Ground-based radar landslide monitoring and early warning method and system

Abstract

The present invention provides a ground-based radar landslide monitoring and early warning method and system, which relates to the technical field of landslide monitoring, including collecting monitoring signals and filtering them; extracting instantaneous phases and performing wavelet packet decomposition; obtaining enhanced features through morphological filtering; constructing monitoring feature vectors; performing nonlinear dimension reduction on feature time series matrices; obtaining evolution trajectories using fractional-order adaptive neural networks; constructing a hierarchical risk reasoning framework; obtaining risk metrics through multi-source information fusion; and using fuzzy evolutionary algorithms to perform dynamic evaluation and determine early warning levels. The method can improve the accuracy and real-time performance of landslide monitoring and early warning.

Description

Technical Field

[0001] The present invention relates to the technical field of landslide monitoring, and in particular to a ground-based radar landslide monitoring and early warning method and system. Background Art

[0002] Ground-based radar is a new type of non-contact landslide monitoring equipment, which can monitor the landslide in real time by emitting electromagnetic waves and receiving the target reflected signals. Compared with traditional monitoring methods, ground-based radar has the advantages of all-weather, large range, high precision and real-time performance, and can effectively obtain key information such as displacement and deformation of the landslide.

[0003] However, the existing ground-based radar landslide monitoring methods still have shortcomings. It is difficult to effectively deal with noise and interference in the signal, which reduces the accuracy of feature extraction; based on a single time domain or frequency domain analysis, it is unable to fully reflect the dynamic evolution characteristics of the landslide body; the fixed threshold method is used, and the dynamic changes of environmental factors and geological conditions are not considered, resulting in insufficient reliability of the early warning results; there is a lack of effective fusion of multi-source information, making it difficult to accurately assess the landslide risk; the description of the evolution law of the landslide body is not precise enough, making it difficult to accurately predict the development trend of the landslide.

[0004] In summary, there is an urgent need for a ground-based radar landslide monitoring and early warning method based on multi-level information fusion to solve the technical problems in the prior art such as inaccurate signal processing, incomplete feature extraction, single early warning model, and insufficient information fusion, and to improve the accuracy and reliability of landslide monitoring and early warning. The present invention can solve the problems in the prior art. Summary of the invention

[0005] The embodiment of the present invention provides a ground-based radar landslide monitoring and early warning method and system, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A ground-based radar landslide monitoring and early warning method is provided, comprising:

[0008] The ground-based radar collects the monitoring signal of the landslide body, obtains the filtered signal through filtering processing with a Butterworth filter, extracts the instantaneous phase through the Hilbert transform of the filtered signal, and performs five-layer decomposition based on the instantaneous phase using a wavelet packet decomposition algorithm to obtain the low-frequency characteristic coefficient and the high-frequency characteristic coefficient of the filtered signal; based on the low-frequency characteristic coefficient and the high-frequency characteristic coefficient, an enhanced feature is obtained through a morphological filter; the enhanced feature is subjected to a discrete Fourier transform to obtain the phase stability coefficient, the amplitude continuity coefficient and the signal-to-noise ratio coefficient, and a landslide body monitoring feature vector is constructed;

[0009] The monitoring feature vector of the landslide body is divided into multiple time windows in time sequence to obtain a landslide body feature time series matrix, a variational autoencoder is used to perform nonlinear dimension reduction on the landslide body feature time series matrix to obtain the landslide body state characteristics, the landslide body state characteristics are input into a fractional-order adaptive neural network, the evolution trajectory of the landslide body state is obtained by solving a fractional-order differential equation, a landslide body state evolution model is established based on the evolution trajectory, and the landslide body evolution law parameters and current state parameters are obtained;

[0010] A hierarchical risk reasoning framework is constructed, and deep Copper network modeling is performed on the evolution law parameters and current state parameters of the landslide body to obtain the probability density function of the landslide body displacement. The probability density function is combined with the pre-acquired geological parameters and environmental parameters, and multi-source information is fused through evidence theory to obtain the landslide risk measurement value. The fuzzy evolutionary algorithm is used to construct an adaptive risk threshold function and embed it into the Monte Carlo tree search optimization to dynamically evaluate the landslide risk measurement value, determine the warning level and generate warning information.

[0011] In an optional embodiment,

[0012] The ground-based radar collects the monitoring signal of the landslide body, and obtains the filtered signal through filtering processing with a Butterworth filter. The filtered signal is subjected to Hilbert transform to extract the instantaneous phase. The wavelet packet decomposition algorithm is used to perform five-layer decomposition based on the instantaneous phase to obtain the low-frequency characteristic coefficient and the high-frequency characteristic coefficient of the filtered signal; based on the low-frequency characteristic coefficient and the high-frequency characteristic coefficient, an enhanced feature is obtained through a morphological filter; the enhanced feature is subjected to discrete Fourier transform to obtain the phase stability coefficient, the amplitude continuity coefficient and the signal-to-noise ratio coefficient, and the landslide body monitoring feature vector is constructed, including:

[0013] A landslide monitoring signal is obtained by using a ground-based radar, and the landslide monitoring signal is filtered by a Butterworth filter to obtain a filtered signal by suppressing high-frequency noise and interference signals;

[0014] Performing phase delay processing on the filtered signal to obtain a delayed signal, constructing an analytical signal according to the filtered signal and the delayed signal, calculating the argument of the analytical signal, and obtaining an instantaneous phase signal;

[0015] Based on the instantaneous phase signal, wavelet packet decomposition is adopted, and five-layer decomposition is recursively performed through high-pass filtering and low-pass filtering to sequentially generate multi-layer frequency band coefficients, and based on the fifth layer of frequency band coefficients in the multi-layer frequency band coefficients, coefficients with frequencies less than a preset dividing frequency are reconstructed into low-frequency characteristic coefficients, and coefficients with frequencies greater than or equal to the preset dividing frequency are reconstructed into high-frequency characteristic coefficients;

[0016] The frequency band coefficients are screened based on energy entropy and dynamic threshold, and optimized low-frequency feature sequences and optimized high-frequency feature sequences are reconstructed;

[0017] Performing an erosion operation on the optimized low-frequency feature sequence and the optimized high-frequency feature sequence to obtain a first intermediate sequence; performing an expansion operation on the first intermediate sequence to obtain an open sequence; performing an expansion operation on the open sequence to obtain a second intermediate sequence, and performing an erosion operation on the second intermediate sequence to obtain an enhanced feature sequence; performing a discrete Fourier transform on the enhanced feature sequence to obtain a frequency domain transform sequence;

[0018] Calculate the phase sequence variance of the frequency domain transformation sequence to obtain a phase stability coefficient, calculate the ratio of the standard deviation to the mean of the amplitude sequence of the frequency domain transformation sequence to obtain an amplitude continuity coefficient; calculate the logarithmic ratio of the signal power to the noise power in the frequency domain transformation sequence to obtain a signal-to-noise ratio coefficient;

[0019] The phase stability coefficient, the amplitude continuity coefficient and the signal-to-noise ratio coefficient are combined to construct a landslide monitoring feature vector.

[0020] In an optional embodiment,

[0021] The frequency band coefficients are screened based on energy entropy and dynamic threshold, and the optimized low-frequency feature sequence and optimized high-frequency feature sequence are reconstructed, including:

[0022] Calculating the sum of the squares of the low-frequency characteristic coefficient and the high-frequency characteristic coefficient, and performing normalization processing to obtain a normalized energy value; calculating the information entropy of the normalized energy value to obtain energy entropy;

[0023] The energy entropy is divided into multiple time segments according to a preset length, and the mean and standard deviation of the corresponding energy entropy in each time segment are calculated; the mean is multiplied by a first weight coefficient to obtain a first weighted value, and the standard deviation is multiplied by a second weight coefficient to obtain a second weighted value, and the first weighted value and the second weighted value are added to construct a dynamic threshold;

[0024] Compare the energy entropy of each frequency band with the dynamic threshold, retain the frequency band coefficients whose energy entropy is greater than the dynamic threshold, and set the frequency band coefficients whose energy entropy is less than or equal to the dynamic threshold to zero;

[0025] The retained frequency band characteristic coefficients are reconstructed by wavelet packets to obtain optimized low-frequency characteristic sequences and optimized high-frequency characteristic sequences.

[0026] In an optional embodiment,

[0027] The monitoring feature vector of the landslide body is divided into multiple time windows in time sequence to obtain a landslide body feature time series matrix. The landslide body feature time series matrix is ​​subjected to nonlinear dimension reduction by a variational autoencoder to obtain the state characteristics of the landslide body. The state characteristics of the landslide body are input into a fractional-order adaptive neural network. The evolution trajectory of the landslide body state is obtained by solving a fractional-order differential equation. A landslide body state evolution model is established based on the evolution trajectory to obtain the landslide body evolution law parameters and current state parameters, including:

[0028] The landslide monitoring feature vector is divided according to a preset time window length and a preset sliding step length to obtain a plurality of time window sequences, the landslide monitoring feature vectors in each time window sequence are combined into a feature vector sequence according to a time series relationship, and all the feature vector sequences are spliced ​​by column to construct a landslide feature time series matrix;

[0029] Constructing a variational autoencoder, setting a multi-layer fully connected network as an encoder in the variational autoencoder, inputting the landslide body feature time series matrix into the encoder, and outputting a mean vector and a variance vector in the last layer of the encoder respectively;

[0030] A noise vector is obtained by randomly sampling a standard normal distribution, and the noise vector is multiplied by an exponential function of the variance vector and then added to the mean vector to obtain a reparameterized landslide state characteristic;

[0031] The re-parameterized landslide state characteristics are input into the fractional-order adaptive neural network, and the fractional-order differential equation is obtained through the synergy of the fractional-order calculation layer and the memory mapping layer.

