Landslide seismic signal precursor identification and early warning method based on machine learning
By using machine learning technology, high-precision landslide seismic signal acquisition and risk assessment in complex environments have been achieved, solving the problems of insufficient accuracy and response speed in existing landslide disaster early warning technologies, and improving the reliability of landslide early warning and the efficiency of emergency response.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-23
- Publication Date
- 2026-04-28
AI Technical Summary
Existing technologies struggle to accurately capture the arrival time of landslide seismic signals in complex geological environments, resulting in insufficient accuracy and response speed for landslide disaster early warning. This is especially true in environments with noise interference and multi-source signal aliasing, which can easily lead to misjudgment or missed detection.
Machine learning methods are employed to denoise seismic signal waveform data using wavelet transform. The STA/LTA algorithm and support vector machine classification are combined to achieve automatic and high-precision acquisition of seismic phase arrival time. Coherence analysis is used to verify the consistency of signals from multiple stations. Based on the least squares method and triangulation algorithm, time difference features and station locations are fused. Convolutional neural networks are used for scale classification to generate risk level labels and integrate real-time data streams for early warning.
It improves the accuracy and response speed of landslide earthquake signal precursor identification, reduces human subjective error, and enhances the reliability of landslide early warning and emergency response efficiency.
Smart Images
Figure CN121938145A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of combining machine learning with seismic signal monitoring, and more specifically, to a method for identifying and warning of landslide seismic signals based on machine learning. Background Technology
[0002] Landslides are among the most destructive geological hazards worldwide, and rapid and accurate early warning of landslides has become a core task in the field of geological disaster prevention and control.
[0003] Large landslides generate special seismic waveform signals before they occur. These signals contain key information such as the time, location, and scale of the landslide, which can provide important information for real-time monitoring and early warning of landslide disasters.
[0004] Currently, the identification of landslide earthquake signals mainly relies on traditional signal processing methods and human experience. In complex geological environments, signals are weak and their characteristics change subtly. Especially in environments with noise interference and multi-source signal aliasing, traditional phase arrival time picking techniques struggle to accurately capture the subtle features of landslide earthquake signals. For example, when background noise levels are high or other vibration sources are present, traditional algorithms are prone to missing true phases or misclassifying noise as valid signals. Therefore, manually marking phase arrival times is not only inefficient but also introduces significant uncertainty due to differences in the operator's subjective experience, further limiting identification accuracy and response speed.
[0005] In practical applications of landslide earthquake identification, multiple seismic stations simultaneously detect landslide signals. It is necessary to correlate these signals to accurately determine their time differences of arrival (TDAs) and then use triangulation to accurately calculate the landslide location. Therefore, the accuracy of the landslide earthquake phase arrival time acquisition directly affects the quality of subsequent signal correlation analysis, which in turn determines the reliability of earthquake location. If even minor deviations in the phase arrival time information are amplified during signal correlation, it can lead to incorrect waveform matching and false event identification, ultimately affecting the accurate location of the landslide. This not only renders early warnings meaningless but may also result in the misallocation of emergency resources, causing serious consequences in emergency situations.
[0006] Therefore, a technical solution is needed to automatically and accurately pick up the arrival time of landslide seismic signals in complex scenarios, so as to achieve accurate and effective identification and early warning of landslide seismic signals. Summary of the Invention
[0007] To achieve the above objectives, this application provides a method for identifying and issuing early warning of landslide earthquake signals based on machine learning, comprising the following steps: The original seismic signal waveform data is acquired, and the original seismic signal waveform data is preprocessed to generate seismic signal waveform data. Based on the seismic signal waveform, the potential phase arrival time point is determined, and a preliminary phase arrival time sequence is generated; Based on the preliminary phase-to-time sequence, the original phase-to-time data is extracted, consistent phase-to-time groups are identified, and the location coordinates of the landslide event are estimated. Based on the estimation of seismic signal waveforms and landslide event location coordinates, the original seismic motion signal waveform data is filtered, landslide earthquake scale is classified, and earthquake risk level is output. If the earthquake risk level is high, the original earthquake signal waveform data, geological displacement and rainfall data are extracted to estimate the location coordinates of the landslide event. The earthquake risk level and the estimated location coordinates of the landslide event are then integrated to obtain the landslide early warning signal output.
[0008] The preprocessing of the raw seismic signal waveform data includes: Sampling is performed on the time series of the original seismic signal waveform data; Determine whether the noise level of the sampled signal exceeds a preset threshold. If it does, perform time series decomposition on the sampled seismic signal data to obtain preliminary filtered waveform data. Extract frequency component information from the preliminary filtered waveform data, determine whether the frequency component information conforms to the characteristic range of the seismic signal, and generate denoised waveform data. The time-domain signal is converted into the frequency domain using Fourier transform, and the seismic signal waveform data is output.
