i-1 ) B i =f2(ω3x i +ω5B i-1 ) Y i =f3(ω4A i +ω6B i ) Where: f1, f2, f3 are activation functions between different layers, x i Input word for time i, A i is the forward LSTM unit output value at time i, B i is the output value of the LSTM unit after time i, Y i is the final output result at time i, ω1 and ω2 are x i , A i-1 Input the weight of the forward LSTM unit at time i, ω3 and ω5 are x i , B i-1 Input the weights of the backward LSTM unit at time i, ω4 and ω6 are A i , B i Input the weight of the output layer at time i; The Attention mechanism assigns weights according to different features, and the attention state conversion process is as follows: S ti =Vtanh(Wh t +Uh i +b),i=1,2,3…,t-1 h′ t =f(F,h t ,y t ) Where: a ti Input value h to the BiLSTM hidden layer t The attention weights for the current input, y1, y2, y3, ..., y t is the input sequence, h1, h2, h3, ..., h t is the corresponding input sequence y1, y2, y3, ..., y t The hidden layer state value, h′ t is the final feature vector V, W, U, b is the learning parameter of the model, S ti To use the SoftMax function to process a ti Required input, F is h t The result of weighted average.

5. A landslide early warning system based on modal decomposition and machine learning, characterized by: The early warning system executes the method described in any one of claims 1 to 4, and the early warning system also includes a discrimination early warning system, which includes a discrimination relationship and a warning notification system. The discrimination relationship is that if the slope of the obtained slope displacement curve exceeds a preset value for a period of time, the warning notification system will be triggered when the discrimination relationship is met, and the information of the slope with landslide risk will be transmitted to relevant units, relevant information will be released to the public, and emergency measures will be taken to deal with it.

Images