[0032] The fractional-order differential equation is solved by using a prediction-correction format to obtain a numerical solution, wherein the prediction-correction format represents the state at a current moment as a linear combination of an initial state and a weighted sum of historical states; and the evolution trajectory of the landslide body state is obtained according to the numerical solution of the fractional-order differential equation;

[0033] Based on the structure of the fractional-order adaptive neural network, the parameters of the landslide evolution law are determined;

[0034] A hidden Markov model is constructed based on the evolution trajectory, and a state transition probability matrix is ​​extracted from the hidden Markov model; current monitoring data is substituted into the hidden Markov model, and the posterior probability distribution of the landslide body in various states is calculated based on the state transition probability matrix, and the state corresponding to the maximum posterior probability is determined as the current state parameter of the landslide body.

[0035] In an optional embodiment,

[0036] The fractional-order adaptive neural network comprises:

[0037] Fractional-order computation layers and memory mapping layers connected in sequence;

[0038] The fractional-order calculation layer receives input of the state characteristics of the landslide body, and constructs a fractional-order derivative operator based on historical state weighting in the fractional-order calculation layer, wherein the fractional-order derivative operator represents the state characteristics of the landslide body at the current moment as a weighted integral of the historical state characteristics, and the weight coefficient in the weighted integral is determined based on the gamma function and the fractional order to be optimized;

[0039] The memory mapping layer receives the weighted integral output by the fractional-order calculation layer, adopts a recursive orthogonal polynomial function in the memory mapping layer, constructs a plurality of memory kernel functions to extract features from the weighted integral, and sets the coefficients of the plurality of memory kernel functions as parameters to be optimized;

[0040] Introducing fractional-order feature neurons in the memory mapping layer, the fractional-order feature neurons receive the output of the memory kernel function, and use fractional-order power series as an activation function for nonlinear mapping, the power of each power term in the fractional-order power series is determined based on the product of the fractional order and an integer, and the coefficient of the fractional-order power series is set as a parameter to be optimized;

[0041] An optimization objective function is constructed, wherein the optimization objective function includes a reconstruction error term between the state characteristics of the landslide body and the landslide body characteristic time series matrix, and a prediction error term between the state characteristics and the actual historical state. The optimization objective function is iteratively optimized by a gradient descent method to obtain optimized values ​​of the fractional orders in the fractional-order calculation layer, optimized values ​​of the memory kernel function coefficients, and optimized values ​​of the fractional-order power series coefficients. Based on the optimized values, the network parameters of the fractional-order adaptive neural network are determined.

[0042] In an optional embodiment,

[0043] The deep Copper network modeling is carried out on the parameters of landslide evolution law and current state parameters to obtain the probability density function of landslide displacement, including:

[0044] The deep Copula network comprises a parameter encoding layer, an edge distribution modeling layer and a joint distribution construction layer;

[0045] The parameter encoding layer receives the input of the landslide body evolution law parameters and the current state parameters, and performs feature extraction and dimensionality reduction processing on the landslide body evolution law parameters and the current state parameters through a multi-layer deep neural network to obtain a low-dimensional feature representation of the parameters;

[0046] The edge distribution modeling layer receives the low-dimensional feature representation, constructs a bandwidth-adaptive non-parametric kernel density estimator for each dimension, and the bandwidth parameter of the non-parametric kernel density estimator is selected from a preset candidate bandwidth set by a cross-validation method to obtain the edge probability distribution of each dimensional feature;

[0047] The joint distribution construction layer receives the marginal probability distribution of each dimension, constructs a parameterized Coppola function family, combines the marginal probability distribution, determines the parameters of the Coppola function family based on historical sample data through the maximum likelihood estimation method, and outputs the joint probability distribution of the landslide body displacement parameters; based on the joint probability distribution, a numerical integration method is used to calculate the probability density function of the landslide body displacement within a preset time interval.

[0048] In an optional embodiment,

[0049] The fuzzy evolutionary algorithm is used to construct an adaptive risk threshold function and embed Monte Carlo tree search optimization to dynamically evaluate the landslide risk metric value, including:

[0050] Based on the preset landslide body evolution index, the key points of landslide body evolution are determined. Based on the key points of landslide body evolution, the pre-collected historical warning data are segmented. For each segment, a piecewise continuous function is fitted using a cubic spline interpolation method. The function coefficient of the piecewise continuous function is used as an initial threshold parameter to construct an adaptive risk threshold function.

[0051] Based on the historical warning data, a fuzzy rule set is extracted to determine the mapping relationship between the risk metric value and the warning level; the warning accuracy of the fuzzy rule set is used as the fitness function, and the parameters of the membership function in the fuzzy rule set are optimized by a genetic algorithm to obtain an optimized fuzzy rule base; the landslide risk metric value is input into a fuzzy inference system based on the optimized fuzzy rule base to obtain a fuzzy decision set of the warning level;

[0052] A multi-layer decision tree is constructed, the warning levels in the fuzzy decision set are set as decision nodes, and the historical risk losses corresponding to each warning level are set as the profit function; Monte Carlo sampling is performed at each decision node to generate state sequence samples, the upper confidence interval value of each decision branch is calculated, and the decision branch corresponding to the maximum value of the upper confidence interval value is selected as the optimal decision path, and the warning decision sequence corresponding to the optimal decision path is extracted to determine the warning level and generate warning information.

[0053] According to a second aspect of the embodiments of the present invention,

[0054] Provided is a ground-based radar landslide monitoring and early warning system, comprising:

[0055] The first unit is used for collecting landslide monitoring signals by ground-based radar, filtering the signals through Butterworth filter to obtain filtered signals, extracting instantaneous phases through Hilbert transform, and performing five-layer decomposition using wavelet packet decomposition algorithm based on the instantaneous phases to obtain low-frequency characteristic coefficients and high-frequency characteristic coefficients of the filtered signals; obtaining enhanced features through morphological filters based on the low-frequency characteristic coefficients and the high-frequency characteristic coefficients; performing discrete Fourier transform on the enhanced features to obtain phase stability coefficients, amplitude continuity coefficients and signal-to-noise ratio coefficients, and constructing landslide monitoring feature vectors;

[0056] The second unit is used to divide the landslide monitoring feature vector into multiple time windows in time sequence to obtain the landslide feature time series matrix, use a variational autoencoder to perform nonlinear dimension reduction on the landslide feature time series matrix to obtain the landslide state characteristics, input the landslide state characteristics into a fractional-order adaptive neural network, obtain the evolution trajectory of the landslide state by solving the fractional-order differential equation, establish a landslide state evolution model based on the evolution trajectory, and obtain the landslide evolution law parameters and current state parameters;

[0057] The third unit is used to construct a hierarchical risk reasoning framework, perform deep Copper network modeling on the landslide evolution law parameters and current state parameters, obtain the probability density function of the landslide displacement, combine the probability density function with the pre-acquired geological parameters and environmental parameters, and fuse multi-source information through evidence theory to obtain the landslide risk measurement value. The fuzzy evolution algorithm is used to construct an adaptive risk threshold function and embed the Monte Carlo tree search optimization to dynamically evaluate the landslide risk measurement value, determine the warning level and generate warning information.

[0058] According to a third aspect of the embodiments of the present invention,

[0059] An electronic device is provided, comprising:

[0060] processor;

[0061] a memory for storing processor-executable instructions;

[0062] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0063] According to a fourth aspect of the embodiments of the present invention,

[0064] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0065] In the embodiment of the present invention, a variety of advanced signal processing methods such as Butterworth filter, Hilbert transform, wavelet packet decomposition algorithm and morphological filter are used to effectively extract the low-frequency and high-frequency characteristics of the landslide monitoring signal. The phase stability coefficient, amplitude continuity coefficient and signal-to-noise ratio coefficient are obtained by discrete Fourier transform, and a comprehensive landslide monitoring feature vector is constructed. This multi-level signal processing and feature extraction method significantly improves the quality and reliability of landslide monitoring data; the variational autoencoder is introduced for nonlinear dimension reduction, and the landslide state evolution model is innovatively established by combining fractional-order adaptive neural network and fractional-order differential equation. This method can not only effectively capture the nonlinear change characteristics of the landslide state, but also accurately predict its evolution trajectory, greatly improving the accuracy and prediction ability of landslide monitoring; a hierarchical risk reasoning framework is constructed, and advanced methods such as deep Copula network, evidence theory and fuzzy evolution algorithm are used to realize the dynamic assessment of landslide risk. By integrating multi-source information and adopting adaptive risk threshold function, combined with Monte Carlo tree search optimization, the present invention establishes an intelligent and precise landslide early warning system. This not only improves the accuracy of the early warning, but also can dynamically adjust the warning level according to real-time conditions, significantly enhancing the practicality and reliability of landslide monitoring and early warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a flow chart of a ground-based radar landslide monitoring and early warning method according to an embodiment of the present invention;

[0067] Figure 2 It is a schematic diagram of the structure of a ground-based radar landslide monitoring and early warning system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only 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.