[0009] The process of generating the preliminary seismic phase to time sequence includes the following steps: Extracting continuous time window data from seismic signal waveforms; The short-time average calculation method is used to calculate the short-time average amplitude value for continuous time window data, while the long-time average calculation method is used to calculate the long-time average amplitude value for continuous time window data. Calculate the ratio of short-time average amplitude value to long-time average amplitude value, and determine whether the location corresponding to the continuous time window data is the potential earthquake phase arrival time point based on the ratio value, so as to obtain a preliminary earthquake phase arrival time point set. The preliminary phase-to-time set is classified and processed, and the locations of the preliminary phase-to-time set that meet the characteristics of the phase are extracted to form a verification phase-to-time sequence; The time interval information of adjacent points is obtained from the verification phase-to-time sequence. The continuity of the time interval information of adjacent points is determined by the pre-set phase interval threshold. The continuous verification phase-to-time sequence is used to construct the preliminary phase-to-time sequence.
[0010] Determining consistent seismic phase arrival time groups includes the following steps: Obtain the original phase arrival time data, and perform noise reduction processing on the original phase arrival time data to generate noise-reduced phase arrival time data; The denoised phase arrival data is normalized to obtain the first time series; Using the first time series, a sliding window tool is used to extract fixed-length time window data from each station, and the average amplitude or peak characteristics are calculated for each window to form a second time series. For the second time series, the Pearson coherence coefficient of each station is calculated. The similarity between the series is quantified by the Pearson correlation coefficient formula. Data with similarity greater than a preset threshold are judged as consistent phase-time pairs and consistent phase-time groups are generated.
[0011] Furthermore, obtaining the estimated location coordinates of a landslide event includes the following steps: Extracting time difference features from consistent seismic phases to time groups; Obtain the arrival times of the filtered P-wave and S-wave; Extract P-wave and S-wave arrival time data from at least two stations and calculate the time difference feature set; The relationship between the fitted time difference feature set and the station location is determined. When the deviation between the time difference feature set and the preset station location model is less than the preset threshold, the least squares fitting parameters are iteratively calculated to obtain the preliminary location coordinate estimate. The preliminary location coordinate estimate is weighted and fused with the time difference feature set. When the accuracy of the fused coordinate estimate meets the preset threshold, the depth of the landslide source is determined to obtain the estimated location coordinates of the landslide event.
[0012] Furthermore, the classification of landslide earthquake magnitude includes the following steps: Based on the estimated coordinates of the landslide event location, the original seismic signal waveform data was obtained from the seismic monitoring equipment of the corresponding station. Wavelet transform was used to denoise the original seismic signal waveform data to obtain the denoised waveform data. Time series features are extracted from the denoised waveform data, and convolutional neural networks are used to classify the time series features to obtain risk scale classification labels for landslide events; the risk scale classification includes: small scale, medium scale, and large scale. If the risk scale is classified as large-scale, determine whether it is a high-risk landslide event; if it is a high-risk landslide event, generate a risk level label.
[0013] Furthermore, integrating earthquake risk levels with the location coordinates of landslide events includes the following steps: Obtain standardized real-time data streams; Based on the standardized real-time data stream and combined with the geographic information database, the location information of the landslide event is estimated by matching the location coordinates of the landslide event according to the earthquake risk level, and the precise location coordinates of the landslide are determined.
[0014] Obtaining standardized real-time data streams includes: Acquire and summarize real-time data of earthquake waveforms, geological displacement, and rainfall. The sliding window algorithm is used to calculate the trend of vibration energy change from the real-time data, identify and remove noise values, remove abnormal high-frequency fluctuations or readings that are outside the normal range, and clean the real-time data in real time to generate a standardized real-time data stream.
[0015] Furthermore, determining the precise coordinates of the landslide location refers to: When a risk level exists, the location coordinates of the landslide event are extracted from the geographic information database to estimate the slope, altitude, and soil type of the corresponding coordinates, thus constructing a slope stability model. The displacement values of unobserved points are estimated through spatial interpolation algorithms to gradually determine the precise landslide location coordinates.
[0016] This invention proposes a machine learning-based method for identifying and warning of landslide earthquake precursors. This method improves signal purity and reduces noise interference by denoising earthquake signal waveform data using wavelet transform. It employs the STA / LTA algorithm combined with support vector machine classification to achieve automatic, high-precision acquisition of seismic phase arrival times, reducing subjective human error. Coherence analysis is used to verify the consistency of signals from multiple stations, avoiding false event identification. The method integrates time difference features with station locations using least squares and triangulation algorithms to improve the accuracy and real-time performance of landslide location estimation. A convolutional neural network is used to classify earthquake signal waveforms by scale, automatically generating risk level labels for rapid risk assessment. Finally, a real-time data stream processing framework is used to integrate location information, landslide features, and risk labels to output warning signals, improving the reliability and response speed of landslide warnings and providing effective support for disaster prevention and control. Attached Figure Description
[0017] Figure 1 This is a step diagram of a machine learning-based landslide earthquake signal precursor identification and early warning method provided in an embodiment of the present invention; Figure 2 This is a step diagram illustrating the process of obtaining the estimated location coordinates of a landslide event according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the data processing steps in the landslide earthquake signal precursor identification and early warning process provided by an embodiment of the present invention. Detailed Implementation
[0018] The landslide earthquake signal precursor identification and early warning method proposed in this invention processes various types of data collected by multiple stations deployed in the landslide area in a hierarchical manner: First, a preliminary phase-to-time sequence is generated based on the original seismic signal waveform data. Based on the time and signal determined by the preliminary phase-to-time sequence, the original seismic signal waveform data of the corresponding stations are screened, and the location coordinates of the landslide event are estimated. Based on the location coordinates estimate, the original seismic signal waveform data, soil movement distance, and rainfall intensity at that location are screened again to accurately locate the landslide and assess the landslide earthquake risk, thereby realizing the identification and early warning of landslide earthquake signals.