[0069] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0070] Figure 1 FIG. 1 is a flow chart of a ground-based radar landslide monitoring and early warning method according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0071] S101. The ground-based radar collects the monitoring signal of the landslide body, obtains the filtered signal through filtering processing with a Butterworth filter, extracts the instantaneous phase through the Hilbert transform of the filtered signal, and performs five-layer decomposition based on the instantaneous phase using a wavelet packet decomposition algorithm to obtain the low-frequency characteristic coefficient and the high-frequency characteristic coefficient of the filtered signal; obtains the enhanced feature through a morphological filter based on the low-frequency characteristic coefficient and the high-frequency characteristic coefficient; performs discrete Fourier transform on the enhanced feature to obtain the phase stability coefficient, the amplitude continuity coefficient and the signal-to-noise ratio coefficient, and constructs the landslide body monitoring feature vector;

[0072] S102. Divide the monitoring feature vector of the landslide body into multiple time windows in chronological order to obtain a landslide body feature time series matrix, use a variational autoencoder to perform nonlinear dimension reduction on the landslide body feature time series matrix to obtain landslide body state characteristics, input the landslide body state characteristics into a fractional-order adaptive neural network, obtain the evolution trajectory of the landslide body state by solving the fractional-order differential equation, establish a landslide body state evolution model based on the evolution trajectory, and obtain the landslide body evolution law parameters and current state parameters;

[0073] S103. Construct a hierarchical risk reasoning framework, perform deep Copper network modeling on the parameters of landslide evolution law and current state parameters, obtain the probability density function of landslide displacement, combine the probability density function with pre-acquired geological parameters and environmental parameters, fuse multi-source information through evidence theory, and obtain the landslide risk measurement value. Use fuzzy evolution algorithm, construct an adaptive risk threshold function and embed Monte Carlo tree search optimization, dynamically evaluate the landslide risk measurement value, determine the warning level and generate warning information.

[0074] The Butterworth filter specifically refers to a signal processing filter with a smooth design, known for its ripple-free and maximum smoothness. Its response curve is flat in the passband, without any ripples ("maximum flatness" characteristics), and gradually decays to zero. In landslide monitoring, the Butterworth filter is used to smooth the monitoring signal, remove high-frequency noise components, retain the main low-frequency components of the signal, and obtain a cleaner "filtered signal" for subsequent analysis.

[0075] The Hilbert transform specifically refers to a mathematical operation used to convert a real signal into a complex signal so that the instantaneous phase and instantaneous frequency information can be extracted from it. Specifically, the Hilbert transform can generate an imaginary signal, which is combined with the original signal (real part) to form a complex signal. The phase of this complex signal is the instantaneous phase. In landslide monitoring, the Hilbert transform helps extract instantaneous phase information, which is used to analyze the phase change and stability of the landslide body signal, and then determine the characteristics of the landslide activity.

[0076] The morphological filter specifically refers to the feature enhancement of the signal through the structural element. Through a series of operations such as corrosion, expansion, opening operation, closing operation, etc., the geometric shape characteristics of the signal are extracted, noise is suppressed and the characteristics are enhanced. In the landslide monitoring signal processing, the morphological filter enhances the signal characteristics based on the low-frequency and high-frequency characteristic coefficients, so that the landslide characteristics can be more easily identified in the subsequent analysis.

[0077] The deep Coppola network specifically refers to a probabilistic modeling tool that combines Coppola theory with deep learning methods. It is used to establish complex dependencies between multiple variables and process multivariate data with nonlinear and non-Gaussian distributions. In Coppola theory, the Coppola function is used to combine marginal distributions into a joint distribution, thereby capturing the dependency structure between variables without relying on the specific marginal distribution shape of each variable. In the application of landslide monitoring, the deep Coppola network captures the dependencies between these variables by modeling the evolution law of the landslide and the current state parameters, thereby generating a probability density function of the landslide displacement. The probability density function represents the probability distribution of the landslide displacement under specific conditions, which helps to quantify the risk level of the landslide and is used for subsequent multi-source information fusion and risk metric evaluation.

[0078] In an optional specific implementation, first, the landslide monitoring signal is collected by ground-based radar. A high-frequency radar system with a frequency range of 10-40 GHz and a sampling rate of 1000 Hz is used to obtain the reflection signal on the surface of the landslide.

[0079] Next, the collected original signal is processed by Butterworth filtering. A 4th-order Butterworth filter is selected, and the cutoff frequency is set to 100 Hz to effectively remove high-frequency noise and obtain the filtered signal.

[0080] Then, the filtered signal is Hilbert transformed to extract the instantaneous phase information. The change of the phase angle over time is obtained by calculating the analytical signal of the filtered signal.

[0081] Based on the extracted instantaneous phase, the wavelet packet decomposition algorithm is used for five-layer decomposition. The db4 wavelet basis is selected to decompose the signal into 32 frequency bands to obtain low-frequency characteristic coefficients and high-frequency characteristic coefficients. The low-frequency coefficient reflects the overall displacement trend of the landslide body, and the high-frequency coefficient reflects the local deformation characteristics.

[0082] The decomposed characteristic coefficients are enhanced by morphological filters. The combination of opening and closing operations is adopted, and the structural element is a 3×3 rectangle, which effectively suppresses noise and retains useful information to obtain enhanced features.

[0083] Perform discrete Fourier transform on the enhanced features and calculate the phase stability coefficient, amplitude continuity coefficient and signal-to-noise ratio coefficient. The phase stability coefficient characterizes the consistency of phase change, the amplitude continuity coefficient reflects the continuity of signal strength, and the signal-to-noise ratio coefficient measures the ratio of useful signal to noise. These three coefficients are combined to construct the landslide monitoring feature vector.

[0084] The landslide monitoring feature vector is divided into multiple time windows in chronological order. Each window contains 100 sampling points and the overlapping rate of adjacent windows is 50%, forming a landslide feature time series matrix.

[0085] The variational autoencoder is used to perform nonlinear dimensionality reduction on the landslide feature time series matrix. Both the encoder and decoder use a three-layer fully connected neural network with 64, 32, and 16 hidden layer neurons and a potential space dimension of 8 to obtain low-dimensional landslide state characteristics.

[0086] The state characteristics of the landslide body after dimensionality reduction are input into the fractional-order adaptive neural network. The network structure contains 5 hidden layers, with 32, 64, 128, 64, and 32 neurons in each layer, and ReLU is used as the activation function. The evolution trajectory of the landslide body state is obtained by solving the fractional-order differential equation.

[0087] Based on the obtained evolution trajectory, the landslide state evolution model is established. The polynomial fitting method is used to obtain the parameters describing the evolution law of the landslide, including displacement rate, acceleration, etc.; at the same time, the current state parameters such as instantaneous displacement and cumulative displacement are extracted.

[0088] A hierarchical risk reasoning framework was constructed to model the parameters of landslide evolution and current state using a deep Coppola network. The network structure contains three hidden layers, with 64, 128, and 64 neurons in each layer, and the t-Coppola function was used as the connection function. Through network training, the probability density function of landslide displacement was obtained.

[0089] The obtained probability density function is combined with the pre-acquired geological parameters (such as rock and soil strength, structural characteristics) and environmental parameters (such as rainfall, groundwater level), and multi-source information is fused through evidence theory. The Dempster combination rule is used to calculate the basic probability distribution of each evidence source and obtain the landslide risk metric.

[0090] The fuzzy evolutionary algorithm is used to construct an adaptive risk threshold function. The triangle membership function is used to divide the risk metric into three levels: low, medium, and high. At the same time, the Monte Carlo tree search optimization algorithm is embedded to optimize the threshold parameters through 1000 iterations.

[0091] Finally, the landslide risk metric is dynamically evaluated and compared with the optimized risk threshold to determine the warning level (such as safety, attention, warning, danger), and generate corresponding warning information, including warning time, location, level and recommended measures.

[0092] In this embodiment, the quality and reliability of landslide monitoring data are effectively improved through multi-level signal processing and feature extraction. By using methods such as Butterworth filtering, Hilbert transform, and wavelet packet decomposition, noise can be effectively removed and key features can be extracted, laying the foundation for subsequent analysis; the introduction of fractional-order adaptive neural networks and deep Copula networks can achieve accurate modeling of the state evolution of the landslide body and accurate estimation of risk probability. This deep learning-based method can capture the nonlinear characteristics and complex dependencies in the landslide process, improving the accuracy of prediction; a hierarchical risk reasoning framework is constructed, combining evidence theory and fuzzy evolutionary algorithms to achieve the fusion of multi-source information and dynamic risk assessment. Through adaptive thresholds and Monte Carlo tree search optimization, the early warning system can be flexibly adjusted according to actual conditions, improving the timeliness and reliability of the early warning.