[0019] The specific implementation of the present invention will now be described in detail with reference to the accompanying drawings.
[0020] The landslide earthquake signal precursor identification and early warning method provided by this invention is as follows: Figure 1 As shown, it includes the following steps: Step S100: Acquire raw seismic signal waveform data collected in real time by seismic sensors from multiple stations, preprocess the raw seismic signal waveform data, and generate seismic signal waveform data. Typically, raw seismic signal waveform data is acquired in real-time by multiple seismic sensor arrays deployed in landslide areas. This seismic signal data undergoes time-series processing using high-speed sampling technology to ensure the capture of even subtle vibration changes. For example, in a landslide monitoring system, the sensor array samples 100 to 200 times per second to effectively cover the main frequency bands of landslide signals, forming continuous time-series data. Receiving and analyzing this continuous time-series data allows for early identification and warning of the risk of landslide instability caused by geological instability in the landslide area. Therefore, if sensor arrays are distributed in a landslide-prone area in a mountainous region, when seismic waves propagate, the data is transmitted wirelessly to a central processing unit, achieving seamless reception, identification, and early warning.
[0021] Throughout the entire process of this invention, the original seismic signal waveform data is processed and used in a hierarchical manner; in this step, the original seismic signal waveform data is processed as follows: Figure 3 As shown in step S310, the original seismic signal waveform data is preprocessed to generate a preliminary phase-to-time sequence.
[0022] During preprocessing, wavelet transform is used to filter noise from the original seismic signal waveform data, resulting in a denoised seismic signal waveform. The specific steps include: 1) After acquiring the raw seismic signal waveform data, real-time acquisition technology is used to sample the time series in the raw seismic signal waveform data; 2) Determine whether the noise level of the sampled signal exceeds a preset threshold. If it does, perform time series decomposition on the sampled seismic signal data to obtain preliminary filtered waveform data. Specifically, the root mean square value of the sampled signal is used as the noise level for judgment. If the noise level exceeds a preset threshold (such as 0.5 units), the denoising process is triggered.
[0023] Wavelet transform is used for time series decomposition. Wavelet transform is a time-frequency analysis tool that can localize signal features, avoiding the limitations of global transforms. During time series decomposition, the mother wavelet function performs multi-scale decomposition of the signal, separating noise from useful signals. For example, in the processing, the Daubechies wavelet basis is selected to perform a three-level decomposition of the time series, retaining low-frequency approximation coefficients and suppressing high-frequency detail coefficients, thus obtaining preliminary filtered waveform data. Especially for signals with high noise levels, such as landslide data affected by wind, wavelet transform can effectively remove random noise, output smoother waveforms, improve signal purity, and lead to higher landslide event detection accuracy and reduced false alarms.
[0024] 3) Extract the frequency component information of the preliminary filtered waveform data and determine whether the frequency component information conforms to the characteristic range of the seismic signal. If it conforms to the characteristic range, the preliminary filtered waveform data can be determined to be denoised waveform data. After obtaining the frequency component information of the preliminary filtered waveform data, the dominant frequency component is extracted using a Fast Fourier Transform (FFT) and compared with a preset seismic signal characteristic range, such as 0.1-10Hz. If they match, the data is confirmed as denoised waveform data. The seismic signal characteristic range is determined based on the typical spectral characteristics of seismic waves. For example, when extracting the frequency components of P-waves and S-waves, if the component information falls within the characteristic range (e.g., around 2Hz), it is considered a valid signal (denoised waveform data); otherwise, it may be residual noise.
[0025] 4) Use Fourier transform to convert the time domain signal to the frequency domain, extract the residual noise spectral features in the denoised waveform data, calculate the high-frequency peak in the spectrum, and output the seismic signal waveform data if it is below the threshold.
[0026] The preprocessed seismic signal waveform data can clearly show the waveform outline of landslide events, which can significantly improve data quality in landslide early warning applications and support accurate disaster prediction.
[0027] Step S110: Determine the potential phase arrival time based on the earthquake signal waveform, and generate a preliminary phase arrival time sequence; In this step, the STA / LTA algorithm is used to calculate the ratio of short-time average to long-time average in the preliminary phase-to-time series. When the ratio exceeds a preset threshold, it is identified as a potential phase-to-time point. Potential phase-to-time points constitute the preliminary phase-to-time series. The specific steps include: 1) Extract continuous time window data from seismic signal waveforms; During the extraction process, the seismic signal waveform is divided into segments of fixed length: assuming the total duration of the seismic signal waveform data is 10 minutes, it can be divided into multiple 1-second windows, which are arranged consecutively to cover the entire signal sequence, ensuring the integrity of subsequent analysis.
[0028] In practical applications, seismic signal waveforms may contain hidden vibration patterns. Windowing can facilitate the capture of local features. Specifically, for a seismic signal waveform with a sampling rate of 500Hz, a series of continuous segments can be generated by extracting a window of data every 0.5 seconds, allowing for precise amplitude calculations for each window of data.
[0029] 2) The short-time average calculation method is used to calculate the short-time average amplitude value for continuous time window data, and the long-time average calculation method is used to calculate the long-time average amplitude value for continuous time window data. Calculating short-time average amplitude values can capture short-term fluctuations in signals. For example, for data points within a 1-second window, first calculate the absolute value of each point, then sum them and divide by the number of points. For instance, if the window has 500 sampling points, the short-time average amplitude value is the sum of these absolute values divided by 500. This method can highlight sudden fluctuations. For example, if the window data shows a sudden amplitude peak, the short-time average value will increase accordingly, indicating that a sudden fluctuation has occurred.
[0030] Calculating the long-term average amplitude value is an extended averaging strategy that involves using a longer reference period to smooth the signal. The principle is to use a larger window or accumulate the average of multiple short-term windows; for example, combining data from 10 windows preceding the current window and calculating their overall amplitude average.
[0031] When processing landslide signals, if the short-time window is 1 second, the long-time average may cover the data from the previous 10 seconds. By averaging the absolute values of these data, a value representing the background level is obtained. This method can effectively distinguish noise from real events. In practical applications, the long-time average amplitude value can reflect the long-term trend in the signal and serve as a benchmark for judging stability, avoiding short-term noise interference.
[0032] 3) The STA / LTA algorithm is used to calculate the ratio of short-time average amplitude value to long-time average amplitude value. Based on the principle that the ratio can amplify burst signals, when the ratio value exceeds the preset threshold (for example, the preset threshold is 3), the corresponding position of the continuous time window data is determined to be the potential phase arrival time point, and a preliminary phase arrival time point set is obtained. During the landslide early warning process, for a given window, if the short-term average is 0.8 units and the long-term average is 0.2 units, the ratio is 4, which exceeds the threshold (the preset threshold is set to 3), it is marked as a potential location.
[0033] 4) Classify the preliminary phase-to-time set, extract the locations of the preliminary phase-to-time set that meet the phase characteristics, and form a verification phase-to-time sequence; This step can be achieved using the vector machine algorithm: a machine learning model that separates data of different categories by constructing a hyperplane based on the principle of maximizing the margin between categories.
[0034] During classification, features such as amplitude and duration are first extracted for each potential location. These features are then input into the model, which classifies the locations based on the training data to determine which locations match the waveform characteristics of landslide rupture signals, thus obtaining the verification phase-to-time sequence. The waveform characteristics that match landslide rupture signals include: sudden increase in signal amplitude, specific frequency components of the signal, and P-wave or S-wave patterns.
[0035] In the actual scenario of landslide monitoring, the preliminary judgment of the waveform characteristics of landslide rupture signals may contain false positives. Therefore, the present invention uses a vector machine algorithm to extract reliable point sequences.
[0036] 5) Obtain the time interval information of adjacent points from the verification phase to time sequence. Determine whether the time interval information of adjacent points is continuous by using a pre-set phase interval threshold (e.g., set the phase interval threshold to 2 seconds). Continuous verification phase to time sequences are used to construct a preliminary phase to time sequence.
[0037] For example, if a sequence of points is located at the 5th, 6th, and 8th seconds, with intervals of 1 second and 2 seconds, and all meet the threshold, the sequence is confirmed to be continuous. This judgment ensures the continuity of potential landslide signal events and provides a reliable basis for landslide risk assessment.
[0038] Step S120: Based on the preliminary phase-to-time sequence, determine a consistent phase-to-time group and obtain the estimated location coordinates of the landslide event; The data processing in this step is as follows: Figure 3As shown in step S320, the original seismic signal waveform data collected by multiple seismic sensors deployed in the landslide area are extracted based on the preliminary phase-to-time sequence. The original phase-to-time data is then extracted from the data. The original phase-to-time data is denoised using a filtering tool to obtain the denoised phase-to-time data, which is used to generate the location coordinate estimation of the landslide event.