[0093] In an optional implementation, a ground-based radar collects a landslide monitoring signal, obtains a filtered signal through filtering with a Butterworth filter, extracts an instantaneous phase through a Hilbert transform, and performs a five-layer decomposition using a wavelet packet decomposition algorithm based on the instantaneous phase to obtain a low-frequency characteristic coefficient and a high-frequency characteristic coefficient of the filtered signal; based on the low-frequency characteristic coefficient and the high-frequency characteristic coefficient, an enhanced feature is obtained through a morphological filter; a discrete Fourier transform is performed on the enhanced feature to obtain a phase stability coefficient, an amplitude continuity coefficient, and a signal-to-noise ratio coefficient, and constructing a landslide monitoring feature vector includes:

[0094] A landslide monitoring signal is obtained by using a ground-based radar, and the landslide monitoring signal is filtered by a Butterworth filter to obtain a filtered signal by suppressing high-frequency noise and interference signals;

[0095] Performing phase delay processing on the filtered signal to obtain a delayed signal, constructing an analytical signal according to the filtered signal and the delayed signal, calculating the argument of the analytical signal, and obtaining an instantaneous phase signal;

[0096] Based on the instantaneous phase signal, wavelet packet decomposition is adopted, and five-layer decomposition is recursively performed through high-pass filtering and low-pass filtering to sequentially generate multi-layer frequency band coefficients, and based on the fifth layer of frequency band coefficients in the multi-layer frequency band coefficients, coefficients with frequencies less than a preset dividing frequency are reconstructed into low-frequency characteristic coefficients, and coefficients with frequencies greater than or equal to the preset dividing frequency are reconstructed into high-frequency characteristic coefficients;

[0097] The frequency band coefficients are screened based on energy entropy and dynamic threshold, and optimized low-frequency feature sequences and optimized high-frequency feature sequences are reconstructed;

[0098] Performing an erosion operation on the optimized low-frequency feature sequence and the optimized high-frequency feature sequence to obtain a first intermediate sequence; performing an expansion operation on the first intermediate sequence to obtain an open sequence; performing an expansion operation on the open sequence to obtain a second intermediate sequence, and performing an erosion operation on the second intermediate sequence to obtain an enhanced feature sequence; performing a discrete Fourier transform on the enhanced feature sequence to obtain a frequency domain transform sequence;

[0099] Calculate the phase sequence variance of the frequency domain transformation sequence to obtain a phase stability coefficient, calculate the ratio of the standard deviation to the mean of the amplitude sequence of the frequency domain transformation sequence to obtain an amplitude continuity coefficient; calculate the logarithmic ratio of the signal power to the noise power in the frequency domain transformation sequence to obtain a signal-to-noise ratio coefficient;

[0100] The phase stability coefficient, the amplitude continuity coefficient and the signal-to-noise ratio coefficient are combined to construct a landslide monitoring feature vector.

[0101] The argument specifically refers to a complex phase angle, which represents the angle of the signal on the complex plane, that is, the phase in the polar coordinate form of the complex number. For a complex signal z(t) = x(t) + jy(t), its argument is θ=arctan(y(t) / x(t)). In signal processing, the argument is used to describe the instantaneous phase change of the signal and help reveal the phase information of the signal. Extracting the argument of the landslide monitoring signal can derive the instantaneous phase, and then analyze the phase stability of the signal, which helps to evaluate the stability and characteristics of the landslide activity.

[0102] The erosion operation specifically refers to a mathematical morphological operation used to "shrink" or "erode" the edge of an image or signal. Specifically, erosion shrinks the high-value area of ​​the signal, eliminates isolated protrusions or noise points, and makes the overall shape smoother. The erosion operation is used in landslide signal processing to suppress noise and optimize the morphology of feature sequences, extract a purer signal contour, and provide a basis for subsequent enhanced feature analysis.

[0103] The dilation operation specifically refers to the inverse operation of the erosion operation, which is also a type of mathematical morphology. Dilation "expands" the high-value area of ​​the signal, fills the low-value area or holes, and makes the overall signal characteristics more prominent. Dilation can amplify the continuous high-value area in the signal and separate it more clearly from other areas. In the processing of landslide monitoring signals, the dilation operation is used to expand the signal characteristics after erosion to enhance the significant features, so as to facilitate more effective extraction of landslide feature information.

[0104] In an optional specific implementation, a ground-based radar is used to obtain a landslide monitoring signal. The ground-based radar monitors the deformation of the landslide by emitting electromagnetic waves and receiving reflected signals. An appropriate radar frequency and bandwidth are selected, and a C-band or X-band radar is usually used, with a frequency range of 4-12 GHz. The radar antenna scans toward the landslide to collect data, the sampling frequency is set to 100 Hz, and each scan time is 60 seconds to obtain an original monitoring signal containing the reflection information of the landslide.

[0105] Next, the original monitoring signal is preprocessed. A 6th-order Butterworth filter is used with a cutoff frequency of 40 Hz to filter the signal to remove high-frequency noise. The signal obtained after filtering is smoother, which is conducive to subsequent processing.

[0106] Then, the filtered signal is Hilbert transformed to extract the instantaneous phase of the signal. First, the filtered signal is delayed by 90 degrees to obtain a delayed signal, and the original signal is used as the real part and the delayed signal as the imaginary part to construct the analytical signal. The instantaneous phase is obtained by calculating the argument of the analytical signal, and the phase range is -π to π.

[0107] Based on the extracted instantaneous phase, the wavelet packet decomposition algorithm is used to perform five-layer decomposition. Using the db4 wavelet basis function, the signal is decomposed into 32 frequency bands in turn. The 32 frequency band coefficients of the fifth layer are reconstructed, and the frequency bands with frequencies below 10Hz are reconstructed into low-frequency characteristic coefficients, and the frequency bands with frequencies higher than or equal to 10Hz are reconstructed into high-frequency characteristic coefficients.

[0108] The low-frequency and high-frequency characteristic coefficients are optimized and screened. The energy entropy of each frequency band is calculated, and the frequency bands with energy entropy less than 0.5 are retained, and the frequency bands with energy entropy greater than 0.5 are discarded. At the same time, the dynamic threshold is set to 1.5 times the signal mean, and the coefficients are binarized. After screening and threshold processing, the optimized low-frequency and high-frequency characteristic sequences are obtained.

[0109] Next, the morphological filter is used to enhance the feature sequence. First, the erosion operation is performed, and the structural element is a 3×3 rectangle to obtain the first intermediate sequence. Then the first intermediate sequence is expanded to obtain the opening sequence, and the expansion and erosion operations are continued to obtain the enhanced feature sequence.

[0110] Perform a 512-point discrete Fourier transform on the enhanced feature sequence to obtain the frequency domain transform sequence. Calculate the phase variance of the frequency domain sequence to obtain the phase stability coefficient. Calculate the ratio of the standard deviation to the mean of the amplitude sequence to obtain the amplitude continuity coefficient. Calculate the logarithmic ratio of the signal power to the noise power to obtain the signal-to-noise ratio coefficient.

[0111] Finally, the calculated phase stability coefficient, amplitude continuity coefficient and signal-to-noise ratio coefficient are combined into a three-dimensional feature vector as the characteristic representation of landslide monitoring. For example, the feature vector obtained in a certain monitoring is [0.85, 0.92, 15.6].

[0112] In this embodiment, the key features of the landslide monitoring signal are extracted through multi-level signal processing, which improves the representativeness and distinctiveness of the features; morphological filtering is used to enhance the useful signal, suppress noise interference, and improve the reliability of the features; the constructed three-dimensional feature vector comprehensively characterizes the stability, continuity and signal-to-noise ratio of the signal, and can effectively reflect the deformation state of the landslide body.

[0113] In an optional implementation, the frequency band coefficients are screened based on energy entropy and dynamic thresholds, and the optimized low-frequency feature sequence and the optimized high-frequency feature sequence are reconstructed, including:

[0114] Calculating the sum of the squares of the low-frequency characteristic coefficient and the high-frequency characteristic coefficient, and performing normalization processing to obtain a normalized energy value; calculating the information entropy of the normalized energy value to obtain energy entropy;

[0115] The energy entropy is divided into multiple time segments according to a preset length, and the mean and standard deviation of the corresponding energy entropy in each time segment are calculated; the mean is multiplied by a first weight coefficient to obtain a first weighted value, and the standard deviation is multiplied by a second weight coefficient to obtain a second weighted value, and the first weighted value and the second weighted value are added to construct a dynamic threshold;

[0116] Compare the energy entropy of each frequency band with the dynamic threshold, retain the frequency band coefficients whose energy entropy is greater than the dynamic threshold, and set the frequency band coefficients whose energy entropy is less than or equal to the dynamic threshold to zero;

[0117] The retained frequency band characteristic coefficients are reconstructed by wavelet packets to obtain optimized low-frequency characteristic sequences and optimized high-frequency characteristic sequences.

[0118] In an optional specific implementation, first, obtain low-frequency characteristic coefficients and high-frequency characteristic coefficients. These characteristic coefficients can be obtained by performing wavelet packet decomposition on the original signal. For example, for a signal with a length of 1024, 3 layers of wavelet packet decomposition can be performed to obtain coefficients of 8 frequency bands.

[0119] Next, the square sum of each frequency band coefficient is calculated and normalized to obtain a normalized energy value. Specifically, for the ith frequency band, its normalized energy value E_i can be expressed as the square sum of the frequency band coefficient divided by the sum of the square sums of all frequency band coefficients. For example, assuming there are 8 frequency bands, the square sum of the coefficients of the first frequency band is 100, and the square sum of the coefficients of all frequency bands is 1000, then the normalized energy value of the first frequency band E_1 = 100 / 1000 = 0.1.

[0120] Then, the information entropy of the normalized energy value is calculated to obtain the energy entropy. For each frequency band i, its information entropy H_i can be calculated by the normalized energy value E_i. For example, assuming E_1 = 0.1, H_1 can be obtained by multiplying E_1 by its natural logarithm.

[0121] The calculated energy entropy is divided into multiple time segments according to a preset length. For example, for a signal with a length of 1024, each 128 points can be divided into a time segment, resulting in 8 time segments. For each time segment, the mean and standard deviation of the corresponding energy entropy are calculated.

[0122] For each time segment, the first weighted value is obtained by multiplying the mean by the first weight coefficient α, and the second weighted value is obtained by multiplying the standard deviation by the second weight coefficient β. The two are added together to construct a dynamic threshold. α and β can be adjusted according to actual needs, for example, α=0.8 and β=0.2. Assuming that the mean value of energy entropy in a time segment is 0.5 and the standard deviation is 0.1, the dynamic threshold of the time segment is 0.8*0.5 + 0.2*0.1 = 0.42.

[0123] Compare the energy entropy of each frequency band in each time segment with the corresponding dynamic threshold. If the energy entropy is greater than the dynamic threshold, the coefficient of the frequency band in the time segment is retained; otherwise, it is set to zero. For example, assuming that the energy entropy of a frequency band in a time segment is 0.45, which is greater than the dynamic threshold of 0.42 for the time segment, the coefficient of the frequency band in the time segment is retained.

[0124] Finally, the retained frequency band coefficients are reconstructed by wavelet packets to obtain optimized low-frequency feature sequences and high-frequency feature sequences. The reconstruction process is the inverse process of wavelet packet decomposition. The selected coefficients are recombined according to the original decomposition structure to obtain the optimized signal.

[0125] In this embodiment, by calculating the energy entropy and comparing it with the dynamic threshold, the important frequency band components in the signal can be effectively identified and retained, and the redundant and noise components can be removed, thereby improving the accuracy and effectiveness of feature extraction; by adopting the dynamic threshold method, it is possible to adaptively screen according to the characteristics of the signal in different time periods, which is more flexible and accurate than the fixed threshold and can better capture the time-varying characteristics of the signal; by reconstructing the filtered frequency band coefficients, an optimized low-frequency and high-frequency feature sequence is obtained, which not only retains the main characteristics of the original signal but also removes useless components, and can significantly improve the efficiency and effect of subsequent signal processing and analysis.

[0126] In an optional implementation, the landslide monitoring feature vector is divided into multiple time windows in chronological order to obtain a landslide feature time series matrix, a variational autoencoder is used to perform nonlinear dimension reduction on the landslide feature time series matrix to obtain landslide state characteristics, the landslide state characteristics are input into a fractional-order adaptive neural network, the landslide state evolution trajectory is obtained by solving a fractional-order differential equation, a landslide state evolution model is established based on the evolution trajectory, and the landslide evolution law parameters and current state parameters obtained include:

[0127] The landslide monitoring feature vector is divided according to a preset time window length and a preset sliding step length to obtain a plurality of time window sequences, the landslide monitoring feature vectors in each time window sequence are combined into a feature vector sequence according to a time series relationship, and all the feature vector sequences are spliced ​​by column to construct a landslide feature time series matrix;

[0128] Constructing a variational autoencoder, setting a multi-layer fully connected network as an encoder in the variational autoencoder, inputting the landslide body feature time series matrix into the encoder, and outputting a mean vector and a variance vector in the last layer of the encoder respectively;

[0129] A noise vector is obtained by randomly sampling a standard normal distribution, and the noise vector is multiplied by an exponential function of the variance vector and then added to the mean vector to obtain a reparameterized landslide state characteristic;

[0130] The re-parameterized landslide state characteristics are input into the fractional-order adaptive neural network, and the fractional-order differential equation is obtained through the synergy of the fractional-order calculation layer and the memory mapping layer.

[0131] The fractional-order differential equation is solved by using a prediction-correction format to obtain a numerical solution, wherein the prediction-correction format represents the state at a current moment as a linear combination of an initial state and a weighted sum of historical states; and the evolution trajectory of the landslide body state is obtained according to the numerical solution of the fractional-order differential equation;

[0132] Based on the structure of the fractional-order adaptive neural network, the parameters of the landslide evolution law are determined;

[0133] A hidden Markov model is constructed based on the evolution trajectory, and a state transition probability matrix is ​​extracted from the hidden Markov model; current monitoring data is substituted into the hidden Markov model, and the posterior probability distribution of the landslide body in various states is calculated based on the state transition probability matrix, and the state corresponding to the maximum posterior probability is determined as the current state parameter of the landslide body.

[0134] Specifically, the monitoring data of the landslide body is first obtained, including characteristics such as displacement, velocity, acceleration, etc. These characteristics are combined into a monitoring feature vector in chronological order.

[0135] Select an appropriate time window length and sliding step, for example, the window length is 30 days and the sliding step is 5 days. According to this rule, the monitoring feature vector is divided into multiple time window sequences. For each time window, the feature vectors inside it are arranged in time sequence to form a feature vector sequence. Then the feature vector sequences of all windows are spliced ​​by column to construct the landslide feature time series matrix.

[0136] Next, a variational autoencoder is constructed, in which the encoder adopts a multi-layer fully connected network structure. For example, three hidden layers can be set, and the number of neurons in each layer is 256, 128, and 64. The feature time series matrix is ​​input into the encoder, and the mean vector and variance vector are output in the last layer, respectively, with a vector dimension of 32.

[0137] A 32-dimensional noise vector is obtained by random sampling from the standard normal distribution. The noise vector is multiplied by the exponential function of the variance vector and then added to the mean vector to obtain the reparameterized 32-dimensional landslide state characteristics.

[0138] A fractional-order adaptive neural network is constructed, which includes a fractional-order computing layer and a memory mapping layer. The order of the fractional-order computing layer is set to 0.5, and the memory length of the memory mapping layer is 10. The reparameterized state characteristics are input into the network, and through the synergy of the two layers, the fractional-order differential equation describing the state evolution of the landslide body is obtained.

[0139] The fractional differential equation is numerically solved using the prediction-correction format. Specifically, the state x(t) at the current time t is expressed as the weighted sum of the initial state x(0) and the historical states x(t-1), x(t-2), etc. The weight coefficient is determined by the definition of fractional differentials. The numerical solution of the equation is obtained through iterative calculation, that is, the evolution trajectory of the landslide state over time.

[0140] Based on the structure of fractional-order adaptive neural network, key parameters describing the evolution law of landslide can be extracted, such as fractional order, memory length, network weight, etc. These parameters describe the nonlinear characteristics of the state evolution of landslide.

[0141] Based on the obtained evolution trajectory, a hidden Markov model is constructed. The trajectory is discretized into a finite number of states, and the transition probability between states is calculated to obtain a state transition probability matrix. For example, the landslide state can be divided into four types: stable, slow deformation, accelerated deformation, and instability, and a 4x4 transition probability matrix is ​​obtained.

[0142] Finally, the latest monitoring data is substituted into the hidden Markov model. Based on the state transition probability matrix, the posterior probability of the landslide body in each state is calculated. The state with the largest posterior probability is selected as the current state parameter of the landslide body. For example, if the posterior probability of the stable state is the largest, which is 0.75, it is determined that the landslide body is currently in a stable state.

[0143] In this embodiment, a variational autoencoder is used to perform nonlinear dimensionality reduction on high-dimensional monitoring data to extract key state characteristics of the landslide body, effectively overcoming the limitations of traditional dimensionality reduction methods; a fractional-order adaptive neural network is introduced to characterize the nonlinear evolution process of the landslide body state through fractional-order differential equations, which can more accurately describe the long-term memory effect of the landslide body; combined with the hidden Markov model, the probabilistic evaluation of the current state of the landslide body is realized, providing a reliable basis for landslide disaster early warning.

[0144] In an optional implementation, the fractional-order adaptive neural network includes:

[0145] Fractional-order computation layers and memory mapping layers connected in sequence;

[0146] The fractional-order calculation layer receives input of the state characteristics of the landslide body, and constructs a fractional-order derivative operator based on historical state weighting in the fractional-order calculation layer, wherein the fractional-order derivative operator represents the state characteristics of the landslide body at the current moment as a weighted integral of the historical state characteristics, and the weight coefficient in the weighted integral is determined based on the gamma function and the fractional order to be optimized;

[0147] The memory mapping layer receives the weighted integral output by the fractional-order calculation layer, adopts a recursive orthogonal polynomial function in the memory mapping layer, constructs a plurality of memory kernel functions to extract features from the weighted integral, and sets the coefficients of the plurality of memory kernel functions as parameters to be optimized;

[0148] Introducing fractional-order feature neurons in the memory mapping layer, the fractional-order feature neurons receive the output of the memory kernel function, and use fractional-order power series as an activation function for nonlinear mapping, the power of each power term in the fractional-order power series is determined based on the product of the fractional order and an integer, and the coefficient of the fractional-order power series is set as a parameter to be optimized;

[0149] An optimization objective function is constructed, wherein the optimization objective function includes a reconstruction error term between the state characteristics of the landslide body and the landslide body characteristic time series matrix, and a prediction error term between the state characteristics and the actual historical state. The optimization objective function is iteratively optimized by a gradient descent method to obtain optimized values ​​of the fractional orders in the fractional-order calculation layer, optimized values ​​of the memory kernel function coefficients, and optimized values ​​of the fractional-order power series coefficients. Based on the optimized values, the network parameters of the fractional-order adaptive neural network are determined.

[0150] Specifically, a fractional-order adaptive neural network is constructed, including a fractional-order calculation layer and a memory mapping layer connected in sequence. The fractional-order calculation layer receives the input of the state characteristics of the landslide body, and a fractional-order derivative operator based on historical state weighting is constructed in the layer. The operator represents the state characteristics of the landslide body at the current moment as a weighted integral of the historical state characteristics, and the weight coefficient in the weighted integral is determined based on the gamma function and the fractional order to be optimized.