[0039] In actual landslide monitoring, a fundamental step is to acquire raw seismic signal waveform data from a dense array of stations deployed around a potential landslide body and extract raw phase arrival time data. These stations are typically distributed along the rear, front, and flank edges of the landslide body, with station spacing ranging from several hundred meters to about one kilometer. They record the arrival time information of seismic waves. Filtering tools are used to process the raw phase arrival time data. For example, low-pass filters are used to remove high-frequency noise (environmental interference in the raw data, such as vehicle vibration or wind noise), and the cutoff frequency of the filter is set to smooth the signal curve, resulting in clearer denoised phase arrival time data. Since the raw phase arrival time data often contains non-seismic-related fluctuations, failure to denoise may lead to misjudgment of the phase arrival time location. For example, three stations, A, B, and C, each acquired a set of raw data sequences. The data from station A shows a clear peak but is accompanied by a large amount of small-amplitude noise. By applying the Butterworth filtering method, the signal spectrum is calculated, retaining frequency components below 20Hz and filtering out noise above this level, ultimately obtaining a smooth, denoised data sequence.
[0040] Then, based on the denoised phase arrival time data, normalization processing is performed to unify the scale differences of data from different stations. For example, due to differences in instrument sensitivity or distance from the landslide source, the amplitude values of different stations are not the same. By using the minimum-maximum normalization method, the data is mapped to the interval between 0 and 1 to obtain the preprocessed first time series.
[0041] Then, using the first time series, a sliding window tool is used to extract fixed-length time window data (e.g., a window length of 10 seconds) from each station. The sequence is traversed in 1-second steps, and the average amplitude or peak characteristics are calculated for each window to form a second time series. This helps to capture the continuity of local seismic phase characteristics.
[0042] For the second time series, the Pearson coherence coefficient of the time series between each station is calculated. The similarity between the series is quantified by the Pearson correlation coefficient formula: for example, for the series of stations A and B, the time axis is first aligned, and then the ratio of covariance to standard deviation is calculated point by point. If the coefficient is greater than the preset threshold of 0.8, it is judged as a consistent phase-time pair, and finally aggregated into a consistent phase-time group.
[0043] For example, in a real landslide event, if the second time series from multiple stations show similar waveform patterns, then the corresponding P-wave arrival can be confirmed through coherence calculation.
[0044] Because signals from multiple stations can be inconsistent, this invention improves the accuracy of landslide event location and the reliability of landslide source parameter estimation by establishing a consistent set of seismic phase arrival times. If multiple stations are involved in a practical application, weighted coherence calculation can be introduced: first, weights are assigned based on station distances, and then an average coefficient is calculated to adapt to data variability in complex terrain.
[0045] The process of extracting time difference features from the consistent seismic phases and time groups, and then using the least squares method to fit the correspondence between the time difference features and the station locations to obtain the estimated location coordinates of the landslide event is as follows: Figure 2 As shown, it includes: 1) Obtain the arrival times of the filtered P-wave and S-wave: Based on the time and station of the consistent phase arrival group, the original seismic wave signal obtained by the corresponding station at the corresponding time is extracted, and bandpass filtering is performed to generate filtered arrival data; at this time, the filtered arrival data is valid data within a specific frequency range. Bandpass filtering can be performed using a Butterworth filter. Low-pass and high-pass cutoff frequencies are set for seismic wave signals with frequencies in the range of 1-10 Hz to remove noise interference and obtain filtered arrival time data. The filtered arrival time data reflects the arrival times of P-waves and S-waves.
[0046] For example, a monitoring network in a mountainous area includes three stations, A, B, and C. When a landslide event occurs, filtered arrival time data is extracted from the seismic wave signals collected by these three stations. For example, the arrival time of the P-wave at station A is t1=5.2 seconds, and the arrival time of the S-wave is t2=8.7 seconds.
[0047] 2) Extract P-wave and S-wave arrival time data from at least two stations from the filtered arrival time data, and calculate the time difference feature set; In specific calculations, the arrival times of P-waves and S-waves from at least two stations are compared: for example, for stations A and B, the P-wave time difference Δt_p = |t1_B - t1_A| and the S-wave time difference Δt_s = |t2_B - t2_A| are calculated. If the P-wave arrival time at station A is 5.2 seconds and at station B it is 6.1 seconds, then Δt_p = 0.9 seconds. Similarly, the S-wave time difference is obtained in the same way, forming a set containing multiple difference pairs. Further, all station pairs are iterated through; for example, if station C is added, the time differences between AC and BC are calculated, thus constructing a comprehensive set of time difference characteristics. This set captures the relative delay of seismic wave propagation and can indicate the location pattern of landslide events.
[0048] 3) The least squares method is used to fit the correspondence between the time difference feature set and the station location. When the deviation between the time difference feature set and the preset station location model is less than the preset threshold, the least squares fitting parameters are iteratively calculated to obtain the preliminary location coordinate estimate. In this step, the landslide source coordinate parameters are estimated by minimizing the squared difference between the observed values and the model predictions using the least squares method: assuming that the preset station location model is a propagation equation based on a velocity model, if the deviation of the time difference set is less than a threshold (e.g., 0.5 seconds), the iterative process is started, and the landslide source coordinate parameters are adjusted in each iteration until convergence is obtained to obtain a preliminary location coordinate estimate.
[0049] Specifically, assuming three monitoring stations A, B, and C are deployed on a potential landslide, with the rear edge of the landslide as the origin of the coordinate system, the three stations are located at coordinates A(0,0,100), B(800,0,90), and C(0,600,95) (unit: meters), respectively. After calculating the time difference set of the three stations, the preliminary estimated coordinates of the landslide location are calculated using the least squares method iteratively as (300,200,50) (unit: meters), which is the upper-middle part of the landslide.