[0151] Specifically, it is assumed that the state characteristics of the landslide body are 10-dimensional vectors, including displacement, velocity, acceleration and other parameters. The order of the fractional derivative operator is initially set to 0.5, and the number of historical states is set to 100. For the state characteristics xt at the current time t, the weighted integral result yt is obtained by weighted summing the first 100 historical states xi:

[0152] yt = Σ(wi * xi), i from t-100 to t-1;

[0153] The weight wi is calculated based on the gamma function Γ(·):

[0154] wi = Γ(order) / [Γ(i+1) * Γ(order-i)];

[0155] The memory mapping layer receives the weighted integral yt output by the fractional-order calculation layer, and uses a recursive orthogonal polynomial function to construct multiple memory kernel functions to extract features from yt. Assume that five memory kernel functions φ1(yt) to φ5(yt) are constructed, and the coefficient αi of each kernel function is set as the parameter to be optimized.

[0156] Next, a fractional-order feature neuron is introduced into the memory mapping layer. The neuron receives the output of the memory kernel function and uses the fractional-order power series as the activation function for nonlinear mapping. The power of each power term in the fractional-order power series is determined based on the product of the fractional order and an integer, and the coefficient βj of the series is set as the parameter to be optimized. For example, when the fractional order is 0.5, the power series is in the form of:

[0157] f(x) = β0 + β1x^0.5 + β2x^1 + β3x^1.5 + β4x^2 + ...;

[0158] Then, the optimization objective function is constructed, which includes two parts: the reconstruction error between the state characteristics of the landslide body and the time series matrix of the landslide body characteristics; the prediction error between the state characteristics and the actual historical state. The objective function is iteratively optimized by the gradient descent method to obtain the optimized values ​​of the fractional order, the memory kernel function coefficient αi and the fractional power series coefficient βj.

[0159] Taking landslide monitoring data as an example, it is assumed that the historical monitoring data contains 10-dimensional state features at 1000 time points. First, the data is divided into a training set (800 samples) and a test set (200 samples). During the training process, 100 consecutive samples are used as a batch in each iteration to calculate the reconstruction error and prediction error, and update the network parameters. After 1000 rounds of iterative optimization, the optimal fractional order, αi and βj are obtained. Finally, the model performance is verified on the test set.

[0160] In this embodiment, by introducing a fractional-order derivative operator, the ability to characterize the long-term memory characteristics of the landslide body is enhanced, and the state prediction accuracy is improved; a recursive orthogonal polynomial is used to construct a memory kernel function, which enhances the ability to extract nonlinear features; and the introduction of fractional-order feature neurons improves the model's ability to express complex nonlinear mapping relationships, which is conducive to capturing the nonlinear characteristics of the landslide evolution process.

[0161] In an optional implementation, deep Copper network modeling is performed on the landslide body evolution law parameters and current state parameters to obtain the probability density function of the landslide body displacement, including:

[0162] The deep Copula network comprises a parameter encoding layer, an edge distribution modeling layer and a joint distribution construction layer;

[0163] The parameter encoding layer receives the input of the landslide body evolution law parameters and the current state parameters, and performs feature extraction and dimensionality reduction processing on the landslide body evolution law parameters and the current state parameters through a multi-layer deep neural network to obtain a low-dimensional feature representation of the parameters;

[0164] The edge distribution modeling layer receives the low-dimensional feature representation, constructs a bandwidth-adaptive non-parametric kernel density estimator for each dimension, and the bandwidth parameter of the non-parametric kernel density estimator is selected from a preset candidate bandwidth set by a cross-validation method to obtain the edge probability distribution of each dimensional feature;

[0165] The joint distribution construction layer receives the marginal probability distribution of each dimension, constructs a parameterized Coppola function family, combines the marginal probability distribution, determines the parameters of the Coppola function family based on historical sample data through the maximum likelihood estimation method, and outputs the joint probability distribution of the landslide body displacement parameters; based on the joint probability distribution, a numerical integration method is used to calculate the probability density function of the landslide body displacement within a preset time interval.

[0166] The nonparametric kernel density estimator specifically refers to a statistical method for estimating data distribution, which does not depend on the specific distribution form of the data, hence the name "nonparametric". The basic idea is to estimate the probability density of the data through the local area around each data point (determined by the kernel function). The bandwidth parameter controls the smoothness of the kernel function and affects the precision of the density estimate. By choosing the appropriate bandwidth, the kernel density estimator can flexibly capture the marginal distribution characteristics of the data without being restricted by specific distribution assumptions.

[0167] The Coppola function family is a mathematical tool used to describe the dependencies between multidimensional random variables. In probability and statistics, the Coppola function connects the marginal distributions of each variable to form a joint distribution, thereby characterizing the dependency structure between variables. The key feature of the Coppola function is that it can separate marginal distributions and dependencies, so that marginal distributions and dependency structures can be selected independently when modeling.

[0168] Specifically, a deep Coppola network is constructed, which includes a parameter encoding layer, a marginal distribution modeling layer, and a joint distribution construction layer. The parameter encoding layer adopts a multi-layer fully connected neural network structure, including 3 hidden layers, with 64, 32, and 16 neurons in each layer. The marginal distribution modeling layer adopts a non-parametric kernel density estimation method, and the kernel function selects a Gaussian kernel. The joint distribution construction layer uses three commonly used single-parameter Coppola function families: Gauss Coppola, Clayton Coppola, and Frank Coppola.

[0169] Next, the historical monitoring data of the landslide body are collected, including the evolution law parameters such as landslide body displacement, rainfall, groundwater level, and current state parameters such as slope, lithology, and vegetation coverage. These parameters are used as input and sent to the parameter encoding layer for processing.

[0170] In the parameter encoding layer, the input parameters are first normalized and preprocessed. Then, through the nonlinear transformation of the three hidden layers, the deep features of the parameters are extracted and the dimension is reduced. Taking a landslide as an example, the input parameter dimension is 15, and after processing in the parameter encoding layer, an 8-dimensional low-dimensional feature representation is obtained.

[0171] The low-dimensional feature representation is input to the marginal distribution modeling layer. For each dimensional feature, a bandwidth-adaptive Gaussian kernel density estimator is constructed. The bandwidth parameter is selected from 10 candidate values ​​ranging from 0.1 to 2.0 through 5-fold cross validation. Taking a certain dimensional feature as an example, the optimal bandwidth is 0.5, and the marginal probability distribution function of this dimension is obtained.

[0172] In the joint distribution construction layer, a parameterized Coppola function family is first constructed based on the marginal probability distribution. Then, the optimal parameters of the Coppola function are determined by the maximum likelihood estimation method using 1,000 sets of historical sample data. Taking Frank Coppola as an example, the optimal parameter is 2.8. Substituting the marginal distribution into the optimal Coppola function, the 8-dimensional joint probability distribution of the landslide displacement parameters is obtained.

[0173] Finally, based on the obtained joint probability distribution, the Monte Carlo numerical integration method is used to calculate the probability density function of the landslide displacement in the next 30 days. Specifically, 10,000 groups of samples are drawn from the joint distribution, and the 30-day cumulative displacement corresponding to each group of samples is calculated. The frequency distribution of the displacement is statistically obtained, which is the required probability density function.

[0174] In this embodiment, nonlinear feature extraction and dimensionality reduction of landslide parameters are realized through deep neural networks, which effectively extracts the complex correlations between parameters and improves the modeling accuracy. The marginal distribution is constructed by non-parametric kernel density estimation method, which avoids the limitations of parameter distribution assumptions and enhances the adaptability and generalization ability of the model. The joint distribution is constructed based on the parameterized Coppola function, which not only retains the marginal distribution information, but also flexibly characterizes the correlation structure between parameters, effectively improving the accuracy of the probability prediction of landslide displacement.

[0175] In an optional implementation, a fuzzy evolutionary algorithm is used to construct an adaptive risk threshold function and embed Monte Carlo tree search optimization to dynamically evaluate the landslide risk metric value, including:

[0176] Based on the preset landslide body evolution index, the key points of landslide body evolution are determined. Based on the key points of landslide body evolution, the pre-collected historical warning data are segmented. For each segment, a piecewise continuous function is fitted using a cubic spline interpolation method. The function coefficient of the piecewise continuous function is used as an initial threshold parameter to construct an adaptive risk threshold function.

[0177] Based on the historical warning data, a fuzzy rule set is extracted to determine the mapping relationship between the risk metric value and the warning level; the warning accuracy of the fuzzy rule set is used as the fitness function, and the parameters of the membership function in the fuzzy rule set are optimized by a genetic algorithm to obtain an optimized fuzzy rule base; the landslide risk metric value is input into a fuzzy inference system based on the optimized fuzzy rule base to obtain a fuzzy decision set of the warning level;

[0178] A multi-layer decision tree is constructed, the warning levels in the fuzzy decision set are set as decision nodes, and the historical risk losses corresponding to each warning level are set as the profit function; Monte Carlo sampling is performed at each decision node to generate state sequence samples, the upper confidence interval value of each decision branch is calculated, and the decision branch corresponding to the maximum value of the upper confidence interval value is selected as the optimal decision path, and the warning decision sequence corresponding to the optimal decision path is extracted to determine the warning level and generate warning information.

[0179] Specifically, the key points of landslide evolution are determined based on pre-set landslide evolution indicators, such as displacement, velocity, acceleration, etc. The key points can be turning points in the landslide evolution process, such as the start point of accelerated sliding, the critical instability point, etc. Then, the pre-collected historical warning data are segmented according to the key points.