[0050] 4) The preliminary location coordinate estimate is weighted and fused with the time difference feature set. When the accuracy of the fused coordinate estimate meets the preset threshold, the depth of the landslide source is determined to obtain the final landslide event location coordinate estimate.
[0051] The weighted fusion takes into account the reliability weight of each time difference. For example, the weight is based on the signal-to-noise ratio. If the accuracy of the fused coordinate estimation is less than a preset threshold of 0.1 km standard deviation, the sliding surface depth (or rupture initiation depth) is finally determined, thereby outputting the location coordinate estimation of the landslide event.
[0052] For example, the initial location coordinates (500, 300, 50) can be optimized to (480, 320, 45) (unit: meters) after fusion. This coordinate system is located in a local coordinate system with the rear edge of the landslide as the origin, and the confirmed rupture depth is approximately 45 meters. This not only controls the positioning error to within tens of meters, but also provides crucial and accurate spatial information for emergency response. In the practical application of landslide early warning, this optimization can clearly indicate the location of the unstable area on the slope (located at the leading edge of the slope, rather than the rear edge or flank), thus making the early warning information and emergency response measures (such as closing the road section at the foot of the corresponding slope and evacuating residents in specific areas) highly targeted, greatly improving emergency response efficiency and reducing the social costs of false alarms.
[0053] Step S130: Based on the earthquake signal waveform and landslide event location coordinates, classify the landslide earthquake magnitude and output the earthquake risk level; The steps for classifying landslide earthquake magnitude include the following: 1) Based on the estimated coordinates of the landslide event location, the original seismic signal waveform data is obtained from the seismic monitoring equipment of the corresponding station. Wavelet transform is used to denoise the original seismic signal waveform data to obtain the denoised waveform data. The data processing in this step is as follows: Figure 3 As shown in step S330, the original ground motion signal waveform data is filtered and processed based on the estimated location coordinates of the landslide event to achieve landslide earthquake scale classification.
[0054] For example, the location coordinates of a monitoring station in a landslide-prone mountain area are (30.5°N, 102.8°E). After collecting the raw seismic signal data at this location, the monitoring system transmits the raw seismic signal data wirelessly to the central processing system. The central processing system uses wavelet transform to denoise the raw seismic signal data to remove environmental noise such as wind interference or clutter caused by human activities, thereby obtaining clearer denoised waveform data. Wavelet transform can select a suitable mother wavelet function (such as the Daubechies wavelet) to decompose the signal, separating the noise components into high-frequency sub-bands, and then reconstructing the signal through thresholding.
[0055] During the decomposition process, the original signal with a duration of 10 seconds is decomposed into multiple scale layers. The noise coefficient in the high-frequency layer is set to zero, and then the inverse transformation is performed to obtain smooth waveform data. This helps to highlight the true characteristics of landslide events and avoid errors in subsequent classification.
[0056] 2) Extract time series features from the denoised waveform data, and use a convolutional neural network to classify the time series features to obtain risk scale classification labels for landslide events; Time-series features, such as peak amplitude, energy spectrum, and autocorrelation function, are extracted from the denoised waveform data. A convolutional neural network (CNN) is then used to classify these features, thereby obtaining a scale classification label for the landslide event. The CNN processes the sequence data through local pattern recognition in convolutional layers, dimensionality reduction in pooling layers, and classification in fully connected layers.
[0057] Specifically, in the convolutional neural network of this invention, multiple layers with a kernel size of 3 are set up, and the network parameters are optimized using a historical earthquake dataset with an input feature vector dimension of 512 to train and form an earthquake scale classification model. The earthquake scale classification model can output risk scale classification labels for the denoised waveform data, including: "small scale", "medium scale" or "large scale", which reflects the potential slip volume and destructive power of the event.
[0058] 3) If the scale classification label indicates a large-scale event, the scale classification label is compared according to a preset threshold and judged as a high-risk landslide event; If the classification label indicates a large-scale event, the displacement characteristics are compared with a preset threshold (e.g., an energy value of 500 joules) to determine whether it is a high-risk landslide event. If the label is "large-scale" and the displacement characteristics exceed the threshold, the system automatically marks it as high-risk, thereby triggering the early warning mechanism.
[0059] 4) Generate risk level labels based on high-risk landslide events; store these labels in a database for easy retrieval of the final landslide risk assessment results during application. The risk level labels include "Level 1 Alert" or "Level 2 Evacuation," reflecting a comprehensive assessment of the event's scale and location.
[0060] Step S140: If the earthquake risk level is high, the location coordinates of the landslide event are estimated, and the risk level and location information are integrated using a real-time data stream processing framework to obtain a landslide early warning signal output.
[0061] The data processing in this step is as follows: Figure 3 As shown in step S340, based on the estimated location coordinates of the landslide event, the original seismic signal waveform data, geological displacement and rainfall data are extracted, and the landslide location is determined and matched in conjunction with the geographic information database. Based on the characteristics of landslide earthquakes, a landslide earthquake early warning signal is generated.
[0062] The specific implementation includes the following steps: 1) Real-time data, including seismic waveforms, geological displacement, and rainfall data, are acquired from multi-source sensors at various monitoring stations. Data stream processing technology is used to preprocess the real-time data, remove noise data, and obtain a standardized real-time data stream.
[0063] The real-time data obtained in this step comes from various devices deployed in landslide-prone mountainous areas, including: seismic sensors installed on slopes to capture changes in underground waveforms, geological displacement gauges to monitor soil movement distances, and rain gauges to record hourly rainfall intensity. This real-time data is transmitted to a central processing system for preprocessing using data stream processing techniques. This includes using a sliding window algorithm to calculate the real-time trend of vibration energy changes, identifying and removing noise values caused by equipment malfunctions or external interference, and removing abnormal high-frequency fluctuations or readings outside the normal range. This results in a standardized real-time data stream. Through preprocessing, the real-time data is cleaned in real time, improving overall monitoring efficiency.
[0064] 2) Based on standardized real-time data streams and combined with a pre-established geographic information database, location coordinates are matched to estimate relevant location information, and spatial interpolation algorithms are used to determine the precise landslide location coordinates.
[0065] When matching the location coordinates of landslide events with location information related to the estimation, information such as slope, elevation, and soil type of the estimated landslide event location coordinates is extracted from the geographic information database and matched with standardized real-time data streams to construct a slope stability model. For example, when an earthquake waveform shows a specific frequency peak indicating an earthquake risk level, the corresponding coordinate data in the database is queried to construct a slope stability model. The displacement values of unobserved points are estimated using spatial interpolation algorithms (such as the Kriging method), gradually determining the precise landslide location coordinates. In this way, the model can be expanded from initial grid points to a centimeter-level precision coordinate system, thus forming a reliable basis for location estimation.
[0066] 3) Based on the risk level of the precise landslide location coordinates, standardized real-time data streams are integrated to generate a risk assessment result. Based on the risk assessment result, an event-triggered mechanism is used to determine whether the risk assessment result exceeds a preset threshold. If it does, an early warning signal is generated and released through the signal output interface.
[0067] If the risk value of the precise landslide location coordinates matches the preset risk level threshold, for example, when the risk value of the coordinate point exceeds 80 points, a comprehensive risk score is generated by fusing earthquake intensity and rainfall data using a Bayesian network. If the comprehensive risk score reaches a certain level, such as exceeding 90 points, an early warning signal is generated and issued through a signal output interface such as an SMS platform or alarm system. For example, an alarm message containing the location coordinates and the expected impact range is sent to local residents and emergency departments to ensure timely response to potential landslide threats.
[0068] This invention provides a machine learning-based method for identifying and warning of landslide earthquake precursors. It utilizes multi-level processing of various types of data collected from multiple stations deployed in landslide areas, tailored to different objectives. At each level, different raw data are selected for preprocessing optimization: wavelet transform is used to denoise the earthquake signal waveform data, improving signal purity and reducing noise interference; the STA / LTA algorithm combined with support vector machine classification achieves automatic, high-precision acquisition of seismic phase arrival times, reducing subjective errors; coherence analysis is used to verify the consistency of signals from multiple stations, avoiding false event identification; time difference features and station locations are fused using least squares and triangulation algorithms to improve the accuracy and real-time performance of landslide location estimation; convolutional neural networks are used to classify the earthquake signal waveform by scale, automatically generating risk level labels for rapid risk assessment; finally, a real-time data stream processing framework is used to integrate location information, landslide features, and risk labels to output warning signals, improving the reliability and response speed of landslide warnings and providing effective support for disaster prevention and control.
[0069] The above-disclosed embodiments are merely a few specific examples of the present invention. However, the present invention is not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.
Claims
1. A method for identifying and issuing early warning of landslide seismic signals based on machine learning, characterized in that, Includes the following steps: The original seismic signal waveform data is acquired, and the original seismic signal waveform data is preprocessed to generate seismic signal waveform data. Based on the earthquake signal waveform, the potential phase arrival time point is determined, and a preliminary phase arrival time sequence is generated; Based on the preliminary phase-to-time sequence, the original phase-to-time data are extracted, consistent phase-to-time groups are determined, and the location coordinates of the landslide event are estimated. Based on the earthquake signal waveform and landslide event location coordinate estimation, the original ground motion signal waveform data is filtered, landslide earthquake scale is classified, and earthquake risk level is output. If the earthquake risk level is high, the original earthquake signal waveform data, geological displacement and rainfall data are extracted based on the estimated location coordinates of the landslide event. The earthquake risk level and the estimated location coordinates of the landslide event are then integrated to obtain a landslide early warning signal output.