[0180] For each segment, a piecewise continuous function is fitted using a cubic spline interpolation method. Specifically, a cubic polynomial can be selected as the interpolation function to ensure the continuity of the function value and the first-order and second-order derivatives at the segmentation points. The function coefficients of the obtained piecewise continuous function are used as the initial threshold parameters to construct an adaptive risk threshold function.

[0181] For example, suppose the displacement data of a landslide is divided into three segments, corresponding to the stable period, acceleration period, and instability period. After performing cubic spline interpolation on each segment of data, three cubic polynomial functions are obtained. The coefficients of these functions are used as initial threshold parameters to construct a segmented adaptive risk threshold function.

[0182] Next, based on the historical warning data, fuzzy rule sets are extracted. First, the risk measurement value and warning level are fuzzified, such as dividing the risk measurement value into three fuzzy sets of "low", "medium", and "high", and the warning level into four levels of "normal", "attention", "warning", and "danger". Then, based on the corresponding relationship of historical data, a fuzzy rule such as "if the risk measurement value is low, the warning level is normal" is generated.

[0183] Taking the early warning accuracy of the fuzzy rule set as the fitness function, the genetic algorithm is used to optimize the parameters of the membership function in the fuzzy rule set. The specific steps include: initializing the population, calculating fitness, selection, crossover, mutation, and iterative optimization until the termination condition is met. Finally, the tuned fuzzy rule base is obtained.

[0184] The current landslide risk measurement value is input into the fuzzy reasoning system based on the tuned fuzzy rule base to obtain the fuzzy decision set of the warning level. The fuzzy reasoning adopts the Mamdani reasoning method, which includes three steps: fuzzification, rule reasoning and defuzzification.

[0185] Construct a multi-layer decision tree, set the warning level in the fuzzy decision set as the decision node, and set the historical risk loss corresponding to each warning level as the profit function. Perform Monte Carlo sampling at each decision node to generate state sequence samples. For each decision branch, calculate the upper confidence interval value. Select the decision branch with the largest upper confidence interval value as the optimal decision path.

[0186] Extract the warning decision sequence corresponding to the optimal decision path, determine the final warning level and generate warning information. The warning information includes key information such as warning level, risk measurement value, and warning time.

[0187] For example, assuming that the current risk measurement value of a landslide is 0.7, the fuzzy set of warning levels obtained through fuzzy reasoning is {normal 0.1, attention 0.3, warning 0.5, danger 0.1}. A 4-layer decision tree is constructed, with each layer corresponding to a warning level. Through Monte Carlo tree search, the optimal decision path is obtained as "warning-attention-normal-normal", and the warning level is finally determined to be "warning", and the corresponding warning information is generated.

[0188] In this embodiment, by constructing an adaptive risk threshold function, the risk assessment standard can be dynamically adjusted, thereby improving the flexibility and adaptability of the assessment; the fuzzy evolutionary algorithm is used to optimize the fuzzy rule set, thereby improving the accuracy and robustness of the mapping of risk metrics to warning levels; and the Monte Carlo tree search is combined for decision optimization, thereby being able to make the optimal warning decision while considering multiple decision sequences, thereby improving the reliability and foresight of the warning.

[0189] Figure 2 FIG. 1 is a schematic diagram of the structure of a ground-based radar landslide monitoring and early warning system according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0190] The first unit is used for collecting landslide monitoring signals by ground-based radar, filtering the signals through Butterworth filter to obtain filtered signals, extracting instantaneous phases through Hilbert transform, and performing five-layer decomposition using wavelet packet decomposition algorithm based on the instantaneous phases to obtain low-frequency characteristic coefficients and high-frequency characteristic coefficients of the filtered signals; obtaining enhanced features through morphological filters based on the low-frequency characteristic coefficients and the high-frequency characteristic coefficients; performing discrete Fourier transform on the enhanced features to obtain phase stability coefficients, amplitude continuity coefficients and signal-to-noise ratio coefficients, and constructing landslide monitoring feature vectors;

[0191] The second unit is used to divide the landslide monitoring feature vector into multiple time windows in time sequence to obtain the landslide feature time series matrix, use a variational autoencoder to perform nonlinear dimension reduction on the landslide feature time series matrix to obtain the landslide state characteristics, input the landslide state characteristics into a fractional-order adaptive neural network, obtain the evolution trajectory of the landslide state by solving the fractional-order differential equation, establish a landslide state evolution model based on the evolution trajectory, and obtain the landslide evolution law parameters and current state parameters;

[0192] The third unit is used to construct a hierarchical risk reasoning framework, perform deep Copper network modeling on the landslide evolution law parameters and current state parameters, obtain the probability density function of the landslide displacement, combine the probability density function with the pre-acquired geological parameters and environmental parameters, and fuse multi-source information through evidence theory to obtain the landslide risk measurement value. The fuzzy evolution algorithm is used to construct an adaptive risk threshold function and embed the Monte Carlo tree search optimization to dynamically evaluate the landslide risk measurement value, determine the warning level and generate warning information.

[0193] According to a third aspect of the embodiments of the present invention,

[0194] An electronic device is provided, comprising:

[0195] processor;

[0196] a memory for storing processor-executable instructions;

[0197] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0198] According to a fourth aspect of the embodiments of the present invention,

[0199] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0200] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0201] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A ground-based radar landslide monitoring and early warning method, characterized in that: include: The ground-based radar collects the monitoring signal of the landslide body, obtains the filtered signal through filtering processing by Butterworth filter, extracts the instantaneous phase through Hilbert transform, and performs five-layer decomposition based on the instantaneous phase by using wavelet packet decomposition algorithm to obtain the low-frequency characteristic coefficient and high-frequency characteristic coefficient of the filtered signal; based on the low-frequency characteristic coefficient and the high-frequency characteristic coefficient, obtains the enhanced feature through morphological filter; Performing discrete Fourier transform on the enhanced features to obtain phase stability coefficient, amplitude continuity coefficient and signal-to-noise ratio coefficient, and constructing a landslide monitoring feature vector; The monitoring feature vector of the landslide body is divided into multiple time windows in time sequence to obtain a landslide body feature time series matrix, a variational autoencoder is used to perform nonlinear dimension reduction on the landslide body feature time series matrix to obtain the landslide body state characteristics, the landslide body state characteristics are input into a fractional-order adaptive neural network, the evolution trajectory of the landslide body state is obtained by solving a fractional-order differential equation, a landslide body state evolution model is established based on the evolution trajectory, and the landslide body evolution law parameters and current state parameters are obtained; A hierarchical risk reasoning framework is constructed, and deep Copper network modeling is performed on the parameters of landslide evolution law and current state parameters to obtain the probability density function of landslide displacement. The probability density function is combined with pre-acquired geological parameters and environmental parameters, and multi-source information is fused through evidence theory to obtain the landslide risk metric. The fuzzy evolution algorithm is used to construct an adaptive risk threshold function and embed Monte Carlo tree search optimization to dynamically evaluate the landslide risk metric, determine the warning level and generate warning information. The ground-based radar collects the monitoring signal of the landslide body, obtains the filtered signal through filtering processing by Butterworth filter, extracts the instantaneous phase through Hilbert transform, and performs five-layer decomposition based on the instantaneous phase by using wavelet packet decomposition algorithm to obtain the low-frequency characteristic coefficient and high-frequency characteristic coefficient of the filtered signal; based on the low-frequency characteristic coefficient and the high-frequency characteristic coefficient, obtains the enhanced feature through morphological filter; The enhanced features are subjected to discrete Fourier transform to obtain phase stability coefficient, amplitude continuity coefficient and signal-to-noise ratio coefficient, and the landslide monitoring feature vector is constructed, including: A landslide monitoring signal is obtained by using a ground-based radar, and the landslide monitoring signal is filtered by a Butterworth filter to obtain a filtered signal by suppressing high-frequency noise and interference signals; Performing phase delay processing on the filtered signal to obtain a delayed signal, constructing an analytical signal according to the filtered signal and the delayed signal, calculating the argument of the analytical signal, and obtaining an instantaneous phase signal; Based on the instantaneous phase signal, wavelet packet decomposition is adopted, and five-layer decomposition is recursively performed through high-pass filtering and low-pass filtering to sequentially generate multi-layer frequency band coefficients, and based on the fifth layer of frequency band coefficients in the multi-layer frequency band coefficients, coefficients with frequencies less than a preset dividing frequency are reconstructed into low-frequency characteristic coefficients, and coefficients with frequencies greater than or equal to the preset dividing frequency are reconstructed into high-frequency characteristic coefficients; The frequency band coefficients are screened based on energy entropy and dynamic threshold, and optimized low-frequency feature sequences and optimized high-frequency feature sequences are reconstructed; Performing an erosion operation on the optimized low-frequency feature sequence and the optimized high-frequency feature sequence to obtain a first intermediate sequence; performing an expansion operation on the first intermediate sequence to obtain an open sequence; performing an expansion operation on the open sequence to obtain a second intermediate sequence, and performing an erosion operation on the second intermediate sequence to obtain an enhanced feature sequence; performing a discrete Fourier transform on the enhanced feature sequence to obtain a frequency domain transform sequence; Calculate the phase sequence variance of the frequency domain transformation sequence to obtain a phase stability coefficient, calculate the ratio of the standard deviation to the mean of the amplitude sequence of the frequency domain transformation sequence to obtain an amplitude continuity coefficient; calculate the logarithmic ratio of the signal power to the noise power in the frequency domain transformation sequence to obtain a signal-to-noise ratio coefficient; The phase stability coefficient, the amplitude continuity coefficient and the signal-to-noise ratio coefficient are combined to construct a landslide monitoring feature vector.