2. The landslide earthquake signal precursor identification and early warning method according to claim 1, characterized in that, The preprocessing of the raw seismic signal waveform data includes: Sampling is performed on the time series of the original seismic signal waveform data; Determine whether the noise level of the sampled signal exceeds a preset threshold. If it does, perform time series decomposition on the sampled seismic signal data to obtain preliminary filtered waveform data. Extract frequency component information from the preliminary filtered waveform data, determine whether the frequency component information conforms to the characteristic range of the seismic signal, and generate denoised waveform data. The time-domain signal is converted into the frequency domain using Fourier transform, and the seismic signal waveform data is output.
3. The method for identifying and issuing early warning of landslide earthquake signals according to claim 1, characterized in that, The generation of the preliminary seismic phase-to-time sequence includes the following steps: Extract continuous time window data from the seismic signal waveform; The short-time average calculation method is used to calculate the short-time average amplitude value for continuous time window data, while the long-time average calculation method is used to calculate the long-time average amplitude value for continuous time window data. Calculate the ratio of the short-time average amplitude value to the long-time average amplitude value, and determine whether the location corresponding to the continuous time window data is the potential earthquake phase arrival time point based on the ratio value, so as to obtain a preliminary earthquake phase arrival time point set. The preliminary phase-to-time point set is classified and processed, and the points of the preliminary phase-to-time point set that meet the characteristics of the phase are extracted to form a verification phase-to-time point sequence; The time interval information of adjacent points is obtained from the verification phase-to-time sequence. The continuity of the time interval information of adjacent points is determined by a pre-set phase interval threshold. A continuous verification phase-to-time sequence is constructed as a preliminary phase-to-time sequence.
4. The landslide earthquake signal precursor identification and early warning method according to claim 1, characterized in that, Determining the consistent seismic phase arrival time group includes the following steps: Obtain the original phase arrival time data, and perform noise reduction processing on the original phase arrival time data to generate noise-reduced phase arrival time data; The denoised phase arrival data is normalized to obtain a first time series; Using the first time series, a sliding window tool is used to extract fixed-length time window data from each station, and the average amplitude or peak characteristics are calculated for each window to form a second time series. For the second time series, the Pearson coherence coefficient of each station is calculated. The similarity between the series is quantified by the Pearson correlation coefficient formula. Data with similarity greater than a preset threshold are judged as consistent phase-time pairs and consistent phase-time groups are generated.
5. The method for identifying and issuing early warning of landslide earthquake signals according to claim 1, characterized in that, The process of obtaining the estimated location coordinates of the landslide event includes the following steps: Extract time difference features from the consistent seismic phases to time groups; Obtain the arrival times of the filtered P-wave and S-wave; Extract P-wave and S-wave arrival time data from at least two stations and calculate the time difference feature set; The relationship between the fitted time difference feature set and the station location is determined. When the deviation between the time difference feature set and the preset station location model is less than the preset threshold, the least squares fitting parameters are iteratively calculated to obtain the preliminary location coordinate estimate. The preliminary location coordinate estimate is weighted and fused with the time difference feature set. When the accuracy of the fused coordinate estimate meets the preset threshold, the depth of the landslide source is determined to obtain the estimated location coordinates of the landslide event.
6. The method for identifying and issuing early warning of landslide earthquake signals according to claim 1, characterized in that, The landslide earthquake magnitude classification includes the following steps: Based on the estimated coordinates of the landslide event location, the original seismic signal waveform data was obtained from the seismic monitoring equipment of the corresponding station. Wavelet transform was used to denoise the original seismic signal waveform data to obtain the denoised waveform data. Time series features are extracted from the denoised waveform data, and convolutional neural networks are used to classify the time series features to obtain risk scale classification labels for landslide events; the risk scale classification includes: small scale, medium scale, and large scale. If the risk scale is classified as large-scale, determine whether it is a high-risk landslide event; if it is a high-risk landslide event, generate a risk level label.
7. The method for identifying and warning of landslide earthquake signals according to claim 1, characterized in that, The integration of the earthquake risk level and the location coordinate estimation of the landslide event includes the following steps: Obtain standardized real-time data streams; Based on the standardized real-time data stream and combined with the geographic information database, the location information of the landslide event is estimated by matching the location coordinates of the landslide event according to the earthquake risk level, and the precise location coordinates of the landslide are determined.
8. The landslide earthquake signal precursor identification and early warning method according to claim 7, characterized in that, The acquisition of standardized real-time data streams includes: Acquire and summarize real-time data of earthquake waveforms, geological displacement, and rainfall. The sliding window algorithm is used to calculate the trend of vibration energy change from the real-time data, identify and remove noise values, remove abnormal high-frequency fluctuations or readings that are outside the normal range, and clean the real-time data in real time to generate a standardized real-time data stream.
9. The method for identifying and warning of landslide earthquake signal precursors according to claim 7, characterized in that, The precise coordinates of the landslide location refer to: When a risk level exists, the location coordinates of the landslide event are extracted from the geographic information database to estimate the slope, altitude, and soil type of the corresponding coordinates, thus constructing a slope stability model. The displacement values of unobserved points are estimated through spatial interpolation algorithms to gradually determine the precise landslide location coordinates.