2. The method according to claim 1, characterized in that: The frequency band coefficients are screened based on energy entropy and dynamic threshold, and the optimized low-frequency feature sequence and optimized high-frequency feature sequence are reconstructed, including: Calculating the sum of the squares of the low-frequency characteristic coefficient and the high-frequency characteristic coefficient, and performing normalization processing to obtain a normalized energy value; calculating the information entropy of the normalized energy value to obtain energy entropy; The energy entropy is divided into multiple time segments according to a preset length, and the mean and standard deviation of the corresponding energy entropy in each time segment are calculated; the mean is multiplied by a first weight coefficient to obtain a first weighted value, and the standard deviation is multiplied by a second weight coefficient to obtain a second weighted value, and the first weighted value and the second weighted value are added to construct a dynamic threshold; Compare the energy entropy of each frequency band with the dynamic threshold, retain the frequency band coefficients whose energy entropy is greater than the dynamic threshold, and set the frequency band coefficients whose energy entropy is less than or equal to the dynamic threshold to zero; The retained frequency band characteristic coefficients are reconstructed by wavelet packets to obtain optimized low-frequency characteristic sequences and optimized high-frequency characteristic sequences.

3. The method according to claim 1, characterized in that The monitoring feature vector of the landslide body is divided into multiple time windows in time sequence to obtain a landslide body feature time series matrix. The landslide body feature time series matrix is ​​subjected to nonlinear dimension reduction using a variational autoencoder to obtain the landslide body state characteristics. The landslide body state characteristics are input into a fractional-order adaptive neural network. The evolution trajectory of the landslide body state is obtained by solving a fractional-order differential equation. A landslide body state evolution model is established based on the evolution trajectory. The landslide body evolution law parameters and current state parameters are obtained, including: The landslide monitoring feature vector is divided according to a preset time window length and a preset sliding step length to obtain a plurality of time window sequences, the landslide monitoring feature vectors in each time window sequence are combined into a feature vector sequence according to a time series relationship, and all the feature vector sequences are spliced ​​by column to construct a landslide feature time series matrix; Constructing a variational autoencoder, setting a multi-layer fully connected network as an encoder in the variational autoencoder, inputting the landslide body feature time series matrix into the encoder, and outputting a mean vector and a variance vector in the last layer of the encoder respectively; A noise vector is obtained by randomly sampling a standard normal distribution, and the noise vector is multiplied by an exponential function of the variance vector and then added to the mean vector to obtain a reparameterized landslide state characteristic; The re-parameterized landslide state characteristics are input into the fractional-order adaptive neural network, and the fractional-order differential equation is obtained through the synergy of the fractional-order calculation layer and the memory mapping layer. The fractional-order differential equation is solved by using a prediction-correction format to obtain a numerical solution, wherein the prediction-correction format represents the state at a current moment as a linear combination of an initial state and a weighted sum of historical states; and the evolution trajectory of the landslide body state is obtained according to the numerical solution of the fractional-order differential equation; Based on the structure of the fractional-order adaptive neural network, the parameters of the landslide evolution law are determined; A hidden Markov model is constructed based on the evolution trajectory, and a state transition probability matrix is ​​extracted from the hidden Markov model; current monitoring data is substituted into the hidden Markov model, and the posterior probability distribution of the landslide body in various states is calculated based on the state transition probability matrix, and the state corresponding to the maximum posterior probability is determined as the current state parameter of the landslide body.

4. The method according to claim 3, characterized in that: The fractional-order adaptive neural network comprises: Fractional-order computation layers and memory mapping layers connected in sequence; The fractional-order calculation layer receives input of the state characteristics of the landslide body, and constructs a fractional-order derivative operator based on historical state weighting in the fractional-order calculation layer, wherein the fractional-order derivative operator represents the state characteristics of the landslide body at the current moment as a weighted integral of the historical state characteristics, and the weight coefficient in the weighted integral is determined based on the gamma function and the fractional order to be optimized; The memory mapping layer receives the weighted integral output by the fractional-order calculation layer, adopts a recursive orthogonal polynomial function in the memory mapping layer, constructs a plurality of memory kernel functions to extract features from the weighted integral, and sets the coefficients of the plurality of memory kernel functions as parameters to be optimized; Introducing fractional-order feature neurons in the memory mapping layer, the fractional-order feature neurons receive the output of the memory kernel function, and use fractional-order power series as an activation function for nonlinear mapping, the power of each power term in the fractional-order power series is determined based on the product of the fractional order and an integer, and the coefficient of the fractional-order power series is set as a parameter to be optimized; An optimization objective function is constructed, wherein the optimization objective function includes a reconstruction error term between the state characteristics of the landslide body and the landslide body characteristic time series matrix, and a prediction error term between the state characteristics and the actual historical state. The optimization objective function is iteratively optimized by a gradient descent method to obtain optimized values ​​of the fractional orders in the fractional-order calculation layer, optimized values ​​of the memory kernel function coefficients, and optimized values ​​of the fractional-order power series coefficients. Based on the optimized values, the network parameters of the fractional-order adaptive neural network are determined.

5. The method according to claim 1, characterized in that The deep Copper network modeling is carried out on the parameters of landslide evolution law and current state parameters to obtain the probability density function of landslide displacement, including: The deep Copula network comprises a parameter encoding layer, an edge distribution modeling layer and a joint distribution construction layer; The parameter encoding layer receives the input of the landslide body evolution law parameters and the current state parameters, and performs feature extraction and dimensionality reduction processing on the landslide body evolution law parameters and the current state parameters through a multi-layer deep neural network to obtain a low-dimensional feature representation of the parameters; The edge distribution modeling layer receives the low-dimensional feature representation, constructs a bandwidth-adaptive non-parametric kernel density estimator for each dimension, and the bandwidth parameter of the non-parametric kernel density estimator is selected from a preset candidate bandwidth set by a cross-validation method to obtain the edge probability distribution of each dimensional feature; The joint distribution construction layer receives the marginal probability distribution of each dimension, constructs a parameterized Coppola function family, combines the marginal probability distribution, determines the parameters of the Coppola function family based on historical sample data through the maximum likelihood estimation method, and outputs the joint probability distribution of the landslide body displacement parameters; based on the joint probability distribution, a numerical integration method is used to calculate the probability density function of the landslide body displacement within a preset time interval.

6. The method according to claim 5, characterized in that The fuzzy evolutionary algorithm is used to construct an adaptive risk threshold function and embed Monte Carlo tree search optimization to dynamically evaluate the landslide risk metric value, including: Based on the preset landslide body evolution index, the key points of landslide body evolution are determined. Based on the key points of landslide body evolution, the pre-collected historical warning data are segmented. For each segment, a piecewise continuous function is fitted using a cubic spline interpolation method. The function coefficient of the piecewise continuous function is used as an initial threshold parameter to construct an adaptive risk threshold function. Based on the historical warning data, a fuzzy rule set is extracted to determine the mapping relationship between the risk metric value and the warning level; the warning accuracy of the fuzzy rule set is used as the fitness function, and the parameters of the membership function in the fuzzy rule set are optimized by a genetic algorithm to obtain an optimized fuzzy rule base; the landslide risk metric value is input into a fuzzy inference system based on the optimized fuzzy rule base to obtain a fuzzy decision set of the warning level; A multi-layer decision tree is constructed, the warning levels in the fuzzy decision set are set as decision nodes, and the historical risk losses corresponding to each warning level are set as the profit function; Monte Carlo sampling is performed at each decision node to generate state sequence samples, the upper confidence interval value of each decision branch is calculated, and the decision branch corresponding to the maximum value of the upper confidence interval value is selected as the optimal decision path, and the warning decision sequence corresponding to the optimal decision path is extracted to determine the warning level and generate warning information.

7. A ground-based radar landslide monitoring and early warning system, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used for collecting landslide monitoring signals by ground-based radar, filtering the signals through Butterworth filter to obtain filtered signals, extracting instantaneous phases through Hilbert transform, and performing five-layer decomposition using wavelet packet decomposition algorithm based on the instantaneous phase to obtain low-frequency characteristic coefficients and high-frequency characteristic coefficients of the filtered signals; based on the low-frequency characteristic coefficients and the high-frequency characteristic coefficients, obtaining enhanced features through morphological filters; Performing discrete Fourier transform on the enhanced features to obtain phase stability coefficient, amplitude continuity coefficient and signal-to-noise ratio coefficient, and constructing a landslide monitoring feature vector; The second unit is used to divide the landslide monitoring feature vector into multiple time windows in time sequence to obtain the landslide feature time series matrix, use a variational autoencoder to perform nonlinear dimension reduction on the landslide feature time series matrix to obtain the landslide state characteristics, input the landslide state characteristics into a fractional-order adaptive neural network, obtain the evolution trajectory of the landslide state by solving the fractional-order differential equation, establish a landslide state evolution model based on the evolution trajectory, and obtain the landslide evolution law parameters and current state parameters; The third unit is used to construct a hierarchical risk reasoning framework, perform deep Copper network modeling on the landslide evolution law parameters and current state parameters, obtain the probability density function of the landslide displacement, combine the probability density function with the pre-acquired geological parameters and environmental parameters, and fuse multi-source information through evidence theory to obtain the landslide risk measurement value. The fuzzy evolution algorithm is used to construct an adaptive risk threshold function and embed the Monte Carlo tree search optimization to dynamically evaluate the landslide risk measurement value, determine the warning level and generate warning information.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images